Feature Extraction and ML Classifier Development FDD

Feature Extraction and Machine Learning Classifier Development

for Fault Detection and Diagnosis

By DOYI JOO, B.ENG

A Thesis Submitted to the School of Graduate Studies in Partial Fulfilment of the

Requirements for the Degree of Master of Applied Science

McMaster University © Copyright by Doyi Joo, June 2020

i
MASTER OF APPLIED SCIENCE (2020) McMaster

University Mechanical Engineering Hamilton,

Ontario, Canada

TITLE: Feature Extraction and Machine Learning Classifier

Development for Fault Detection and Diagnosis

AUTHOR: Doyi Joo, B.Eng

SUPERVISOR: Professor Saeid Habibi

NUMBER OF PAGES: xvii, 139

ii
Abstract

In this research, a fault detection and diagnosis strategy for internal combustion

engine is developed using measurements that are readily available in engine testing

environment to monitor abnormal combustion. The FDD strategy is designed to monitor

the engine on a cycle-by-cycle basis using measurements that are accessible on a running

vehicle. Pressure measurements are easily accessible in a testing facility that provide useful

insight into the quality of the combustion occurring inside the engine. However, due to its

cost and complex installation procedures, it is not feasible to obtain in-cylinder pressure

measurements from an in-vehicle engine. Faults of a mechanical system are often

investigated using vibration. Due to the low cost and non-invasive nature of

accelerometers, vibration measurement is used to monitor the in-vehicle engine. However,

as vibration behaviors of complex system such as an engine is hard to characterize, in-

cylinder pressure measurement is used during the development of the FDD strategy to

assist in characterizing the vibration measurement. Upon data acquisition, features are

extracted from the vibration measurements using Extended-MSPCA for better

characterization and data reduction with a multi-baseline technique. Pressure

measurements are analyzed using thermodynamic theories to assess the combustion quality

of each cycle. The vibration measurements are labelled corresponding to the pressure

analysis. An artificial neural network classifier is developed using the extracted and

labelled features. Developed classifier detected the fault and its location with an overall

accuracy of 96.3%.

iii
To my parents, my sister & R.T for their endless love and support

iv
Acknowledgements

I would like to thank my supervisor, Dr. Habibi for his guidance throughout this

research. I would also like to thank Dr. Tjong and his team at Ford Powertrain Engineering

Research and Development Centre for their support during my time at PERDC.

v
Table of Contents

Abstract .............................................................................................................................. iii

Acknowledgements ..............................................................................................................v

Table of Contents ............................................................................................................... vi

List of Figures .................................................................................................................... xi

List of Tables .....................................................................................................................xv

List of Abbreviations ....................................................................................................... xvi

List of Symbols ............................................................................................................... xvii

Chapter 1: Introduction ........................................................................................................1

1.1 Overview ....................................................................................................................1

1.2 Research Motivation ..................................................................................................2

1.3 Research Objectives ...................................................................................................4

1.4 Thesis Structure .........................................................................................................4

Chapter 2: FDD Strategies for Internal Combustion Engines .............................................6

2.1 Background ................................................................................................................6

2.1.1 Internal Combustion Engine .............................................................................. 6

2.1.2 Fault Detection & Diagnosis ............................................................................ 10

2.2 Fault Detection & Diagnosis for Internal Combustion Engines ..............................13

vi
 2.2.1 FDD Approach ................................................................................................. 14

2.2.2 Measurements for FDD.................................................................................... 15

Vibration Signals .................................................................................................. 15

Pressure Signals .................................................................................................... 17

2.3 Feature Extraction Methods .....................................................................................17

2.3.1 Time Domain Analysis .................................................................................... 17

2.3.2 Crank Angle Domain Analysis ........................................................................ 18

2.3.3 Frequency Domain Analysis ............................................................................ 19

2.3.4 Order Analysis ................................................................................................. 20

2.3.5 Time-Frequency Analysis ................................................................................ 21

2.3.5.1 Short Time Fourier Transform .................................................................. 22

2.3.5.2 Wavelet Transform ................................................................................... 26

2.3.6 Principal Component Analysis ........................................................................ 36

2.3.7 Multi-Scale Principal Component Analysis..................................................... 48

2.3.8 Mod-MSPCA ................................................................................................... 49

Chapter 3: Proposed FDD Strategy....................................................................................51

3.1 Proposed Experimental Set-up & Data Acquisition ................................................55

3.1.1 Proposed Instrumentation ................................................................................ 55

vii
 3.1.2 Proposed Fault Induction Method .................................................................... 57

3.2 Proposed Data Pre-Processing Methods ..................................................................57

3.2.1 Proposed Crank Angle Conversion Method .................................................... 58

3.2.2 Proposed Cycle Segmentation Method ............................................................ 58

3.2.3 Proposed Pressure Analysis Method ................................................................ 59

3.2.4 Proposed Data Labeling Method ..................................................................... 60

3.3 Proposed Feature Extraction Method.......................................................................60

3.4 Classifier Development ............................................................................................65

Chapter 4: Data Acquisition...............................................................................................71

4.1 Experimental Setup ..................................................................................................71

4.1.1 Engine .............................................................................................................. 72

4.1.2 Test Cell ........................................................................................................... 74

4.1.2.1 Engine Assembly ...................................................................................... 74

4.1.2.2 Dynamometer ............................................................................................ 74

4.1.3 Encoder ............................................................................................................ 75

4.1.4 Pressure Transducer ......................................................................................... 77

4.1.5 Accelerometers ................................................................................................ 79

4.1.6 Data Acquisition Hardware.............................................................................. 83

viii
 4.2 Engine Test Procedure .............................................................................................84

4.2.1 Operating Conditions ....................................................................................... 84

4.2.2 Testing Procedures ........................................................................................... 85

4.2.3 Raw Data Collected ......................................................................................... 85

Accelerometer Data .............................................................................................. 85

Encoder Data ......................................................................................................... 88

Pressure Data ........................................................................................................ 91

4.3 Proposed Data Pre-Processing Strategy Implementation ........................................93

Step 1: Filtering of Accelerometer Data ................................................................... 94

Step 2: Encoder Measurement Analysis ................................................................... 94

Step 3: Resampling of Transducer Measurements.................................................... 99

Step 4: Engine Cycle Segmentation ........................................................................ 100

Step 5: Data Labeling.............................................................................................. 101

Pre-Feature Extraction Data Summary ................................................................... 105

Chapter 5: Proposed FDD Strategy Implementation and Results ....................................112

5.1 Proposed Feature Extraction Strategy Implementation .........................................112

Step 1: Wavelet Function Selection ........................................................................ 113

Step 2: Wavelet Level Selection ............................................................................. 117

ix
 Step 3: Defining Baseline ....................................................................................... 118

Feature Summary .................................................................................................... 119

5.2 Results and Observations .......................................................................................124

Classifier 1: Wavelet Function db4..................................................................... 126

Classifier 2: Wavelet Function: db10 ................................................................. 128

Classifier 3: Wavelet Function: db16 ................................................................. 130

Chapter 6: Conclusion......................................................................................................136

6.1 Research Contributions ..........................................................................................136

6.2 Future Work ...........................................................................................................138

References ............................................................................................................................ I

x
List of Figures

Figure 1: Four-stroke cycle of internal combustion engine with piston and crankshaft

position [5] .......................................................................................................................... 7

Figure 2: PV diagram of Otto cycle [6] .............................................................................. 8

Figure 3: Transformations in FDD systems [11] .............................................................. 11

Figure 4: Resampling for order analysis [44] ................................................................... 20

Figure 5: Signal windowing technique [12]...................................................................... 22

Figure 6: Spectrogram and spectrum of a non-stationary signal [12] ............................... 23

Figure 7: High frequency resolution [12] ......................................................................... 25

Figure 8: High time resolution [12] .................................................................................. 26

Figure 9: Frequency and Time Resolution of analysis in different domains [46] ............ 29

Figure 10: Discrete wavelet transform filter bank [12] .................................................... 33

Figure 11: DWT decomposition and synthesis overview [49] ......................................... 34

Figure 12: Wavelet Packet Transform filter banks [12] ................................................... 35

Figure 13: Frequency distributions of DWT and WPT [22] ............................................. 36

Figure 14: Principal component analysis [50] .................................................................. 38

Figure 15: MSPCA schematic [12] ................................................................................... 49

Figure 16: Mod-MSPCA schematic [12] .......................................................................... 50

xi
Figure 17: Development overview of the proposed strategy: ........................................... 54

Figure 18: Mod-MSPCA signal transformations [12] ...................................................... 62

Figure 19: Extended MSPCA signal transformations [12] ............................................... 62

Figure 20: MLP with two hidden layers [57].................................................................... 66

Figure 21: Model of a neuron [57] .................................................................................... 67

Figure 22: Function and error signals [57] ....................................................................... 69

Figure 23: Overhead valve vs overhead cam engine design [57] ..................................... 73

Figure 24: Different styles of engines [58] ....................................................................... 73

Figure 25: Engine cylinder numbering for the V8 test engine [59] .................................. 73

Figure 26: Encoder components [61]................................................................................ 76

Figure 27: The mechanical load on the crystal producing electrical charge through electric

dipole [63] ......................................................................................................................... 77

Figure 28: Piezoelectric pressure sensor components [64] ............................................... 78

Figure 29: IEPE accelerometer components [66] ............................................................. 80

Figure 30: Knock sensor locations on sample V8 engine [68] ......................................... 81

Figure 31: Accelerometer locations on sample V8 engine block [69].............................. 82

Figure 32: Raw accelerometer data from front valley, axis parallel with front-back

orientation of the engine ................................................................................................... 86

xii
Figure 33: Raw accelerometer data from front valley, axis parallel with right-left

orientation of the engine ................................................................................................... 87

Figure 34: Raw accelerometer data from front valley ...................................................... 88

Figure 35: Raw accelerometer data from front left knock sensor ..................................... 88

Figure 36: Raw encoder ticks @ 600rpm ......................................................................... 89

Figure 37: Raw encoder trigger @ 600rpm ...................................................................... 90

Figure 38: Raw Cylinder 1 Pressure Trace @600RPM .................................................... 92

Figure 39: Overlay of all of the raw pressure traces from each cylinder @600RPM....... 93

Figure 40: Filtering of accelerometer data using Butterworth Filter ................................ 94

Figure 41: Encoder tick signal .......................................................................................... 95

Figure 42: Calculated engine speed @ 600RPM .............................................................. 96

Figure 43: Pressure trace of a cylinder [62] ...................................................................... 97

Figure 44: Overlay of trigger data and Cylinder 1 pressure trace @ 600RPM ................ 99

Figure 45: Filtered accelerometer data in time domain .................................................... 99

Figure 46: Filtered accelerometer data in crank angle domain ....................................... 100

Figure 47: All pressure traces for 8 cylinder engine cycle ............................................. 102

Figure 48: Cylinder 2 Pressure Trace of 2 Consecutive Cycles ..................................... 104

Figure 49: Cylinder 5 Pressure Trace of 2 Consecutive Cycles ..................................... 104

xiii
Figure 50: Cylinder 8 Pressure Trace of 2 Consecutive Cycles ..................................... 105

Figure 51: Cylinder 1 Pressure Traces for Faulty and Healthy sample cycles ............... 107

Figure 52: Front-Back Axis Accelerometer data from Front Valley for Faulty and Healthy

sample cycles .................................................................................................................. 108

Figure 53: Right-Left Axis Accelerometer data from Front Valley for Faulty and Healthy

sample cycles .................................................................................................................. 109

Figure 54: Up-Down Axis Accelerometer data from Front Valley for Faulty and Healthy

sample cycles .................................................................................................................. 110

Figure 55: Axis Accelerometer data from Front Left Knock Sensor for Faulty and Healthy

sample cycles .................................................................................................................. 111

Figure 56: Frequency response of Daubechies wavelets [74] ........................................ 114

Figure 57: Low and high pass filter for db16 ................................................................. 115

Figure 58: Low pass and high pass filter for db10.......................................................... 116

Figure 59: Low pass and high pass filter for db4............................................................ 116

Figure 60: Features extracted from valley accelerometers ............................................. 122

Figure 61: Features extracted from knock accelerometers ............................................. 123

Figure 62: Confusion matrices for Classifier 1 ............................................................... 127

Figure 63: Confusion matrices for Classifier 2 ............................................................... 129

Figure 64: Confusion matrices for Classifier 3 ............................................................... 131

xiv
List of Tables

Table 1: Dynamometer Specifications .............................................................................. 75

Table 2: Summary of labeled data .................................................................................. 106

Table 3: Max wavelet level for corresponding wavelet .................................................. 118

Table 4: Data breakdown ................................................................................................ 120

Table 5: Time duration for each feature extraction ........................................................ 124

Table 6: Performance Scores .......................................................................................... 135

xv
List of Abbreviations

ADC Analog to Digital Converter

ANN Artificial Neural Network
BDC Bottom Dead Center
CAD Crank Angle Domain
CMHT Center for Mechatronics and Hybrid Technologies
CWT Continuous Wavelet Transform
DWT Discrete Wavelet Transform
FDD Fault Detection and Diagnosis
ICE Internal Combustion Engine
IEMSPCA Industrial Extended Multi-Scale Principal Components Analysis
MSPCA Multi-Scale Principal Component Analysis
OBD On-Board Diagnostics
PCA Principal Components Analysis
PERDC Powertrain Engineering Research and Development Center
RBC Reconstruction Based Charts
RMS Root Mean Square
SPE Squared Prediction Error
STFT Short Time Fourier Transform
SVD Singular Value Decomposition
TDC Top Dead Center
WPT Wavelet Packet Transform

xvi
List of Symbols

𝜓(∙) Mother Wavelet Function

𝐶𝜓 Wavelet Admissibility Condition Parameter

𝑋𝑤 (𝑎, 𝑏) Wavelet Transform

𝑇 Principal Component Scores Matrix

𝑃 Principal Component Loadings Matrix

Λ Eigenvalue Diagonal Matrix

𝑋𝑇 Transpose of X

𝒯2 Hotteling’s T-Squared Index

𝒬 Squared Prediction Error Index

𝜑 Combined Index

Σ𝑇 Diagonal Matrix of Principal Components

𝛿 Upper Limit of SPE Index

𝜏 Upper Limit of Hotteling’s Index

Φ Upper Limit of the Combined Index

𝜁 Chi-Squared Distribution Upper Limit

𝑔𝜑 Chi-Squared Distribution Weights

ℎ𝜑 Chi-Squared Distribution Degree of Freedom

𝐶𝑂𝑉(∙) Covariance Function

𝐹𝑐 Fault Isolation Index

xvii
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Chapter 1: Introduction

1.1 Overview

Fault detection and diagnosis has been a topic of interest for many years in the

industrial sector such as in the oil and gas, the power generation and the automotive

industries. The origins of FDD is rooted in the field of control systems and more

specifically, in automation technology which was enabled by the rise of computers. An

early form of automation is called regulatory control, which allows for low level control

actions to be performed without a human operator [1]. More recently, due to great

improvements in computational capabilities, the realm of automation has surpassed

regulatory control and advanced well into more complex tasks such as condition

monitoring (including fault detection and diagnosis) of very complex processes and

systems.

Fault Detection and Diagnosis (FDD) has mostly been applied to industrial

components and systems to prevent events that may bring forth tremendous amount of

financial burden if a malfunction were to occur. However, as service and quality of

consumer products become more and more in demand and as sophisticated data analysis

methods arise such as machine learning algorithms, FDD technologies are in demand by a

variety of industries that aim to provide the best quality for service and product.

1
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Currently, the automotive sector is at the verge of a very important change that may

redefine the automotive sector itself and the way consumers perceive transportation and

vehicles. The factor that are initiating such monumental changes in the automotive sector

is the advancement of computing technology that were mentioned previously. Machine

learning algorithms have proven that there is very useful information embedded in data

that are being collected by sensors that are currently not being extracted. In the automotive

industry large quantities of data are being collected every day. One of the areas that have

great potential for extracting useful information through tools like machine learning is in

testing, where tremendous amounts of valuable data is being collected on important

components of the vehicle such as the engine. In this research, the potential opportunities

for technological advancement through the access of unused data or unused data source is

explored along with development of advanced data analytics methods.

1.2 Research Motivation

The motivation for this research can be viewed from different perspectives. For

example, the condition monitoring or FDD technology for internal combustion engines can

be developed and utilized in a variety of different sectors of the automotive industry such

as in the manufacturing lines, fleet operations, engine diagnostics, and engine controls.

For a car manufacturer, a high productivity and high-quality assurance can be

achieved with the help of FDD technology, which could ensure the detection of faulty

engine components. Automated FDD strategy can greatly reduce the number of human

operators that are required in quality assurance. Automation of quality assurance would

2
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

reduce the time required to assess the quality of the components and increase the accuracy

rates of the quality assessment, which allows for human operators to attend to more

challenging tasks such as mitigating issues around faulty components discovered.

Productivity of the manufacturer and the quality of their products can be improved

significantly through an automated FDD strategy.

For a fleet operator, in-vehicle FDD technology would be useful in monitoring the

health of the engine. Although On-Board Diagnostics (OBD) exists in every consumer

vehicle, it can be argued that there is room for improvement in the sophistication of the

data analysis methods to broaden the range of faults and the levels of faults that are

monitored.

In an auto-shop or dealership, diagnosis of faults in engines can greatly be

expedited with the use of the automated FDD strategy.

For engine controls, an FDD technology that is capable of monitoring the engine’s

combustion quality in real time would be of great significance as control systems could be

adjusted in real time to improve engine emissions and performance.

FDD strategies are being research internationally, including at McMaster’s Centre

for Mechatronics and Hybrid Technologies (CMHT). This research is a continuation of the

FDD research that was conducted by the following researchers at CMHT:

1. S. R. Haqshenas – who introduced CAD-MSPCA for its application to

internal combustion engines,

3
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

2. M. Ismail – who introduced the Industrial Extended Multi-Scale Principal

Components Analysis (IEMSPCA) for its application to alternators and starters

using sound and vibration, and

3. A. Doghri – who successfully applied an FDD strategy for in-vehicle

application using vibration measurements.

1.3 Research Objectives

The objective of this research is to develop a FDD strategy for internal combustion

engines that is intended to monitor the engine on a cycle-by-cycle basis to detect

intermittent faults. The development of a FDD strategy involves three main stages as

follows, [2]:

1. data collection and pre-processing of time series data,

2. feature extraction of dataset, and

3. classifier development using machine learning algorithms.

This research was performed in collaboration with Ford Powertrain Engineering

Research and Development Center (PERDC).

1.4 Thesis Structure

This thesis consists of seven chapters that are organized as follow. Chapter 2 presents

the literature review of FDD and feature extraction methods. Chapter 3 present the theory

behind the feature extraction and classification strategy that is implement in this research.

Chapter 4 presents the experimental set-up, data collection, and data pre-processing

4
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

implementations. Chapter 5 presents the implementations of the feature extraction strategy,

the results, and the observations. Chapter 6 presents the noteworthy findings of the research

and the respective recommendations for future research.

5
Chapter 2: FDD Strategies for Internal Combustion Engines

2.1 Background

In this section, a review of the internal combustion engine and associated FDD

strategies is provided. The objective of this section is not to provide an exhaustive overview

of each topic, but only the important factors that influence their interaction and integration.

2.1.1 Internal Combustion Engine

An internal combustion engine functions as a prime mover for a wide range of

applications by converting chemical energy that are contained in the fuel into useful

mechanical power. The type of internal combustion engine that is most often used in

consumer vehicles is the spark-ignition four stroke engine using the Otto Cycle [3]. The

Otto cycle has four strokes that consist of intake, compression, power and exhaust. These

occur over two full revolutions of the crank shaft [4]. As shown in Figure 1, one rotation

of the crank shaft results in linear motion of the piston from top-dead center (TDC) to

bottom-dead center (BDC) then back to TDC. This repetitive linear motion, which is also

known as reciprocating motion, is achieved by the utilization of a crank that converts

rotational motion of the shaft into linear motion of the piston and vice versa.

6
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 1: Four-stroke cycle of internal combustion engine with piston and crankshaft position [5]

Each stroke in the Otto cycle can be expressed as a thermodynamic process as shown in

Figure 2.

7
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 2: PV diagram of Otto cycle [6]

1. Intake stroke: The intake stroke starts with the piston at top-dead center where the

volume of the chamber is at its minimum and ends with the piston at (BDC) where the

volume of the chamber is at its maximum [3]. As the piston travels down the chamber,

the intake valve opens. During this process, the air-fuel mixture enters the chamber

while the pressure remains relatively constant [3].

2. Compression stroke: The compression stroke begins with the intake valve closing and

the piston traveling up the chamber from BDC. This compression is described to be

adiabatic and reversible which is also known as isentropic [7]. During this compression,

the piston does work on the mixture inside the chamber by decreasing the volume and

increasing the pressure inside the chamber [4].

3. Power stroke: In between the compression and the power strokes, a thermodynamic

process called combustion occurs where the air-fuel mixture is ignited. In the case of a

spark ignition engine, which will be the focus of this research, the mixture is ignited by

using an electrical discharge. During the very rapid combustion process, the volume

stays constant while the pressure and temperature are significantly increased due to the

substantial amount of heat that is released inside the chamber. The power stroke, which

can be described as an isentropic expansion follows the combustion process [7]. During

the power stroke, the piston is pushed from top-dead center to BDC by the gas that is

expanded adiabatically upon the combustion of the air-fuel mixture [3]. In the power

stroke, pressure decreases while the volume increases and the work is said to be done

onto the piston by the gas.

8
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

4. Exhaust stroke: The exhaust stroke is where the residual gas is removed from the

chamber as the piston travels back to top-dead center with the exhaust valve open. The

piston movement decrease the volume while the open exhaust valve maintains a

constant pressure [8].

The analysis of each stroke shows that the power generated from the cylinder does

not occur constantly throughout the cycle instead it occurs in bursts, specifically during the

power stroke. This means that a single cylinder engine would deliver power in a pulsating

form. Therefore, most engines in a vehicle are equipped with multiple cylinders to provide

more consistent power output by staggering the occurrences of the power stroke for each

cylinder. Although this results in more consistent power output, it does not eliminate the

presence of the pulsation. The fluctuations observed in the multi-cylinder engine occurs

more rapidly with smaller amplitude variations. This fluctuation in power delivery is

reflected in the rotation of the crankshaft which is never constant in a given engine cycle.

Heywood describes the combustion process that occurs inside the chamber as a fast-

exothermic gas-phase reaction [4]. During a normal engine operation, the fuel is mixed

together with air and inducted into the chamber through the open intake valve during the

intake stroke. As the air-fuel mixture enters the chamber, it gets mixed with the residual

gas that remains in the chamber after the exhaust stroke. This mixture of fuel, air, and

residual gas from the previous combustion process is then compressed and ignited by an

electric discharge of a spark plug. Upon ignition, a flame develops and propagates through

the chamber until it reaches the chamber walls where the flame is extinguished [4]. Flames

9
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

in a spark-ignition engine are classified as premixed unsteady turbulent flame that

propagate in gaseous state [4]. In addition, flame development and propagation vary cycle-

by-cycle and cylinder-to-cylinder. And these variations in combustion process is caused

by:

1. variations in the mixture motion within the cylinder,

2. variations in the amount of air-fuel mixture in the cylinder, and

3. variations in the ratio of fresh mixture to residual gases in the cylinder [4].

This leads to the fact that no two engine cycles, even for a given cylinder, are identical.

Due to this inconsistent nature of the engine, the combustion may not occur optimally for

a given cycle since most engines are controlled to run at an “optimal” condition that was

configured for an average cycle [4]. Further information regarding internal combustion

engines that has not been covered in this section can be found in Heywood’s text [4].

2.1.2 Fault Detection & Diagnosis

Fault detection is defined as the capability to recognize that a fault has occurred

and fault diagnosis is defined as the ability to locate the fault and the cause of the fault [9].

Such technology was in demand to prevent potentially detrimental failures of machinery,

which are arguably the greatest assets of an industrial company, and to prevent extreme

costs associated with unexpected machinery or facility downtime. FDD technology poses

great benefits as it eliminates the escalation of minor faults through their prompt

identification and initiation of timely mitigation measures. Lately, Fault Detection and

10
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Diagnosis (FDD) technology is being pursued by a wide range of industries that are simply

looking to provide more reliable products and processes.

Fundamental elements of FDD include detection of malfunctions or abnormalities

in the system, isolation and localizing of faults and identification of the fault condition

[10]. Before delving into the different concepts of FDD, a few key terms should be

explained and defined. A fault can be broadly defined as a state where a measured variable

or a calculated parameter within a process or a system is out of its acceptable range [11].

Malfunction of a component can be defined as the root cause for the symptoms exhibited

as a fault and leading to failures [11].

Venkakasubramanian suggest that FDD systems commonly follow a series of

measurement transformations to arrive at a diagnosis [11]. Figure 3 shows three separate

transformations that are involved in determining a fault or failure class given a set of

measurements.

Figure 3: Transformations in FDD systems [11]

The three stage transformations in FDD systems include:

1. Transformation of the input to the overall FDD system called measurements into

features through a process called feature extraction. These methods use a priori

11
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

knowledge to analyze measurements and obtain useful features, which usually

consists of fewer dimensions than measurements [11].

2. Mapping of features to decision through a search or complex learning algorithms

without the use of a priori process knowledge. This transformation is commonly

performed to minimize misclassification of faults [11].

3. Transformations of the decisions into classes, which indicate different, fault

categories including the healthy category. This transformation can be executed

using a threshold function accompanied by symbolic logic discriminant [11].

After reviewing the series of transformations that measurements go through in a

given FDD system, it becomes evident that feature extraction and diagnostic search

strategies are important elements of a FDD system.

When developing a FDD system, there are two different strategies called model-

based and signal-based approaches. As the name suggests, the model-based approach

utilizes a mathematical model that formulates an expected behavior of the system. The

variations in the parameters that results in change of the system behavior is monitored using

observers or filters [12]. On the other hand, signal based strategies rely purely on historical

measurement data. For complex systems that are challenging to model, signal based FDD

may be better suited. This eliminates the need for a mathematical model and the difficulties

and limitations of modeling [1]. Instead of mathematical models, signal based approaches

attempt to find patterns in the historical measurements. Numerous techniques such as

12
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

artificial neural network algorithms and statistical methods are utilized to obtain patterns

that is assumed to exist in the data [10].

2.2 Fault Detection & Diagnosis for Internal Combustion Engines

FDD strategies for internal combustion engines have been explored extensively in

the past. Earlier FDD strategies for internal combustion engines were motivated by the

need for improved productivity and quality in production such presented by Miller and

Tjong [13], [14]. These researchers envisioned a system that is capable of detecting

manufacturing and assembly defects through an online engine monitoring and diagnostic

system [13], [14]. Utilization of such monitoring and diagnostic systems were intended to

be integrated in the manufacturing line through a process that was later popularized as cold

tests and hot tests [13]. During these tests, the engine is put on a test stand where it is

operated in cold or hot modes while measurements are collected at very high sampling

rates [15]. Cold test refers to the operation of the engine without combustion and hot test

refers to the operation of the engine with combustion. The obtained measurements are

analyzed to determine whether or not the engine should proceed to its next manufacturing

process [16]. Due to its simplicity and low costs, more research efforts have been found in

the development and improvement of cold tests. Following Miller and Tjong, a variety of

measurement analysis methods were studied to improve the performance of the cold test

procedures [15], [17], [18], [16], [19], [20].

Currently, automation companies are providing fully automated FDD technologies

that can be used in the engine manufacturing line. These strategies involve running the

13
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

engine in different modes such as cold testing, hot testing with load, and hot testing without

load while obtaining measurements such as speed, torque, and pressure. Although it is not

clear as to how the measurements are analyzed, it is noted that a variety of defects such as

missing piston rings, leaky head gaskets, leaky intake and exhaust valves along with many

other defects can be detected using this FDD strategy [21].

More recently, FDD algorithms for in-vehicle applications were explored, where

with the appropriate hardware, online monitoring of the engine can be achieved within the

vehicle [22]. Currently, vehicles are equipped with On-Board Diagnostic (OBD) systems

that monitor the engine closely with a variety of measurements. However, as concepts of

autonomous and connected vehicles are becoming closer and closer to reality, it is

becoming increasingly important for the automotive industry to broaden the diagnostics

for a wider range of components in the vehicle [22], [23]. With ever increasing computation

power, more sophisticated methods such as wavelet analysis, Principal component analysis

and artificial neural networks (ANN) are becoming tangible for automotive applications

[2], [20], [22], [24], [25], [26], [27].

2.2.1 FDD Approach

In the literature, the widely accepted FDD approach for internal combustion

engines is the signal-based approach. Model-based strategies have been used but with very

limited scope. Wei applied model-based FDD to diesel engines by modeling the in-cylinder

pressure and torque produced by each cylinder in crank angle domain based on

thermodynamic and dynamics theory [17]. However, in Wei’s research, combustive

14
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

behaviors were not within the scope of the model as it was intended for the cold test

application [17]. Chandroth confirms that signal-based strategies are preferred over the

model-based strategies as it is very difficult to accurately model an internal combustion

engine, especially the combustive process, given its turbulent and variable nature that was

discussed in the previous section [27].

2.2.2 Measurements for FDD

The selection of measurements is very important for signal-based strategies as the

correlations that are assumed to exist in the measurements selected act as the a priori

knowledge. The type of the measurement such as speed, pressure and/or voltage, the sensor

used to obtain the selected type of measurement and the location of the measurement are

example of all the decisions that must be made in selecting the measurements for a signal-

based strategy. In the following section, the types of measurements that should be

considered for the development of FDD for internal combustion engines are discussed.

Vibration Signals

Sound and vibration measurements are commonly used in FDD systems for

mechanical systems [28], [29], [30], [31] . Vibration is inherent in any mechanical system

that consists of moving components [32]. Every system has its own specific vibration

modes that can be observed during its normal operating conditions. These modes are

dependent on system parameters such as mass, system stiffness, fitting tolerances and

frictions. When changes occur in these system parameters due to a faulty component,

observable vibration behaviors may deviate from its normal signature. In addition,

15
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

vibration signals can be obtained through non-intrusive methods that are relatively cheap

and easy to install. In the literature, a wide range of vibration analysis for FDD of rotating

machines are present [1], [28], [31], [33], [34], [35], [36].

In comparison to vibrations of rotating machines, vibrations of internal combustion

engines are very complex. Vibration of an engine may arise from a variety of components

such as crank train, valve train, piston assembly, fuel system and exhaust system [37]. In

addition, the turbulent combustion process contributes to the overall vibration of the

engine, which further complicates the task analyzing the vibration. Tjong categorized three

main types of faults that may contribute uniquely to vibration:

1. Impulse impact at a specific crank angle due to incorrect clearance between

components such as valve train, gears, connecting rod, pistons and bearing;

2. Imbalance in rotating components such as crankshaft, camshaft, balance shaft

and gears; and

3. Abnormal combustion [13].

In addition to the complexity that arise from having numerous components,

vibration analysis of an engine is especially difficult as vibration contributed by these

components are exhibited in different forms. For example, the engine is comprised of

components that function in reciprocating motion and rotating motion. Vibration due to

rotating components are exhibited harmonically whereas vibrations due to reciprocating

motion are exhibited in impulses in the time domain or relative to the crank position.

16
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Furthermore, engine vibrations also comprise of vibration components that are induced by

combustion which are often described as stochastic.

Pressure Signals

In the literature, the in-cylinder pressure of an engine is shown to provide a great

amount of information on the quality of the combustion in a cycle by cycle basis [4].

Researchers such as Sharkey had performed FDD studies using vibration and in-cylinder

pressure signal trace to detect faults such as leaking intake and exhaust valves [30]. Miller,

Tjong, Jones and Chandroth presented valuable studies that exhibit great potential for

accurate FDD strategies using pressure traces. However, it should be noted that obtaining

in-cylinder pressure traces require retrofitting of engine to allow for installation of very

expensive pressure transducers.

2.3 Feature Extraction Methods

This section presents a review of a variety of feature extraction methods that are

commonly utilized for signal based FDD strategies involving vibrational measurements.

2.3.1 Time Domain Analysis

Time domain analysis is the simplest and cheapest approach to analyzing signals

for FDD strategies as measurements are often obtained in time domain. These time series

signals are usually referred to as raw signal. A variety of statistical indicators such as crest

factor, kurtosis root mean square and standard deviation have been applied to FDD

applications for detecting bearing damages [33].

17
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Time domain analysis has been performed on diesel engines for cold test

applications, which were aimed at determining the reliable threshold values for pass or fail

decision for the cold test given vibrational measurements with very low computational

costs [15]. The research concluded that kurtosis coefficients and the RMS value of the

vibration signal in time domain was successful at characterizing and detecting faults such

as inverted piston, overpressure valve, out of housing exhaust equalizer, improperly

tightened oil pump screw and improperly tightened oil jet during cold test operating

conditions. Cold test operating conditions refer to conditions at which the engine is

operated without combustion. However, time domain analysis is not the most ideal

vibration analysis method for FDD strategies for engines at normal operating conditions,

which includes combustion, as time domain measurements of engines are very difficult to

understand and analysis in time domain yield information in reference to time, which is not

very meaningful for the application of FDD on ICE. Time referenced information is not

very valuable as engines operate in various different rotating speeds, which would indicate

that a fault condition analyzed at one speed would not be able to correlate its learned

features at another speed if time domain analysis is applied.

2.3.2 Crank Angle Domain Analysis

As mentioned in the previous section, time domain is not a good reference domain

for reciprocating machinery such as internal combustion engine. Instead, measurements

obtained in reference to the angle of the crank shaft is widely used. Miller, Tjong and

Delvecchio applied statistical indicators from time domain analysis on crank angle domain

[13, 15, 14]. Analysis in crank angle domain allows for occurrences of abnormal behaviors
18
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

of a signal to be trace to a specific location in an engine cycle which gives insight into to a

range of components that may have induced such abnormal behavior. This is because

movement of a certain component within the engine is directly related to the position of

the crank shaft. Sophisticated methods of converting time domain signal to crank angle

domain are explored by Arasaratnam [38]. It is important to note that conversion methods

cannot assumed that the crankshaft of the internal combustion engines has a constant

rotational speed. If such assumptions were to be made during the conversion process, the

reference of the measurements with respect to specific crank angles will be skewed

significantly as the rotational speed of an internal combustion engine is not constant. In

addition, analysis in crank angle domain eliminates the need for FDD strategy to operate

only under the same operating speeds that it was previously trained on. This characteristic

is especially important for the development of in-vehicle condition monitoring systems for

internal combustion engines.

2.3.3 Frequency Domain Analysis

Frequency domain analysis is a reversible transformation that decomposes the

given time-domain signal into their frequency components. These frequency components

can be used as features for a given condition. Frequency analysis are most appropriate for

stationary signals that are periodic as time information is not available in Fourier transform.

In the literature, frequency analysis such as Fourier transform has been used for

FDD of machinery [39], [40], [41]. However, it is often concluded that the event of interest

must be periodic for frequency analysis to be effective [42]. Therefore, frequency domain

19
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

analysis is not an ideal method to be applied for internal combustion engines where signals

are time varying. If internal combustion engines were stationary signals such as in the case

of motoring, frequency domain analysis may be able to detect the presence of a fault and

narrow down the list of potential culprits. However, if multiples of a component that share

the same frequency component exist, the features obtained from frequency domain analysis

will not be able to tell which of those components are faulty. This is because the frequency

analysis provides information on which frequency appears in the signal but not when the

frequency appears in the signal [43].

Figure 4: Resampling for order analysis [44]

2.3.4 Order Analysis

Order analysis is a form of frequency domain analysis that is very specific to

rotating and reciprocating machineries such as engines, turbines and pumps [45]. Instead

20
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

of performing FFT on time-domain signals, order analysis describes a method where FFT

is performed on crank angle domain signals. This allows frequency components of non-

stationary signals to be extracted in reference to the primary rotating component, which in

the case of engines would be the crankshaft. Order analysis provide behavior of harmonic

orders of the crankshaft speed with each order corresponding to the multiple of the

crankshaft’s rotational speed. This method is especially useful for an application where a

wide range of speeds are being observed. However, like frequency domain analysis, order

analysis is unable to provide information that will aid in identification of a faulty

component unless the harmonic order associated to the faulty component is unique in the

whole system.

2.3.5 Time-Frequency Analysis

Retrieving the frequency components and the occurrences of those components for

a given signal is essential for FDD of rotating and reciprocating machines or for any

application that requires the analysis of a non-stationary or transient signal. However, time-

domain analysis is only able to provide information on when a component of the signal

occurred with respect to time, and frequency analysis is only able to provide information

on the frequency components that the signal consists of. To mitigate this constraint and

obtain time localization, a variety of approaches that are known as time-frequency analysis

was introduced. These approaches include but are not limited to Short-Time-Fourier-

Transform, Wavelet Transform, Wigner Transform, and Hilbert-Huang Transform. These

approaches map a one-dimensional time-domain signal into a two-dimensional function of

21
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

time and frequency. In this literature review, Short Time Fourier Transform (STFT) and

Wavelet Transform will be discussed in detail.

2.3.5.1 Short Time Fourier Transform

Short-Time Fourier Transform is a time-frequency analysis method that results

from a compromise between time and frequency analysis. STFT achieves the two-

dimensional representation of a transient or non-stationary signal in time and frequency

through the utilization of windowing technique and Fourier Transform [45]. The

windowing technique divides the given signal into short equal time-framed segments,

which is individually analyzed using Fourier transform. This windowing technique is

applied with the assumption that signals can be considered to be stationary if they are

segmented into small enough segments. Figure 5 illustrates the basics of the windowing

technique. As shown in Figure 5, the windowing function is first located at time, t = 0, then

the window is shifted throughout the signal. The most notable characteristic of STFT is

that its window size is kept constant throughout the whole transformation.

Figure 5: Signal windowing technique [12]

22
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

The STFT is defined as:

𝑆𝑇𝐹𝑇(𝜏, 𝜔) = 𝑋(𝜏, 𝜔) = ∑ 𝑥(𝑡)𝑤(𝑡 − 𝜏)𝑒 −𝑖𝜔𝑡 (2.1)

𝑡= −∞

where 𝜏 is the time, 𝜔 is the frequency, 𝑡 is the offset, 𝑥(𝑡) is the signal to be transformed,

and 𝑤(𝑡) is the windowing function. And the energy density of the signal at a given time

and frequency can be described as:

𝑆𝑝𝑒𝑐𝑡𝑜𝑔𝑟𝑎𝑚(𝜏, 𝜔) = |𝑋(𝜏, 𝜔)|2 (2.2)

Figure 6: Spectrogram and spectrum of a non-stationary signal [12]

23
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

As shown in Figure 6, STFT is capable of simultaneously capturing both the

frequency components and its occurrences given a non-stationary signal. However, the

compromise between time and frequency domain analysis has its limitation, which is

described by the Heisenberg uncertainty Principal. It infers that the increase in resolutions

cannot be achieved concurrently due to the inability to know the exact time frequency

representation of a signal. Figure 7 and Figure 8 depict the two extreme cases of high

frequency and high time resolution. It can be observed that when the resolution of one

domain is increased, the resolution of the other decreases. For example, when the size of

the window is decreased the time resolution improves while frequency resolution decreases

and vice versa when the size of the window is increased. This phenomenon is often referred

to as the resolution problem.

24
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 7: High frequency resolution [12]

A FDD strategy using STFT was applied to bearings by Yazici, where it was

reported that high accuracy was achieved [46]. However, the application of STFT on

engine measurements may not be enough for the development of an effective FDD strategy

for engines. Localization of frequency components in the crank angle is an essential

characteristic of the feature extraction methods for FDD of engines due to their non-

stationary and transient behavior. STFT has been applied to the analysis of engine vibration

by Chen, where different vibration sources are attempted to be identified and characterized.

The research concluded that STFT is capable of identifying events such as strong valve

impact and piston slap but unable to separate closely-overlapping or weak events such as

valve closures of different cylinders. This finding supports the prior judgment that STFT

is simply not sophisticated enough for the objective of the FDD strategy that is being

explored in this research [50]. Due to the fixed resolution in time and frequency in STFT,

other Time-Frequency techniques such as wavelet analysis are suggested in the literature

for the analysis of engine vibrations [47]. [48], [49].

25
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 8: High time resolution [12]

2.3.5.2 Wavelet Transform

Wavelet transform is another form of time-frequency analysis that has been

developed as an alternative to STFT that enables analysis of signals consisting of transient,

non-stationary and time-varying characteristics. Unlike STFT, where the resolution of the

signal is fixed in both time and frequency domains throughout the whole transformation of

the signal, wavelet utilizes different window sizes to develop a multi-resolution analysis

(MRA) in the time and frequency domain of a signal. In this section, the fundamental

review of continuous wavelet transform (CWT), discrete wavelet transform (DWT), and

wavelet packet transform (WPT) will be discussed.

26
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

The word wavelet originates from the equivalent French word ondelette, which

translates to small wave. It is defined as continuous functions that decay to zero at both

ends at different rates, which provide compact support in comparison to continuous

sinusoids. The rate of decay is altered by the dilation or compression of the wavelet. The

wavelet is then translated to capture localized events along with the time domain of the

signal that is to be analyzed. These characteristics allow wavelet transform to highlight

details of the signal being analyzed, which other forms of analysis may not be able to.

Therefore, the ability to detect such minor details of a signal that is very favorable in

condition monitoring and FDD application.

The method is established by very simple intuition regarding frequency

components. It is known that low frequencies can be observed throughout longer time

periods and that higher frequencies can be observed in very short time periods. With this

intuition, it was suggested that a large window size be used to obtain information regarding

low-frequency components and very small window sizes to obtain information regarding

high-frequency components. The ability to capture both the time and frequency

components of a signal with high resolution makes wavelet transform a promising method

for analyzing transient signals that consist of both high and low-frequency components. It

is important to note that the Heisenberg uncertainty Principal still applies to wavelets,

however, wavelet transform is a very clever work around method. The resolution problem

is mitigated by prioritizing the resolution that is more important at a given frequency range.

27
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

The difference between STFT and wavelet analysis is best presented in Figure 9.

The wavelet transform’s multiresolution property is illustrated by the different sizes of the

boxes. The boxes indicate the frequency and time resolution at which each transform can

analyze the signal. As illustrated in Figure 9, time and frequency domain analysis are only

able to provide information on the signal at their own domain. This is indicated by their

rectangular shape that has a fixed width in the frequency domain for time-domain analysis

and time domain for frequency domain analysis. STFT is shown to have a square shape

with the same width for time and frequency. This infers that STFT has a fixed resolution

in both time and frequency throughout the entire frequency ranges. On the other hand, the

wavelet transform has both rectangular and square shapes in different sizes. Every box in

the wavelet transform has the same area due to the Heisenberg uncertainty Principal;

however, the widths of both frequency and time are altered, providing different proportions

of resolutions at different frequency levels. For example, at low frequencies, the boxes

have larger heights with thinner widths, which indicates that it has better frequency

resolution with lower time resolution. And at high frequencies, the boxes are shorter and

wider, indicating that it has lower frequency resolution but higher time resolution.

28
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 9: Frequency and Time Resolution of analysis in different domains [46]

Continuous Wavelet Transform

Similar to STFT, wavelet transform multiplies the signal with the wavelet function,

which acts as STFT’s window function, at different time segments in the signal. The

continuous wavelet transform and the inverse continuous wavelet transform are described

in Equation 2.3 and 2.4, respectively.

∞
1 𝑡−𝑏
𝑋𝑤 (𝑎, 𝑏) = 1⁄∫ 𝑥(𝑡)𝜓 ( ) 𝑑𝑡 , 𝑎 𝜖ℝ + 𝑎𝑛𝑑 𝑏 𝜖ℝ (2.3)
|𝑎| 2 −∞ 𝑎

∞
1 𝑑𝑎 𝑑𝑏
𝑥(𝑡) = ∬ 𝑋𝑤 (𝑎, 𝑏) 𝜓𝑎,𝑏 (𝑡) 2 (2.4)
𝐶𝜓 −∞ 𝑎

29
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

where a is the scale factor, b is the translational value and 𝜓(𝑡) is the mother wavelet where

children wavelets are derived from given the following equation.

1 𝑡−𝑏 (2.4)
𝜓𝑎,𝑏 (𝑡) = 1 𝜓( )
|𝑎| ⁄2 𝑎

Wavelet analysis is a process where different scaled and time-shifted versions of the mother

wavelets are compared with the signal that is being analyzed to measure the similarity

between the signal and the wavelet function. The scale parameter corresponds to frequency

resolution of the analysis where higher scale provides more global behaviors of the signal

and lower scale provide more detailed information that may only last for a short time span.

Scaling parameter achieved varied frequency resolution analysis through dilation and

compression of the wavelet function [47].

Unlike Fourier transform, the wavelet functions are localized in both time and

frequency [47]. Due to this localization, the wavelet functions must be translated

throughout the time duration of the signal to ensure characterization of the entire signal.

Similarly, wavelets must be scaled in time and shifted in frequency domain due to the

localization of wavelets in frequency. For these reasons, the continuous wavelet is

described as a process where the signal being analyzed is multiplied with analysis windows

that are changed in scale and shifted in time then integrated over time. The square root of

the scaling factor is an energy normalization term which ensures that wavelets of different

scales have the same amount of energy.

30
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

All mother wavelets and its offspring have two main characteristics that are very

important for normalizing the transformation. First characteristic is described as,

∞
∫ 𝜓𝑎,𝑏 (𝑡) = 0 (2.5)
−∞

which basically states that all wavelets must have zero average. Second characteristic is

described as,

‖𝜓‖ = 1 (2.6)

which states that their energy must equal to unity. The ability to obtain the transformation

inverse is very useful in Fourier transform. In wavelet transform, dual wavelet function

which is a synthesis function for a given wavelet function is utilized to perform the inverse

wavelet transformation. To ensure that inverse wavelet transformations are possible,

mother wavelets are chosen under admissibility condition which is described in the

following equations.

∞
|𝜓(𝜔)|
𝐶𝜓 = ∫ 𝑑𝜔 < ∞ (2.7)
−∞ 𝜔

The admissibility condition requires that 𝜓(𝜔) is differentiable. There are numerous

amounts of wavelets that have been developed for different applications to analyze

transient, time-variant, and/or non-stationary characteristics. The selection of wavelet is

very important for the success of the application. Although the concept of wavelet

transform is very powerful, application of wavelet transform in a continuous form is

impractical on digitized computers. In the following section, a discrete form of wavelet

that can be applied on computers is introduced.

31
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Discrete Wavelet Transform

Discrete wavelet transform is implemented using multi-resolution analysis method

to reduce computation time through elimination of redundancy while still ensuring that an

appropriate amount of information is provided for analysis and synthesis of the signal.

Multi-resolution analysis method suggests the decomposition of the input signal into two

different signals using two channel filter bank and down sampling of the filtered signal. In

order to analyze the signal at different scales, the cut-off frequencies of the filter are altered.

The signal is essentially being decomposed into different frequency bins using a series of

half band high and low pass filters. This process is illustrated in Figure 10 where g[n] and

h[n] are filters which are associated with the mother wavelet and scaling function [48]. The

following equation describes the relation between the two filters.

𝑔[𝑁 − 1 − 𝑛] = (−1)𝑛 ∙ ℎ[𝑛] (2.8)

The filtering of signals is described by a convolution operation of the signal and the impulse

response of the filter, as described in the following equation.

𝑥[𝑛] ∗ ℎ[𝑛] = ∑ 𝑥[𝑘]ℎ[𝑛 − 𝑘] (2.9)

𝑘→ −∞

The low pass filter removes all the frequencies that are higher than the half of the full

frequency range of the signal and the result is known as approximation. The high pass filter

removes all the frequencies that are lower than the half of the full frequency range of the

signal and the result is known as detail. Through this process, half of the information from

the original signal is lost which results in the reduction of the resolution in half. The

32
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

filtering operation varies the frequency of the signal. The down-sampling process upon

filtering alters the scale through the reduction of the sampling rate, which is achieved

through the removal of samples. Down-sampling process eliminates redundancies that

exist in the filtered signal according to the Nyquist sampling theorem and doubles the scale.

The following equations describes how approximate and detail representations are

obtained.

𝑐𝐴 = 𝑥𝑙𝑜𝑤 [𝑘] = ∑ 𝑥[𝑛] ∗ ℎ[2𝑘 − 𝑛] (2.10)

𝑛
𝑐𝐷 = 𝑥ℎ𝑖𝑔ℎ [𝑘] = ∑ 𝑥[𝑛] ∗ 𝑔[2𝑘 − 𝑛] (2.11)
𝑛

This process of filtering and downsampling to obtain detailed and approximate

representations of a given signal can be performed iteratively to break down the original

signal into smaller frequency ranges. The decomposition of a signal reduces the time

resolution in half as the number of sample has been reduced by half. On the other hand, the

decomposition of a signal doubles the frequency resolution as the frequency band of a

signal spans only half of the original frequency band [47].

Figure 10: Discrete wavelet transform filter bank [12]

33
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Reconstruction of the original signal can be obtained using Equation 2.12, where 𝑦ℎ𝑖𝑔ℎ and

𝑦𝑙𝑜𝑤 are given by Equations 2.12a and 2.12b. The summary of the decomposition and

synthesis of filter bank is shown in Figure 11.

𝑥[𝑛] = ∑(𝑦ℎ𝑖𝑔ℎ [𝑘] ∙ 𝑔[−𝑛 + 2𝑘]) + (𝑦𝑙𝑜𝑤 [𝑘] ∙ ℎ[−𝑛 + 2𝑘]) (2.12)

𝑦ℎ𝑖𝑔ℎ [𝑘] = ∑ 𝑥[𝑛] ∙ 𝑔[−𝑛 + 2𝑘] (2.12a)

𝑛
𝑦𝑙𝑜𝑤 [𝑘] = ∑ 𝑥[𝑛] ∙ ℎ[−𝑛 + 2𝑘] (2.12b)
𝑛

Figure 11: DWT decomposition and synthesis overview [49]

34
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 12: Wavelet Packet Transform filter banks [12]

Wavelet Packet Transform

Wavelet packet transform is very similar to the discrete wavelet transform in that

signal is transformed by passing it through a series of high-pass and low-pass filters.

However, the wavelet packet transform differs from the discrete wavelet transform as it

passes both approximation and detail coefficients through low and high pass filters. Figure

12 shows the wavelet packet decomposition. The main benefit of wavelet packet transform

is that the entire frequency spectrum is highlighted equally unlike DWT, which mainly

highlights the high frequency components. The difference in distribution of the frequency

bands between DWT and WPT can be observed in Figure 13.

35
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 13: Frequency distributions of DWT and WPT [22]

2.3.6 Principal Component Analysis

Measurements from multiple sensors, either from the same type or different types

of sensors, are required as the complexity of the system that is being observed is increased.

For fault detection and diagnosis applications, behaviors that relate to faults must be

extracted from these multi-sensory measurements using multivariate analysis. Multivariate

36
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

analysis refers to statistical methods in which more than one statistical variable is observed

and analyzed. Multivariate analysis aims to obtain relationships that exist among the

different variables which effect the system’s behavior and to obtain dominant variables that

contribute the most to the system’s behavior through mathematical conversion of

measurements into statistical indices. One of the key factors that must be accounted for is

the amount of cross-correlation that exists in the entire dataset. The level of cross-

correlation would indicate that, although many different measurements were obtained, very

little information can be extracted from it. Depending on the situation, an additional

measurement may reach a point of diminishing return due to the lack of uniqueness that

exist in the supplementary measurement. Principal component analysis is an example of a

popular statistical method for multivariate analysis in the field of engineering including

fault detection and diagnosis.

Additionally, Principal Component Analysis (PCA) assumes existence of

correlations among a set of multiple variables that are provided for a given system. With

that assumption, PCA aims to convert the dataset into principal components using

orthogonal transformation. In other words, a dataset with many variables that are assumed

to be correlated are transformed into a new space composed of principal components,

which are a set of linearly uncorrelated orthogonal axes. The contribution of each principal

component on the system’s behavior is dictated by the variance that exists in the data along

the principal component axes. A given principal component’s contributions to a system’s

behavior is high if large variance is observed along its axes. For example, the first principal

component has the largest variance of the dataset, the second principal component has the

37
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

largest variance of the dataset given the constraint that it must be orthogonal to the first

component and so on. This concept is illustrated in Figure 14.

Figure 14: Principal component analysis [50]

During this transformation the number of variables is significantly reduced by

eliminating the data points that dependent on principal components with very low variances

which are considered to have very little contribution to the system’s behavior. This

reduction in data set is often referred to as dimensionality reduction. In addition, PCA

performs statistical analysis on the compressed dataset to understand the dynamics of the

system. To summarize, PCA is a simple and non-parametric form of feature extraction

method that extracts the dominant dynamics that exist in the complex dataset, which is

38
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

very applicable in the field of fault detection and diagnosis.

𝑇 = 𝑋𝑃 (2.12)

PCA transform is described by Equation 2.12, where 𝑋 is the data matrix of size

𝑛 𝑥 𝑚 , 𝑇 is the principal component scores matrix of size 𝑛 𝑥 𝑚 , 𝑃 is the principal

component loadings matrix of size 𝑚 𝑥 𝑚, 𝑛 is the number of measurements, and 𝑚 is the

number of variables. Principal component loadings represent the basis vectors of the new

principal component dimension. Principal component scores correspond to the

uncorrelated signal representation of the dataset and is organized from highest to low

variance components. This indicates that first couple of columns of the principal

components scores hold the most information regarding the data set when compared to the

last couple of principal components. This implies that PCA can identify which

measurement contains the fault signature of interest for the application.

Every column of P represents a basis vector in the new principal component dimension.

There are two conditions that must be met by the transformation matrix P. These conditions

states that the basis vectors must be orthogonal and that the magnitude of variance

influences the principal direction [24]. In addition, the principal component scores matrix

must be a diagonal. A diagonal matrix is defined as a matrix whose entries outside of the

main diagonal equals to zero. As described by Equation 2.12, the aim of PCA is to

transform the data that are described by correlated variables into data that are described by

uncorrelated variables. In other words, the main purpose of PCA is to remove cross

correlation that exist in the raw dataset. Mathematical, this would translate to obtaining the

orthogonal transform matrix P that maps the raw dataset into transformed diagonal
39
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

covariance matrix [24]. The derivation of the solution is introduced in the following series

of equations with the calculation of covariance of the transformation matrix 𝑇, where 𝐴 =

𝑋 𝑇 𝑋 is a square 𝑚 𝑥 𝑚 symmetric matrix, which can be factorized using the spectral

decomposition of a matrix formulation as shown in Equation 2.14.

1
∑𝑇 = 𝑇𝑇𝑇
𝑛−1

1
∑𝑇 = (𝑋𝑃)𝑇 𝑋𝑃
𝑛−1

1
∑𝑇 = 𝑃𝑇 𝑋 𝑇 𝑋𝑃
𝑛−1

1
∑𝑇 = 𝑃𝑇 (𝑋 𝑇 𝑋)𝑃
𝑛−1

1 (2.13)
∑𝑇 = 𝑃𝑇 𝐴𝑃
𝑛−1

𝐴 = 𝑉Λ𝑉 −1 (2.14)

Matrix 𝑉 consists of eigenvectors of 𝐴 and Λ is a diagonal matrix which consists of

eigenvalues that correspond to eigenvectors in 𝑉 [24]. Due to the symmetric nature of

matrix 𝐴, orthogonality of its eigenvectors can be inferred. A matrix property which states

that a transpose of an orthogonal matrix is equal to its inverse is applied to Equation 2.14

to arrive at Equation 2.15. When Equation 2.15 is substituted into Equations 2.13, Equation

2.16 is obtained and 𝑃 is set to 𝑉 to arrive at a diagonalized covariance of the

transformation matrix 𝑇. Equation 2.17 shows the final form of 𝛴𝑇.

40
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

𝐴 = 𝑉Λ𝑉 𝑇 (2.15)

1
∑𝑇 = 𝑃𝑇 (𝑉𝛬𝑉 𝑇 )𝑃 (2.16)
𝑛−1
1
∑𝑇 = 𝛬 (2.17)
𝑛−1

A factorization method of matrices called Singular Value Decomposition (SVD) theorem

is utilized in obtaining the eigenvectors to reduce exposure to numerical errors. Equation

2.18 describes the SVD theorem, where 𝑋 is any arbitrary matrix with size 𝑛 𝑥 𝑚, 𝑈 is a

unitary matrix of size 𝑛 𝑥 𝑛 called left singular matrix, 𝑉 is a unitary matrix of size 𝑚 𝑥 𝑚

called right singular matrix, and 𝛴 is a rectangular diagonal matrix of size 𝑛 𝑥 𝑚 called

singular values matrix.

𝑋 = 𝑈𝛴𝑉 𝑇 (2.18)

SVD is applied to calculate PCA as shown in Equation 2.19 given that 𝑈 is a unitary matrix

and 𝑈 𝑇 𝑈 = 𝐼. The relation provided in Equation 2.20 relates the SVD and eigenvalue

deposition which infers that transformation matrix P is equal to the right singular matrix V

since 𝑃 = V.

𝑋 𝑇 𝑋 = 𝑉Σ 𝑇 𝑈 𝑇 𝑈𝛴𝑉 𝑇

= 𝑉Σ 𝑇 𝛴𝑉 𝑇 (2.19)

Σ = 𝑋𝑇 𝑋 (2.20)

The application of SVD ensures that columns of 𝑉 , the eigenvectors and the

corresponding elements of Σ, the eigen values are organized in the descending order which

means that the first principal components will have the highest variance. The small

41
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

eigenvalues are often eliminated in practice as these components of the measurement have

small signal-noise-ratio. Therefore, the elimination of negligible eigenvalues may reduce

noise or error that exist in the dataset.

To summarize the procedure of PCA transformation is provided in the following:

1. Transformation of the given data set to the principal component dimension by

calculating the principal loadings, scores and eigenvalues;

2. Setting of a minimum threshold for the eigenvalues;

3. Elimination of the corresponding loadings and scores of those eigenvalues that

do not meet the minimum threshold;

4. Reconstruction of the remaining matrix to original dimension using the inverse

PCA transform described in Equation 2.21, where 𝑇̂ and 𝑃̂ are the remaining

scores and loadings upon elimination.

𝑋̂ = 𝑇̂𝑃̂𝑇 (2.21)

The method of defining the threshold limit is a field of its own that is being

researched extensively as it is crucial for the performance of PCA. Heuristic rule and

Kaiser’s rule are two most common methods of thresholding.

The heuristic rule states that components that consists of 95% of the total variance

should be considered which infers that any principal component with an eigenvalue higher

than 0.05 should remain. The Kaiser’s rule states that eigenvalues that are greater than the

average eigenvalue must be considered.

42
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

In addition to thresholding methods, different pre-processing methods available for

the application of PCA should be considered. Non-centered PCA is the method of PCA

application with no prior processing of the raw dataset. Centered or covariance PCA is

method where the mean value of each variable is subtracted from each instance. And

correlation PCA is a method where dataset is scaled and normalized to have zero mean and

unit standard deviation by subtracting the mean value and dividing it by the variable’s

standard deviation. For applications where the variables are in different units or where the

variances differ significantly correlation PCA is often utilized.

PCA has been widely utilized in signal based FDD applications due to its ability to

quantify the correlation between the components of a multivariate dataset and the system’s

behavior. This quantification, which provides measure of the covariance, allows for the

ability to highlight the components that have the most significant influence on the system’s

behavior without having in depth technical knowledge or a mathematical model of the

system. PCA technique is especially useful in signal based FDD approaches where system

is complex, and dataset is composed of high dimensional, noisy and correlated data [51].

For the application of PCA on FDD, Cherry and Qin suggests decomposing the

multidimensional data collected for a given system into two smaller sets as shown in

Equation 2.21 [52]. 𝑋̂ represents the main features that consists of high varying principal

components and 𝑋̃ represents the residuals that consists of low varying principal

components. Equations 2.22 shows how the principal component scores and loading

43
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

matrixes are decomposed in a similar manner. The method of decomposition is performed

through a thresholding technique.

𝑋 = 𝑋̂ + 𝑋̃ (2.21)

𝑋 = 𝑇̂𝑃̂ + 𝑇̃𝑃̃ (2.22)

= [𝑇̂ 𝑇̃][𝑃̂ 𝑃̃ ]

= 𝑇𝑃𝑇

The thresholding is applied to the variance of the entire dataset by setting a criterion

on the variance, where meeting the criterion means that the principal component in

question belongs to 𝑋̂ subset and inability to meet the criterion means that the principal

component in question belongs to 𝑋̃ subset. If 𝑝 is the total number of principal

components, 𝑧 number of principal components belong to the 𝑋̂, feature subset while 𝑝 −

𝑧 number of principal components belong to the 𝑋̃ , residual subset. The selection of

thresholding techniques that differ in the criterion that it uses to decompose the dataset into

main features and residual subsets. Most popular methods of thresholding are Heuristic

rule, Kaiser rule and manual selection. Heuristic rule categorizes the principal components

that collectively represent 95% of the variance of all principal components into the feature

subset and the remaining into the residual subset. Kaiser rule categorizes the principal

components that have higher variance than the average variance of all the principal

components into the feature subset and the remaining into the residual subset.

44
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Upon decomposition of the dataset into features and residuals subsets, the

information that are embedded in the principal components must be expressed in terms of

statistical indices to proceed with fault detection and diagnosis methods. Indices

popularized by Qin consist of Hoteling’s 𝒯 2 , Squared Prediction Error (SPE) 𝒬 , and

combined index 𝜑. These indices provide a measure of how different a given measurement

is compared to the defined baseline measurement. Therefore, defining the baseline

measurement, which can be described as a healthy measurement, is a very important

component in executing a FDD strategy using PCA.

Hoteling’s index is calculated using Equation 2.23, where 𝑥𝑛𝑒𝑤 is the new

̂ is 𝒯 2 projecting matrix. 𝐷
measurements, and 𝐷 ̂ is
̂ is defined in Equation 2.24, where Λ

defined in Equations 2.25 and 𝑛 is the size of the measurements. Hoteling’s index

calculates the distance between the feature subset of the baseline and the new measurement

in the baseline’s Principal component feature space.

̂−1 𝑃̂𝑡 𝑥𝑛𝑒𝑤

𝒯 2 = 𝑥𝑛𝑒𝑤 𝑃̂Λ 𝑡

̂ 𝑥𝑛𝑒𝑤
= 𝑥𝑛𝑒𝑤 𝐷 𝑡 (2.23)

𝐷 ̂−1 𝑃̂𝑡
̂ = 𝑃̂ Λ (2.24)

1 (2.25)
̂=
Λ 𝑇̂𝑇̂ 𝑡
𝑛−1

SPE index is calculated using Equation 2.26, where 𝐶̃ is the residual subspace that

the new measurement is projected on. 𝐶̃ is defined in Equation 2.27. SPE index calculates

45
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

the distance between the new measurement and the residuals subset of the baseline

measurement in the residual subspace.

2 (2.26)
𝑆𝑃𝐸 = 𝒬 = ‖𝑥̃𝑛𝑒𝑤 ‖2 = ‖𝑥̃𝑛𝑒𝑤 𝐶̃ ‖

𝐶̃ = 𝑃̃𝑃̃𝑡 (2.27)

The combined index introduced by Yue and Qin is obtained using Equation 2.28

and 2.29, where 𝛿 is the upper limit of 𝒬, 𝜏 is the upper limit of 𝒯 2 , and Φ is the projecting

matrix. Equations 2.30 defines Φ. As shown in Equation 2.28, the combined index is a

uniform combination of the previous indices.

𝒬 𝒯2
𝜑= 2+ 2 (2.28)
𝛿 𝜏
𝑡
𝜑 = 𝑥𝑛𝑒𝑤 Φ𝑥𝑛𝑒𝑤 (2.29)

𝑃̃𝑃̃𝑡 𝑃̂𝜆̂−1 𝑃̂𝑡

Φ= + (2.30)
𝛿2 𝛿

The calculated indices are compared to an upper limit where exiting the limit

indicates the presence of a fault condition in the system. Upper control limit of index 𝒯 2 ,

which is proportional to the F-distribution, is calculated using Equation 2.31 or 2.32 [53],

[18]. ∝ is the assigned confidence level, 𝑛 is the sample size of the original data, and 𝑙 is

the number of principal components. Equation 2.31 is utilized for larger sample size

applications whereas, Equation 2.32 is more suitable for smaller sample size applications.

𝑙(𝑛 − 𝑙) (2.31)
𝜏 2 = 𝜏𝛼2 = Ƒ (𝑙, 𝑛 − 𝑙), 𝑤ℎ𝑒𝑟𝑒 ∀𝒯 2 , 𝒯 2 ≤ 𝜏 2
𝑛−𝑙 𝛼

46
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

𝑙(𝑛2 − 𝑙) (2.32)
𝜏 2 = 𝜏𝛼2 = Ƒ (𝑙, 𝑛 − 𝑙), 𝑤ℎ𝑒𝑟𝑒 ∀𝒯 2 , 𝒯 2 ≤ 𝜏 2
𝑛(𝑛 − 𝑙) 𝛼

For index 𝒬, which is quadratic in nature, the upper control limit is calculated using

the weighted chi-squared distribution as shown in Equation 2.33, where 𝒳𝛼2 is the chi-

square distribution, 𝑚 is the sample mean, 𝑣 is the sample variance, 𝑔 is the weight of the

chi-squared distribution, and ℎ is the degrees of freedom of the chi-squared distribution

[54]. Equations 2.34 and 2.35 provide definitions of 𝑔 and ℎ.

𝛿 = 𝛿𝒬,𝛼 = 𝑔𝒳𝛼2 (ℎ), 𝑤ℎ𝑒𝑟𝑒 ∀𝒬, 𝒬 ≤ 𝛿 2 (2.33)

𝑣 (2.34)
𝑔=
2𝑚

2𝑚2 (2.35)
ℎ=
𝑣

Similar to index 𝒬, combined index, 𝜑 has been shown to be approximated using

the chi-square distribution by Yue and Qin [15 ismail]. The upper control limit is described

in Equation 2.36, where 𝑔𝜑 is the weight of distribution, ℎ𝜑 is the degree of freedom and

∝ is the assigned confidence level. Definitions of the weight of distribution and degree of

̃), 𝜃2 =
freedom are shown in Equation 2.37 and 2.38, respectively, where, 𝜃1 = 𝑡𝑟𝑎𝑐𝑒 (Λ

1
̃2 ), and Λ
𝑡𝑟𝑎𝑐𝑒 ( Λ ̃= 𝑇̃𝑇̃ 𝑇 .
𝑛−1

𝜁 2 = 𝜁𝛼2 = 𝑔𝜑 𝒳𝛼2 (ℎ𝜑 ), 𝑤ℎ𝑒𝑟𝑒 ∀𝜑, 𝜑 ≤ 𝜁 2 , (2.36)

𝑙 𝜃2 𝑙 𝜃1 (2.37)
𝑔𝜑 = ( 4
+ 2 )⁄( 2 + 2 )
𝜏𝛼 𝛿𝛼 𝜏𝛼 𝛿𝛼

47
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

𝑙 𝜃1 𝑙 𝜃2 (2.38)
ℎ𝜑 = ( 2
+ 2 )2 ⁄( 4 + 4 )
𝜏𝛼 𝛿𝛼 𝜏𝛼 𝛿𝛼

2.3.7 Multi-Scale Principal Component Analysis

As discussed in previous sections, PCA and wavelet transforms are very powerful

tools that can be applied to FDD applications. More specifically, PCA is specialized in its

ability to analyze multiple sensor measurements and identifying the relationships between

those measurements whereas wavelet transform is specialized in its ability to decompose a

time series measurement into time and frequency components. However, FDD strategy for

complex systems such as an internal combustion engine requires the specialities of both

PCA and wavelet transform simultaneously. In other words, for a successful FDD strategy,

the decomposition of every measurement in time-frequency components and the

identification of the correlation between all of the measurements at each time-frequency

component is vital.

Bakshi introduced a new analysis method called Multi-Scale Principal Components

Analysis (MSCPA), which can be described as a hybrid between PCA and wavelet

transform. MSPCA consists of decomposing every measurement signal using discrete

wavelet transform then performing PCA on the wavelet coefficients [55]. The concept of

MSPCA is presented in Figure 15 where the signal is shown to be decomposed using DWT

and the resulting wavelet coefficients become inputs to PCA where the loading and scores

matrices are calculated for each wavelet coefficient. Upon loading and scores calculation,

upper limits of 𝒯 2 , 𝒬, and 𝜑 are calculated for fault detection and diagnosis that were

48
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

introduced in Section 2.3.6. The difference between PCA and MSPCA is Principal

components and the upper limits are calculated for each wavelet bin.

Figure 15: MSPCA schematic [12]

2.3.8 Mod-MSPCA

Yoon and MacGregor introduced a modified version of MSPCA called Mod-

MSPCA [56]. The schematic of Mod-MSPCA is provided in Figure 16, which shows that

the wavelet coefficients are not directly inputted into PCA upon decomposition through

DWT. The wavelet coefficients in Mod-MSPCA are reconstructed using wavelet synthesis

method then inputted into PCA at each level. The purpose of this modification is to utilize

Reconstruction Based Charts for better detection and diagnosis of faults. RBC required

sample size closer to the measurement sample size but due to the decomposition step prior

to the PCA original MSPCA was not compatible with RBC. The reason for applying RBC

for fault detection and diagnosis was to mitigate the issues of the upper limit technique

such as inability to quickly detect changes in states of the operating conditions. Such

limitation would manifest itself as false alarm indicating a presence of a fault even when
49
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

the fault has been resolved. Yoon and MacGregor states that RBC guarantees a higher

accuracy in fault diagnosis. Further information regarding RBC can be found in Yoon and

MacGregor’s research [56].

Figure 16: Mod-MSPCA schematic [12]

50
Chapter 3: Proposed FDD Strategy

This section will provide the details on the proposed fault detection and diagnosis

strategy that will be developed and tested using real engine test-data. An overview of the

purpose, constraints and assumptions involved in the development of the proposed FDD

strategy is presented to provide a relevant summary of the information discussed in Chapter

2: FDD Strategies for Internal Combustion Engines as an introduction.

The purpose of the proposed FDD strategy is to monitor the internal combustion

engine in a cycle-by-cycle basis to detect intermittent misfire faults that may occur in

different cylinders of the engine. In this context, intermittent faults are faults that do not

exhibit in every engine cycle. The application of the FDD strategy is aimed at detecting the

presence of a misfire and the location of the fault condition, in other words, in which

cylinder the misfire has occurred.

Due to the complexity of engine dynamics discussed in Section 2.2.1, the proposed

FDD strategy utilizes the signal-based approach to enable rapid development of a robust

strategy. The measurements used in the development of the FDD strategy includes

pressure, vibration and angular position. The data acquisition for the development of the

FDD strategy is envisioned to be incorporated smoothly into the existing engine

development testing procedures. However, it is very important to note that the in-vehicle

51
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

application of the proposed FDD strategy is intended to only consider the vibration and

angular position measurement. This is due to the very expensive cost associated with

pressure sensors and the retrofitting of the pressure sensors in the engine as discussed in

Section 2.2.2. The ability to detect fault conditions using only the vibration signal is

hypothesized with an important assumption. This assumption is that there exists a

correlation between the pressure and vibration measurement in the crank angle domain.

With the knowledge of the pressure measurement’s ability to assess the combustion quality,

this assumption hypothesizes that the characterization of faulty combustion conditions is

plausible with just the vibration measurement if the knowledge of the combustion quality

obtained through the analysis of the pressure measurement is transferred through data

labeling. Another important concept that will be utilized in the proposed FDD strategy is

the analysis of data in the crank angle domain to correlate signal behaviors to specific

component or engine event as discussed in Section 2.3.2.

The development method of the proposed strategy is outlined in Figure 17. It

outlines four major components of the development method: equipment set-up & data

acquisition, data pre-processing, feature extraction and classifier development. This

chapter will introduce detailed concept of each component of the strategy. Chapter 4 and 5

will provide the detailed implementation of the concepts discussed in this chapter on real

data set.

52
53
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 17: Development overview of the proposed strategy:

54
3.1 Proposed Experimental Set-up & Data Acquisition

As mentioned before, the proposed FDD strategy utilizes the signal-based approach

due to the complex nature of the system being monitored. From the literature review, it can

be concluded that modeling of engines requires tremendous amount of effort and resources

that is not feasible for the development of a FDD strategy. Therefore, the use of engine

testing process is suggested to develop a signal based FDD strategy for internal combustion

engines. The benefit in utilizing the testing environment for the data acquisition component

of the FDD development process is the amount of quality measurement that is available.

These measurements have potential to provide valuable insight into the engine conditions

and behaviors. The specifics of the engine and test cell set-up utilized in this research is

provided in Section 4.1.1 and 4.1.2.

3.1.1 Proposed Instrumentation

The main measurement being proposed to be monitored in this FDD strategy is

vibration in crank domain due to its low cost and effectiveness in characterizing faults in

mechanical systems as discussed in Section 2.2.2. Another measurement of interest in the

development of the proposed FDD strategy is the pressure traces of each cylinder. Pressure

measurement is an example of measurement that is conveniently available in an engine

testing environment. However, pressure measurement is not a measurement that is to be

monitored in the application of the FDD strategy as pressure sensors are extremely costly

and require complex retrofitting of the engine. Therefore, the pressure traces are proposed

to be incorporated only in the development of the FDD strategy. The role of the pressure

55
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

traces during the development of the FDD strategy is to provide valuable information,

specifically on the combustion quality of each engine cycle. This high quality information

is to be incorporated into the development of the FDD strategy through date labeling

method, where each vibration cycle is labeled as healthy or faulty through the analysis of

the pressure measurement. The specifications and the locations of the accelerometers is

further elaborated in Section 4.1.5. Similarly, the specifications of the pressure transducer

are provided in Section 4.1.4. The pressure analysis method proposed is explained in

Section 3.2 and the implementation of the method in this research is presented in Section

4.3.

The position measurement of the crankshaft of the engine is utilized to segment the

continuous vibration and pressure measurement for the development process and just the

vibration measurement for the application of the developed FDD strategy into aligned

individual engine cycles. Also, the position measurement is to be utilized to convert the

vibration and pressure measurement into crank angle domain from time domain. The

rotational speed of an internal combustion engine is variable and never constant which

means that events that are referenced in time provide little information regarding what

process or component caused that event. However, when events are referenced using crank

angle, simple analysis can infer useful information regarding the components or the process

that caused the observed event. Detailed information regarding the optical encoder utilized

is provided in Section 4.1.3. The theoretical concepts involved in the analysis of the

encoder signals will be discussed in Section 3.2 and the implantation of the method is

provided in Section 4.3.

56
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

3.1.2 Proposed Fault Induction Method

The purpose of data acquisition is to obtain engine behavioral data on healthy and

faulty conditions under the same operating conditions to ensure that the fault condition is

not being induced due to the change in the operating condition. The proposed method of

fault induction during data acquisition process is to test the engine at parametric conditions

that increases the possibility of a misfire. This method is proposed to obtain data that

closely resembles the behavior that would be observed from an engine under normal

operating conditions. Previously in Dohgri’s research, misfires were induced by physically

unplugging the spark plug from the engine [22]. This provides conditions of the engine that

does not provide a realistic representation of how a misfire may occur in an engine that is

operating within a vehicle. The details on the test equipment such as the engine,

dynamometer, and the data acquisition hardware are provided in Section 4.1 and the testing

conditions and test procedures are provided in Section 4.2.

3.2 Proposed Data Pre-Processing Methods

The purpose of the proposed data pre-processing method is to transform the raw

signal that has been acquired through testing into a format that will allow the feature

extraction method to extract the most information it can infer from the raw signals. This

transformation includes the conversions of raw time series measurement into crank angle

domain measurement and the segmentation of continuous measurement into individual

engine cycles.

57
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

3.2.1 Proposed Crank Angle Conversion Method

As discussed in Section 2.3.2, representing the vibration and pressure

measurements in crank angle domain is very important for reciprocating machines such as

the internal combustion engine, where all events of the engine are intricately timed in terms

of the crank position. The raw measurements that are obtained during the data acquisition

is in the time domain and engine events are difficult to accurately be inferred in this domain

due to wide operating speed ranges and its inability to hold a constant speed, which was

explained in Section 2.1.1. The encoder measurement provides occurrences of a defined

angular displacement of the crank shaft in time domain. The time measurements are

proposed to be converted into crank angle domain using the crank angle and time

relationship provided by encoder measurement through linear interpolation. Linear

interpolation is suggested as the method of encoder measurement analysis as it is

computationally very efficient compared to other interpolation methods. The detailed

calculation of the proposed conversion method is provided in Section 4.3.

3.2.2 Proposed Cycle Segmentation Method

In order for the FDD strategy to monitor the engine in cycle-by-cycle manner, the

continuous crank angle domain measurements need to be segmented into individual engine

cycles. Engine cycle is defined as two revolution of the crankshaft which is equivalent of

720 degrees. Using the measurements in crank angle domain, 720 degrees of the crank

shaft can easily be obtained. However, each individual cycle must be aligned to a specific

engine event to ensure that alignment of the entire data set of cycles. Alignment is

58
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

especially important for feature extraction where time-frequency information is extracted.

If the engine cycles of the same fault condition are not aligned at the same engine event,

the time information that is extracted from the feature extraction process will differ

between the features extracted from the two separate cycles. This means that the

representation of the fault condition using both of these engine cycles can be considered

invalid. For successful characterization of each fault condition, it is crucial that aligned

engine cycles are utilized for fault extraction. The engine event that is utilized for this

alignment is the top-dead-center event of cylinder 1. The detailed application of this

proposed cycle segmentation method is provided in Section 4.3.

3.2.3 Proposed Pressure Analysis Method

The combustion quality information of the engine cycles inferred from the pressure

traces using thermodynamic theories are to be incorporated in the development of the

proposed FDD strategy. The method of pressure analysis that is proposed in this research

is the calculation of Indicated Mean Effective Pressure (IMEP). The purpose of this

analysis is to define a misfire mathematically using thermodynamic theories. IMEP

represents the average pressure that is acting on the piston during the engine cycle. IMEP

is a very accurate and an easy to calculate indicator of misfires. For the pressure trace of a

combustion chamber when an engine is motoring, which means that there are no

combustion occurring inside the combustion chambers, the IMEP value is zero. A

calculated value of IMEP below a set threshold limit will be defined as a misfire Therefore,

if an IMEP of a combustion chamber during normal operating conditions of an engine is a

value close to the threshold limit, it would indicate that a misfire has occurred in that
59
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

combustion chamber. The steps involved in calculating the IMEP values of each engine

cycle is outlined in Section 4.3.

3.2.4 Proposed Data Labeling Method

The information regarding the combustion quality obtained through the pressure

analysis is to be incorporated into the development strategy through a process called

labeling. Labeling of vibration data for each engine cycle is required as the fault conditions,

which are differentiated by its combustion quality, are not easily inferred using vibration

signals. With the assumption that correlation exists between the vibration measurements

and the corresponding fault condition, the label is aimed at guiding the characterization of

each condition using vibration measurements. The true state of the combustion quality is

to be determined through the pressure analysis discussed in the previous section. Different

fault labels were given to misfires occurring in different cylinders to allow the classifier to

be able to differentiate a misfire that has occurred in one cylinder or another, for fault

diagnosis component of the FDD strategy. The main purpose of the labeling process is to

eliminate the need for using pressure transducers in vehicles for the application of the FDD

strategy. The process of the labeling method is provided in Section 4.3.

3.3 Proposed Feature Extraction Method

Feature extraction of vibration measurements of each segmented and aligned

engine cycle that are labeled using pressure measurements are proposed to be performed

using Extended Multi-Scale Principal Component Analysis (Extended-MSPCA).

60
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Extended-MSPCA, which was introduced by Ismail, is a feature extraction method

that utilizes PCA and wavelet packet transform (WPT). This method is used due to its

ability to analyze non-transient behaviors of a complex system that have been captured

through multiple sensors. In short, the Extended-MSPCA decomposes a given signal using

WPT, which provides corresponding wavelet coefficients then applies PCA on those

coefficients.

Extended-MSPCA is very similar to Mod-MSPCA, which was discussed in Section

2.3.7. However, the feature that differentiates the Extended-MSPCA from Mod-MSPCA

is improved performance in the characterization of the fault conditions. The main

difference between the two methods is that Extended-MSPCA utilizes Wavelet Packet

Transform (WPT) while Mod-MSPCA utilizes Discrete Wavelet Transform (DWT). The

mathematical theories regarding WPT and DWT were discussed in Section 2.3.5.2. The

two transforms differ in the way they decompose the signal being transformed. DWT

decomposes the signals in a way that highlights the higher frequency contents whereas,

WPT decomposes the signal equally along the signal’s frequency spectrum. Therefore,

WPT provides better granularity in terms of frequency in comparison to the DWT. The

Extended-MSPCA is proposed, as opposed to the Mod-MSPCA, due to its ability to

highlight all of the frequency bands equally. As faults may occur in both high and low

frequencies, it is very important to utilize the feature extraction method that is capable of

capturing both those cases.

61
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

In addition to the type of wavelet transform utilized, Extended-MSPCA differs

from Mod-MSPCA in the way normalization is achieved. In Mod-MSPCA the raw signal

is immediately decomposed using DWT and the output coefficients are normalized for

PCA. However, Extended-MSPCA normalizes the raw measurements before decomposing

them using WPT. Therefore, the outputs of WPT, which are the wavelet coefficient, are

considered to be normalized which means that normalization upon WPT is unnecessary for

PCA. The summary of the proposed normalization method is provided in Figure 18 and

Figure 19. The proposed order of normalization allows for features extraction process to

minimize the effects of varying scale and variances of measurements due to difference in

the sensor’s calibration or sensitivity.

Figure 18: Mod-MSPCA signal transformations [12]

Figure 19: Extended MSPCA signal transformations [12]

Given the difference between Mod-MSPCA and Extended-MSPCA, the detailed

steps involved in executing the Extended-MSPCA is as follows. Upon data acquisition and

necessary data pre-processing that were introduced previously in this chapter, the baseline

measurement is to be first defined. The Extended-MSPCA generates the features for a

given measurement by comparing how different the new measurement is in comparison to

a defined baseline measurement. This baseline measurement is proposed to represent a

62
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

normal or healthy measurement. However, due to the variability that is inherent in engine

operation, which was discussed in Section 2.1.1, the selection of the baseline measurement

is not trivial. It is clear that the attempt to represent what normal engine cycle looks like

with a single-engine cycle measurement is a contradiction in itself. If a single-engine cycle

is to be plotted in the feature domain, the normal cycle will be represented as a single point.

If all of the normal cycles were to be plotted in the feature domain, the normal cycle will

be represented as a cluster of points. Given that cycle-to-cycle variations exists even in

normal conditions; the cluster representation is a more realistic depiction of what normal

conditions look like as opposed to the single point representation. Therefore, the baseline

measurement is proposed to be represented by multiple healthy measurements as opposed

to a single healthy measurement. The process of baseline selection is described in Section

5.1. Multiple baseline measurements are not proposed to be manipulated to provide a

calculated baseline data such as an average of the multiple baseline measurement. Instead,

the utilization of the multiple baseline measurements is proposed to extract multiple

representations of the measurement being observed. For example, if 10 individual cycles

are selected to represent the baseline measurements of the given system.

Upon selection of the baseline measurements, the measurement, that the feature

will be extraction from, is proposed to be normalized using Equations 3.1 and 3.2 where

normalization is performed based on the parameters of the assigned baseline measurement.

The assigned baseline measurement is also normalized using its own parameters.

63
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

𝑠𝑖𝑔𝑛𝑎𝑙𝑟𝑎𝑤 − 𝑚𝑒𝑎𝑛 𝑣𝑎𝑙𝑢𝑒 (𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒 𝑐𝑒𝑛𝑡𝑒𝑟)

𝑠𝑖𝑔𝑛𝑎𝑙𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 = (3.1)
𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 (𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒 𝑟𝑎𝑤 𝑠𝑖𝑔𝑛𝑎𝑙)

1
𝑥𝑟𝑎𝑤 − 𝑛 ∑𝑛1 𝑥𝑟𝑎𝑤,𝑖
𝑥𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 = 2 (3.2)
1 1
∑𝑛1 (𝑥𝑟𝑎𝑤,𝑖 − ∑𝑛1 𝑥𝑟𝑎𝑤,𝑖 )
𝑛−1 𝑛

Once the measurements are normalized, it is decomposed into time-frequency domain

using WPT. The PCA component of the proposed feature extraction method is then applied

to calculate the principal component matrices for the healthy and the faulty measurements

using Equations 3.3 and 3.4 given the relationship provided in Equation 3.5.

Σ 𝑇,ℎ𝑒𝑎𝑙𝑡ℎ𝑦 = 𝐶𝑂𝑉(𝑇ℎ𝑒𝑎𝑙𝑡ℎ𝑦 ) (3.3)

Σ 𝑇,𝑓𝑎𝑢𝑙𝑡𝑦 = 𝐶𝑂𝑉(𝑇𝑓𝑎𝑢𝑙𝑡𝑦 ) (3.4)

𝐶𝑂𝑉(𝑇) = 𝐸(𝑇𝑇 𝑡 ) − 𝐸(𝑇)𝐸(𝑇)𝑡 (3.5)

A statistical index was introduced by Ismail to enable fault diagnosis capabilities of the

Extended-MSPCA [12]. This index, which is defined by Equation 3.6, provides a

quantitative measure of a fault condition.

𝐹𝑐,𝑗 = 𝑠𝑖𝑔𝑛 (𝐿𝑗 ) ∘ √∑ 𝑠𝑖𝑔𝑛(𝐶𝑂𝑉(𝐹))𝑗 ∘ [𝐶𝑂𝑉(𝐹) ∘ 𝐶𝑂𝑉(𝐹)]𝑗 (3.6)

𝑖=1

Further to the above detailed overview of the theory related to Extended-MSPCA,

specific application method is proposed. This application method is heavily focused around

64
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

how to perform feature extraction using multiple baselines. The proposed feature extraction

process for the development of the FDD strategy aims to produce multiple representations

of the fault signature based off of multiple baseline data as shown in Figure 17. This will

lead to a clustering of healthy and faulty data in the feature domain. More representation

of one faulty data with respect to multiple baseline data allows for the classification

algorithm to account for the variable nature of the engine cycle. It should be noted that

these baseline data are not used to represent a condition. In other words, a feature

representation of baseline data is not extracted or used in the training process. It is also

important to have healthy data represented in comparison to the multiple baseline data to

account for the variability even within the healthy condition.

3.4 Classifier Development

In this FDD strategy, a supervised machine learning algorithm is proposed to

develop a classifier algorithm that would classify features into fault categories. Supervised

learning algorithms take labelled data set that provide input-output examples and learns the

relationship between the input features and the output labels [57]. During the learning

process, the algorithm refines the mapping of the inputs and the outputs by minimizing the

error between the desired outputs and the classifier algorithm’s output. The errors are

represented by a cost function and it is minimized by updating the weights.

65
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 20: MLP with two hidden layers [57]

A type of artificial neural network (ANN) called a multilayer perceptron (MLP) is

proposed to be the classifier algorithm for this research due to its compatibility with

nonlinearity and ease of application. Similar to the human brain, ANN are composed of

fundamental information processing units called neurons. MLP consists of multiple

clustering or a row of neurons called a layer, as shown in Figure 20. The simplest MLP

consists of three layers: input, hidden and output layer. Figure 20 depicts MLP architecture

with two hidden layers. It is important to note that every neuron from a layer is connected

to every neuron in the previous and the following layer, this type of architecture is known

as a fully connected neural network. Figure 20 also shows that the information flows from

the input to the hidden layers then to the output layer, this type of architecture is known as

a feedforward neural network.

66
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Model of a neuron, which consist of m number of inputs, a bias, an activation

function and an output, is depicted in Figure 21. As shown in the figure, each neuron

receives m number of inputs which are multiplied with their respective weights. The

weights correspond to the connection between a neuron from the previous layer and a

neuron in the following layer. The weights are unique to each connection and are updated

during the training process. Figure 21 also shows that each neuron consists of its own bias

which also gets multiplied with its corresponding weight value. The values obtained

through the weight multiplications are summed at the junction as 𝑣𝑘 = ∑𝑚

𝑗=0 𝑤𝑘𝑗 𝑥𝑗 , which

is fed into the activation function as an input to calculate the output of the neuron, 𝑦𝑘 =

𝜑(𝑣𝑘 ) [57]. A variety of activation functions are available such as sigmoid, tangent

hyperbolic, softmax and rectified linear units. The nonlinear behavior of these activation

functions enables artificial neural networks to be nonlinear. As this research involves a

multi-class classification problem, a softmax activation function is used in the output layer.

Figure 21: Model of a neuron [57]

67
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

A MLP is trained using a technique called backpropagation. During training there

are two types of signals that flow through the network which are called function and error

signals as shown in Figure 22 [57]. The main distinction between the two signals are the

direction in which they travel through the network. The forward propagation is carried out

by the function signal and the back propagation is carried out by the error signal [57].

During forward propagation, the function signal travels through each layer from one neuron

to another to transform the input into an output value. During back propagation, an error

signal is obtained by comparing the output value of the neural network to the desired

output. This error signal travels backwards through the network layer by layer. As the error

signal travels, the weights associated with each neuron to neuron connection is adjusted to

minimize the error [57]. As mentioned before, errors are represented by loss functions

which are minimized by a gradient method. The feedforward and backpropagation is

repeated until the loss function is minimized to a satisfactory value. For this research a

cross-entropy loss function is proposed to be minimized using the scaled conjugate gradient

backpropagation methods.

68
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 22: Function and error signals [57]

As mentioned before, the multiple baseline strategy is proposed in this research to

take into consideration the variable engine dynamics. Given a single engine cycle, n

individual features are extracted if n number of baselines are used. These individual

features are different feature representations of a single engine cycle. During the training

of the classifier, it is suggested that each representation of a given fault is to function as an

individual training sample. The classifier algorithm will not know that these

representations were obtained from one engine cycle. This method is proposed to provide

classifier with enough variations of each fault condition to learn from. Exposing the

classifier to as many variations possible allows the classifier to understand the variable

nature of the engine behavior. This is especially useful if the dataset obtained is small.

Once the classifier is fully trained, the FDD strategy will utilize a majority voting

method to detect a fault. For example, if n number of baselines are used to extract features

from a new engine cycle, n number of feature representations will be extracted. Each of the

feature representations will separately be classified using the trained classifier, which will

69
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

provide n different assessments of the fault condition. A majority voting method is utilized

to make the final prediction on the fault condition of the new engine cycle. The concept of

the majority voting method is to select, among all the of the different predictions that were

provided by individual features, the prediction that had the most votes. The schematic

process of this method is shown in Figure 17 under the classifier development subset. This

majority voting method is utilized to reduce detection error caused by uncertainty of the

fault condition dimensions in the feature space.

70
Chapter 4: Data Acquisition

In this section, the details of the proposed strategies regarding data acquisition are

provided. These subsections include the experimental setup, the engine test procedures and

data pre-processing.

4.1 Experimental Setup

This chapter provides detailed descriptions of the experimental setup utilized for

the development of the FFD algorithm for internal combustion engines. The FDD

algorithm monitors the engine on a cycle by cycle basis. It is a signal-based approach that

detects intermittent combustion related to faults in different cylinders.

The experimental setup was used to acquire data for feature extraction of different

conditions, training of the machine learning classification tool, and testing of the classifier.

The main elements of the experimental setup include a test cell, an engine, an engine

dynamometer, a data acquisition system, and sensors (including an encoder, pressure

transducers and accelerometers). The data acquired includes measurements from the

encoder, accelerometers, and pressure sensors in time domain.

71
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

4.1.1 Engine

All the experiments for data acquisition were performed on a Ford’s V8 spark

ignition, pushrod developmental engine. The engine was provided by Ford’s Powertrain

Engineering Research and Development Center (PERDC). Detailed specifications of the

engine are omitted due to confidentiality. However, the focus of this research and the

resulting conclusions are not dependent on the specifics of this particular engine. Instead,

the measurements acquired and the strategy implemented consider phenomena that are

universal to engines of this type, at the least; V-style, 8 cylinder, and overhead valve model.

Figure 23 illustrates the difference between an overhead valve and an overhead cam engine

design and Figure 24 illustrates the difference between an inline engine and a V-style

engine.

72
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 23: Overhead valve vs overhead cam engine design [58]

Figure 24: Different styles of engines [59]

Figure 25 shows the cylinder numbering for the test engine utilized in this research.

The firing order of the engine is cylinders 1-5-4-8-6-3-7-2 as shown in Figure 22.

Figure 25: Engine cylinder numbering for the V8 test engine [60]

73
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

4.1.2 Test Cell

The testing facility was provided by Ford Powertrain Engineering Research and

Development Center. The test cell is composed of an engine dynamometer, engine

assembly and monitoring instrumentations. The test cell that was utilized for this research

was unique amongst all of the test cells that are located at PERDC. The test cell is in a

semi-anechoic chamber which ensures the elimination of surrounding noise and absorption

of the energy created within the chamber. In addition, only the engine assembly, including

the engine, was housed in the test cell to ensure that vibrations from the engine alone are

obtained during testing.

4.1.2.1 Engine Assembly

The engine assembly consisted of the clutch, transmission assembly, fan belt,

power steering pump, alternator, compressor, powertrain mounts, and mount brackets. The

purpose of the engine assembly is to simulate the in-vehicle conditions that the engine is

to operate within.

4.1.2.2 Dynamometer

The dynamometer was housed outside the chamber in a separate utility room. This

separation ensures that only the noise and vibration from the engine assembly is being

captured during data acquisition without external noise. The following table summarizes

the specifications of the dynamometer that was utilized in this research.

74
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Table 1: Dynamometer Specifications

Manufacturer Meidensha Inc

Manufacturer Date 2001
Type FREC Dynamometer
Model S95-355L
Number of Poles 2P
Base Speed 2500/2647 RPM
Max Speed 8000 RPM
Absorbing Power
370 kW
(Continuous)
Motoring Power
250 kW
(Continuous)

4.1.3 Encoder

The encoder is used to measure the rotational position of a shaft [61]. An encoder

is equipped with an optical sensor, light source and rotating disk with either opaque and

transparent sections or slits. The optical encoder is mounted on the rotating shaft to be

studied and the optical sensor outputs “on” or “off” pulses. depending on whether or not it

sensed the light that was emitted by the light source. The optical sensor would sense the

light in the case where the transparent or the opening of the slit is passing through the light

source and the sensor which are in line with each other with the disk rotating in between

the two as shown in Figure 26. The output will be “off” if the light is blocked due to the

opaque section or the solid portion of the disk. The output of the encoder is a digital signal

which means that it will be in the form of a square wave with the high values indicating

the “on” state and the low values indicating the “off” state. With the knowledge of the

angular width of the sections or the angular distances between slits, the angular position of

75
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

the shaft can be obtained with respect to time. The angular speed and acceleration of the

shaft can also be derived from the position and time data obtained from the encoder.

Figure 26: Encoder components [62]

In this experiment, the AVL 365 angle encoder was used as it provides high

precision and high resolution angle information that is required for converting the

transducer signals from time domain to crank angle domain. This encoder has a resolution

of 0.5-degree crank angle which means that each engine cycle, which involves two

rotations of the crankshaft, will consist of 1440 data points. A trigger signal was also

obtained from this encoder which indicates the occurrence of the top dead centre (TDC) of

Piston 1 with an offset. The offset of the trigger signal to TDC of Piston 1 is 131.662 degree

which means that the trigger signal gets switched “on” 131.662 degrees after the

occurrence of TDC of Piston 1. The trigger signal is crucial in resampling of the transducer

signals as it ensures that every cycle is synchronized properly in crank angle domain and

76
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

allows for the data set to be divided into cycles in the time domain. The AVL 365 angle

encoder has the RS422 transistor-to-logic (TTL) output driver, output range of 0 – 5V and

with a standard threshold being 3.5V for “on” signal [63].

4.1.4 Pressure Transducer

Pressure traces provide useful insight into the combustion happening inside the

engine. To correctly label the data, pressure traces were used to calculate values such as

Indicated Mean Effective Pressure (IMEP) to determine if a combustive fault has occurred

or not in a given engine cycle.

Two main types of transducers for sensing pressure are piezoelectric or piezo-

resistive. The application determines which type of sensor should be used. These two types

of transducers differ in that piezoelectric pressure sensors measure the electrical charge

produced by the piezoelectric material, whereas piezo-resistive pressure sensors measure

the change in electrical resistance of a silicon semiconductor.

Figure 27: The mechanical load on the crystal producing electrical charge through electric dipole [64]

77
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

For the piezoelectric sensor, the electrical charge is produced by the piezoelectric

material, such as quartz, and occurs when the surfaces of the crystal are loaded

mechanically. This load causes the shift of positive and negative crystal lattice elements in

relation to each other resulting in the formation of an electric dipole [64]. These sensors

have a very high natural frequency which makes them ideal for measuring fast pressure

rise times. However, one of the downsides of piezoelectric sensors is that the measurement

drifts upon application of a static load.

Figure 28: Piezoelectric pressure sensor components [65]

For the data acquisition portion of the research, pressure traces were obtained

through Kistler’s PiezoStar Type 6125C pressure sensors which are designed for in-

cylinder pressure measurement of combustion engines. Ford PERDC provided cylinder

heads with machined holes for each cylinder that are compatible with the pressure

transducers for testing. PiezoStar pressure transducers are piezoelectric sensors with

PiezoStar crystals which are produced by Kistler instead of quartz. It has a measuring range

78
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

of 0 – 300 bar and a sensitivity of -33 pC/bar [66]. The negative sensitivity of the pressure

transducer is due to the fact that pressure applied to a piezoelectric sensor produces

negative charge signal. This phenomenon is not found in the output signal as external

charge amplifier converts the negative going charge into a positive voltage signal [64].

Piezoelectric transducers were utilized instead of the piezo-resistive transducers as the

nature of combustive pressure signals are highly dynamic in nature.

4.1.5 Accelerometers

. Accelerometers are used to capture vibrational behaviors by measuring the

dynamic acceleration of a mechanical component or assembly, [67]. Similar to pressure

transducers that were discussed in the previous section, accelerometers use piezoelectric

material to measure acceleration. An accelerometer consists of a seismic mass,

piezoelectric material and preload bolt. The acceleration experienced by the component

being studied is transmitted to the accelerometer and causes the seismic mass inside it to

also accelerate. A proportional force generated by the acceleration of the seismic mass is

induced on the piezoelectric material that is clamped between the seismic mass and the

base of the accelerometer by a preloaded bolt. The proportional force applied to the

piezoelectric material by the seismic mass results in an electric dipole formation due to the

shift in the lattice structure of the piezoelectric material. This electric dipole then generates

electrical charge proportional to the acceleration that can be acquired by the data

acquisition hardware.

79
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

There are two types of accelerometers, charge type piezoelectric sensors and

integrated electronic piezoelectric (IEPE) sensor and they differ in the way they convert

the charge from the piezoelectric crystal. In the charge type accelerometers, the charge is

converted through an external amplifier or inline charge converter. However, an IEPE

sensor does not require an external amplifier as it has a built-in charge amplifier. The IEPE

sensors require a data acquisition hardware that provides current excitation for the built-in

amplifier.

Figure 29: IEPE accelerometer components [67]

In this research, Dytran IEPE tri-axial accelerometers were used to obtain the

vibration from various parts of the engine. The accelerometer has a range of +/- 5000g,

frequency response of 10 kHz and a sensitivity of 1mV/g [68]. It should be noted that some

may argue that charge type accelerometer would have been preferred over the IEPE as they

are suited for high temperature applications. However, given the fact that the normal

operating temperature for an engine is anywhere between 195 to 220 degrees Fahrenheit
80
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

and that the operating temperature of the IEPE accelerometers are rated up to 250 degrees

Fahrenheit, IEPE sensors are used with cautionary measures in place such as wrapping the

exhaust and sensor cables with heat shield wraps.

Figure 30: Knock sensor locations on sample V8 engine [69]

The accelerometers were mounted on the engine with epoxy or with bolts when

possible to ensure that the true vibrational readings are being obtained. Two accelerometers

were placed on top of two knock sensors in the left bank of the V8 engine. From the

accelerometer placed on top of the knock sensors, only the z-axis measurement was

acquired. The z-axis is parallel to the vibrational measurement obtained by the knock

sensor. Therefore, this measurement should be indicative of the vibrational components

that knock sensors observe. This location is strategically selected as the success of the

proposed FDD strategy using sensors and sensor locations that already exist in the

manufacturing process would reduce cost of implementing such strategy. The locations of

81
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

the knock sensors for the experimental engine are shown on a similar V8 engine in Figure

30. Two other accelerometers were placed in the valley of the V-style engine as shown in

Figure 31. The placement of these sensors was solely explorative which aims to capture

the overall vibrational behaviors of the engine. All 3 axes from these accelerometers were

acquired to observe the up and down, left to right and front to back vibrations of the engine.

Y Y

X X

Figure 31: Accelerometer locations on sample V8 engine block [70]

All the accelerometers were calibrated to 100 Hz at 1g-rms as specified by Dytran

prior to mounting. The calibrations were performed using The Modal Shop 9100D Portable

Shaker Table.

82
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

4.1.6 Data Acquisition Hardware

During a data acquisition process, sensors convert physical phenomenon to

electrical signals which are collected and digitized by a data acquisition hardware [71].

Depending on the data acquisition hardware, the module may be capable of providing

either analog input/output, digital input/output or counter operation functionality on a

single device or all of the listed functionalities on a single multifunction device. The type

of sensor dictates the functionality of the data acquisition hardware required to acquire the

signal [71]. For example, the encoder signal which outputs a digital signal would require

the digital input/output functionality whereas, the accelerometer signal which outputs an

analogue signal would require the analogue input functionality and Analog-to-Digital

Conversion (ADC) to acquire the data.

For this research experimentation, signals from the encoder, accelerometers and

pressure transducers were acquired by multifunction devices called PROSIG P8020 and

P8012. These two modules allow for fixed channel counts of 24 low speed 24-bit ADC

channels and 16 high speed 16-bit ADC channels to be recorded simultaneously. These

multifunction devices have a range of +/- 10 V and an accuracy of +/- 0.1% of the range.

The high speed and low speed modules are able to support a maximum sampling frequency

of 400k and 100k samples/sec/channel, respectively. Given the maximum engine speed

that the data will be acquired at, frequency component of the encoder signal was calculated.

Although Nyquist Theorem suggests that a signal can be reconstructed with a

sampling rate equal to twice the highest frequency of the interested component, National

83
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Instruments advise that a sampling rate equal to ten times the highest frequency of the

interested component should be used [72]. The sampling rate of the encoder was carefully

selected to ensure that the angular position of the crankshaft is being captured as precisely

as possible without any loss of information. All the vibration and pressure signals were

recorded at 100kHz/channel and the encoder data was recorded at 400kHz/channel. The

encoder signal was sampled at a much higher rate compared to the transducers as the

encoder’s data will be used to resampled the transducer data in crank angle domain. It is

very important to have data point that is as close to the timestamp as possible to when the

switch occurred. Therefore, in this application a high sampling frequency of the acquisition

system is preferred.

Prosig data acquisition system performs filtering on the signals that are to be obtain

through a built-in anti-aliasing filter. The cut-off frequency for the filtering is adjusted

automatically according to the sampling rate of the signals to ensure that aliasing effects

are not captured in the signals acquired.

4.2 Engine Test Procedure

4.2.1 Operating Conditions

The fault of interest for this research is intermittent misfire. The associated FDD

experiment involves obtaining both healthy and faulty cycles under a same operating

condition and by running the engine in a condition where the chances of the fault occurring

are high. Under the normal operating range of the engine, the idle condition is considered

to be the most unstable, with significant fluctuations in combustion quality. Heywood

84
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

specifies that at idle conditions, chances of partial-burning cycles are higher, [4].

Therefore, the data was collected at the idle condition where the engine is running without

a load at 600 rpm. It is also mentioned that the chances of misfires can be reduced at the

idle condition by retarding the spark timing [4]. Therefore, spark timing was advanced

incrementally to increase the severity of the engine's instability. The following list

summarizes the operating condition:

 idle speed of 600 RPM

 no load

 engine oil temperature below 265 F

 coolant temperature below 195 F

 spark timing advance varied from 35 to 55 degree

4.2.2 Testing Procedures

The engine was warmed up at 1500 rpm with a load of 2 bar until the oil temperature

reached 160 F, and a stable condition is reached. Once the engine is warmed up, the engine

speed was lowered to 600 rpm, the load was removed and the appropriate spark timing

advance was applied. The data acquisition was then triggered when steady state was

reached further to changes made to spark timing advance.

4.2.3 Raw Data Collected

Accelerometer Data

Measurement signals from eight accelerometers were recorded in this research.

Raw accelerometer data from a three axes accelerometer located on the front valley is

85
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

shown in Figure 32 to Figure 34. Raw accelerometer data from an accelerometer located

on the front left knock sensor is show in Figure 35.

Figure 32: Raw accelerometer data from front valley, axis parallel with front-back orientation of the

engine

86
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 33: Raw accelerometer data from front valley, axis parallel with right-left orientation of the engine

87
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 34: Raw accelerometer data from front valley

Figure 35: Raw accelerometer data from front left knock sensor

Encoder Data

Two signals were obtained from the encoder: encoder ticks and trigger signal.

88
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 36: Raw encoder ticks @ 600rpm

Encoder ticks indicate the 0.5-degree travel of the crankshaft. Figure 36 is a plot of

the raw encoder ticks data at an engine speed of 600 rpm. To verify the data, two points

that indicate the start of an on switch on consecutive occasions are displayed. The time

differences between these two points indicate the time it took for the crankshaft of the

engine set to 600 rpm to travel 0.5 degrees. Assuming that the engine is running at a

constant speed of 600 rpm the following calculation obtains the theoretical time difference

between the two points selected in Figure 33. The difference between the theoretical and

the actual time durations can be explained by the fact that actual engines are not capable of

holding the speed constant. Therefore, the encoder ticks data obtained is considered to

89
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

accurately represent the change in the rotational position of the crankshaft with respect to

time.

𝑟𝑒𝑣 𝑚𝑖𝑛 60 𝑠
∆𝑡𝑡ℎ𝑒𝑜𝑟𝑒𝑡𝑖𝑐𝑎𝑙 = 0.5° ∙ ∙ ∙ = 1.3889 × 10−4 𝑠 (4.1)
360° 600 𝑟𝑒𝑣 𝑚𝑖𝑛

∆𝑡𝑎𝑐𝑡𝑢𝑎𝑙 = 𝑡2 − 𝑡1 = 0.0006875 − 0.00054 = 1.4750 × 10−4 𝑠 (4.2)

Figure 37: Raw encoder trigger @ 600rpm

Trigger signal indicates the occurrence of the Top-Dead-Center (TDC) in Cylinder

1, which occurs once per revolution of the crankshaft. Figure 37 is a plot of the raw encoder

trigger data at an engine speed of 600 rpm. To verify the data, two points that indicate the

start of an on switch on consecutive occasions are displayed as shown in Figure 34. With

similar assumptions and method, the data is verified as shown in the calculations below.

90
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

With similar reasoning for the differences observed in the two time durations, the trigger

data is considered to accurately represent the occurrence of TDC in Cylinder 1.

𝑚𝑖𝑛 60 𝑠 (4.3)
∆𝑡𝑡ℎ𝑒𝑜𝑟𝑒𝑡𝑖𝑐𝑎𝑙 = 1 𝑟𝑒𝑣 ∙ ∙ = 0.1000𝑠
600 𝑟𝑒𝑣 𝑚𝑖𝑛

∆𝑡𝑎𝑐𝑡𝑢𝑎𝑙 = 𝑡2 − 𝑡1 = 0.1605 − 0.06001 = 0.1005𝑠 (4.4)

Pressure Data

Raw pressure data of Cylinder 1 is plotted in Figure 38. An overlay of pressure

traces from all the cylinders are plotted in Figure 39 to validate the firing order of the

engine. The order of the pressure peaks is shown to align with the firing order indicated in

Section 4.1.1 which verifies that the pressure transducers were operating and placed

properly.

91
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 38: Raw Cylinder 1 Pressure Trace @600RPM

92
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 39: Overlay of all of the raw pressure traces from each cylinder @600RPM

4.3 Proposed Data Pre-Processing Strategy Implementation

In this section, the implementation of data pre-processing to transform the raw data

into a compatible form for feature extraction is described in detail. As shown in Section

4.2.3, raw data consists of three different types of measurements; encoder, pressure and

acceleration in the time domain. The final form required for feature extraction is

acceleration measurement in the crank angle domain that is segmented into individual

93
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

cycles with corresponding fault class labels. The following sections describe in detail, the

steps taken to transform the data.

Step 1: Filtering of Accelerometer Data

The accelerometer measurements were filtered using a Butterworth low-pass filter

to remove noise. The sampling frequency of the vibration measurement was 100 kHz and

the frequency response of the accelerometer is 1.2 Hz to 10 kHz. Therefore, the cutoff

frequency was set to 10 kHz with a 12th order filter with phase compensation to eliminate

phase shift. Figure 40 shows the raw accelerometer signal and the filtered signal.

Figure 40: Filtering of accelerometer data using Butterworth Filter

Step 2: Encoder Measurement Analysis

The purpose of the encoder measurement analysis is to obtain crank angle versus

time relationship and to obtain the occurrences of the Top-Dead-Center (TDC) of

Cylinder1. The crank angle versus time relationship is utilized to resample the pressure and

94
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

acceleration measurements in crank angle domain. The timestamps at which TDC occurs

in Cylinder 1 is utilized to segment the signals into aligned individual cycles.

Figure 41: Encoder tick signal

The relationship between crank angle and time was determined by obtaining the

timestamps of the instances where the light indicator switch of the encoder was flipped

from “off” to “on”. More specifically, the instances at which the voltage was equal to the

threshold value were obtained on the rising edge. The threshold value of 3.5 V was utilized

as specified by the manufacturer. The exact timestamp was estimated through linear

interpolation of two points immediately before and after the threshold value on the rising

edge as shown in Figure 41.

95
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Rotational speed of the crankshaft is derived from this crank position versus time

relationship as shown in Figure 42 to validate the alignment of the set engine speed and the

actual engine speed. It can be observed that the average engine speed aligns with the set

speed of 600RPM. However, significant fluctuations also exist in the speed profile, which

aligns with the unstable behavior described by Heywood at idle conditions [4]. The

fluctuations in engine speed is a direct reflection of the fluctuations in combustion quality

as mentioned by Heywood [4].

Figure 42: Calculated engine speed @ 600RPM

The occurrences of TDC in Cylinder 1 referenced in time were obtained using the

trigger signal. A similar approach was used for encoder tick signal analysis to retrieve the

96
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

time at which the switch was flipped as an indicator for the occurrence of the TDC in

Cylinder 1. It is important to note that TDC occurs twice in a given engine cycle as shown

in Figure 43. Therefore, timestamps obtained by the trigger signal are not sufficient enough

to synchronize individual cycles in the data set. Figure 43 illustrates that a piston of an

engine reaches TDC once at the end of the compression stroke and again at the end of an

exhaust stroke. With this knowledge, the cycle can be synchronized with additional

information provided by Cylinder 1 pressure trace.

Figure 43: Pressure trace of a cylinder [62]

Figure 44 shows an overlay of trigger signal and Cylinder 1 pressure trace. Given

the timestamp of a specific occurrence of TDC of Cylinder 1, pressure trace of Cylinder 1

97
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

can be referenced to see if this particular occurrence was at the end of a compression stoke

or the exhaust stroke. If the pressure trace at the given timestamp is close to 0 bar, the

current TDC can be concluded to have occurred upon an exhaust stroke. On the other hand,

if the pressure value is well above 0, it can be concluded that the current TDC occurred

upon a compression stroke. The timestamp associated with the exhaust stroke was chosen

as the engine event at which the cycles will be segmented at. Therefore, the timestamps at

which the TDC occurred at upon the exhaust stroke was obtained and was defined as the

start timestamps indicating the start of a new cycle. It should be noted that there exists an

offset between the occurrence of the TDC in Cylinder 1 presented by the trigger signal and

the TDC that can be inferred by using the pressure signal. This offset is as expected as

specified in Section 4.1.3.

98
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 44: Overlay of trigger data and Cylinder 1 pressure trace @ 600RPM

Step 3: Resampling of Transducer Measurements

The crank position versus time relationship was used to resample the pressure and

vibration measurements. The voltage of the transducer measurements at each crank angle

– timestamps were obtained through linear interpolation. Figure 45 and Figure 46 show the

plot of the transducer measurement in time and crank angle domain.

Figure 45: Filtered accelerometer data in time domain

99
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 46: Filtered accelerometer data in crank angle domain

Step 4: Engine Cycle Segmentation

The start timestamps obtained in Section 4.3.2 is used to segment the transducer

measurements into individual cycles. The start timestamps are referenced by a specific

engine event to ensure the alignment of all the cycles obtained from different test files. The

alignment is required as the data acquisition was randomly triggered which means that each

test file starts at different positions of the crank angle that relate to different events of the

engine. In this research, an engine cycle is defined as a 720-degree rotation of the crank

angle that is aligned at 131.662 degree after the occurrence of TDC in Cylinder 1 upon an

exhaust stroke event. Each measurement per cycle consists of 1440 points with a crank

angle resolution of 0.5 degrees. Due to the fact that the offset of 131.662 degree is not

100
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

divisible by the crank angle resolution of 0.5 degrees, the start timestamp and the crank

angle timestamp never shared a common timestamp. Therefore, first crank angle timestamp

that occurs after the start timestamp is used as the start of the cycle. And the end of the

cycle is marked by the first crank angle timestamp that occurs after next start timestamp.

Step 5: Data Labeling

Pressure trace provides very accurate and insightful information regarding the

quality of the combustion occurring inside the cylinder. It would be ideal to develop a

condition monitoring tool that utilizes in-cylinder pressure data from the engine. However,

it is very unrealistic to have in-cylinder pressure transducers in operating vehicles due to

their extremely high cost. Therefore, information obtained through pressure traces in test

engines is correlated with more convenient and less costly vibration measurements. This

correlation is performed through labeling the engine conditions of the vibration

measurements using information extracted from the pressure traces.

For this experiment, the analysis of the pressure trace was performed by calculating

the Indicated Mean Effective Pressure (IMEP). IMEP was calculated by integrating the

pressure signal from 180-degree crank angle before TDC to 180-degree crank angle after

TDC with respect to displacement volume as shown by the following equation, where 𝑝(𝑖)

is the cylinder pressure at crank angle position 𝑖, 𝑉(𝑖) is the cylinder volume at crank angle

position 𝑖, 𝑉𝑠 is the cylinder swept volume, 𝜃0 is the bottom dead center (BDC) induction

integer crank angle position, and 𝜃𝑛 is the BDC integer crank angle position [73].

101
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

𝜃𝑛
∆𝜃 𝑑𝑉(𝑖)
𝐼𝑀𝐸𝑃 = ∑ 𝑝(𝑖) (4.5)
𝑉𝑠 𝑑𝜃
𝑖=𝜃0

If the IMEP value is below the specified lower limit, it is considered to be a misfire which

is the fault condition of interest in this research.

Figure 47 shows a plot of all the pressure traces in one, previously defined, engine

cycle. The plot shows that at a given engine cycle, a total of 360-degree pressure trace

segment consisting of compression and power strokes required for IMEP calculation for

every cylinder is not available within the specified engine cycle.

Figure 47: All pressure traces for 8 cylinder engine cycle

102
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 48 to Figure 50 are plots from two consecutive cycles for Cylinders 2, 5, and

8, respectively. Cylinder 2 is an example of those cylinders that have their compression

and power strokes fully represented in the segmented engine cycle. On the other hand,

Cylinders 5 and 8 are examples of those cylinders that do not have their compression and

power strokes fully represented in the segmented engine cycle as shown in Figure 47. Also,

it can be observed that the pressure trace of Cylinder 5 flows into the next engine cycle

while the pressure trace of Cylinder 8 flows from the previous engine cycle. This infers

that in order to calculate the IMEP of the combustion that is happening in Cylinder 5 for

the specified engine cycle, the pressure trace for the following cycle is required. In a similar

manner, the pressure trace of the previous cycle is required for IMEP calculation of the

combustion happening in the Cylinder 8 for the same engine cycle. This concept was

adopted for the IMEP calculations of those cylinders that were affected by the segmentation

of the engine cycles. In addition, the very first and last cycles of every test were discarded

as accurate calculation of the IMEP values for all of the cylinder were not obtainable as

full compression and power strokes are not available.

103
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 48: Cylinder 2 Pressure Trace of 2 Consecutive Cycles

Figure 49: Cylinder 5 Pressure Trace of 2 Consecutive Cycles

104
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 50: Cylinder 8 Pressure Trace of 2 Consecutive Cycles

It should be noted that the vibration measurements were segmented without

carrying over the vibrations that occurred during the full compression and power strokes

of the cylinders. The vibration measurements were not carried over due to the definition of

an engine cycle that was established earlier. This lack of data representation may lead to

difficulty in correlating the pressure to the vibration relationship of faults occurring in the

affected cylinders.

Pre-Feature Extraction Data Summary

The following table summarizes the labeled cycles in the dataset obtained through

engine testing. It is very important to note the imbalance of the different conditions in the

data collected. This is primarily due to the way the faults were simulated. In previous

research conducted, misfire faults were simulated by disconnecting the spark plug which

105
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

ensures that every cycle of the specified cylinder misfires [22]. This method of simulating

the fault allows for a balanced distribution of the data to be obtained. However, in this

research the faults were induced by simulating a condition of the engine where the

probability of incomplete combustion is high. Therefore, the occurrence of faults or the

distribution of faults were not guaranteed.

Table 2: Summary of labeled data

600 RPM Cycles

Healthy 8002
Fault 1 Misfire in Cylinder 1 66
Fault 2 Misfire in Cylinder 2 9
Fault 3 Misfire in Cylinder 3 188
Fault 4 Misfire in Cylinder 4 7
Fault 5 Misfire in Cylinder 5 369
Fault 6 Misfire in Cylinder 6 81
Fault 7 Misfire in Cylinder 7 167
Fault 8 Misfire in Cylinder 7 76
Three sample cycles that consist of one faulty cycle and two different healthy cycles

that were labeled in the previous section using IMEP calculations are plotted in the

following series of figures. The fault class of the faulty cycle is 1 which indicates that the

fault has occurred in Cylinder 1. Figure 51 is an overlay of pressure traces from Cylinder

1 for three different cycles. Figure 52 to Figure 54 are the vibration data from the three-

axis accelerometer located on the front valley. Figure 55 is the vibration data from the

accelerometer located on the front left knock sensor.

106
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

When observing the pressure signals, the maximum pressure level and the area

under the pressure trace of faulty cycle is noticeably smaller than that of the healthy cycles

which would numerically be translated into the IMEP values. Therefore, it can be

concluded that faulty cycles are easily identifiable amongst a mixture of healthy and faulty

cycles for a given cylinder.

Figure 51: Cylinder 1 Pressure Traces for Faulty and Healthy sample cycles

When observing all the vibration data, the difference between healthy and faulty

data is not as clear. The distinguishable characteristics such as big spikes and locations of

those big spikes do not seem to be unique for healthy cycles. When observing the spikes,

it can be seen that there does not seem to be a commonality in the locations of the spike

107
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

with respect to crank angle. In terms of amplitude, it can be noticed that Healthy 2 cycle

shows lower amplitude than the Faulty cycle in all of the vibration plots. Therefore, it can

be concluded that faulty cycles are not easily identifiable amongst a mixture of healthy and

faulty cycles without further analysis or without advanced FDD tools.

Figure 52: Front-Back Axis Accelerometer data from Front Valley for Faulty and Healthy sample cycles

108
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 53: Right-Left Axis Accelerometer data from Front Valley for Faulty and Healthy sample cycles

109
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 54: Up-Down Axis Accelerometer data from Front Valley for Faulty and Healthy sample cycles

110
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 55: Axis Accelerometer data from Front Left Knock Sensor for Faulty and Healthy sample cycles

111
Chapter 5: Proposed FDD Strategy Implementation and

Results

5.1 Proposed Feature Extraction Strategy Implementation

In this section, the implementation of the feature extraction strategy is described in

detail. The purpose of the feature extraction strategy is to transform the pre-processed data,

represented in Section 4.3, into features. The theory pertaining to the proposed strategy

was provided in Chapter 3. This section focuses on elaborating on the parameters chosen

to implement the strategy.

The proposed feature extraction method takes the user-defined baseline data and

faulty data and generates features that represent how different the faulty data is in

comparison to the baseline data. The proposed strategy utilizes the Extended-MSPCA to

decompose both the baseline and faulty data through wavelet packet transformation, then

transforms them into main components using PCA and generates features using statistical

indices. The implementation of the strategy consists of a selection of wavelet function,

wavelet level, and baseline data.

In this research, the algorithm is tasked to differentiate the misfiring cycle from the

healthy data and differentiate the locations of the misfire. The difference between a misfire

112
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

occurring in one cylinder from the other is hypothesized to primarily depend on the

occurrence of the misfire with respect to the crank angle domain. This hypothesis assumes

that the differentiating characteristics of the misfire condition in comparison to the healthy

condition will occur within the compression and power strokes of a given cylinder.

Depending on the firing order for a given engine, each cylinder will fire at a different crank

angle range. This infers that the wavelet function must have enough localization in the

crank angle to be able to differentiate the same frequency components occurring 90 degrees

apart. On the other hand, the differentiation of faulty and healthy cycles may be highly

dependent on frequency. Therefore, a balance of frequency and time resolution must be

achieved for the success of accurate characterization of faults.

Step 1: Wavelet Function Selection

A systematic process of selecting an optimal wavelet function for a given

application does not exist. Consequently, the wavelet function that behaves similarly to the

signal of interest is advised to be utilized for the wavelet transform. Therefore, in the case

of fault detection and diagnosis, the ideal wavelet function to be utilized would be one that

resembles the faulty components of the signal. However, the faulty components are not

easily identifiable in the signals obtained, given the resources and time allowed for further

analysis of the signals. Therefore, a wavelet family called Daubechies, which is commonly

used for vibration in FDD applications, is utilized in this research. Daubechies wavelets

have been utilized in the previous studies conducted by Ismail, Doghri, and Narendiranath

for FDD applications on alternators and starters, internal combustion engines, and journal

bearings, respectively with successful outcomes [12], [22], [74].

113
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Daubechies wavelets are said to be compactly supported, which means that it will

provide results that are more localized as opposed to global results. Daubechies wavelets

consist of numerous individual wavelet functions that are differentiated by its number of

vanishing moments. Wavelet db4 represents Daubechies wavelet with vanishing moments

of 4. The increase in vanishing moments implies that more complex behaviors of the signal

will be captured. In addition, Figure 56 shows that higher-order mother wavelets provide

superior frequency resolution with sharper roll-off. However, the improvements made on

frequency resolution results in a reduction in time resolution.

Figure 56: Frequency response of Daubechies wavelets [75]

For the feature extraction process component of this research, different wavelet

functions are utilized to study their effects on the performance of the FDD strategy. These

114
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

wavelet functions consist of db4, db10, and db16 were utilized. Figure 57, 55, and 56 show

the corresponding db16, db10, and db4 low pass and high pass filters that were utilized

during Extended-MSPCA.

Figure 57: Low and high pass filter for db16

115
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 58: Low pass and high pass filter for db10

Figure 59: Low pass and high pass filter for db4

116
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Step 2: Wavelet Level Selection

Wavelet levels dictate the number of frequency bins the given signal will be

decomposed to. As wavelet levels increase, the number of frequency bins increases while

the range of each frequency is decreased. In other words, the increase in wavelet levels

leads to finer frequency resolution information. The increase in frequency resolution results

in decreased crank angle resolution as crank angle resolution is halved at each subsequent

decomposition level. The crank angle information is predicted to be crucial in the

characterization of faults in different cylinders due to the cyclic operation of the engine. In

addition, there is a maximum number of wavelet levels that can be used for a given sample

signal and the wavelet being used for the analysis. The maximum number of wavelet levels

can be obtained through the following expression:

𝑁𝑠𝑎𝑚𝑝𝑙𝑒
𝐿𝑚𝑎𝑥 < 𝑙𝑜𝑔2 ( − 1) (5.1)
𝑁𝑤𝑎𝑣𝑒𝑙𝑒𝑡

Given this relationship, the maximum wavelet level for each wavelet function is

obtained and summarized in the table below. For this research, a wavelet level of 4 will be

utilized for all three wavelets which is limited by db16’s maximum wavelet level. The

maximum level of wavelet is restricted in this research due to the short length of the dataset.

In previous studies performed by Ismail and Doghri, longer sample lengths were utilized,

which meant that higher wavelet levels could be utilized. For cycle-by-cycle monitoring,

the only method of increasing the sample length is through the increase in angular

resolution of the dataset.

117
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Table 3: Max wavelet level for corresponding wavelet

Wavelet Max Level

db16 5
db10 6
db4 7

Step 3: Defining Baseline

As discussed in Chapter 3, it was proposed that multiple baseline data approach is

to be taken as opposed to the single baseline data approach. For the application of this

strategy, the method of selection and the method of determining the number of baselines

required are important components of the feature extraction strategy that must be

considered further for a thorough study of the feature extraction strategy. However, in this

research, a simplified method of selection is pursued for the interest of time. The simplified

method of selection involves randomly selecting a specified number of individual cycles

from all the healthy data set that have been collected. In this experiment, 10 different

baseline data will be utilized for the feature extraction component.

In the previous study performed by Doghri where the FDD strategy was applied to

detect constant fault conditions, averaging method of features that belong to the same fault

condition was proposed [22]. The constant fault of an internal combustion engine can be

defined as a condition that is known to manifest regardless of the cycle-to-cycle variations.

The purpose of the moving average is to capture the dominant characteristics of a given

fault signature. However, the application of a moving average is not recommended for

118
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

cycle-by-cycle monitoring, which is required for the detection of intermittent faults. This

is due to the implementation methods of the developed FDD strategy and the type of fault

the strategy is aimed to monitor. If averaging was performed during the development of

the FDD strategy, averaging would also be required to guarantee the success of the

classifier. If the unknown measurements are not averaged, the measurement will consist of

noise and other cycle-specific variations that the training algorithm has not seen during the

training process. The presence of noise and cycle-to-cycle variation of the unknown

measurement may affect the results of the classification tool. Therefore, implementation

methods must be taken into consideration during the development process of the FDD

strategy.

Feature Summary

From all the cycles obtained and labeled, 100 randomly selected healthy cycles and

the rest of the faulty cycles were defined as a finalized dataset, which consists of 1063

cycles, which equates to 10630 features. The number of healthy cycles was reduced to 100

cycles to even out the imbalance of data. Aside from 100 healthy cycles, 10 additional

healthy cycles were randomly selected to form the baseline dataset. From the dataset of all

the cycles, ~70%, ~20%, and ~10% of each class was randomly selected to form the

training, validation and testing subsets. The testing subset was formed by randomly

selecting ~10% of the total cycles in the dataset. For example, for the healthy condition

there are a total of 100 cycles which equate to 1000 features. To select the testing set, 10

cycles from 100 cycles were randomly chosen and the testing set was formed by 100

features that have been generated from the 10 cycles that were selected. However, to divide
119
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

the remaining data into training and validation, 70% of the features were randomly chosen

per condition to form the training set and the remainder formed the validation set. The

breakdown of total cycles and features are shown in Table 4. The method chosen to divide

the total data into subsets were chosen to ensure that the testing set will be executed in the

way that the classifier is to be used in real application. Further explanations of each dataset

and its importance will be elaborated in the classifier section.

Table 4: Data breakdown

Cycles / Features Training Validation Testing

Healthy 100 / 1000 - / 630 - / 180 90 / 900
Fault 1 66 / 600 - / 460 - / 130 7 / 70
Fault 2 9 / 90 - / 60 - / 20 1 / 10
Fault 3 188 / 1880 - / 131 - / 380 19 / 190
Fault 4 7 / 70 - / 50 - / 10 1 / 10
Fault 5 369 / 3690 - / 2580 - / 740 - / 370
Fault 6 81 / 810 - / 570 - / 160 8 / 80
Fault 7 167 / 1670 - / 1170 - / 330 17 / 170
Fault 8 76 / 760 - / 530 15 / 150 8 / 80
Total 1063 / 10630 - / 7430 - / 2120 108 / 1080

The features extracted consist of 16 coefficients for each sensor signal utilized,

which results in a feature dimension of 16 x 8 or a total of 128 coefficients. This is a great

reduction in dimensionality from the raw signal, which consists of 1440 data points for

each sensor signal, resulting in 11520 data points. The reduction in dimensionality will

reduce the time and computational resources required for the classification algorithm.

Figure 60 and Figure 61 show the features that were extracted from accelerometers

on the valley and knock sensors, respectively. A general trend can be observed from the

120
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

features extracted from the valley accelerometers across all the fault conditions, including

the healthy condition, and subtle differences that can be observed from one condition to

the other. However, these differences are not easily translated into quantifiable

relationships to characterize each fault condition. The difficulties found in differentiating

each fault class from the visual representations of the feature sets validate the need for a

sophisticated classifier algorithm that can infer subtle nuances of the features that are not

easily identifiable.

121
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 60: Features extracted from valley accelerometers

122
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 61: Features extracted from knock accelerometers

123
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

The table below summarizes the time it took for the generation of features for each

wavelet type and wavelet level used. The time of feature extraction is very important for

both the training and utilization of the FDD strategy. The Extended-MSPCA algorithm

used in this research is sufficient for the purposes of evaluating the analysis method and its

ability to perform analysis on a cycle by cycle basis. However, the algorithm will require

extensive improvement in order for it to be capable of monitoring the engine on a cycle by

cycle basis.

Table 5: Time duration for each feature extraction

Wavelet Type Wavelet Level Time per Cycle/Baseline

db4 4 1.031360s
db10 4 1.199280s
db16 4 2.161937s

5.2 Results and Observations

In this section, the results of the developed FDD strategy are presented. The results

will review the performances of three different classifiers, Classifier 1, 2 and 3, that have

been developed using features extracted by three different wavelet functions, db4, db10,

and db16. The hyper parameters such as number of hidden nodes, learning rate and

stopping criteria, were tuned individually for each classifier to yield the most optimal

classifier for the given dataset. For each classifier, the results are presented through four

different confusion matrices. These include the training confusion matrix, validation

confusion matrix, testing confusion matrix, and detection confusion matrix as shown in

Figure 59 – 61.
124
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

The training confusion matrix shows how well the classifier was able to learn from

the training data set. The validation confusion matrix shows how well the classifier is able

to generalize. The testing confusion matrix shows the classifier’s ability to classify the

features that the classifier has not seen during training. And the detection confusion matrix

shows the ability to correctly classify the given cycle using majority voting method.

125
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Classifier 1: Wavelet Function db4

126
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 62: Confusion matrices for Classifier 1

127
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Classifier 2: Wavelet Function: db10

128
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 63: Confusion matrices for Classifier 2

129
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Classifier 3: Wavelet Function: db16

130
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Figure 64: Confusion matrices for Classifier 3

131
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Table 6 presents the performance scores of each classifier for training, validation,

testing and detection including precision, recall and F1-sores. The following are the notable

observations from the classifier’s and their performances.

1. The best overall detection accuracy of 96.3% was achieved by Classifier 3 which

was trained on features extracted from db16 wavelet function.

2. All three classifiers were not able to classify any fault representations of fault

classes 2 and 4 during testing.

3. Majority voting technique improved the overall accuracy rate for all classifiers

which is indicated by the accuracy improvement from testing to detection.

The best detection accuracy of Classifier 3 indicates that the features extracted with

db16 wavelet function had the most information that could be inferred to characterize the

fault conditions. These results prove that enough information exists in the vibration

measurements that can be correlated using in-cylinder pressure measurement to monitor

the engine’s combustion quality. The performance of Classifiers 1 and 2 indicate that

although the information was not as readily available for correlation as it were for Classifier

3, there still exist information that can be utilized for FDD strategy. Even though Classifiers

1 and 2 showed overall accuracy rate that is about 10% lower than that of Classifier 3, more

research should be performed to improve their performance rather than ruling them out as

potential options for future FDD strategy. The reason is due to the benefits of utilizing db4

or db10 wavelet functions for feature extraction. And the most appealing benefit is the

reduction in processing time. As mentioned in Section 5.1, the processing time for

132
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

extracting features using db16 is double the amount of time it takes using db4. This could

be crucial for when the FDD strategy is to be applied in real-time.

Despite having the high accuracy rates during training and validation across all

three classifiers, as shown in Table 6, none of the classifiers were able to detect faults from

fault classes 2 and 4. The root cause of this phenomenon is the lack of data for fault classes

2 and 4. The lack of representation of a given fault class in the dataset results in classifier’s

inability to learn the characteristics that are universal to that specific fault class. The

explanation for the very high performance observed during training and very low

performance observed during testing could be that the classifiers were able to learn the

characteristics of the fault classes but the extent of the engine’s variability of behavior was

not accounted for due to the limited dataset. Another explanation could be that the classifier

was overfitting the dataset that belong to fault classes 2 and 4 due to the imbalance of the

dataset. The difference between the two explanations is that the first explanation is

reflective of the dataset’s inability to provide the classifier the overall picture while the

second explanation is reflective of the classifier’s neglect of the fault classes 2 and 4. The

classifier is trained on the training dataset by updating its weights and biases to reduce the

overall accuracy. Therefore, given that fault classes 2 and 4 had very small datasets, they

were set as low priority for the classifier to learn as learning the correlation of fault class

5, which has a very high dataset, would yield better accuracy.

The effects of the majority voting technique are reflective in the improvement of

overall accuracy observed from testing to detection. As discussed in Section 5.1, a total of

133
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

10 baselines were utilized for feature extraction of cycles for both training and testing

datasets. The testing confusion matrix accuracy indicates the classifier’s ability to detect

each feature representation of a single cycle. The majority voting method’s function is to

mitigate the variations that exist among the 10 feature representations extracted. It can be

argued that not all 10 baselines utilized for training ought to be utilized for testing as

utilizing fewer baselines for testing would mean shorter processing time of new cycles.

However, all 10 baselines were utilized with majority voting method in this research in

case the dataset utilized for the development was too small. In the case where the testing

accuracy is on par with the accuracy level required to for the FDD strategy the majority

voting technique can be omitted.

134
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

Table 6: Performance Scores

Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9

Validation Training Precision 91.1 100.0 95.4 100.0 91.4 98.9 96.9 97.4 91.4
Recall 84.6 93.3 95.9 98.0 95.0 97.9 96.5 93.2 86.7
87.7 96.5 95.7 99.0 93.2 98.4 96.7 95.3 89.0
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Precision 85.3 100.0 92.1 100.0 87.4 96.2 94.3 92.4 80.7
Classifier 1

Recall 71.5 80.0 92.3 70.0 91.9 95.6 95.5 88.7 77.5
77.8 88.9 92.2 82.4 89.6 95.9 94.9 90.5 79.1
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Testing

Precision 73.8 0.0 85.7 0.0 84.4 84.7 89.1 98.2 80.8
Recall 64.3 0.0 88.4 0.0 86.5 90.0 95.9 68.8 97.0
68.7 0.0 87.0 0.0 85.4 87.3 92.4 80.9 88.2
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Detection

Precision 57.1 0.0 88.9 0.0 84.6 80.0 94.4 100.0 90.9
Recall 57.1 0.0 84.2 0.0 89.2 100.0 100.0 62.5 100.0
57.1 0.0 86.5 0.0 86.8 88.9 97.1 76.9 95.2
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Validation Training

Precision 88.9 94.0 96.0 100.0 92.2 98.6 97.4 97.1 93.3
Recall 83.5 78.3 96.4 100.0 95.3 98.6 96.3 94.2 90.0
86.1 85.4 96.2 100.0 93.7 98.6 96.9 95.6 91.6
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Precision 87.0 94.4 92.2 100 87.6 96.8 93.8 91.2 84.0
Classifier 2

Recall 72.3 85.0 93.4 100 90.5 94.4 95.8 89.3 81.5
79.0 89.5 92.8 100.0 89.0 95.6 94.8 90.2 82.7
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Testing

Precision 51.4 0.0 82.5 0.0 86.9 94.6 93.6 78.8 92.8
Recall 51.4 0.0 89.5 0.0 89.5 87.5 94.7 78.8 90.0
51.4 0.0 85.9 0.0 88.2 90.9 94.2 78.8 91.4
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Detection

Precision 50.0 0.0 85.0 0.0 89.7 100.0 94.1 85.7 100
Recall 57.1 0.0 89.5 0.0 94.6 87.5 94.1 75.0 100
53.3 0.0 87.2 0.0 92.1 93.3 94.1 80.0 100.0
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Validation Training

Precision 96.1 98.4 98.0 100.0 96.3 99.8 98.8 98.5 96.0
Recall 91.5 100.0 98.2 96.0 97.8 99.5 99.1 98.3 93.4
93.7 99.2 98.1 98.0 97.0 99.7 99.0 98.4 94.7
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Precision 90.4 100.0 94.0 100.0 90.3 96.3 96.6 94.5 86.3
Recall 79.2 90.0 94.5 90.0 94.7 96.3 95.2 92.0 82.0
.Classifier 3

84.4 94.7 94.3 94.7 92.5 96.3 95.9 93.2 84.1

Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Testing

Precision 87.7 0.0 86.5 0.0 92.6 97.5 98.8 95.7 97.8
Recall 91.4 0.0 91.1 0.0 98.1 98.8 99.4 83.8 91.0
89.51 0.0 88.74 0.0 95.27 98.15 99.10 89.36 94.28
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Fault 7 Fault 8 Fault 9
Detection

Precision 87.5 0.0 90.0 0.0 97.4 100.0 100.0 100.0 100.0
Recall 100.0 0.0 94.7 0.0 100.0 100.0 100.0 87.5 100.0
93.3 0.0 92.3 0.0 98.7 100.0 100.0 93.3 100.0

135
Chapter 6: Conclusion

The purpose of this research was to develop a FDD strategy that can monitor a

combustion related fault of an internal combustion engine on a cycle-by-cycle basis. This

research was performed in collaboration with Ford Powertrain Engineering Research and

Development Center as a continuation of the ongoing FDD research conducted at CMHT.

This section will summarize the research, highlight the notable contributions, and provide

suggestions for future work.

6.1 Research Contributions

In this research, a FDD strategy consisting of data acquisition, data pre-processing,

feature extraction and classifier development specific sub strategies was proposed. The

dataset required for the development of the FDD strategy was proposed to be acquired from

currently existing testing procedures and testing facilities. During this research, the dataset

consisting of information rich sensor measurements were obtained. These measurements

function to provide insight into what is occurring inside the engine. In this research, the

combustion quality was the characteristic being monitored by the FDD strategy. Therefore,

the state of the engine’s combustion quality was obtained through the analysis of the in-

cylinder pressure measurement. The combustion quality indicated by analysis of the

pressure measurement functioned as the ground truth. However, due to the invasive and

136
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

expensive nature of the pressure transducers, they are only available in the testing

environment. Therefore, a vibration measurement was proposed to be the monitored

measurement for the FDD strategy due to the non-invasive and cheaper cost of

accelerometers. The difficulty in inferring the combustion quality using vibration

measurement was proposed to be mitigated during the classifier development through

labeling of vibration measurement using information rich pressure measurement. To help

with the dimensionality reduction and correlation of vibration data to the combustion

quality, feature extraction using Extended-MSPCA was performed on vibration

measurements in crank angle domain. Given the features extracted and its corresponding

combustion quality obtained through pressure measurement analysis, the correlation

between the features and the combustion quality was developed using the machine learning

algorithm through supervised learning.

In summary, the following are the major contributions of this research:

1. Development of new data acquisition procedures for combustion related faults

that seamlessly integrates into engine testing processes to obtain high quality

engine data.

2. Development of data pre-processing techniques unique to internal combustion

engines that allows analysis in crank angle domain on a cycle by cycle basis.

3. Development of vibration measurement labeling procedure through analysis of

in-cylinder pressure measurement.

137
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

4. Application of Extended-MSPCA on vibration measurement using multiple

baseline techniques customized for internal combustion engines to mitigate

variable nature of the dataset.

5. Development a of classifier using ANN and majority voting technique that has

shown to be able to classify new dataset for combustion quality with 96.3%

accuracy.

This research demonstrated that there are creative ways to utilized the data that are

being collected during testing processes to enable advancements in technology through

data analysis. More specifically, the research has shown that complex relationship between

pressure and vibration of an engine can be extracted to develop a FDD strategy that is

capable of monitoring the engine’s combustion quality on a cycle by cycle basis.

6.2 Future Work

This research was conducted to assess the feasibility of the proposed FDD strategy

in monitoring the engine’s combustion quality. The development of the proposed FDD

strategy was focused more on the successful application rather than the fine tuning of

parameters in each of the sub-strategies. Therefore, there are many aspects of the proposed

FDD strategy that can be researched to understand the capabilities of the strategy itself and

how it can be fine-tuned to improve the its performance or to be applied to different types

of faults.

Given the set of data that was collected in this research, the following are the

recommendations for future research:

138
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

1. Study of different wavelet families on the existing dataset for feature extraction.

2. Study of different pressure analysis methods for data labeling of combustion

related fault conditions.

3. Study of different baseline selection methods and determination of the optimal

number of baselines for feature extraction.

4. Development of data imbalance mitigation methods.

5. Development of data processing procedures for continuous monitoring in real

time applications.

In order to bring this technology to life, a continuous research approach is

recommended where a set of data collection protocols are developed to acquire coherent

data sets from different types of engines. This protocol is envisioned to standardize data

collection properties such as the types of measurements being collected and the preferred

location of the transducers. Data collection protocols may differ for different types of faults

being monitored. In addition, the protocol should consist of generalized rules to enable

application to different types of engines. With data from different types of engines, a

research can be conducted to see if a FDD strategy developed for one engine can be applied

to a different engine or different type of engine.

139
References

[1] R. Isermann, Fault-Diagnosis Systems: An Introduction from Fault Detection to

Fault Tolerance, Springer, 2006.

[2] P. K. K. S. K. Yadav, "Condition Monitoring of Internal Combustion Engine Using

EMD and HMM," in Intelligent Autonomous Systems, Springer, 2010, pp. 167-185.

[3] R. Stone, Introduction to Internal Combustion Engines, London: Palgrave, 1999.

[4] J. Heywood, Internal Combustion Engine Fundamentals, McGraw-Hill Education,

1988.

[5] S. Fallah, Electric and Hybrid Vehicles - Technologies, Modeling and Control: A

Mechatronic Approach, Wiley-Blackwell, 2014.

[6] Nuclear Power contributors, "Otto Cycle – Otto Engine," Nuclear Power , [Online].

Available: https://www.nuclear-power.net/nuclear-

engineering/thermodynamics/thermodynamic-cycles/otto-cycle-otto-engine/.

[Accessed 2019].

I
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[7] J. A. Caton, "The Thermodynamics of Internal Combustion Engines: Examples of

Insights," Inventions, vol. 3, no. 2, p. 33, 2018.

[8] N. Hall, "Ideal Otto Cycle," National Aeronautics and Space Administration, 05

May 2015. [Online]. Available: https://www.grc.nasa.gov/www/k-

12/airplane/otto.html. [Accessed 2019].

[9] S. K. Y. a. P. K. Kalra, "Condition Monitoring of Internal Combustion Engine

Using EMD and HMM," in Intelligent Autonomous Systems, Springer, 2010, pp.

167-185.

[10] V. Venkatasubramanian, "A review of process fault detection and diagnosis Part

III: Process history based methods," Elsevier - Computers & Chemical

Engineering, pp. 327-346, 2003.

[11] V. Venkatasubramanian, "A Review of Process Fault Detection and Diagnosis Part

1: Quantitative Model-Based Methods," Elsevier - Computers & Chemical

Engineering , 2003.

[12] M. Ismail, "Industrial Extended Multi-Scale Principle Components Analysis for

Fault Detection and Diagnosis of Car Alternators and Starters," McMaster

University , 2015.

II
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[13] J. Tjong, "Engine dynamic signal monitoring and diagnostics," University of

Windsor, 1993.

[14] H. W. Miller, "Piston slap detection in combustion engines using unique signal

analysis techniques," University of Windsor, 1989.

[15] S. Delvecchio, "Condition Monitoring in Diesel Engines for Cold Test

Applications. Part I: Vibration Analysis for Pass/Fail Decision," 2007.

[16] R. Tailony, "Internal Combustion Engine Cold Testing: Modeling, Analysis, and

Development," University of Toledo, 2019.

[17] L. Wei, "Modeling and Simulation of System Dynamics for Cold Test," Second

Inernational Symposium on Intelligent Information Technology Application, 2008.

[18] B. Ray, "Engine defect defection using wavelet analysis," University of Windsor ,

2007.

[19] E. Leitzinger, "Development of in-process engine defect detection methods using

NVH indicators," University of Windsor , 2002.

[20] I. Arasaratnam and S. Habibi, "Engine Fault Detection Using Vibration Signal

Reconstruction in the Crank Angle Domain," SAE International, 2011.

III
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[21] ATS Automation , "Powertrain Final Test," ATS Automation, [Online]. Available:

https://www.atsautomation.com/en/Transportation/Testing%20Automation/Power

train%20Final%20Test.aspx#cold. [Accessed December 2019].

[22] A. Doghri, "A Real Time Fault Detection and Diagnosis System for Automotive

Applications," McMaster University, 2019.

[23] S. Hodgins, "A Wireless Sensor for Fault Detection and Diagnosis of Internal

Combustion Engines," McMaster University, 2017.

[24] S. R. Haqshenas, "Multiresolution - multivariate analysis of vibration signals:

application in fault diaagnosis of internal combustion engines," McMaster

University, 2012.

[25] N. Gunnemann, "Predicting Defective Engines using Convolutional Nueral

Networks on Temporal Vibration Signals," Proceedings of Machine Learning

Research, vol. 74, p. 92:102, 2017.

[26] Y. Shantnawi, "Fault diagnosis in Internal Combustion Engines Using Extension

Neural Network," IEEE Transactions on Industrial Electronics, pp. 1434- 1443,

2014.

IV
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[27] A. S. G. Chandroth, "Vibration Signatures, Wavelets and Principal Components

Analysis in Diesel Engine Diagnostics," WIT Transactions on The Built

Environment, vol. 45, p. 10, 1999.

[28] Q. Gao, C. Fan and Q. Meng, "Rotating machine fault diagnosis using empirical

mode decomposition," Mechanical System and Signal Processing, no. 22, pp. 1072-

1081, 2008.

[29] J. Wu and C. Liu, "Investigation of engine fault diagnosis using discrete wavelet

transform and neural network," Expert Systems with Applications, vol. 35, pp. 1200-

1213, 2008.

[30] A. Sharkey, O. Gopinath and Chandroth , "Acoustic emission, cylinder pressure

and vibration: a multisensor approach to robust fault diagnosis," University of

Sheffield, 2000.

[31] D. Mba and R. Rao, "Development of Acoustic Emission Technology for Condition

Monitoring and Diagnosis of Rotating Machines; Bearings, Pumps, Gearboxes,

Engines and Rotating Structures," The Shock and Vibration Digest , vol. 38, pp. 3-

16, 2006.

[32] J. Ben-Ari, "Vibration Signature Analysis as a Fault Detection Method for SI

Engines," SAE , 1993.

V
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[33] R. B. W. Heng, "Statistical Anlysis of Sound and Vibration Signals for Monitoring

Rolling Element Bearing Condition," Elsevier: Applied Acoustics, vol. 53, pp. 211-

226, 1998.

[34] S. Zhang, Machine Learning and Deep Learning Algorithms for Bearing Fault

Diagnostics - A Comprehensive Review, 2019.

[35] W. Zhang, "A New Deep Learning Model for Fault Diagnosis with Good Anti-

Noise and Domain Adaptation Ability on Raw Vibration Signals," MDPI, 2017.

[36] M. D. Prieto, "Bearing Fault Detection by a Novel Condition-Monitoring Scheme

Based on Statistical Time Features and Neural Networks," IEE Transactions on

Industrial Electronics, vol. 30, no. 8, 2013.

[37] S. Hunt, "Engine Vibration of a Compressed Natural Gas Engine as an effect of

Combustion," University of Windsor, 2016.

[38] I. Arasaratnam, "Engine Fault Detection Using Vibration Signal Reconstruction in

the Crank-Angle Domain," SAE International, 2011.

[39] S. E. Pandarakone, "A Comparative Study between Machine Learning Algorithm

and Artificial Intelligence Neural Network in Detecting Minor Bearing Fault of

Induction Motors," MDPI, 2019.

VI
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[40] I. Amihai, "An Industrial Case Study Using Vibration Data and Machine Learning

to Predict Asset Health," IEEE, 2018.

[41] M. A. Lilo, "Identify and Classify Vibration Fault Based on Artificial Intelligence

Techniques," Journal of Theoretical and Applied Information Technology, vol. 94,

no. 2, 2016.

[42] S. Seidlitz, "An Optimization Approach to Valvetrain Design," SAE Technical

Paper, 1990.

[43] N. Politis, "Advanced Time-Frequency Analysis Applications in Earthquake

Engineering," 2003.

[44] Muller-BBM VibroAkustik System, Rotational Analysis in the Angular Domain,

Planegg: Muller-BBM VibroAkustik System, 2016.

[45] National Instruments, "Getting Started with Order Analysis," National Instruments

Corporation, 2007.

[46] G. Sosa, "Automatic detection of wheezes by evaluation of multiple acoustic

feature extraction methods and C-weighted SVM," Progress in Biomedical Optics

and Imaging - Proceedings of SPIE, vol. 9287, 2014.

VII
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[47] R. Polikar, "The Engineers Ultimate guide to Wavelet Analysis: The Wavelet

Tutorial," 11 5 2006. [Online]. Available:

https://cseweb.ucsd.edu/~baden/Doc/wavelets/polikar_wavelets.pdf. [Accessed

2019].

[48] I. Daubechies, "Where do Wavelets Come From?," IEEE, vol. 84, no. 4, pp. 510-

513, 1996.

[49] V. Dhandapani and R. Seshasayanan, "Power-optimized log-based image

processing system," EURASIP Journal on Image and Video Processing , p. 37,

2014.

[50] Hackernoon, "Principal Component Analysis - Unsupervised Learning Model,"

Hackernoon, Jan 2019. [Online]. Available: https://hackernoon.com/principal-

component-analysis-unsupervised-learning-model-8f18c7683262. [Accessed Dec

2019].

[51] T. Villegas, M. Fuente and M. Rodriguez, "Principal Component Analysis for Fault

Detection and Diagnosis. Expereience with a pilot plant," Advances in

Computerational Intelligence, Mac-Machine Systems and Cybernetics , pp. 147-

152, 2010.

VIII
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[52] G. A. Cherry and S. J. Qin, "Monitoring Non-normal Data with Principal

Component Analysis and Adaptive Density Estimation," IEEE Conference on

Decision and Control, pp. 352-359, 2007.

[53] J. Jackson, in A User's guide to principal components, New York, John Wiley and

Sons Inc., 1991.

[54] G. Box, "Some theorems on quadratic forms applied in the study of analysis of

variance problems, i. effect of inequality of variance in the one-way classification,"

The Annals of Mathematical Statistics , vol. 2, no. 25, pp. 290-302, 1954.

[55] B. Bakshi, "Multiscale PCA with application to multivariate statistical process

monitoring," AlChE, vol. 44, no. 7, 2004.

[56] S. Yoon and M. J, "Principal-Component Analysi sof Multiscale Data for Process

Monitoring and Fault Diagnosi," AlChE Journal, vol. 50, no. 11, pp. 2891-2903,

2004.

[57] S. Haykin, Neural Networks and Learning Machines, Hamilton: Pearson, 1999.

[58] Z. Benedek, "Pushrod vs. Overhead cam engines," 19 Jan 2018. [Online].

Available: https://allaboutdieselz.blogspot.com/2018/01/pushrod-vs-overhead-

cam-engines.html. [Accessed Dec 2019].

IX
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[59] E. Webb, "Introduction to Internal Combustion Engines," 2019. [Online].

Available: https://slideplayer.com/slide/13027058/. [Accessed Dec 2019].

[60] R. Muscoplat, "4.6L V8 Ford firing order," Rick's Auto Repair Advice, 2009.

[Online]. Available: https://ricksfreeautorepairadvice.com/4-6l-v8-ford-firing-

order/. [Accessed 2019].

[61] National Instruments, "Encoder Measurements: How-To Guide," National

Instruments, 9 Oct 2019. [Online]. Available: http://www.ni.com/tutorial/7109/en/.

[Accessed 2019].

[62] S. B. Quang, "Rotary Encoder Training Material," Maruei Vietnam Precision, 2013.

[63] AVL, "Crank Angle Encoder 365-Series Prodcut Description," AVL, Feb 2010.

[Online]. Available: https://www.avl.com/documents/10138/885965/angle-

encoder-365C-PS-2010-ENG+neu.pdf. [Accessed 2019].

[64] Kistler, "Test & Measurement Pressure," Kistler, 2018. [Online]. Available:

https://www.kistler.com/?type=669&fid=541&model=download&callee=frontend

. [Accessed 2019].

[65] AVNET ABACUS, "Piezoelectric pressure sensors," AVNET ABACUS, [Online].

Available:

X
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

https://www.avnet.com/wps/portal/abacus/solutions/technologies/sensors/pressure

-sensors/core-technologies/piezoelectric/. [Accessed Dec 2019].

[66] Kistler, "PiezoStar Pressure Sensor," Kistler, 2018. [Online]. Available:

https://www.kistler.com/?type=669&fid=102821. [Accessed Dec 2019].

[67] National Instruments, "Measuring Vibration with Accelerometers," National

Instruments, 14 Mar 2019. [Online]. Available: https://www.ni.com/en-

ca/innovations/white-papers/06/measuring-vibration-with-accelerometers.html.

[Accessed Dec 2019].

[68] Dytran, "Performance Specification Triaxial IEPE Acceleromter Model Number

3313A1," 2010.

[69] AutoZone, "Knock Sensor Repair Guide," AutoZone, [Online]. Available:

https://www.autozone.com/repairguides/Silverado-2008/Components-

Systems/Knock-Sensor-KS/_/P-0996b43f81b3c351. [Accessed 2019].

[70] I. Dibrani, "The new V8 POwer Unit for the BMW M3," BMW Heaven, 8 April

2007. [Online]. Available:

http://www.bmwheaven.com/index2.php?option=com_content&task=view&id=1

19&pop=1&page=6&Itemid=3. [Accessed 2019].

XI
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[71] National Instruments, "Introduction to Data Acquisition," National Instruments, 14

Mar 2019. [Online]. Available: https://www.ni.com/en-ca/innovations/white-

papers/06/introduction-to-data-acquisition.html#section-1450863631. [Accessed

Dec 2019].

[72] National Instruments, "How to Choose the Right DAQ Hardware for Your

Measurement System," National Instruments, 5 Mar 2019. [Online]. Available:

https://www.ni.com/en-ca/innovations/white-papers/11/how-to-choose-the-right-

daq-hardware-for-your-measurement-system.html. [Accessed Dec 2019].

[73] W. C. Chang, "An Improved Method of Investigation of Combustion Parameters in

a Natural Gas Fuelled SI Engine with EGR and H2 as Additives," University of

Birmingham, 2002.

[74] B. T. Narendiranath, "Journal Bearing Fault Detection Based on Daubechies

Wavelet," Archives of Acoustics, vol. 42, no. 3, pp. 401-414, 2017.

[75] C. S. Burrus, "Fast Approximate Fourier Transform via Wavelets Transform,"

Proceedings of SPIE - The International Society for Optical Engineering , 1997.

[76] W. Ribbens, "Chapter 4 - The Basics of Electronic Engine Control," Understanding

Automotive Electronics , pp. 135-182, 2017.

XII
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[77] P. Niedenfuhr, "The Piezoresistive Effect and Measuring Pressure," FirstSensor,

Feb 2019. [Online]. Available: https://blog.first-sensor.com/en/piezoresistive-

effect. [Accessed Dec 2019].

[78] E. Walter and R. Walter, "On-Board Diagnostics Background and Standards," in

OBD and CAN data acquisition for light-duty vehicles, SAE, 2018, pp. 25-36.

[79] AutoTap, "OBDII: Past, Present & Future," AutoTap, [Online]. Available:

http://www.autotap.com/techlibrary/obdii_past_present_future.asp. [Accessed

DEc 2019].

[80] J. Faiz, "A Survey on Condition Monitoring and Fault Diagnosis in Line-start and

Inverter-fed Broken Bar Induction Motors," IEEE International Conference on

Power Electronics, Drives and Energy Systems , 2012.

[81] B. Yazici, "An Adaptive Statistical Time-Frequency Method for Detection of

Broken Bars and Beraing using the Stator Current," IEEE Transaction on Industry

Applications, vol. 35, no. 2, pp. 442-452, 199.

[82] G. Yen, "Conditional Health Monitoring using Vibration Signatures," IEEE

Proceedings of the 38th Conference on Decision and Control , pp. 4493-4498,

1999.

XIII
M.A.Sc Thesis McMaster University
D. Joo Mechanical Engineering

[83] S. Goumas, "Classification of Washing Machine Vibration Signals Using Discrete

Wavelet Analysis for Feature Extraction," IEEE Transaction on Instrumentation

and Measurements, vol. 51, no. 3, pp. 497-508, 2002.

[84] I. Djurovic, "A Virtual Instrument for Time-Frequency Analysis," IEEE

Transcation on Instrumentation and Measurements, vol. 48, no. 6, pp. 1086-1092,

1999.

[85] A. Chen, "Internal combustion engine vibration analysis with short-term Fourier-

transform," IEEE 2010 3rd International Congress on Image and Signal

Processing, 2010.

XIV