2018 5th International Conference on Industrial Engineering and Applications

A Research Study on Unsupervised Machine Learning Algorithms for Early Fault

Detection in Predictive Maintenance

Nagdev Amruthnath Tarun Gupta

Department of IEE and EDMM Department of IEE and EDMM
Western Michigan University Western Michigan University
Kalamazoo, Michigan, USA Kalamazoo, Michigan, USA
e-mail: [email protected] e-mail: [email protected]

Abstract—The area of predictive maintenance has taken a lot subcomponents to not only collect the process data and its
of prominence in the last couple of years due to various reasons. parameters but also to collect the physical health aspects of
With new algorithms and methodologies growing across the machine such as vibration, pressure, temperature,
different learning methods, it has remained a challenge for acoustics, viscosity, flow rate and many as such. This
industries to adopt which method is fit, robust and provide information is widely used for early fault detection, fault
most accurate detection. Fault detection is one of the critical identification, health assessment of the machine and predict
components of predictive maintenance; it is very much needed the future state of the machine. Some of this is made possible
for industries to detect faults early and accurately. In a due to machine learning algorithms available across different
production environment, to minimize the cost of maintenance,
learning domains.
sometimes it is required to build a model with minimal or no
Machine learning is a subsection of Artificial Intelligence
historical data. In such cases, unsupervised learning would be
a better option model building. In this paper, we have chosen a Figure 1. Machine learning can be defined a program or an
simple vibration data collected from an exhaust fan, and have algorithm that is capable of learning with minimum or no
fit different unsupervised learning algorithms such as PCA T2 additional support. Machine learning helps in solving many
statistic, Hierarchical clustering, K-Means, Fuzzy C-Means problems such as big data, vision, speech recognition, and
clustering and model-based clustering to test its accuracy, robotics [7]. Machine learning is classified into three types.
performance, and robustness. In the end, we have proposed a In supervised learning, the predictors and response variables
methodology to benchmark different algorithms and choosing are known for building the model, in unsupervised learning, ,
the final model. only response variables are known, and in reinforced
learning, the agent learns actions and consequences by
Keywords-predictive maintenance; fault detection; interacting with the environment. In this research, the main
manufacturing; machine learning; just in time focus will be on unsupervised learning methodology. One of
the most commonly used approaches in unsupervised
I. INTRODUCTION learning is clustering where, response variables are grouped
The concept of predictive maintenance (PdM) was into clusters either user-defined or model based on the
proposed a few decades ago. PdM is also a subset of planned distance, model, density, class, or characteristic of that
maintenance. PdM did not gain prominence until the recent variable. For this research, vibration data has been used. Data
decade. This rapid advance is mainly due to emerging collection, feature selection, and extraction will be described
internet technologies, connected sensors, systems capable of in the later sections.
handling big data sets and realizing the need to use these
techniques. The abrupt growth can also be theorized due to
the demand for high-quality products, at the least cost and
with shortest lead time. Every year, it is estimated that U.S.
industry spends $200 billion on maintenance of plant
equipment and facilities and the result of ineffective
maintenance leads to a loss of more than $60 billion [1]. In
food and beverage industry it was estimated that failures and
downtime accounted for 18% of OEE [2]. Over the years,
different architecture, algorithms, and methodologies have
been proposed. One of the most prominent methods is
watchdog agent, a design enclosed with various machine Figure 1. Structure of learning methods.
learning algorithms [3] [11]. Some of the other architectures
are an OSA-CBM architecture [4], SIMAP Architecture [5], All the programming in this research is performed in a
and predictive maintenance framework [6]. Emerging statistical tool called as R- Programming. R- Program is
technologies such as the Internet of things (IoT) devices have open source software and was designed by Ross Ihaka and
formed a gateway to connect to machines and its Robert Gentleman in August 1993. As of today, there are

978-1-5386-5748-5/18/$31.00 ©2018 IEEE 355

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
over 10,000 packages which include thousands of different process of identifying the abnormal behavior of a subsystem.
algorithms contributed by various authors for different Any deviation from a standard behavior can be categorized
applications. as a failure. In this section, we will discuss different
algorithms such as Principle Component Analysis (PCA) T2
II. LITERATURE REVIEW statistic, Hierarchical clustering, K- Means clustering, C-
The primary goal of PdM is to reduce the cost of a Means, and Model-based clustering for fault detection and
product or service and to have a competitive advantage in the benchmark its results for vibration monitoring data.
market to survive. Today business analytics are embedded A. Data Collection
across PdM to realize the need for it and to make appropriate
decisions. Business analytics can be viewed in three different Vibration data is one of the most commonly used
prospective (i) Descriptive analytics (ii) Predictive analytics technique to detect any abnormalities in a submachine. In
and (iii) Prescriptive analytics [16]. Descriptive analytics is a this research paper, a vibration monitor sensor was set up on
process of answering questions like what happened in the an exhaust fan. The vibration was collected every 240
past? This is done by analyzing historical data and minutes for 12 days at a sampling frequency of 2048 Hz on
summarizing them in charts. In maintenance, this step is both X and Y axis. From the following data, different
performed using control charts. Predictive analytics is an features were extracted such as peak acceleration, peak
extension to descriptive analytics where historical data is velocity, turning speed, RMS Velocity, and Damage
analyzed to predict the future outcomes. In maintenance, it accumulation. Figure 2 is the time series plots of the data.
is used predict type of failure and time to complete failure.
Finally, prescriptive analytics is a process of optimization to
identify the best alternatives to minimize or maximize the
objective. This also answers the questions such as what can
be done? In maintenance, this can be used to optimize the
maintenance schedules to minimize the cost of maintenance.
In this paper, our primary focus will be on descriptive and
predictive analytics to detect the faults.
Predictive analytics has spread its applications into
various applications such as railway track maintenance,
vehicle monitoring [23], automotive subcomponents [8],
utility systems [19], computer systems, electrical grids [13],
aircraft maintenance [21], oil and gas industry,
computational finance and many more.
Fault detection is one of the concepts in predictive Figure 2. Feature data plot.
maintenance which is well accepted in the industry. Early
Failure detection could potentially eliminate catastrophic In Figure 2, we can see a trend line generating closer to
machine failures. In one of the recent research studies, this index 60th observation. In this paper, we will test to see how
process is classified into different methods such as different algorithms help in detecting this fault earlier.
quantitative model-based methods, qualitative model-based
methods, and process history based methods [25]. B. Feature Selection Using PCA
Principle component analysis (PCA) is one of the oldest Not all features extracted provide a true correlation. If
and most prominent algorithms that are widely used today. It right features are not selected, then a significant amount of
was first invented by Karl Pearson in 1901. Since then, they noise would be added to the final model and hence, reduce
have been many hybrid approaches to PCA for fault the accuracy of the model. One of the most prominent
detection such as using Kernel PCA [17], adaptive threshold algorithms for that is used for dimensionality reduction is
using Exponential weight moving average for T2 and Q Principle component analysis. Principal component analysis
statistic [9], multiscale neighborhood normalization-based (PCA) is a mathematical algorithm that reduces the
multiple dynamic principal component analysis (MNN- dimensionality of the data while retaining most of the
MDPCA) method [27], Independent Component Analysis. variation (information) in the data set [18]. In a simple
Another common method used for fault detection is context, it is an algorithm to identify patterns in data and
clustering method. Similar to PCA, there are various expressing such a way to showcase those similarities and
algorithms such as neural net clustering algorithm neural differences [29].
networks and subtractive clustering [28], K-means [10], Algorithm:
Gaussian mixture model [15], C-Means, Hierarchical Step 1: Consider a data matrix X
Clustering [22], and Modified Rank Order clustering
(MROC) [33]. [X]mxn (1)
where, X is the matrix, m is a row, and n is a column
III. FAULT DETECTION
Step 2: Subtract the mean from each dimension
Fault detection is one of the most critical components of [ ] −[ ] (2)
predictive maintenance. Fault detection can be defined as a

356

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
Step 3: Calculate the covariance matrix Based on the results from T2 statistic in Figure 5, we can
[ ] (3) observe that the faults can be detected as early as 41
Step 4: Calculate the eigenvectors and eigenvalues of the observations. Hence, this early detection would help the
covariance matrix maintenance teams to monitor these process changes and
([ ] − ){ } = {0} (4) take corrective actions accordingly.
Step 5: Store the eigenvector in a matrix D. Cluster Analysis
[ ] = [{ }{ }{ } … . . { }] (5)
Clustering analysis is one of the unsupervised learning
Step 6: Store eigenvalues in a diagonal matrix methods. In cluster analysis, similar data are grouped into
[ ] (6) different clusters. Some of the most prominent cluster
where [Eigen] is the eigenvalues corresponding to the analyses are K-Means clustering, C-Means clustering, and
principal components, and P contains the loading vectors hierarchical clustering. There are various merging principles
Step 7: Rank eigenvalues in decreasing order and choose top in hierarchical clustering. They are iterative, hierarchical,
“r” vectors to retain density based, Metasearch controlled and stochastic. In this
[ ] (7) paper, we will be discussing one of the commonly used
Step 8: Retain “r” eigenvectors hierarchical clusterings.
[ ] = [{ }{ }{ } … . . { }] (8) E. Optimal Number of Clusters
Step 9: Calculate the principal components [U] which is
projected in data matrix In cluster analysis, we need to know the optimal number
[ ] [ ] =[ ] (9) of clusters that can be formed. Although we know that, we
Summary of the PCA indicates that the first two principal have healthy data and faulty data, identifying the number of
components show 95.65% of variance compared to rest of optimal cluster formations in our data would help in
the components. understanding different states in the data and representing the
A scree plot can be plotted for Eigenvalues versus data more accurately. To identify the number of clusters,
principle components as shown in Figure 4. This plot can be there are many procedures available such as elbow method,
used to define the components that show significant variance Bayesian Inference Criterion method and nbClust package in
in the data. R. The results for elbow method is shown in Figure 6 and
From summary data and scree plot, we can conclude that using nbClust [30] is shown in Figure 7.
the first two principal components present maximum
variation compared to the rest of the principal components.
C. T2 Statistic
T2 Statistic is a multivariate statistical analysis. The T 2
statistic for the data observation x can be calculated by [12]
= ∑ (10)
2
The upper confidence limit for T is obtained using the
F-distribution:
( )
, ,∝ = , ,∝ (11)

Figure 4. Scree plot to determine the variation between principal

components.

Figure 3. Summary of PCA.

where n is the number of samples in the data, a is the number

of principal components, and α is the level of significance
[24]. This statistic can be used to measure the values against
the threshold and any values above the threshold; can be
concluded as out of control data. In this case, it is going to be
faulty data. The results for the vibration data are shown the
Figure 5. Figure 5. T2 statistic results for training dataset and testing dataset.

357

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
 From both the procedures shown in Figure 6 and Figure 7, just one item. Let the distances (similarities) between the
we can identify that 3 clusters are the optimal number of clusters equal the distances (similarities) between the items
clusters. For fault detection, we can use three clusters and they contain [24].
theorize each cluster represents a normal condition, warning
condition, and faulty condition. In the next section of cluster Algorithm:
analysis, we can observe how each of the clustering Step 1: Find the closest (most similar) pair of clusters and
algorithms provides the results. merge them into a single cluster, so that now you have one
From both the procedures shown in Figure 6 and Figure 7, less cluster.
we can identify that 3 clusters are the optimal number of Step 2: Compute distances (similarities) between the new
clusters. For fault detection, we can use three clusters and cluster and each of the old clusters.
theorize each cluster represents a normal condition, warning Step 3: Repeat steps 2 and 3 until all items are clustered into
condition, and faulty condition. In the next section of cluster a single cluster of size N.
analysis, we can observe how each of the clustering In Figure 8, the cluster is formed based on the feature
algorithms provides the results. data using Ward's method. Irrespective of feature data and
Principle components, the results were identical. Three
clusters were formed, where the first cluster includes
observations from 1 to 40, the second cluster includes
observations 41 to 67 and finally, the third cluster includes
observations from 68 to 71. Based on the domain knowledge,
we can represent cluster 1 as healthy dataset, cluster 2 as
warning dataset and finally cluster 3 as faulty data set.
G. K-Means and Fuzzy C-Means Clustering
K-means is one of the most common unsupervised
learning clustering algorithms. This most straightforward
algorithm’s goal is to divide the data set into pre-determined
clusters based on distance. Here, we have used Euclidian
distance. The graphical results as shown in Figure 9.
C-means is a data clustering technique where each data
point belongs to every cluster at some degree. Fuzzy C
means was first introduced by Bezdek [14]. Fuzzy C-Means
has been applied in various applications such as agricultural,
engineering, astronomy, chemistry, geology, image analysis
Figure 6. Determining the optimal number of clusters based on elbow [14], medical diagnosis, and shape analysis and target
method. recognition [26]. The graphical results for C-Means is as
shown in Figure 9.
Summary of K-Means and C-Means Clustering
TABLE I. CLUSTER MEANS OF K-MEANS ALGORITHM

1 2
1 -9.665 -1.609
2 -0.497 1.856
3 1.301 -1.092

Within cluster sum of squares by cluster:

[1] 16.758705 39.575966 8.823486
(between_SS / total_SS = 90.2 %)
TABLE II. FUZZY C-MEANS CLUSTER CENTERS WITH 3 CLUSTERS
1 2

1 1.275 -1.071
Figure 7. Determining the number of clusters using nbClust package.
2 -0.289 1.920
F. Heirarchical Clustering
Start by assigning each item to its own cluster, so that if 3 -9.935 -1.723
you have N items, you now have N clusters, each containing

358

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
 From Table III summary of K-means and C-means The results are summarized in Table 3. The results from
clustering, we can observe that clusters of sizes 4, 27 and 40 Gaussian finite mixture model fitted by EM algorithm
are formed. Observation 1 to 40 formed one cluster, 41 to 67 Classification, there was a total of 5 groups of components
formed second cluster and the third cluster with 68 to 71 are formed. Component 1 and two are assigned to
observations. These results are same as hierarchical observation 1 to 40, component group 3 consists of
clustering. observation 41 to 63, component group 4 consist of
observations 64 to 67 and finally component 5 consists of
observations 68 to 71. It is interesting to note that, the critical
fault detection which is accurately predicted similarly to
other clustering algorithms as well.
IV. RESULTS
In this research, initially, we were hypothesized that two
states in data. One is healthy data set, and the other is
unhealthy data set. Using PCA and T2 statistic, we were able
to fit our hypothesis states and able to detect the faults 31
observations ahead. Whereas, without a tool and just based
on data plots we could observe the trends only 11
observations ahead. As we moved on to fitting different
unsupervised clustering algorithms, we found most of the
Figure 8. Hierarchical clustering solution for fault identification.
clustering algorithms provided much more than the T2
statistic.
H. Model-Based Clustering Using elbow method and nbClust package, we were able
to identify that the most optimal number of clusters that
A Gaussian mixture model (GMM) is used for modeling could be formed was three. Based on these results, when data
data that comes from one of the several groups: the groups was fitted in hierarchical clustering, K-means, and C-means,
might be different from each other, but data points within the the results were nearly identical. Based on the previous
same group can be well-modeled by a Gaussian distribution knowledge of the data, we were able to identify each of three
[20]. Gaussian finite mixture model fitted by EM algorithm states. The first state was identified as healthy state (since it
is an iterative algorithm where some initial random estimate was calibrated for healthy data), second state was identified
starts and updates every iterate until convergence is detected as the warning state and finally the third state was identified
[31] [32]. Initialization can be started based on a set of initial as faulty state. It would not be surprising to obtain the
parameters and start E-step or set of initial weights and following results as all these algorithms were based on a
proceed to M-step. This step can be either set randomly or distance measure.
could be chosen based on some method.
Summary of Classification
Mclust EVV (ellipsoidal, equal volume) model with five
components:
log.likelihood n df BIC ICL
-57.23501 71 25 -221.037 -222.0734

Figure 10. Gaussian finite mixture model fitted by EM algorithm

classification.

For our final model, Gaussian finite mixture model fitted

by EM algorithm was used. Unlike providing the number of
clusters, this model identifies optimal clusters and
accordingly classifies the observations into groups. Here, the
model recognized a total of 5 components. Although with
Figure 9. K-Means and C-Means clustering for fault identification.
five components, upon closer investigation, we could

359

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
observe that, there is an overlap of component 1 and 2 and But if fault detection needs to be performed under different
component 3 and 4. When these components are reorganized levels then, clustering algorithms would be a better choice.
we can observe much similar pattern to the previous cluster
analysis. VI. FUTURE SCOPE OF WORK
Fault detection is one of the preliminary analytics for
V. CONCLUSION predictive maintenance. Hence, detecting the fault accurately
This research started out as a test bed to benchmark is regarded important. This work is currently performed for
different machine learning algorithms for early fault vibration data. The scope of this research can be extended
detection using unsupervised learning. In our results, T2 out to other physics-based parameters and combination of
statistic provided more accurate results compared to GMM these parameters. It would also be interesting to observe the
method, and no hypothesis was required to identify the detection accuracy for bigger sample size and multiple fault
relationship between cluster and state. One of the main states.
benefits of this method is that, even when this is deployed to
the manufacturing environment, with minimum or no REFERENCES
domain knowledge, one can identify fault or critical [1] Mobley, R Keith, “An Introduction to predictive maintenance”, 2002,
condition when compared to clustering analysis. On the other 2nd ed, ISBN 0-7506-7531-4
hand in clustering, some information about the data is needed [2] Battini, D., Calzavara, M., Persona, A., and Sgarbossa, F. (2016)
to name the clusters as healthy, warning or critical. “Sustainable Packaging Development for Fresh Food Supply Chains.
Package.” Technol. Sci., 29: 25–43. doi: 10.1002/pts.2185.
Clustering methodology is undoubtedly a better tool in
[3] Jay Lee, Hung-An Kao, Shanhu Yang, (2014) “Service Innovation
detecting different levels of faults where T2 statistic would and Smart Analytics for Industry 4.0 and Big Data Environment”,
be challenging after certain levels. To emphasize this, when Procedia CIRP Volume 16, 2014, Pages 3-8
the cost machine maintenance is expensive, clustering would [4] Lebold M, Thurston M. “Open standards for condition-based
be a flexible option where machine health can be monitored maintenance and Prognostic systems”. In: Proceedings of MARCON
continuously until a critical level is reached. 2001—fifth annual maintenance and reliability conference,
Gatlinburg, USA, 2001.
TABLE III. SUMMARY RESULTS OF ALL MODELS [5] Garcia E, Guyennet H, Lapayre J-C, Zerhouni N. “A new industrial
cooperative tele-maintenance platform”. Comput Ind Eng 2004;46(4):
Heirarchi Model- Heirarchi Model-
Obs Actual T2
cal
K-Means C-Means
Based
Obs Actual T2
cal
K-Means C-Means
Based
851–64.
1 H 0 1 1 1 1 37 H 0 1 1 1 1 [6] Groba. C, Cech. S, Rosenthal. F., Gossling. A, “Architecture of the
2 H 0 1 1 1 1 38 H 0 1 1 1 1 predictive maintenance framework”, 6th International Conference on
3 H 0 1 1 1 1 39 H 0 1 1 1 1 Computer Information Systems and Industrial Management
4 H 0 1 1 1 1 40 H 0 1 1 1 1
Applications, 2007, IEEE
5 H 0 1 1 1 1 41 F 1 2 2 2 3
6 H 1 1 1 1 1 42 F 1 2 2 2 3 [7] Ethem Alpaydin, “Introduction,” in Introduction to Machine
7 H 0 1 1 1 1 43 F 1 2 2 2 3 Learning,3rd ed. Cambridge
8 H 0 1 1 1 1 44 F 1 2 2 2 3
[8] Ahmed, M., Baqqar, M., Gu, F., Ball, A.D., 2012. “Fault detection
9 H 0 1 1 1 1 45 F 1 2 2 2 3
10 H 1 1 1 1 1 46 F 1 2 2 2 3
and diagnosis using principal component analysis of vibration data
11 H 0 1 1 1 1 47 F 1 2 2 2 3 from a reciprocating compressor”, in: Proceedings of the UKACC
12 H 0 1 1 1 1 48 F 1 2 2 2 3 International Conference on Control, 3-5 September 2012, IEEE
13 H 0 1 1 1 1 49 F 1 2 2 2 3 Press.
14 H 0 1 1 1 1 50 F 1 2 2 2 3
[9] Azzeddine Bakdi, Abdelmalek Kouadri, Abderazak Bensmail, “Fault
15 H 0 1 1 1 1 51 F 1 2 2 2 3
16 H 0 1 1 1 1 52 F 1 2 2 2 3
detection and diagnosis in a cement rotary kiln using PCA with
17 H 0 1 1 1 1 53 F 1 2 2 2 3 EWMA-based adaptive threshold monitoring scheme”, Control
18 H 0 1 1 1 1 54 F 1 2 2 2 3 Engineering Practice, Volume 66, September 2017, Pages 64-75
19 H 1 1 1 1 1 55 F 1 2 2 2 3
[10] C. T. Yiakopoulos, K. C. Gryllias. I. A. Antoniadis, “Rolling element
20 H 0 1 1 1 1 56 F 1 2 2 2 3
bearing fault detection in industrial environments based on a K-means
21 H 0 1 1 1 1 57 F 1 2 2 2 3
22 H 0 1 1 1 1 58 F 1 2 2 2 3
clustering approach”, Expert Systems with Applications Volume 38,
23 H 0 1 1 1 1 59 F 1 2 2 2 3 Issue 3, March 2011, Pages 2888-2911
24 H 0 1 1 1 1 60 F 1 2 2 2 3 [11] Djurdjanovic D, Lee J, Ni J. “Watchdog agent, an infotronics-based
25 H 0 1 1 1 1 61 F 1 2 2 2 3
prognostics approach for product performance degradation
26 H 0 1 1 1 1 62 F 1 2 2 2 3
assessment and prediction”. Advance Engineering Informatics 2003;
27 H 0 1 1 1 1 63 F 1 2 2 2 3
28 H 0 1 1 1 1 64 F 1 2 2 2 3
17(3–4): 109–25.
29 H 0 1 1 1 1 65 F 1 2 2 2 4 [12] Hotelling, H. (1933). “Analysis of a complex of statistical variables
30 H 0 1 1 1 1 66 F 1 2 2 2 4 into principal components.” Journal of Educational Psychology, 24,
31 H 0 1 1 1 1 67 F 1 2 2 2 4
417–441.
32 H 0 1 1 1 1 68 F 1 2 2 2 4
33 H 0 1 1 1 1 69 F 1 3 3 3 5 [13] IBM Israel, “Israel Electric corporation moves towards smarter
34 H 0 1 1 1 1 70 F 1 3 3 3 5 maintenance”, 2013, retrieved from www.IBM.com
35 H 0 1 1 1 1 71 F 1 3 3 3 5
[14] J. C. sridek, “Pattern Recognition with Fuzzy Objective Function
36 H 0 1 1 1 1 72 F 1 3 3 3 5
Algorithms”, New York: Plenum Press, 1981.
[15] Jacob Goldberger, Sam Roweis, “Hierarchical Clustering of a
In conclusion of this study, although most algorithms Mixture Model”, Neural Information Processing Systems Conference
provided nearly similar results, each algorithm provided [16] James R. Evans, Carl H. Lindner (2012), “Business Analytics: The
deeper insight into the data. Hence, if the application is just Next Frontier for Decision Sciences,” College of Business, University
to detect the faults, T2 statistic would be an excellent tool. of Cincinnati, Decision Science Institute

360

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.
[17] Jingli Yang, Yinsheng Chen, Zhen Sun, “A real-time fault detection [26] Y. Yong, Z. Chongxun and L. Pan, “A Novel Fuzzy C-Means
and isolation strategy for gas sensor arrays”, Instrumentation and Clustering Algorithm for Image Thresholding”, Measurement Science
Measurement Technology Conference (I2MTC), 2017 IEEE Review, vol. 4, no.1, 2004.
International, 22-25 May 2017, 10.1109/I2MTC.2017.7969906 [27] Yajun Wang, Fuming Sun, Bo Li, “Multiscale Neighborhood
[18] Jolliffe, “I.T. Principal Component Analysis”, Springer, New York, Normalization-Based Multiple Dynamic PCA Monitoring Method for
2002. Batch Processes With Frequent Operations”, IEEE Transactions on
[19] P. Liggan and D. Lyons, “Applying Predictive maintenance Automation Science and Engineering ( Volume: PP, Issue: 99 )
techniques to Utility Systems, Retrieved from Pharmaceutical [28] Zhimin Du, .Bo Fan..Xinqiao Jin, Jinlei Chi, “Fault detection and
Engineering”, Official Magazine of ISPE, Nov/Dec 2011, Vol 31 diagnosis for buildings and HVAC systems using combined neural
No.6 networks and subtractive clustering analysis ”, Building and
[20] Ramesh Sridharan, “Gaussian mixture models and the EM algorithm”, Environment Volume 73, March 2014, Pages 1-11
retrieved from https://people.csail.mit.edu/rameshvs/content/gmm- [29] Lindsay I Smith, “A tutorial on Principal Components Analysis”,
em.pdf February 26, 2002, page 2-8.
[21] Samaranayake, P. & Kiridena, S. (2012). “Aircraft maintenance [30] Malika Charrad, Nadia Ghazzali, Veronique Boiteau, Azam Niknafs
planning and scheduling: An integrated framework”. Journal of (2014). ”NbClust: An R Package for Determining the Relevant
Quality in Maintenance Engineering, 18 (4), 432-453. Number of Clusters in a Data Set. Journal of Statistical Software”,
[22] Stephen P. Borgatti, “How to Explain Hierarchical Clustering”, 1994, 61(6), 1-36. URL http://www.jstatsoft.org/v61/i06/
Connections 17(2):78-80 Copyright 1994 INSNA [31] Chris Fraley, Adrian E. Raftery, T. Brendan Murphy, and Luca
[23] T. R¨ognvaldsson, S. Byttner, R. Prytz, S. Nowaczyk and M. Scrucca (2012) “mclust Version 4 for R: Normal Mixture Modeling
Svensson, “Wisdom of Crowds for Self-organized Intelligent for Model-Based Clustering, Classification, and Density Estimation
Monitoring of Vehicle Fleets” 2014, IEEE Technical Report No. 597”, Department of Statistics, University of
Washington
[24] Thamara Villegas, María Jesús Fuente, Miguel Rodríguez,”Principal
Component Analysis for Fault Detection and Diagnosis. Experience [32] Chris Fraley and Adrian E. Raftery (2002) “Model-based Clustering,
with a pilot plant”, Advances in Computational Intelligence, Man- Discriminant Analysis and Density Estimation” Journal of the
Machine Systems and Cybernetics, ISBN: 978-960-474-257-8 American Statistical Association 97:611-631
[25] Venkat Venkatasubramanian, Raghunathan Rengaswamy b, [33] Nagdev Amruthnath, Tarun Gupta (2016), “Modified Rank Order
KewenYinc, .Surya N. Kavurid, “A review of process fault detection Clustering Algorithm Approach by Including Manufacturing Data”,
and diagnosis: Part I: Quantitative model-based methods”, Computers 4th IFAC International Conference on Intelligent Control and
& Chemical Engineering Volume 27, Issue 3, 15 March 2003, Pages Automation Sciences, Reims, France, June 1-3, 2016
293-311

361

Authorized licensed use limited to: TU Ilmenau. Downloaded on September 16,2021 at 14:08:18 UTC from IEEE Xplore. Restrictions apply.