See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/277723565

Machine Learning for Predictive Maintenance: A Multiple Classifier Approach

Article  in  IEEE Transactions on Industrial Informatics · June 2015

DOI: 10.1109/TII.2014.2349359

CITATIONS READS
168 5,826

5 authors, including:

Gian Antonio Susto Simone Pampuri

University of Padova University of Pavia
67 PUBLICATIONS   649 CITATIONS    22 PUBLICATIONS   407 CITATIONS   

SEE PROFILE SEE PROFILE

Sean Mcloone Alessandro Beghi

Queen's University Belfast University of Padova
183 PUBLICATIONS   1,821 CITATIONS    214 PUBLICATIONS   1,657 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Smart Street View project

PhD Maynooth View project

All content following this page was uploaded by Gian Antonio Susto on 17 February 2018.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 1

Machine Learning for Predictive Maintenance:

a Multiple Classifier Approach
Gian Antonio Susto, Andrea Schirru, Simone Pampuri, Seán McLoone Senior Member, IEEE,
Alessandro Beghi Member, IEEE

Abstract—In this paper a multiple classifier machine and newly introduced state-of-the-art algorithms. At the
learning methodology for Predictive Maintenance (PdM) is same time the efficient management of maintenance
presented. PdM is a prominent strategy for dealing with activities is becoming essential to decreasing the costs
maintenance issues given the increasing need to minimize
downtime and associated costs. One of the challenges with associated with downtime and defective products [14],
PdM is generating so called ’health factors’ or quantitative especially in highly competitive advanced manufactur-
indicators of the status of a system associated with a given ing industries such as semiconductor manufacturing.
maintenance issue, and determining their relationship to op- Approaches to maintenance management can be
erating costs and failure risk. The proposed PdM methodol-
grouped into three main categories which, in order of
ogy allows dynamical decision rules to be adopted for main-
tenance management and can be used with high-dimensional increasing complexity and efficiency [22], are as follows:
and censored data problems. This is achieved by training i. Run-to-Failure (R2F) - where maintenance interventions
multiple classification modules with different prediction are performed only after the occurrence of failures. This
horizons to provide different performance trade-offs in terms is obviously the simplest approach to dealing with main-
of frequency of unexpected breaks and unexploited lifetime
tenance (and for this reason it is frequently adopted),
and then employing this information in an operating cost
based maintenance decision system to minimise expected but it is also the least effective one, as the cost of
costs. The effectiveness of the methodology is demonstrated interventions and associated downtime after failure are
using a simulated example and a benchmark semiconductor usually much more substantial than those associated
manufacturing maintenance problem. with planned corrective actions taken in advance. ii.
Index Terms—Classification Algorithms, Data Mining, Ion Preventive Maintenance (PvM) - where maintenance ac-
Implantation, Machine Learning, Predictive Maintenance, tions are carried out according to a planned schedule
Semiconductor Device Manufacture. based on time or process iterations. With this approach,
also referred to as scheduled maintenance, failures are
I. I NTRODUCTION usually prevented, but unnecessary corrective actions are
often performed, leading to inefficient use of resources
The increasing availability of data is changing the and increased operating costs.
way decisions are taken in industry [17] in important iii. Predictive Maintenance (PdM)1 - where maintenance is
areas such as scheduling [15], maintenance management performed based on an estimate of the health status of
[24] and quality improvement [6], [23]. Machine learning a piece of equipment [7]. PdM systems allow advance
(ML) approaches have been shown to provide increas- detection of pending failures and enable timely pre-
ingly effective solutions in these areas, facilitated by the failure interventions, thanks to prediction tools based
growing capabilities of hardware, cloud-based solutions, on historical data, ad hoc defined health factors, statistical
inference methods, and engineering approaches.
Manuscript received March 17, 2014. Accepted for publication Au-
gust 7, 2014. Among statistical inference based methods, those
Copyright 2014 IEEE. Personal use of this material is permitted. based on ML are the most suitable for dealing with mod-
However, permission to use this material for any other purposes
must be obtained from the IEEE by sending a request to pubs-
eling of high-dimensional problems, such as those aris-
[email protected] ing in semiconductor manufacturing where hundreds
G.A. Susto (corresponding author) is with the Department of In- or thousands of physical variables (pressures, voltages,
formation Engineering, University of Padova, Italy and with Statwolf
LTD, Ireland. E-mail: [email protected].
currents, flows, etc.) act on the process [10], [26].
A. Schirru and S. Pampuri are with the National University of In this paper a new PdM methodology based on mul-
Ireland, Maynooth and with Statwolf LTD, Ireland. tiple classifiers is introduced for integral type faults (the
S. McLoone is with Queen’s University Belfast, United Kingdom and
with the National University of Ireland, Maynooth. most frequent in semiconductor manufacturing), a term
A. Beghi is with the Department of Information Engineering, Uni- which describes the failures that happen on a machine
versity of Padova, Italy. due to the accumulative ’wear and tear’ effects of usage
This work has been done within the framework of INTEGRATE (In-
tegrated Solutions for Agile Manufacturing in High-mix Semiconduc-
tor Fabs), an European Nanoelectronics Initiative Advisory Council’s 1 Other authors [8] refer to this class of maintenance approaches as
project. Condition-Based Maintenance
0000–0000/00$00.00©2014 IEEE
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 2

and stress on equipment parts. Even if no direct evidence From a ML perspective, supervised approaches re-
of process/machine degradation is available, PdM tools quire the availability of a dataset S
exploit process and logistic variables collected during n
S = {xi , yi }i=1 , (1)
production to identify the ’footprint’ of this degradation
in the data. The proposed methodology, referred to as where a couple {xi , yi } (called observation) contains the
Multiple Classifier (MC) PdM, can effectively deal with information related to the i-th process iteration. Here,
the unbalanced datasets that arise in maintenance classi- vector xi ∈ R1×p contains information related to the p
fication problems, that is datasets where the observations variables associated with available process or logistic
relating to normal production greatly outnumber the ob- information. Depending on the type of output y two
servations associated with abnormal/faulty production classes of supervised problem are possible:
[22]. It also allows planning of maintenance schedules i. if y assumes continuous values a regression problem is
using a statistical cost minimization approach. obtained;
The remainder of the paper is structured as follows: ii. if y assumes categorical values a classification problem
Section II provides a brief literature review and an results.
introduction to ML-based approaches to PdM, while In PdM problems, regression-based formulations gen-
Section III is dedicated to presenting the proposed MC erally arise when predicting the Remaining Useful Life
PdM methodology. Tuning guidelines and classification of a process/equipment, directly [2] or indirectly (eg.
approaches are discussed in Section IV. Then in Section through the computation of the conditional reliability
V, to demonstrate the principles and efficacy of MC [5]), while classification-based PdM formulations occur
PdM, results are presented comparing MC PdM with when seeking to discriminate between healthy and un-
PvM for a benchmark semiconductor manufacturing healthy conditions of the system being monitored [1].
maintenance problem, namely, changing of filaments in While classification tools are a natural choice for dis-
ion implantation tools. Finally, in Section VI, concluding tinguishing between faulty and non-faulty process itera-
remarks and possible extensions of the work are dis- tions based on observed process data, they do not map
cussed. naturally to health factors that can be extrapolated for
maintenance related decision making, unlike for exam-
ple, regression models of Remaining Useful Life. In the
II. L ITERATURE R EVIEW following a methodology based on multiple classifiers is
presented to address this limitation.
Maintenance issues can be completely different in
nature and the predictive information to be fed to the
III. M ULTIPLE C LASSIFIER P D M
PdM module has in general to be tailored to the par-
ticular problem at hand, thus justifying the presence A. Classification for PdM
in the literature of many different approaches to PdM Let us suppose that data regarding N maintenance
PN
[12]. However, PdM-related solutions based on ML tech- cycles are available, for a total of n = i=1 ni machine
niques seem to be among the most popular (see e.g. [5] runs. We define the matrix X ∈ Rn×p containing all the
and [21] with reference to semiconductor manufacturing collectible information
applications).  T
X = x1 x2 . . . x n (2)
ML-based PdM can be divided into two main classes:
i. supervised - where information on the occurrence of regarding p physical variables or logistic information on
failures is present in the modeling dataset; the process or on the tool that are used as inputs to
ii. unsupervised - where logistic and/or process informa- the PdM module. Information on maintenance events is
tion is available, but no maintenance related data exists. contained in the variable Y . If during the run the fault
The availability of maintenance information mostly under consideration takes place, then the observation
depends on the nature of the existing maintenance man- is marked as faulty (F), and not faulty (NF) otherwise.
agement policy: in the case of R2F policies the data Accordingly:
related to a maintenance cycle (the production activity 
F if iteration i is faulty
between two successive failure events) is available and Y (i) = yi = . (3)
N F if iteration i is not faulty
therefore supervised approaches can readily be adopted;
on the other hand, when PvM policies are currently in When dealing with R2F data, each maintenance cycle
place the full maintenance cycle may not be observable ends with a failure, hence the available data contains N
given the fact that maintenance is generally performed samples for class F and n − N samples for class N F .
well in advance of any potential failure, and hence only Based on X and Y , a classifier learns a decision rule f (·)
unsupervised learning approaches are feasible. When that assigns one of the two classes {F, N F } to each point
possible, supervised solutions are evidently preferable: in the input space Rp .
given the wide diffusion of R2F maintenance policies in The formulation presented above is weak from a PdM
industry, and hence the availability of suitable datasets, perspective in two aspects:
we consider in this work a supervised approach to PdM. 1. only the current process iteration is classified, i.e. no
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 3

fault prevention policy can be implemented. x2 x2 x2

2. no operating cost optimization policy is enforced.
As far as total operating cost minimization is con-
cerned, two key metrics can be defined, namely2 :
1) Frequency of Unexpected Breaks (ρUB ) - percentage
of failures not prevented; x1 x1 x1
2) Amount of Unexploited Lifetime (ρUL ) - average
number of process iterations that could have been y (1) y (2) y (k)
run before failure if the preventative maintenance
suggested by the maintenance management mod-
Classifier 1 Classifier 2 Classifier k
ule had not been performed.
Different costs, cUB and cUL , can be associated with ρUB
and ρUL , respectively, where generally cUB >> cUL , given Fig. 1. 2-dimensional example for the k classifiers of the MC PdM
that cUB relates to unplanned interruption of produc- methodology: the labels N F (orange circle) and F (blue plus) change
tion. It is not possible to simultaneously minimize both with failure horizon m resulting in a different classification problem
for each classifier
metrics, rather an optimal trade-off solution is sought
through minimization of the total operating costs, as
defined by the weighted sum cycle (of length ni ) we have
J = ρUB cUB + ρUL cUL . (4)

(j) N F if t ≤ ni − m(j)
yt = , (5)
F otherwise
As formulated above, the classification problem is not
suitable for maintenance managements purposes, be- where m(j) ∈ N+ ; the associated PdM performance of
cause it doesn’t facilitate identification of policies that (j) (j)
the j-th classifier is indicated by ρUB and ρUL .
minimise J in (4). In the presented approach the k classifiers work in
From a classification perspective, the formulation de- parallel and a maintenance event is triggered by the
scribed in Eq. 3 also suffers from the fact that the dataset decision making logic based on the following operating
is very unbalanced or skewed, with N samples in class F costs minimization philosophy: given the current costs
and n − N samples in class N F and N  n. Classifica- cUB (t) and cUL (t) at time t, the MC PdM suggests a
tion using unbalanced datasets generally results in poor corrective action when the j ∗ -th classifier
prediction accuracy and generalization performance [9].
These issues are addressed by the MC PdM method- j ∗ = arg min J (j) (t)
j=1,...,k
ology presented in the next subsection.
(j) (j)
= arg min ρUB cUB (t) + ρUL cUL (t) (6)
j=1,...,k
B. MC PdM Concept outputs a label F classification.
A possible approach to preventing unexpected failures The training procedure for MC PdM is sketched in
is to consider a different classification problem where, Algorithm 1. The performance of each classifier is eval-
instead of only labeling the last iteration of a mainte- uated using Repeated Random Sub-Sampling Validation
nance cycle as F , we label as F the last m iterations. [16], also known as Monte Carlo crossvalidation (MCCV): Q
From a PdM perspective, this approach allows us to simulations are performed by randomly splitting the N
provide more conservative maintenance recommenda- maintenance cycles into a training dataset of NTR = bN qc
tions by choosing larger values for the failure horizon maintenance cycles and a validation dataset of NVL =
m. Moreover, by assigning more samples to the F class dN (1 − q)e maintenance cycles, with 0 < q < 1. It has
we reduce the skewness of the dataset (N m samples in been shown [20] that MCCV is asymptotically consistent
class F and n − N m samples in class NF). resulting in more pessimistic predictions of the test
This can be repeated for k different values of the hori- data compared with full crossvalidation. Algorithm 1 is
zon m. In the proposed MC PdM approach, k different applicable to any choice of classification algorithm. In
classifiers run on the module, each one facing a different general (as in the case of the two algorithms presented
classification problem and therefore providing differ- in Section IV-B) to set up the classifiers the tuning of one
ent maintenance management performance outcomes in or more hyper-parameters is required. This is achieved
terms of ρUB and ρUL . through crossvalidation with the Missclassification Rate
The MC PdM scheme is depicted in Fig. 1. The j-th (MCR)
classifier is associated with the labels Y (j) , where, con-
sidering the t-th process iteration of the i-th maintenance MCR[%] = Percentage of missclassified samples

2 The
employed as a performance indicator.
R2F approach clearly guarantees a total of ρUB = 1 and ρUL =
0, while the performance of PvM and PdM approaches depend on Fab integration and on-line operation of the MC PdM
some tuning parameters as will be illustrated in Section IV module are described in Fig. 2. Estimates of ρUB and
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 4

Algorithm 1: MC PdM Training IV. I MPLEMENTATION D ETAILS

 k
Data: X, Y , k, q1 , q2 , Q1 , Q2 , m(j) j=1 A. Failure Horizon Selection
 k Performance of the MC PdM methodology increases
Result: Classifiers f (j) j=1 and PdM performances
with the number of classifiers, k, since each classifier pro-
1. Let ρUB = [ · ] and ρUL = [ · ] (empty vectors) vides more information on the health status of the pro-
for j = 1 to k do cess. Unfortunately k cannot be considered as a degree
2. Compute Y (j) as in Eq. 5 for i = 1 to Q1 do of freedom in the MC PdM design since it is constrained
3. Randomly split the maintenance cycles
by the available computational and storage capabilities
between training and validation samples,
and by the algorithm chosen for the classification. In
keeping the ratio q1
the following we propose two strategies for defining the
4. Compute the MCR of the classifier with
failure horizons of the k classifiers.
different hyper-parameters
The k classifiers are distributed in order to have
5. Chose the hyper-parameters based on the equally spaced failure horizons in terms of the number
averaged MCR over the Q1 simulations of process iterations; thus the k horizons employed in
6. Compute f (j) with the selected (5) are assigned as follows
hyper-parameters  
for i = 1 to Q2 do (j − 1)(M1 − 1)
m(j) = + 1 for j = 1, . . . , k (7)
7. Randomly split the maintenance cycles k−1
between training and validation, keeping the where bue is the closest integer to u and M1 > 0 specifies
ratio q2 the maximum failure horizon and is chosen based on the
8. Compute f (j) with the selected lengths of the maintenance cycles in the training dataset.
hyper-parameters An alternative approach for selecting the failure hori-
9. Compute ρUB and ρUL of f (j) zons is to express them as the percentage of a main-
10. Compute ρ∗UB and ρ∗UL as averaged over the tenance cycle, rather than a fixed number of process
Q2 simulations iterations. Thus, in this second approach we have
11. ρUB = [ρUB ; ρ∗UB ] and ρUL = [ρUL ; ρ∗UL ]  

(j) N F if t ≤ ni 1 − m(j)
yt = , (8)
F otherwise
with
(j − 1)M2
m(j) = for j = 1, . . . , k (9)
k−1
where M2 ∈ (0, 1].
We remark that the choice of criteria used to define
the failure horizons depends on the nature of the fault
under investigation and, as such, there is no a-priori best
solution.

B. Classification Algorithms
The MC PdM methodology presented does not im-
Fig. 2. Overview of how the MC PdM module integrates with the pose any restrictions on the classification algorithm that
Fab
can be adopted for the individual classifiers. Here, we
consider two well known and widely used classification
techniques, namely: Support Vector Machines (SVMs)
ρUL are provided by the historical and simulation per- and k-Nearest Neighbors (k-NN). This choice is moti-
formances of the PdM module, while cUB and cUL are vated by the fact that it provides a comparison between
provided by the user, and may be changed at each two contrasting types of classifier; a powerful, but com-
evaluation of the PdM module (i.e. unexpected breaks putationally complex to estimate, model based classifier
are much more costly during highly intense production (SVM); and a low-complexity non-parametric classifier
periods than during normal production periods). (k-NN). In the following the two selected classifiers are
While undesirable from a operating cost perspective, briefly presented.
when an unexpected break occurs during operation of a 1) Support Vector Machines: SVMs are probably the
PdM module it means that a full maintenance cycle has most popular approach to classification, thanks to their
been observed, and hence valuable new data is available high classification accuracy, even for non-linear prob-
with which to update the MC PdM module and related lems, and to the availability of optimized algorithms for
performance metrics. Therefore, as illustrated in Fig. 2, it their computation [3].
is important to retain the facility to update the MC PdM Consider the problem where two classes of data (F
module following deployment. and N F in the problem at hand) have to be classified;
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 5

we initially consider the separable case, where the two

classes can be linearly separated. Suppose that a training
dataset S as in Eq. (1) is available and the values {−1, 1}
are assigned to the two classes which the data belong to,
for example

1 if i-th sample ∈ F
yi =
−1 if i-th sample ∈ N F
We define the hyperplane F0 in the Rp space as
F0 = {x |f (x) = xβ + β0 = 0 } , (10)
Fig. 3. Schematic of a generic Ion Implanter tool. The tool can be
p divided into three main parts: the Source , the Beamline Area , the
where β ∈ R with norm kβk = 1. The classification is
End Station 
then based on the choice of f (x) (and consequently of
F0 ): for a new sample xnew ∈/ S, we classify
if f (xnew ) > 0 2) k-Nearest Neighbours: k-NN is probably the simplest

new = 1 (Class F )
yd (11) approach to classification as it requires just computation
−1 (Class N F ) if f (xnew ) < 0
of distances between samples.
Since, by assumption, the two classes are separable,
In the k-NN procedure each point of the input space is
then it is possible to find a function f (x) s.t.
labelled according to the labels of its k closest neighbour-
yi f (x) > 0 ∀i; ing samples (where distances are computed according to
The hyperplane yielding the largest margin Π between a given metric, often the Euclidean norm).
the two classes is chosen: this can be rephrased as the The only parameter that requires tuning in k-NN is k,
maximization problem the value of the number of samples in the considered
neighbourhood: the choice of k is usually data-driven
max Π (often decided though cross-validation). Larger values
β,β0 ,kβk (12)
subject to yi f (x) ≥ Π, i = 1, . . . , n. of k reduce the effect of noise on the classification, but
make decision boundaries between classes less distinct
We refer the interested reader to [4] for details on how
[28].
(12) is solved.
Optimized algorithms for computing k-NN have a
In real classification problems, classes are often non-
computational cost of O(log n).
separable, i.e., the two categories overlap: SVMs can still
be used in this case by allowing some samples to be V. E XPERIMENTAL R ESULTS
on the ’wrong’ side of the separation line. With respect
to the separable case the optimization problem (12) is A. Use Case
changed by modifying the constraint to: We test the proposed MC PdM methodology for re-
placing tungsten filaments used in Ion Implantation [13],
y (xi β + β0 ) ≥ Π − ξi∗ = Π(1 − ξi ) ∀i, (13)
one of the most important processes in semiconductor
where the slack variables ξi have to satisfy the conditions: manufacturing fabrication. Ion Implantation is used to
ξi ≥ 0 (the points on the wrong side of their margin are modify the electrical properties of wafers by injecting
labeled with ξi∗ > 0,Pwhile the points on the correct side doping atoms. It is often considered a ’bottleneck’ in
n
have ξi∗ = 0) and i=1 ξi ≤ C ∈ R. The optimization production lines due to the high cost of the tool, making
problem is now (see [4] for a detailed solution): it a critical operation for throughput [11].
X n The components of a typical Implanter tool are illus-
min 12 kβk + γ ξi trated in Fig. 3. The filament is part of the Ion Source
β,β0
 i=1 (14) section of the tool. During the process, the filament is
y (xi β + β0 ) ≥ 1 − ξi , ∀i heated and electrons are ’boiled’ off the heated filament.
subject to ,
ξi ≥ 0 The electrons are then accelerated in the beamline area
where the regularization parameter γ governs the trade- and impinge on the target wafers in the End Station.
off between the margin width and the sum of values of The tungsten filament must be frequently replaced. Ev-
the slack variables. ery time a filament is changed, the tool is down for
SVMs are usually employed in combination with approximately 3 hours making this the most important
Kernel Methods to further enhance the classification maintenance issue for process engineers working on the
performance by allowing non-linear solutions. The tool.
computational cost for training a nonlinear SVM is Several factors may influence the working duration of
generally between O(n2 ) and O(n3 ), depending on a filament. For example, it is known that high values
the algorithm employed for its computation. For more of pressure, voltage and filament current can drastically
detailed presentations on SVMs we refer interested reduce the lifetime of the filament. The operations of
readers to [19] and [25]. cleaning, installation and degasification can also have a
fundamental impact on filament ’health’ and duration.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 6

Physical/Electrical Variable Quantity

Current 9 The maintenance strategies investigated are
Current Transfer Ratio 1 summarized in Table II. A number of approaches
Deceleration 1 are compared to MC PdM as follows:
Flow 1
Number of Scans 1
Plasma Gun Emission 1 1) A simulated PvM policy: PvM policies are usually
Position 2 based on the mean µ or median η of the maintenance
Pressure 3
Quantity of Electric Charge 2 cycle lengths and on the definition of an optimal action
Scan Speed 1 threshold τµ∗ or τη∗ . Once τµ∗ is computed, then the PvM
Tilt Angle 2 maintenance is triggered if
Voltage 7
TABLE I f µ = iterations since last maintenance > µ − τµ∗ , (15)
R ECORDED TOOL VARIABLES
and similarly for the median. The PvM policy based on
the mean (PvM-µ) is outlined in Algorithm 2 (the version
Type Based on: Acronym based on the median, PvM-η, is a natural extension).
Mean µ PvM-µ
PvM
Median η PvM-η
Linear SVM PdM-lin Algorithm 2: PvM module
PdM
Gaussian Kernel SVM PdM-rbf Data: Y , cUB , cUL
k-NN MC PdM-knn
MC PdM
SVM MC PdM-svm Result: Maintenance Rule (defined by τµ ∗), ρµUB , ρµUL
and J µ (T µ )
TABLE II
T ESTED MAINTENANCE MANAGEMENT TECHNIQUES 1. Compute µ, the mean number of process
iterations in a maintenance cycle
2. Define a set of threshold values T µ ∈ Rd
3. Let ρµUB = [ · ] and ρµUL = [ · ] (empty vectors)
B. Data Description for j = 1 to N do
4. Compute ρµUB and ρµUL for all entries in T µ
The available data consists of N = 33 maintenance
5. ρµUB = [ρµUB ; ρµUB ] and ρµUL = [ρµUL ; ρµUL ]
cycles, with the filament maintained using a R2F policy
(i.e. each maintenance cycle consists of the data for the 6. ρµUB = Mean(ρµUB ) and ρµUL = Mean(ρµUL )
implanter tool from the installation of a new filament to 7. Compute J µ (τ µ ) for all entries in T µ : J µ (T µ )
the point that the filament breaks and the tool is stopped 8. Let τ µ ∗ = minτ µ ∈T µ J µ (τ µ )
for maintenance), for a total of n = 3671 batches.
A total of 31 variables, as listed in Table I, were 2) An SVM Classification distance-based PdM system:
recorded from the ion-implantation tool during the n This is based on the PdM approach developed in [24]
runs. These variables have a time series evolution during and involves exploiting the distance from a computed
each run, with some presenting with a non-uniform SVM decision boundary to define a metric for the dis-
sampling rate from observation to observation. In or- tance an observed sample is from a faulty situation.
der to construct a design matrix to feed the classifiers, The assumption is that the more stress on a machine,
features must be extracted. A classical approach [18] to the closer statistically we get to a faulty situation and
extracting such features is to rely on statistical moments. hence the distance from the separation boundary of the
In this work we extracted 6 features for each of the two SVM classes decreases. The resulting PdM module
31 time series, namely: maximum, minimum, average, recommends that a corrective action be performed when
variance, skewness and kurtosis. After constant variables
are discarded a total of p = 125 input variables are f (x) < τ ∈ R+ , (16)
retained in the dataset. and different choices of τ affect the performance of the
PdM in terms of ρUB and ρUL . We consider here both
C. Comparison of SVM and kNN Based Approaches linear SVM (PdM-lin) and RBF kernel SVM (PdM-rbf)
classifiers.
The MC PdM approach has been tested on the prob-
lem described above using the classification algorithms
D. Settings and Results
described in Section IV-B. We use MC PdM-knn and
MC PdM-svm to denote MC PdM implemented with The parameters employed in the simulations of the
k-NN and SVM, respectively. In the case of SVM, we MC PdM methodology are
have employed Radial Basis Function (RBF) kernels to • k = 10;
implement the non-linear classification boundaries in • M1 = 85.
the linear framework presented in Section IV-B (at the These values have been selected manually for illustra-
expense of a more complicated tuning procedure, see tive purposes. Setting M1 = 85 ensures that the PdM
[4]). A linear SVM implementation, denoted PdM-lin, is modules produce no unexpected breaks, while selecting
also considered. k = 10 gives sufficient diversity in performance in terms
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 7

90
PvM−µ PvM−µ
80 PvM−η 0.5 PvM−η
PdM−lin PdM−lin
Unexploited Lifetime

Unexpected Breaks
70
PdM−rbf 0.4 PdM−rbf
60 MC PdM−knn MC PdM−knn
MC PdM−svm 0.3 MC PdM−svm
50

40 0.2

30
0.1

20

0
5 15 25 35 45 55 65 75 85 5 15 25 35 45 55 65 75 85
m m

Fig. 4. Average ρUL over Q2 Monte Carlo simulations with the various Fig. 5. Average ρUB over Q2 Monte Carlo simulations with the various
maintenance management approaches maintenance management approaches

0.6
of ρUL and ρUB to be effective. The criterion described PvM−µ
in Eq. (5) and Eq. (7) was used to select the failure
0.55 PvM−η
0.5 PdM−lin
horizon for each classifier as this was found to provide a

J [Arbitrary Unit]
PdM−rbf
better distribution of results in terms of ρUL and ρUB than 0.45
MC PdM−knn
the one described in Eq. (8) and Eq. (9). In the interest 0.4
MC PdM−svm
of compactness, the results of this investigation are not 0.35

included here. 0.3

As described in Algorithm 1, prior to application 0.25

the hyper-parameters of the MC PdM classifiers must 0.2

be tuned using cross-validation techniques. This also 0.15

applies to the SVM classification distance based PdM 2 3 4 5 6 7 8 9 10
approach. In the case of linear SVMs, only one parameter Costs Ratio x 10
−3

has to be optimised, namely, the variable γ in Equation

Fig. 6. Optimal value of J for the various maintenance strategies as
(14), while for RBF SVMs two parameters need to be a function of the ratio of the costs cUL /cUB (averaged results over Q2
tuned (γ and the variance of the Gaussian kernels). In Monte Carlo simulations)
the case of k-NN the parameter k (the size of the neigh-
bourhood) has to be set. The parameters we employed
in the cross-validation procedure are: classifiers are estimated using more balanced (i.e. less
• q1 = q2 = 0.7; skewed) datasets. In the case of ρUB , MC PdM-svm
• Q1 = Q2 = 1000. consistently achieves the best results. Overall when con-
Cross-validation outcomes are not detailed here due to sidering both ρUL and ρUB MC PdM-svm guarantees
space constraints, but we report the fact that the optimal better performance than MC PdM-knn, but this is at
value for the parameter k in MC PdM-knn is consistently the expense of a much more time consuming tuning
equal to 1 in the simulations. This may be due to the high procedure.
dimensionality of the problem. In Table III the average classification accuracy (100% -
In Fig. 4 and 5 the average performance of the various MCR), precision and recall are reported for the various
maintenance approaches is reported in terms of ρUL and
ρUB for different values of m over Q2 Monte Carlo simu- kNN SVM
lations. The thresholds of the PvM and PdM approaches m Accuracy Precision Recall Accuracy Precision Recall
are scaled to facilitate comparison with the MC PdM 1 98.51 69.27 61.69 98.52 69.34 62.96
10 98.47 64.09 67.85 98.47 63.14 70.84
approaches. The linear scaling has been applied after ρUL 20 98.39 58.00 75.42 98.37 56.61 77.77
and ρUB has been computed with the various approaches 29 98.29 53.22 81.72 98.33 54.46 85.79
so that their performances are in a similar range when 38 98.15 48.27 87.03 98.16 48.85 88.95
48 97.95 43.31 91.09 97.95 43.36 92.94
plotted in Fig. 4 and Fig. 5. 57 97.74 39.44 95.30 97.74 39.61 98.15
It can be appreciated how PdM and MC PdM outper- 66 97.52 35.96 97.76 97.57 36.96 99.63
form PvM approaches for both metrics. In the case of ρUL 76 97.26 32.74 99.36 97.26 32.86 100.00
85 97.03 30.23 99.94 97.08 30.08 100.00
the MC PdM classifiers achieve similar performance to
the PdM approaches, with PdM-rbf marginally superior TABLE III
AVERAGE ACCURACY, PRECISION AND RECALL FOR THE MC P D M
for low values of m and MC PDM-svm marginally MODULES ( PERCENTAGE RESULTS ) OVER Q2 M ONTE C ARLO
superior for higher values. The improves performance SIMULATIONS
with higher m can be ascribed to the fact that these
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 8

classifiers that make up MC PdM-knn and MC PdM- classifier distance based PdM alternative. It has also been
svm. In computing these values a process iteration with shown in this case study that SVMs offer superior per-
a broken filament is considered as the positive condition, formance to k-NN classifiers when implementing MC-
while an uninterrupted process iteration is considered PdM, and that in general MC-PdM-knn also consistently
as the negative condition. It can be observed that as the outperforms PvM approaches.
fault horizon m increases, precision decreases (more false The high dimensionality of datasets typical of PdM
positives, i.e. unexploited lifetime) while recall increases problems encountered in semiconductor manufacturing
(less false negatives, i.e. unexpected breaks). and other manufacturing industries, suggests that par-
In Fig. 6 the performance of the six maintenance ap- simonious (or sparse) classification approaches such as
proaches is plotted as a function of the ratio of the costs, Relevance Vector Machines [27] could provide competi-
cUL /cUB . Here, J is the minimum total cost achieved with tive results. This will be the subject of a future extension
each approach and is computed based on the average of the present work.
Monte Carlo results reported in Fig. 4 and 5. The values We have seen how the choice of the fault horizon m
of the costs ratio cUL /cUB were chosen with the aid of strongly affects the performance of the corresponding
process experts in order to have plausible results for the classifier in MC PdM. It may be the case that some areas
problem at hand. of the [ρUB , ρUL ] space are more ’interesting’, given the
These results highlight that MC PdM-svm consistently associated costs, than others in the context of minimis-
guarantees better performance than any of the other ing total operating costs. One approach to cope with
approaches considered. This is the most important out- this issue is to employ a two steps procedure, where
come of the present work, since it demonstrates that classifiers assigned to an uninteresting area of [ρUB , ρUL ]
the proposed maintenance methodology achieves the are reallocated in the second training iteration to more
lowest operating costs while at the same time being relevant areas of the cost space.
robust to different choices of ρUL and ρUB . MC PdM-knn
consistently delivers lower operating costs than classical R EFERENCES
PvM techniques and exhibits comparable performance to [1] R. Baly and H. Hajj. Wafer classification using support vector
PvM-svm, especially for low values of the ratio cUL /cUB . machines. IEEE Transactions on Semiconductor Manufacturing,
25(3):373–383, 2012.
[2] S. Butler and J. Ringwood. Particle filters for remaining useful
VI. C ONCLUSIONS life estimation of abatement equipment used in semiconductor
A novel multiple classifier PdM system for integral manufacturing. In IEEE Conference on Control and Fault-Tolerant
Systems, pages 436–441, 2010.
type faults has been presented. The multiple ML clas- [3] C.-C. Chang and C.J. Lin. Libsvm: a library for support vector
sifiers work in parallel to exploit the knowledge of the machines. ACM Transactions on Intelligent Systems and Technology
tool/logistic variables at each process iteration in order (TIST), 2(3):27, 2011.
[4] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical
to enhance decision making. The proposed tool guar- Learning. Data Mining, Inference and Prediction. Springer, 2009.
antees improved maintenance management decisions in [5] A. Heng, A. Tan, J. Mathew, N. Montgomery, D. Banjevic, and
terms of minimising operating cost and can be applied A. Jardine. Intelligent condition-based prediction of machinery
reliability. Mechanical Systems and Signal Processing, 23:1600–1614,
to any maintenance problems characterised by integral 2009.
type faults provided R2F historical data is available or [6] G. Koksal, I. Batmaz, and M.C. Testik. A review of data mining
can be collected. applications for quality improvement in manufacturing industry.
Expert Systems with Applications, 38(10):13448 – 13467, 2011.
The MC PdM module is robust to variations in the [7] L. Krishnamurthy, R. Adler, P. Buonadonna, J. Chhabra, M. Flani-
operating costs associated with unexpected breaks and gan, N. Kushalnagar, L. Nachman, and M. Yarvis. Design and
unexploited equipment part lifetime that can occur over deployment of industrial sensor networks: Experiences from a
semiconductor plant and the north sea. In SenSys, 2005.
time. It also allows the user to dynamically change main- [8] R. Langone, C. Alzate, B. DeKetelaere, and J.A.K. Suykens. Kernel
tenance policy based on current costs and needs of man- spectral clustering for predicting maintenance of industrial ma-
ufacturing production. Process engineers are provided chines. In Computational Intelligence and Data Mining (CIDM), 2013
IEEE Symposium on, pages 39–45, April 2013.
with a tool that enables them to adjust the performance [9] S.S. Lee. Noisy replication in skewed binary classification. Com-
and the action policy so that the total operating cost is putational statistics & data analysis, 34(2):165–191, 2000.
minimized. Moreover, the presence of several classifiers [10] B. Lenz and B. Barak. Data mining and support vector regression
machine learning in semiconductor manufacturing to improve
allows the proposed tool to be employed as a health virtual metrology. In System Sciences (HICSS), 2013 46th Hawaii
factor indicator in a more natural and straightforward International Conference on, pages 3447–3456. IEEE, 2013.
fashion that with other PdM approaches. Specifically, the [11] S.-Y. Lin and S.-C. Horng. A classification-based fault detection
and isolation scheme for the ion implanter. IEEE Transactions on
k classifiers provide an estimate of the current status of Semiconductor Manufacturing, 19(4):411–424, 2006.
the maintenance issue, enhancing process understand- [12] M. Luo, Z. Xu, H.L. Chan, and M. Alavi. Online predictive
ing and allowing cost aware maintenance management maintenance approach for semiconductor equipment. In Industrial
Electronics Society, IECON 2013-39th Annual Conference of the IEEE,
decisions at any time. pages 3662–3667. IEEE, 2013.
The proposed methodology has been demonstrated for [13] C.M. McKenna. A personal historical perspective of ion implan-
a semiconductor manufacturing Ion-Implanter related tation equipment for semiconductor applications. In Conference
on Ion Implantation Technology, pages 1 – 19, 2000.
maintenance task and shown to guarantee better perfor- [14] R.K. Mobley. An Introduction to Predictive Maintenance.
mance than classical PvM approaches and a single SVM Butterworth-Heinemann, 2002.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. 9

[15] L. Mönch, J.W. Fowler, S. Dauzère-Pérès, S.J. Mason, and O. Rose. Andrea Schirru received the M.S. degree (cum
A survey of problems, solution techniques, and future challenges laude) in computer science and the Ph.D. in
in scheduling semiconductor manufacturing operations. Journal Information Engineering from the University of
of Scheduling, 14(6):583–599, 2011. Pavia, Pavia, Italy, in 2008 and 2011. He is cur-
[16] R.R. Picard and R.D. Cook. Cross-validation of regression models. rently a post-doctoral associate at National Uni-
Journal of the American Statistical Association, 79:575–583, 1984. versity of Ireland, Maynooth and Chief Technol-
[17] General Electrics Intelligent Platforms. The rise of industrial big ogy Officer at Statwolf Ltd.
data, August 2012. He has been an Intern Researcher at Infi-
[18] A. Schirru, G.A. Susto, S. Pampuri, and S. McLoone. Learning neon Technologies Austria AG, Villach, Austria
from time series: Supervised aggregative feature extraction. In (2010-2011), and during his studies he has been
51st IEEE Conference on Decision and Control, Proceedings, pages awarded with the ICINCO Best Student Paper
5254–5259. IEEE, 2012. Award (2011) and the Best Student Presentation at the Intel European
[19] B. Scholkopf and A.J. Smola. Learning with Kernels. MIT Press, Research and Innovation Conference (2010).
2002.
[20] J. Shao. Linear model selection by cross-validation. Journal of the
American Statistical Association, 88:486–494, 1993.
[21] Y.-C. Su, F.-T. Cheng, M.-H. Hung, and H.-C.Huang. Intelligent
prognostics system design and implementation. IEEE Trans. on
Semiconductor Manufacturing, 19:195–207, 2006. Simone Pampuri received the M.S. degree in
[22] G.A. Susto, A. Beghi, and C. DeLuca. A predictive maintenance computer science and the Ph.D. in Informa-
system for epitaxy processes based on filtering and prediction tion Engineering from the University of Pavia,
techniques. IEEE Trans. on Semiconductor Manufacturing, 25:638– Pavia, Italy, in 2009 and 2012. He is currently
649, 2012. a post-doctoral associate at National University
[23] G.A. Susto, S. Pampuri, A. Schirru, A. Beghi, and G. DeNicolao. of Ireland, Maynooth and Chief Executive Offi-
Multi-step virtual metrology for semiconductor manufacturing: a cer at Statwolf Ltd.
multilevel and regularization methods-based approach. Comput- He has been an Intern Researcher at Infi-
ers & Operations Research, (0):–, 2014. neon Technologies Austria AG, Villach, Austria
[24] G.A. Susto, A. Schirru, S. Pampuri, D. Pagano, S. McLoone, and (2011). His research interests include manu-
A. Beghi. A predictive maintenance system for integral type facturing data analytics, machine learning and
faults based on support vector machines: An application to ion sentiment analysis.
implantation. In Automation Science and Engineering (CASE), 2013
IEEE International Conference on, pages 195–200. IEEE, 2013.
[25] J. Suykens and J. Vandewalle. Least squares support vector
machine classifiers. Neural Processing Letters, 9(3):293–300, 1999.
[26] K. Tamaki and S. Kaneko. Multiparametric virtual metrology
model building by job-shop data fusion using a markov chain Seán McLoone (S’94, M’96, SM’02) received a
monte carlo method. 2013. Master of Engineering degree with Distinction
[27] M.E. Tipping. Sparse bayesian learning and the relevance vector and a Ph.D. degree in Control Engineering
machine. The journal of machine learning research, 1:211–244, 2001. from Queens University Belfast (QUB), Belfast,
[28] M.-L. Zhang and Z.-H. Zhou. Ml-knn: A lazy learning approach Ireland, in 1992 and 1996, respectively. He is
to multi-label learning. Pattern recognition, 40(7):2038–2048, 2007. currently Professor and Director of the Energy
Power and Intelligent Control (EPIC) Research
Cluster at QUB.
His research interests lie in the general area
to data based modelling and analysis of dy-
namical systems. This encompasses techniques
ranging from classical system identification, fault diagnosis, and sta-
tistical process control to modern computational intelligence based
adaptive learning algorithms and optimization techniques. He is a Past
Chairman of the UK and Republic of Ireland (UKRI) Section of the
IEEE.

Gian Antonio Susto (S’11) received the M.S.

degree (cum laude) in control systems engineer-
ing and the Ph.D. in Information Engineering
from the University of Padova, Padova, Italy, in Alessandro Beghi (M’06) received the laurea
2009 and 2013. He is currently a post-doctoral degree cum laude in Electrical Engineering in
associate at University of Padova and Chief 1989 and the Ph.D. degree in Control Systems
Operating Officer at Statwolf Ltd. Engineering in 1993, both from the University
He has been a visiting student at the Uni- of Padova, Italy. In 1994 he joined the Depart-
versity of California, San Diego (2008-09), and ment of Information Engineering, University of
at National University of Ireland, Maynooth Padova, Italy, where he is currently an Asso-
(2012) and an Intern Researcher at Infineon ciate Professor of Control Systems Theory.
Technologies Austria AG, Villach, Austria (2011). During his studies He held visiting positions at universities and
he has been awarded with IEEE-CASE Best Student Conference Paper international research centers both in Europe
Award (2011), IEEE/SEMI-ASMC Best Student Paper Award (2012) and in the USA. He is a member of the IEEE
and IEEE-MSC Best Student Paper Award (2012). His research inter- Control System Society, serving also as Associate Editor for the Con-
ests include manufacturing data analytics, machine learning, gesture ference Editorial Board and co-chair of the Technical Committee on
recognition and partial differential equations control. Power Generation. He has worked on a wide range of control applica-
tions, from control of fusion devices, guidance algorithms for virtual
vehicles, to control of HVAC&R systems, Adaptive Optics systems, and
Semiconductor Manufacturing processes. He is the author of more than
130 publications in journals, books, and conference proceedings. He
is co-inventor of 5 International Patents on applications of advanced
control techniques.

View publication stats