CHAPTER 1

INTRODUCTION

Air compressors operate by converting power into potential energy stored in

pressurized air. They achieve this through the compression of atmospheric air,
leading to its reduction in volume and subsequent increase in pressure. Among the
various types of air-compression technologies, piston compressors, also known as
reciprocating compressors, hold a prominent position. These compressors function
on the principle of positive displacement, where a piston reciprocates within a
cavity to compress air. This technology offers reliability, simplicity, and versatility,
making it well-suited for a wide range of industrial applications.

Since the inception of the first compressor plant by Austrian engineer Viktor
Popp, air compressors have emerged as a transformative force in modern industry.
Compressed air, often referred to as the 'fourth utility' alongside electricity, water,
and gas, has become indispensable across a myriad of sectors. This report aims to
explore the significance, functionality, and diverse applications of piston
compressors within the broader context of industrial operations.

The versatility of compressed air renders it integral across numerous

industries. Compressed air serves two primary functions: as energy for mechanical
work (Energy Air) and as an active component in various processes (Active Air).
In Energy Air applications, compressed air powers pneumatic tools, operates
machinery, and facilitates material handling processes. Conversely, Active Air finds
application in industries demanding clean, dry air, such as pharmaceuticals, food
and beverage, and semiconductor manufacturing.

The industrial significance of piston compressors extends beyond traditional

1
applications like tire inflation. These compressors play a vital role in
manufacturing processes, ranging from automotive production to metal fabrication
and beyond. Moreover, they are crucial components in safety systems, powering
air brakes and ensuring timely deployment of airbags in vehicles.

Compressed air usage, storage, leakage, and efficiency are several factors
that influence the efficiency of a compressed air system. Usually for those
industrial plants having multiple types and sizes of air compressors, the efficiency
of each one is quite different, even for those of the same type and size. It greatly
influences overall system performance when demand is matched with the air
compressor or group operated at or near maximum efficiency levels. The air
compressor provides not only the required pressure air for the cooling equipment
and refrigeration plant, but also the raw material required for the air separation
plant.

Utilizing the air compressor is extremely extensive in many significant

departments such as the metallurgy, oil and chemical industry. The air compressor
production process is a time-varying, delay and nonlinear complex system and gas
consumption in the industrial field is irregular. Defects in the traditional control
method lead to the pressure instability of the compressed air outputted from the air
compressor in volatile gas-consumption situations, sometimes a wide range of
fluctuation arising. Therefore, the working state of the air compressor becomes
unstable. It results in both a waste of resources and greater energy consumption.
Therefore, guaranteeing the output pressure of the air compressor system steady is
extraordinarily significant for improving efficiency and saving energy for the
overall air separation system as discussed by Gongfa Li et al.

As stated, Air compressors (ACs) play a crucial role in various industries,

including aerospace, thermal plants, power generation, vehicle propulsion, and

2
pipeline systems. In aircraft design, ACs are integral for functions like fuel
injection

3
and metal finishing. They are also essential for quality control systems, particularly
in evaluating tolerable pressure for different aircraft components during transit,
ensuring operational success. Aircraft manufacturers rely heavily on high-quality
compressors at every production phase to maintain operational efficiency.

Contamination prevention is another critical aspect in aircraft ACs, as

components must avoid mixing air with lubricants. Compressor simulation is
essential for assessing the tolerable pressure of aircraft components, ensuring
safety and performance. Figure 1.1 illustrates common applications of ACs in
aircraft.

To design a complete line of compressor products, compress air systems are

utilized. Two main types of air compressors are widely employed: Rotary screw
ACs and Reciprocating ACs or Piston ACs, chosen based on specific application
requirements. These compressors are vital for maintaining functionality and
efficiency across various industrial sectors, highlighting their significance in
modern engineering and manufacturing processes.

Therefore, piston compressors shown in Figure 1.1, represent a cornerstone

of modern industrial operations, facilitating the widespread utilization of
compressed air across diverse sectors. Their reliability, efficiency, and adaptability
have propelled technological advancements and economic growth. As industries
continue to evolve, the role of piston compressors remains indispensable, driving
innovation and progress in the era of compressed air utilization.

4
Figure 1.1 A Brief Overview on Common Faults that occurs in Air Compressor

The following faults are extensively discussed by Divya et al:

 Bearing Fault: A malfunction in the compressor bearings, essential for

ensuring smooth operation of the compressor wheels, can lead to various
issues. This may involve bearing breakage or imbalance due to wear and
tear, resulting in increased friction within the machine and the emergence of
noise.

 Piston Fault: The piston, a critical component responsible for converting

rotary motion to linear motion and vice versa, plays a pivotal role in the
machine's performance. Any fault in the piston can lead to a reduction in the
machine's RPM and a decrease in overall noise levels.

 Flywheel Dysfunction: The flywheel serves as the primary storage unit for
kinetic energy within a machine. When wear and tear affect the flywheel, it

5
 may spin faster but store less kinetic energy. Consequently, an increase in
spinning speed may elevate the frequency of noise emitted.

 Inlet Valve Leakage: When leakage occurs in the compressor's inlet valve, it
results in a significant reduction in cylinder pressure, compromising the
compressor's efficiency. As a consequence, noise levels decrease due to
reduced piston speed and frequency.

 Outlet Valve Leakage: Conversely, leakage in the outlet valve leads to

heightened noise levels, characterized by high frequency and loud sounds.
Despite maintaining consistent cylinder pressure and piston speed, the
escaping air through the outlet valve contributes to additional noise
frequencies.

 Non-Return Valve (NRV) Fault: A malfunction in the non-return valve

occurs when air flow reverses, typically due to blockage or damage. This
imposes an impulsive load on the valve, resulting in the emergence of low-
frequency noises alongside other auditory disturbances.

 Rider Belt Faults: Rider belts, integral in transmitting power within the
compressor, require vigilant monitoring. Signs of faults, such as slippage,
wear, or improper tension, necessitate immediate attention. A healthy rider
belt maintains optimal tension, remains free from visible wear, and operates
silently, ensuring seamless power transmission.

As discussed by Chao Cui, once a fault happened, the economy and security
of pneumatic production will be negatively affected. In the aspect of economy, the
faults of air compressor may lead the production equipment lose power, thus
resulting in halting production and then a big decrease of corporate profits.

6
Meanwhile, the faults of air compressor will add an extra cost for companies
because of high repair cost. In the aspect of security, air compressor is a kind of
mechanical equipment with high-speed spinning motors, its gas tank has high
pressure for storing large quantities of compressed air. The faults of air compressor
will result in serious security accidents such as parts flying out at high speed or
cylinder explosion and so on, which are seen as serious threats to the safety of field
workers.

In essence, proactive monitoring and swift resolution of these identified

faults are imperative for sustaining optimal air compressor performance and
mitigating costly operational disruptions.

1.1 MACHINE LEARNING AND DEEP LEARNING

Machine Learning (ML) and Deep learning (DL) as discussed by Christian

et al is a revolutionary aspect of Artificial Intelligence (AI), allowing computers to
learn from data and improve performance without explicit programming. Coined
by Arthur Samuel et al, ML represents a shift in computing, automating learning
processes based on experiences. ML involves feeding high-quality data into
machines, training them using various algorithms tailored to the task.

ML has diverse applications, including image recognition, speech

processing, fraud detection, and recommendation systems like those used by
Netflix. Reinforcement learning enhances recommendation systems by allowing
agents to make decisions based on environmental feedback, improving user
satisfaction.

The ML process involves problem analysis, data collection, preparation,

model selection, training, evaluation, and deployment. ML encompasses

7
supervised, unsupervised, and reinforcement learning, handling vast datasets and
dynamic environments across various domains.

8
 Deep Learning (DL), a subset of ML, mimics human brain decision-making
using multi-layered neural networks. DL algorithms automate feature extraction,
expanding data analysis possibilities. Convolutional Neural Networks (CNNs)
excel in image recognition, while Recurrent Neural Networks (RNNs) process
sequential data like natural language.

DL's impact spans industries like finance, healthcare, and customer service.
In finance, DL powers algorithmic trading and risk assessment, while in healthcare,
it aids medical image analysis and disease diagnosis. DL-driven chatbots enhance
customer service experiences, and in cybersecurity, DL algorithms detect
anomalies to safeguard networks.

In industrial machinery like air compressors, fault diagnosis is crucial for

reliability and safety. ML and DL techniques automate fault analysis using vast
datasets and algorithms to detect subtle deviations from normal operations. These
techniques enable predictive maintenance, pre-emptively addressing issues to
prevent downtime.

ML and DL also optimize maintenance schedules and resource allocation,

extending equipment lifespan and enhancing operational efficiency. By harnessing
data-driven insights, industries ensure operational continuity and mitigate risks in
the era of smart manufacturing and industrial automation.

In summary, ML and DL revolutionize fault diagnosis processes, ensuring

operational continuity, reliability, and safety across industries. These techniques
leverage data-driven insights to pre-emptively address issues, optimize
maintenance, and enhance operational efficiency in the age of industrial
automation.

9
 CHAPTER 2

LITERATURE SURVEY

Verma N.K et al in year 2011 were discussed about data from reciprocating
air compressors, representing various operating conditions and potential faults, is
compiled into a tabulated dataset. Support Vector Machine (SVM) is a machine
learning algorithm commonly employed for classification tasks. In this study, four
SVM-based methods - one-against-one (OAO), one-against-all (OAA), fuzzy
decision function (FDF), and decision directed acyclic graph (DDAG) - are
compared alongside the proposed SVM method. SVMs classify data points by
identifying the optimal hyperplane that maximally separates different classes. The
proposed method focuses on parameter selection, particularly the radial basis
function (RBF) parameter σ, which significantly influences kernel computations.
The results obtained through implementation of all five techniques are thus
compared as per their accuracy rate in percentages and the performance of the
proposed method with 98.03 percent accuracy rate was found to be better than all
other classification methods. SVMs exhibit robustness against overfitting and are
adept at handling complex datasets. They demonstrate the ability to accurately
classify unseen data beyond the training set. However, their major drawback lies in
their high time complexity, making them less efficient for large datasets or real-
time applications. Additionally, proper tuning of SVM parameters and selection of
appropriate kernel functions can be challenging, requiring careful optimization.
Moreover, SVM performance may exhibit sensitivity to the choice of parameters,
such as the RBF parameter σ, necessitating meticulous adjustment for optimal
results. Due to the diverse nature of faults, the vibration signals indicating the
condition of air compressor components are inherently complex. Conventional
time-

1
0
frequency analysis and multi-resolution analysis often fail to capture a
comprehensive range of fault features, thus limiting their effectiveness.

In year 2015, Verma N.K et al explained about acoustic signals serve as the
foundation for fault diagnosis in reciprocating air compressors, recorded via
strategically positioned sensors on the compressor. The data undergoes meticulous
analysis to identify the most sensitive sensor placement locations, ensuring optimal
signal capture for fault diagnosis. Subsequently, the raw acoustic signals are
cleaned and prepared to eliminate noise and irrelevant data, facilitating further
analysis. Relevant features indicative of various fault conditions in the air
compressor are then identified and extracted from the pre-processed signals. A
classification model, employing machine learning techniques, categorizes the
extracted features to diagnose different fault states of the air compressor, enabling
early fault detection and mitigating costly breakdowns while minimizing
downtime. Remarkably, faults are accurately detected using acoustic signals from
just a single sensor, streamlining the monitoring process and reducing equipment
costs. Successfully implemented in real-time on an air compressor, the approach
showcases its practical applicability and reliability in industrial settings. However,
identifying the most sensitive sensor placement locations may require expertise
and experimentation, adding complexity to the implementation process. Currently
focused on diagnosing individual faults, adapting it to recognize multiple faults
simultaneously may demand additional research and development efforts. Further
optimization may also be necessary to ensure efficient online fault recognition,
potentially increasing the complexity of the model and implementation process.

The method described by Aravinth. S et al in year 2018 study about the

dataset comprises vibration signals captured from the compressor system using an
accelerometer sensor, aimed at fault detection. Statistical features are then
extracted from these vibration signals for each distinct fault condition, aiding in

1
1
the

1
2
characterization of unique fault patterns. To classify the vibration signals into
different fault conditions, a random forest classifier is employed. Random forest, a
machine learning algorithm, constructs numerous decision trees during training and
outputs the mode of the classes for classification tasks. The developed model
undergoes testing using a 10-fold cross-validation technique to evaluate its
performance in accurately categorizing fault conditions. This data modelling
approach holds promise for application across various machinery and industrial
equipment for fault diagnosis, with the potential for expansion beyond air
compressors through meticulous parameter tuning. Despite achieving high
accuracy, interpreting the underlying patterns and features identified by the random
forest classifier may pose challenges, diminishing the transparency of the diagnosis
process. Continuous monitoring and periodic updates may thus become necessary
to uphold the model's effectiveness over time, thereby increasing overall
maintenance overheads. Additionally, while the model demonstrates adaptability to
different compressor setups, fine-tuning of initial parameters may prove essential,
potentially requiring additional effort and resources.

The method described by Cui .C et al, in year 2019 explained about real-
time data from the air compressor's operation are collected and pre-processed to
eliminate outliers and dynamic operating points. The model is constructed using
Principal Component Analysis (PCA) and Back Propagation (BP) methods, which
are useful for selecting important parameters from the pre-processed dataset.
Additionally, a Multivariate State Estimation Technique (MSET) is employed to
develop a fault early warning model under normal conditions. For fault detection,
the deviation between observed and estimated values is calculated using a
deviation function. A sliding window statistical method is utilized to establish a
fault warning threshold. When the deviation exceeds this threshold, it triggers a
fault warning, indicating a potential fault in the air compressor. The advantages of
this approach include timely detection, increased safety, automation, and

1
adaptability. However, its drawbacks

1
include dependence on real-time data, complexity, high initial setup costs,
sensitivity to the threshold, limited scalability, and potential challenges in
integrating with existing systems.

In response to this limitation Yang W. S et al in year 2019 proposes an air

compressor fault diagnosis algorithm based on the lifting wavelet transform and
probabilistic neural network (PNN). This algorithm aims to overcome the
deficiency by decomposing the vibration signal from the air compressor using the
lifting wavelet transform, which effectively dissects the signal into various
frequency components. Subsequently, feature extraction is conducted, where
statistical measures such as peak value and kurtosis are computed from each
decomposition layer. These extracted fault features are then fed into a Probabilistic
Neural Network (PNN) for fault classification, leveraging PNN's adeptness in
pattern recognition tasks. The proposed method is compared with alternative fault
extraction techniques like Wavelet Packet Decomposition (WPT) and Continuous
Wavelet Transform (CWT), demonstrating its superiority in fault diagnosis
accuracy and efficiency. The lifting wavelet transform proves to capture fault
features more comprehensively, while the proposed method exhibits reduced
training time. However, it's essential to note that the performance of the proposed
method may vary when applied to different air compressor models or operating
conditions, potentially necessitating retraining or adaptation to suit diverse
scenarios. Furthermore, the efficacy of the model heavily relies on signal quality
and parameter tuning. Although the training time is efficient, deploying the PNN
model may require substantial computational resources, especially for large-scale
applications.

The method described by Wang. J et al in year 2019, focuses on the Fault

Diagnosis of Rotating Machines utilizing the Empirical Mode Decomposition
(EMD) Manifold (EMDM). EMD serves as a signal processing technique designed

1
to decompose intricate signals into a series of intrinsic mode functions (IMFs),
each

1
representing distinct oscillatory modes or trends within the signal. The proposed
EMDM method specifically addresses the challenge of mode mixing inherent in
EMD, wherein decomposed modes contain residual noise, by employing a noise-
assisted approach to bolster fault diagnosis. EMDM nonlinearly and adaptively
integrates fault-related modes harboring various noise sources via a manifold
learning algorithm, preserving genuine fault-related transients while suppressing
non-fault-related components and residual noise. Moreover, a novel criterion,
termed the TESK criterion, is introduced to automatically identify sensitive modes
among the IMFs generated by the EMD method. The method's efficacy is validated
through experimental studies conducted on gearbox and bearing faults, showcasing
superior performance compared to traditional noise-assisted EMD methodologies.
EMDM effectively segregates fault-related transients from noise, thereby
enhancing the accuracy of fault diagnosis in rotating machines. It achieves this by
adaptively fusing fault-related modes, facilitating nonlinear integration of diverse
noise sources while preserving authentic fault-related components. Additionally,
EMDM adjusts the local data distribution of the high-dimensional matrix using
random noise, ensuring satisfactory symmetry of the EMDM feature and
augmenting its reliability. However, the iterative process involving the
construction of high-dimensional matrices and adjustment of data distribution
using random noise may demand substantial computational resources, potentially
hindering real-time application or necessitating efficient optimization techniques.

In the realm of rotating machinery, multi-fault vibration signals often present

a higher level of complexity compared to single faults, posing a challenge for
human-designed fault feature sets to adequately address fault diagnosis tasks. In
response, Xue. Y et all in year 2020 proposes a method based on deep
convolutional neural network (DCNN) and support vector machine (SVM). The
DCNN is employed to automatically extract fault features from vibration signals of
rotating machinery, learning to discern subtle differences in local fault signals

1
across various

1
time scales and frequency ranges. Initially, the envelope spectrum derived from the
vibration signal using the Hilbert transform serves as input to the DCNN for fault
feature extraction. These extracted fault features are then combined with an SVM,
serving as the classifier for diagnosing multiple faults in bearings and rotors.
Furthermore, two time-domain features, handcrafted by human experts, are
incorporated into the model, resulting in the semi-DCNN method, which
amalgamates machine-derived insights with human experience in fault feature
design. The method demonstrates high diagnostic accuracy rates, with the semi-
DCNN method outperforming both DCNN and SVM individually, particularly in
diagnosing compound faults in bearings and rotors. Advantages of the approach
include the integration of human expertise in designing time-domain features,
which reduces preprocessing requirements. Moreover, fault feature extraction is
automated and adaptable to complex signals. However, drawbacks encompass the
complexity of the model and potential overfitting, which may lead to poor
generalization and reduced performance on unseen data.

The method described by Wang et al in the year 2020 deal with Data
Collection involves gathering vibration acceleration and pressure signals from the
reciprocating compressor air valve, while Signal Processing includes wavelet noise
reduction and decomposition of the vibration acceleration signal to eliminate
unwanted components. Time-Varying Singular Spectrum introduces time
information into the analysis, constructing a spectrum that offers insights into the
vibration signal over time. Fault Feature Extraction captures parameters indicating
specific fault types, used to construct fault feature vectors for Fault Diagnosis
using an SVM classifier, enabling accurate diagnosis. Advantages include Detailed
Analysis aiding in identifying specific fault types and Effective Fault Diagnosis
with high accuracy. However, Complexity in mathematical analysis and Signal
Processing, along with Data Collection Requirements, may pose challenges.
Additionally, Parameter Optimization and the method's Limited Applicability to air

1
valve faults may require extensive resources and limit direct adaptation to other
faults.

In order to detect potential bearing faults in rotating machines, Choudhary. A

et al in year 2021 utilizes thermal images. Thermal images of rolling-element
bearings under various conditions, including healthy and faulty states, are captured
using an infrared camera. Features are then extracted from these thermal images,
capturing temperature variations and patterns associated with different bearing
conditions. Both shallow artificial neural networks (ANNs) and deep learning,
specifically CNNs based on the LeNet-5 architecture, are trained using the
extracted features to classify different bearing conditions. The trained models serve
as classifiers to automatically identify the bearing conditions based on the thermal
images. CNNs, in particular, demonstrate superior performance in classifying
bearing faults compared to ANNs. This method offers a non-invasive and non-
contact approach to bearing fault diagnosis, facilitating early fault detection and
minimizing system shutdowns caused by bearing failures. However, while
effective for detecting bearing faults, this method may not capture other types of
faults or mechanical issues present in rotating machinery, necessitating
complementary diagnostic approaches. Additionally, the effectiveness of thermal
imaging-based fault diagnosis hinges on the availability and proper functioning of
infrared cameras, which may pose logistical challenges and incur additional costs.

The method described in the Kim. W et al, in year 2021 considered the
methodology involves utilizing virtual sensors, which are software-based
algorithms simulating the behaviour of physical sensors, to identify and diagnose
faults in air conditioners. These virtual sensors analyze data collected from a
limited number of physical sensors installed in the air conditioner, offering insights
into the system's performance and potential faults. One of the main advantages is
cost-effectiveness, as virtual sensors eliminate the need for additional physical

1
sensors, thereby

1
reducing initial setup costs. Moreover, they enhance fault detection by being able
to identify and isolate specific faults, even when multiple faults occur
simultaneously. Additionally, virtual sensors offer adaptability, as they can be
easily adjusted or reconfigured to accommodate different operating conditions and
fault scenarios. Furthermore, the fault detection and diagnostics process can be
automated, allowing for timely identification and resolution of issues without
human intervention. However, virtual sensors may face challenges accurately
detecting faults under certain conditions, such as condenser fouling, potentially
affecting the reliability of fault diagnosis. Moreover, there is a risk of false alarms,
as inaccurate estimations by virtual sensors may trigger unnecessary maintenance
or downtime.

Table 2.1 Summary of Literature Survey

Improvements that can be

Title of the paper Methodology used done
Utilizes MSET, PCA, and BP
A Novel Fault methods for fault prediction.
Measure and Early Employs AHP for weighting Enhance model generalization
Warning System for variables and sliding window and expand applicability to
Air Compressor for threshold settings. other equipment types.
Employs SVM-based
techniques including OAO,
An Optimized Fault OAA, FDF, and DDAG for
Diagnosis Method fault diagnosis. Uses optimized Improve time efficiency,
for Reciprocating Air SVM with RBF kernel particularly in reducing the
Compressors Based parameter tuning for high complexity of the DDAG
on SVM classification. method.
Air compressor fault
diagnosis based on
lifting wavelet Enhance feature extraction to
transform and minimize the training time
probabilistic neural Uses lifting wavelet transform while maintaining high
network for signal decomposition and accuracy, particularly in
feature extraction followed by differentiating faults with
Probabilistic Neural Network similar frequency
(PNN) for fault classification. characteristics.
1
 Improvements that can be
Title of the paper Methodology used done

Uses Empirical Mode

Decomposition (EMD) with a Further refinement in noise
Fault diagnosis of manifold learning algorithm to management to improve mode
rotating machines adaptively fuse noise-filtered separation and minimize
based on the EMD sensitive modes for fault interference from residual
manifold diagnosis of rotating machines. noises.

Combines Deep Convolutional

Multi-fault diagnosis Neural Network (DCNN) with
of rotating machinery Support Vector Machine (SVM) Further improvements in
based on deep for automatic and semi- distinguishing similar fault
convolution neural automatic feature extraction and types and reducing noise in
network and support classification of multi-faults in vibration signals for enhanced
vector machine rotating machinery. diagnosis accuracy.
Convolutional
neural network- CNN (using the LeNet-5
based bearing fault Utilizes thermal imaging and structure) outperforms ANN in
diagnosis of rotating machine learning models (ANN classifying bearing conditions
machine using and CNN) to diagnose bearing with a high accuracy of
thermal images faults in rotating machines. 99.80%.
Improve algorithms for
handling simultaneous faults,
Use of virtual sensors and fault increase validation and
impact models to detect and reliability of virtual sensors
Fault detection and diagnose faults in air under diverse conditions, and
diagnostics analysis conditioners, including field further evaluate the cost-
of air conditioners testing and data analysis to set effectiveness of virtual sensor
using virtual sensors diagnostic thresholds. systems.
The methodology involves
using acoustic signal analysis
for fault diagnosis in Improve recognition of multiple
reciprocating air compressors, simultaneous faults, develop
Intelligent Condition which includes data acquisition, adaptive models for ongoing
Based Monitoring sensor positioning, signal learning and changes in the
Using Acoustic preprocessing, feature machine's acoustic signature,
Signals for Air extraction, feature selection, and and optimize models for online
Compressors classification approaches. fault recognition.

1
 Improvements that can be
Title of the paper Methodology used done

The methodology includes The paper could explore

collecting vibration acceleration improvements in the efficiency
and pressure signals from an air of data collection and
Reciprocating valve, performing wavelet noise processing, enhance the
compressor air valve reduction, constructing a time- accuracy of the SVM classifier,
fault diagnosis varying window function for and possibly integrate more
method based on multi-fractal analysis, and advanced or different machine
time-varying singular employing SVM for fault learning techniques for better
spectrum diagnosis. fault detection.

CHAPTER 3

EXISTING METHOD

3.1 TRADITIONAL METHODS USED FOR FAULT DIAGNOSIS

In Rangarajan M. Tallam et al, Inseok Hwang et al and Andre Lemos et al,

deliberate upon a multitude of traditional methodologies of fault detection

 Visual Inspection: Visual inspection serves as an initial step in fault

diagnosis, involving a thorough examination of the compressor system for
any visible signs of wear, damage, or abnormalities. Maintenance personnel
visually inspect components such as belts, pulleys, hoses, fittings, and
connections for signs of wear, corrosion, leaks, or loose connections.

 Listening Tests: Listening tests involve carefully listening to the operating

sounds of the air compressor to detect any abnormal noises that may indicate
potential faults. Maintenance technicians pay close attention to sounds such
1
1
 as grinding, knocking, squealing, or rattling, which can signal issues with
bearings, pistons, valves, or other internal components.

 Vibration Analysis: Vibration analysis is a technique used to assess the

mechanical condition of rotating machinery, including air compressors.
Maintenance personnel use handheld vibration measurement devices or
sensors to detect and analyse vibrations in the compressor system. Excessive
vibration levels can indicate misalignment, unbalance, worn bearings, or
other mechanical issues.

 Pressure Testing: Pressure testing involves measuring and analysing the

pressure levels within the compressor system to identify any abnormalities
or deviations from expected values. Maintenance technicians use pressure
gauges, transducers, or data loggers to monitor the inlet and outlet pressures,
as well as the pressure within various components such as valves, cylinders,
and tanks.

 Temperature Monitoring: Temperature monitoring is critical for detecting

overheating or abnormal temperature levels in different parts of the
compressor system. Maintenance personnel use infrared thermometers,
thermocouples, or thermal imaging cameras to measure and monitor
temperatures at key locations, such as bearings, cylinders, valves, and motor
windings.

 Oil Analysis: Oil analysis involves sampling and analysing the lubricating oil
used in the compressor to assess its condition and detect any contaminants,
wear particles, or degradation products. Maintenance technicians collect oil
samples at regular intervals and send them to a laboratory for analysis. Oil

1
 analysis can provide valuable insights into the condition of internal
components such as bearings, pistons, and cylinders.

 Performance Testing: Performance testing involves evaluating the overall

performance and efficiency of the air compressor under various operating
conditions. Maintenance personnel conduct tests to measure parameters such
as airflow rate, power consumption, compression efficiency, and duty cycle.
Deviations from expected performance metrics can indicate underlying
faults or inefficiencies in the compressor system.

3.2 DRAWBACKS OF EXISTING METHODS

Traditional fault diagnosis methods have been the cornerstone of air

compressor maintenance practices for decades, offering valuable insights into the
condition of these crucial systems. However, as technology advances and industrial
environments become increasingly complex, the limitations of these traditional
methods have become more apparent.

One significant drawback of traditional fault diagnosis methods is their

reliance on manual inspection and testing. Visual inspections, listening tests, and
manual measurements are inherently prone to human error and subjectivity. This
introduces the risk of inaccuracies in fault detection and diagnosis, potentially
leading to ineffective maintenance decisions. Moreover, the specialized training
and expertise required for these methods make them time-consuming and resource-
intensive, posing challenges for maintenance personnel.

Additionally, traditional methods typically provide only a snapshot of the

compressor's condition at a specific point in time. This snapshot-based analysis

1
limits their ability to continuously monitor and track changes in performance or
detect early warning signs of faults. Consequently, maintenance efforts may
become reactive rather than proactive, resulting in increased downtime, repair
costs, and safety risks for industrial operations.

Furthermore, traditional fault diagnosis methods may struggle to analyze

complex data and identify subtle or intermittent faults. Vibration analysis, pressure
testing, and temperature monitoring techniques may overlook hidden issues or
transient abnormalities that could compromise compressor performance over time.
Without comprehensive data analysis tools and techniques, maintenance personnel
may find it challenging to uncover the root causes of recurring problems or
systemic issues within the compressor system.

In the context of modern high-performance air compressor systems equipped

with advanced control systems, sensors, and predictive analytics capabilities,
traditional methods face further challenges. These advanced systems generate vast
amounts of real-time data that traditional methods may struggle to effectively
process or interpret. Consequently, maintenance personnel may miss critical
insights or fail to leverage predictive maintenance strategies to prevent downtime
and optimize equipment performance.

To overcome these limitations, organizations must embrace innovative

technologies such as predictive maintenance, condition monitoring systems, and
remote monitoring solutions. These advanced tools enable proactive identification
and resolution of faults, minimizing downtime and maximizing the lifespan of air
compressor systems. By leveraging predictive analytics, maintenance personnel
can anticipate issues before they occur, optimizing maintenance schedules and
resource allocation to ensure the reliability, efficiency, and longevity of air
compressor systems in today's fast-paced industrial landscape.

2
 CHAPTER 4

PROPOSED METHODOLOGY

The proposed methodology conducts a comparative study of two different

deep learning models, long short-term memory (LSTM) and convolutional neural
network (CNN), using acoustic recordings from a single stage reciprocating air
compressor to classify and predict the various faults.

4.1 DATASET DESCRIPTIONS

Acoustic recordings were captured using a single-stage reciprocating-type

air compressor. The data is sampled at 16 kHz. The specifications of the air
compressor are as follows:

 Range of Air Pressure: 0 to 500 lb/m2, 0 to 35 kg/cm2.

 5 HP, 415 V, 5 Amp, 50 Hz, 1440 rpm induction motor.
 Type PR-15 Pressure Switch with a 100 – 213 PSI Range

Table 4.1 List of Audio Recordings in the Dataset

FAULT\ HEALTHY CONDITION NO OF RECORDINGS

Leakage inlet valve 225
Leakage outlet valve 225
Non-return valve 255
Piston ring 255

2
 Rider belt 255
Flywheel 255
Bearing 255
Healthy 255
Total Recording 1980

Table 4.1 contains a list of eight states, one of which is healthy while the
other seven are faulty. The recordings are stored in separate folders for each faulty
and healthy condition.

Figure 4.1 Overview of Subfolder Organization within the Dataset

Figure 4.1 shows that the dataset is divided into subfolders, each of which
contains 225 recordings, for a total of 1980 acoustic recordings. The recordings are
in waveform file format, which the software can readily process. Figure 4.2 shows
the waveforms of numerous faults, from which we can see the different variations
accessible in each method of fault recording.

2
 Figure 4.2 Time Domain Waveform of Various Faults

The waveform analysis provides critical insights, yet the intricate details
within the graphs are beyond the resolution of the human eye, making it
challenging to identify errors manually. Deep learning models have revolutionized
this process by offering an automated and highly effective solution for detecting a
wide range of issues in reciprocating air compressors. The dataset is divided as
described in Figure
4.3 for further processing.

Figure 4.3 Splitting of Dataset

2
4.2 GENERAL WORKING OF MACHINE LEARNING MODEL

The general working of the machine learning model can be well understood
from Figure 4.4.

Figure 4.4 Workflow of Machine Learning

In the workflow of a machine learning model, the process begins with the
retrieval of data, where relevant data is collected from various sources to form a
comprehensive dataset. This stage is crucial as the quality and quantity of data
directly impact the performance of the model. Following data retrieval, the next
step involves cleaning and exploring the data. This includes handling missing
values, removing outliers, and conducting exploratory data analysis (EDA) to
understand the patterns, correlations, and distributions within the data. These initial
steps set a strong foundation for building effective machine learning models.

2
 Once the data is clean and well-understood, the next phase is preparation and
transformation. Here, the data is formatted and engineered to better suit the
algorithms, for instance normalizing or scaling numerical inputs and encoding
categorical variables. After the data is prepared, the development and training of
the model commence. Algorithms are selected based on the problem type (e.g.,
regression, classification) and trained using the processed data. Following the
training, the model is validated and evaluated using metrics such as accuracy,
precision, and recall to ensure it meets the desired performance criteria. If the
model performs satisfactorily, it is then deployed into a production environment
where it can make predictions on new data. Post-deployment, continuous
monitoring and updates are essential to maintain its relevance and accuracy over
time, adjusting to new data and feedback to refine and improve the model's
performance.

4.3 WORKING PRINCIPLE OF PROPOSED METHODOLOGY

Deep learning is a subset of machine learning that uses neural networks with
multiple layers to process large and complex datasets effectively. These networks
mimic the human brain's structure, allowing deep learning models to perform
advanced tasks. For fault diagnosis of air compressor, two different models are
trained and tested in order to find the most suitable model. Figure 4.5 shows the
workflow of the methodology implemented.

2
 Figure 4.5 Block Diagram of Proposed Model Workflow

4.3.1 WORKING OF LSTM MODEL

 Dataset and Preprocessing:

The workflow begins with loading audio files from a specified directory.
Each file is expected to be a WAV file, indicating that the raw audio format
is used for input. The functions from librosa library are employed to read
these audio files, specifying a uniform sampling rate and duration for each
file to ensure consistency in audio quality and length across all samples.
Once loaded, the audio data undergoes feature extraction where Mel-
Frequency Cepstral Coefficients (MFCCs) are computed. MFCCs are chosen
because they effectively represent the power spectrum of sound based on a
linear cosine transform of a log power spectrum on a nonlinear mel scale of
frequency, making them highly suitable for audio recognition tasks as they
capture the timbral and textural aspects of the sound. The extracted MFCC

2
 features are then either padded or truncated to a fixed length to match the
expected input requirements of the LSTM model, ensuring that all input data
has uniform dimensions. The librosa library has a function that converts
audio signals to MFCC features, internally using the following formula:

𝑚 = 2595𝑙𝑜𝑔 𝑓
10(1+ ) (4.1)
100

Where,

m = frequency in Mels

f = frequency in Hertz

After the audio data is processed, the next step involves handling the
labels associated with each audio file, which denote the type of fault or
condition the compressor audio represents. The directory structure is used to
determine these labels, with each subdirectory name serving as the label for
all files within it. These labels are then encoded using a `LabelEncoder`
from the scikit-learn library, transforming them from textual descriptions
into a numerical format suitable for modeling. The encoded labels are
further converted into a one-hot encoded format using `to_categorical`,
which is necessary for the classification model to predict the probability of
each class as output. Lastly, the dataset is split into training and testing sets,
allowing for the validation of the model on unseen data, which is a crucial
step for assessing the model’s ability to generalize beyond the data it was
trained on.

 Model Building:

2
 In the architecture of the neural network designed using Keras Sequential
model, each layer is constructed to perform distinct and critical roles,

2
processing the input data linearly from one layer to the next. This model is
particularly effective due to its simplicity and versatility, making it suitable
for both simple and complex neural network tasks.

Figure 4.6 Block Diagram of LSTM Model

LSTM layer - The LSTM (Long Short-Term Memory) layer, a type of

recurrent neural network, is pivotal for handling sequence data such as time
series or text. It is adept at learning long-term dependencies within the data,
thereby mitigating common issues like vanishing gradients that plague
traditional RNNs. In this specific implementation, the LSTM is configured
with 128 units, where each unit outputs a 128-dimensional vector per time
step, encapsulating significant learned features from the sequence data. This
capability allows the LSTM to discern intricate patterns that are crucial for
tasks like classification or predictive analysis.

Dense layer - Subsequent to the LSTM, a Dense layer, or fully connected

layer, integrates these features across all units and maps them to the final
output classes using a SoftMax activation function. This function outputs a
probability distribution over the classes, indicating the predicted likelihoods
2
 of the input sequence belonging to each class, based on the features learned
by the LSTM.

The LSTM layer effectively captures the temporal dynamics and

dependencies within the input sequences, while the Dense layer interprets
these features to make final class predictions. The model is then trained with
80% of the dataset and 20% is used for testing the built model. The model is
then evaluated and predictions of unseen data is done.

4.3.2 WORKING OF CNN MODEL

 Data Loading and Processing

The initial stage in utilizing Convolutional Neural Networks (CNNs) for
audio classification is to load and process the audio files, which are
commonly in the WAV format. This is undertaken using the librosa library,
a powerful tool for audio analysis, to extract Mel-Frequency Cepstral
Coefficients (MFCCs). These coefficients are paramount as they distill the
audio signal into a form that mirrors the human auditory system’s response
and succinctly represents the power spectrum of the sound. To maintain
uniformity across the dataset, each MFCC feature set is padded to a fixed
length, ensuring the model receives inputs of consistent dimensions. This
meticulous process involves not only standardizing the length of each
feature set but also normalizing the data to aid in the model's learning
efficiency. Additional features such as Chroma and spectral contrast may
also be extracted depending on the complexity required by the application,
enhancing the model's ability to differentiate between diverse acoustic
environments.

2
 This pre-processing phase also involves translating the categorical labels,
indicative of the air compressor's operational state, from text to a numerical
form using one-hot encoding. Such encoding is essential for the subsequent
classification process within the CNN model. This meticulous preparation of
features and labels sets the stage for effective model training and accurate
fault detection. Ensuring the balance of classes within training batches is
critical to avoid biases in the model's predictions, necessitating techniques
like stratified sampling or synthetic data generation to manage unbalanced
datasets effectively.

 Model Building

Figure 4.7 Block Diagram of LSTM Model

The model's architecture is built on a Sequential framework provided

by Keras, starting with convolutional layers (Conv2D) that apply learnable
filters to the MFCCs. These layers are adept at recognizing local patterns in
the input data, making them particularly advantageous for tasks like acoustic
signal classification. The max-pooling layers (MaxPooling2D) that follow
are crucial for reducing feature map sizes, thereby condensing the
information and minimizing computational complexity. This architectural
choice facilitates the extraction of salient features from the audio
signals while
3
 maintaining a manageable computational load. Subsequent layers, such as
batch normalization, can be included to stabilize the learning process by
normalizing the activations of previous layers, enhancing the model's overall
performance and efficiency.

A flatten layer converts the pooled feature maps into a one-

dimensional vector, suitable for processing via densely connected layers.
These dense layers (Dense) are the decision-makers of the network,
leveraging the extracted features to infer the class probabilities. To combat
overfitting, dropout layers randomly omit a subset of features during
training, which promotes the development of a more robust and generalized
model. The inclusion of regularization techniques like L2 regularization can
further penalize large weights in the model, ensuring smoother decision
boundaries. The CNN is perfected by a SoftMax activation function in the
output layer, providing a probability distribution across the classes. This
model, armed with these characteristics and compiled with appropriate
performance metrics, is then trained on the processed data. Through training
and evaluation on a held- out test set, the CNN demonstrates its capacity to
discern intricate patterns in acoustic data, which is essential for pinpointing
faults in air compressors, thereby underscoring the transformative potential
of CNNs in predictive maintenance.

4.4 EVALUATION OF MODELS

To assess the effectiveness of the designed models, we have considered four

metrics that help us gauge the model's performance. They are:

3
Accuracy: This metric measures the overall correctness of the model and is defined
as the ratio of correctly predicted observations to the total observations. The
formula is,

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)+𝑇𝑟𝑢𝑒

Accuracy =
𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠(𝑇𝑁) (4.2)
𝑇𝑜𝑡𝑎𝑙 𝑂𝑏𝑠𝑒𝑟𝑣𝑎𝑡𝑖𝑜𝑛𝑠

Sensitivity/Recall: Sensitivity measures the proportion of actual positives that are

correctly identified as such and is particularly important in medical testing. The
formula for sensitivity is,

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)

Sensitivity = 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝐹𝑁) (4.3)

Precision: Precision assesses the accuracy of the positive predictions made by the
model, i.e., the proportion of positive identifications that were actually correct.

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)

Precision = 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝑇𝑃)+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝐹𝑃) (4.4)

Specificity: Specificity measures the proportion of actual negatives that are correctly
identified and is crucial for tests that need to rule out diseases.

𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝑇𝑁)

Specificity = 𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (𝑇𝑁)+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 (𝐹𝑃) (4.5)

Each of these metrics provides valuable insights into different aspects of a

model's performance, particularly in classifications tasks where the balance
between types of errors can be critical.

3
 CHAPTER 5

RESULTS AND DISCUSSION

Our research involved a thorough analysis into the effectiveness of LSTM

and CNN models for the crucial task of predicting faults within air compressor
systems. Through meticulous stages of training, rigorous testing, and thorough
validation, we scrutinized the performance of these models using a suite of
fundamental metrics: accuracy, recall, precision, specificity, and loss. Our
overarching objective was to discern the most optimal combinations of parameters
governing each model's architecture. These parameters encompassed a wide array
of factors, including the number of filters, activation functions, sizes of pooling
windows, spatial dimensions, neuron quantities, optimizers, types of loss functions,
epochs, batch sizes, and LSTM units. The resultant tables stand as comprehensive
repositories, offering a detailed comparative analysis across diverse configurations.
This wealth of data not only highlights the strengths and weaknesses of each model
but also provides actionable insights aimed at refining fault prediction accuracy
within air compressor systems.

Table 5.1 Optimal Parameters Tested in LSTM Model

Development

3
 Table 5.2 Optimal Parameters Tested in CNN Model Development

Following meticulous evaluation shown in table 5.1 and 5.2, we identified

the most optimal parameter combinations for both LSTM and CNN models in
predicting faults within air compressor systems. Subsequently, we present detailed
analyses encapsulated within confusion matrices and accuracy-versus-number-of-
epochs graphs for each model.

5.1 RESULTS

The confusion matrices provide a visual depiction of the models' proficiency

in identifying the seven distinct faults present within air compressor operations.
The intensity of the blue color within the matrices reflects the accuracy of fault
detection. Darker shades of blue indicate higher accuracy, ranging from 30 for the
CNN model to 50 for the LSTM model, while lighter shades suggest lower
accuracy, ranging from 0 to 20 for the CNN model and 0 to 30 for the LSTM
model, indicating areas for potential improvement. In Figure 5.1, we obtain a high
score of 31 for the number of correct predictions made by the model and a lowest
score of 25 for the LIV fault. Considering Figure 5.2, we achieve a high score of
51 for the number of correct predictions made by the model and a lowest score of

3
31 for the LIV and NRV faults.

3
By scrutinizing these matrices, stakeholders can gain insights into the models'
strengths and weaknesses in classifying specific fault types. This nuanced
understanding facilitates targeted enhancements to the models' predictive
capabilities, thereby fostering more reliable fault detection mechanisms.

In summary, the presented confusion matrices and accuracy-versus-number-

of-epochs graphs provide comprehensive insights into the LSTM and CNN models'
fault prediction capabilities within air compressor systems. Leveraging these
visualizations, stakeholders can refine model parameters, optimize training
protocols, and enhance fault detection accuracy. Ultimately, this iterative
refinement process culminates in the development of robust predictive models
capable of mitigating operational risks and maximizing system reliability.

Figure 5.1 Confusion Matrix of CNN Model

3
 Figure 5.2 Confusion Matrix of LSTM Model

Furthermore, the accuracy-versus-number-of-epochs graphs offer a dynamic

portrayal of the models' performance trajectory during the training process. As the
number of epochs, or iterations, increases, the graphs demonstrate a corresponding
improvement in accuracy. For instance, we have achieved a very high accuracy of
100% with 35 epochs in Figure 5.3 for the CNN Model and 99.68% with 25
epochs in Figure 5.4 for the LSTM Model. This incremental enhancement
underscores the iterative learning process inherent within neural network training.
By visualizing the convergence of accuracy over epochs, stakeholders can discern
the models' learning dynamics and ascertain convergence points where further
training may yield diminishing returns. Moreover, these graphs serve as valuable
tools for monitoring model performance and gauging the efficacy of parameter
adjustments over successive training cycles.

3
 Figure 5.3 Accuracy vs Epochs Plot for CNN

Figure 5.4 Accuracy vs Epochs Plot for LSTM

In our project, both LSTM and CNN models have demonstrated remarkable
strengths in predicting faults within air compressor systems, underscoring their
critical importance. The LSTM model excels in capturing long-term dependencies
in sequential data, making it particularly adept at analyzing time-series data
common in fault prediction scenarios. Its ability to retain memory over time
enables it to effectively learn complex patterns within the operational data of air
compressors, thereby enhancing fault detection accuracy. On the other hand, the
3
CNN model

3
showcases exceptional feature extraction capabilities, leveraging convolutional
layers to identify spatial patterns within sensor data. This spatial awareness is
invaluable for detecting spatially distributed faults, such as anomalies in
compressor components.

Figure 5.5 Test Results of CNN Model

Figure 5.5 shows the results obtained for LSTM model. The LSTM
model demonstrated outstanding performance with a test loss of 0.0020 and
an accuracy of 99.68%. It achieved a precision of 99.67%, a recall of
99.35%, and a specificity of 99.93%, indicating its high reliability and
accuracy in fault detection for air compressors, which underscores its
potential as a predictive maintenance tool.

3
 Figure 5.6 Test Results of LSTM Model

Figure 5.6 shows the results obtained of CNN model. The CNN model
showcased excellent results with perfect test accuracy and specificity, both
achieving a score of 1.0000. The model also recorded high precision of 1.0000 and
a recall of 0.9957, alongside a test loss of 0.0749. These metrics indicate the
model's exceptional capability in accurately identifying and classifying faults in air
compressors, affirming its effectiveness for use in predictive maintenance
scenarios.

5.2 FUTURE SCOPE

The integration of both LSTM and CNN models in our project holds
significant promise for future improvements. One potential avenue for
enhancement is the utilization of real-time acoustic signals for fault detection. By
incorporating audio data from the compressor system, we can leverage the models'
strengths in sequential and spatial data analysis to detect anomalous sounds
indicative of potential faults. This expansion into multi-modal data fusion could

3
greatly enhance

3
the robustness and accuracy of fault prediction systems, providing early warnings
for maintenance interventions.

Furthermore, there is a compelling opportunity to extend the capabilities of

our models beyond fault detection to predict the overall health status of air
compressors. By incorporating additional sensor data related to temperature,
pressure, and vibration, we can develop comprehensive health monitoring systems.
These systems can not only identify faults but also assess the overall condition of
the compressor and predict potential failure trajectories. Such predictive
maintenance capabilities would enable proactive interventions, minimizing
downtime and extending the lifespan of critical equipment.

CHAPTER 6

CONCLUSION

The objective of this project on fault diagnosis of air compressors is to

reduce machine downtime, decrease cost-related losses, and enhance productivity
in various industries where air compressors are critical. The proposed system
utilizes two different models, CNN and LSTM, to classify various faults and also
to detect the healthy state of the air compressor. Both models achieve an accuracy
greater than 99.5%, making them highly suitable for classification tasks and
demonstrating strong predictive performance on unseen data. In conclusion, this
project employs deep learning models for predictive maintenance, significantly
improving the productivity of industries that rely on air compressors. With the
necessary hardware setup, this model can be implemented for real-time fault
detection.
4
 REFERENCES

1. Aravinth, S., & Sugumaran, V. (2018). Air compressor fault diagnosis

through statistical feature extraction and random forest classifier. Progress in
Industrial Ecology, an International Journal, 12(1-2), 192-205.

2. Choudhary, A., Mian, T., & Fatima, S. (2021). Convolutional neural

network- based bearing fault diagnosis of rotating machine using thermal
images. Measurement, 176, 109196.

3. Cui, C., Lin, W., Yang, Y., Kuang, X., & Xiao, Y. (2019). A novel fault
measure and early warning system for air compressor. Measurement, 135,
593-605.

4. Cui, C., Lin, W., Yang, Y., Kuang, X., & Xiao, Y. (2019). A novel fault
measure and early warning system for air compressor. Measurement, 135,
593-605

5. Divya, M. N., C. K. Narayanappa, and S. L. Gangadharaiah (2021).

"Modeling a Fault Detection Predictor in Compressor using Machine
Learning Approach based on Acoustic Sensor Data." International Journal of
Advanced Computer Science and Applications 12.9.

6. I. Hwang, S. Kim, Y. Kim and C. E. Seah (2010), "A Survey of Fault

Detection, Isolation, and Reconfiguration Methods," in IEEE Transactions
on Control Systems Technology, vol. 18, no. 3, pp. 636-653.

7. Janiesch, C., Zschech, P. & Heinrich, K(2021). Machine learning and deep
learning. Electron Markets 31, 685–695.

4
8. Kim, W., & Lee, J. H. (2021). Fault detection and diagnostics analysis of air
conditioners using virtual sensors. Applied Thermal Engineering, 191,
116848.

9. Lemos, A., Caminhas, W., & Gomide, F. (2013). Adaptive fault detection
and diagnosis using an evolving fuzzy classifier. Information Sciences, 220,
64- 85.

10.Li, Gongfa, et al. "Intelligent control of air compressor production process."

Applied Mathematics & Information Sciences 7.3 (2013): 1051.

11.R. M. Tallam et al., "A Survey of Methods for Detection of Stator-Related

Faults in Induction Machines," in IEEE Transactions on Industry
Applications, vol. 43, no. 4, pp. 920-933, July-aug. 2007, doi:
10.1109/TIA.2007.900448.

12.Verma, N. K., Roy, A., & Salour, A. (2011, June). An optimized fault
diagnosis method for reciprocating air compressors based on SVM. In 2011
IEEE International Conference on System Engineering and Technology (pp.
65-69). IEEE.

13.Verma, N. K., Sevakula, R. K., Dixit, S., & Salour, A. (2015). Intelligent
condition-based monitoring using acoustic signals for air compressors. IEEE
Transactions on Reliability, 65(1), 291-309.

14.Wang, J., Du, G., Zhu, Z., Shen, C., & He, Q. (2020). Fault diagnosis of
rotating machines based on the EMD manifold. Mechanical Systems and
Signal Processing, 135, 106443.

4
15.Wang, Jindong., Li, Yanyang., Zhao, Haiyang., Yi-qi, Gao., Yu, Delong.,
Chen, Xin. (2020). Reciprocating compressor air valve fault diagnosis
method based on time-varying singular spectrum.

16.Xue, Y., Dou, D., & Yang, J. (2020). Multi-fault diagnosis of rotating
machinery based on deep convolution neural network and support vector
machine. Measurement, 156, 107571.

17.Yang, W. S., Su, Y. X., & Chen, Y. P. (2019, October). Air compressor fault
diagnosis based on lifting wavelet transform and probabilistic neural
network. In IOP conference series: materials science and engineering (Vol.
657, No. 1,
p. 012053). IOP Publishing.