An Artificial Neural Network For Automated Fault Detection

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

An Artificial Neural Network for Automated Fault Detection

Julian B itterw olf E vgenia Rusak Sebastian Reiter


FZI R esearch Center for FZI Research Center for FZI Research Center for
Information T ech n ology (FZI) Information T echnology (FZI) Information T echnology (FZI)
Karlsruhe, Germany Karlsruhe, Germany Karlsruhe, Germany
e-mail: bitterw o@ fzi.de e-mail: rusak@ fzi.de e-mail: sreiter@ fzi.de

A lexander V iehl O liver Bringmann


FZI Research Center for U niversity o f Tubingen
Information T ech n ology (FZI) Tubingen, Germany
Karlsruhe, Germany e-mail: bringman@ i nformatik .uni-tuebingen.de
e-mail: vieh l@ fzi.d e

Abstract
Intelligent and interconnected cyber physical systems are Predictive maintenance aims to ensure the maximum
a key enabler for future cost-efficient, automated and deployment efficiency o f production systems and thereby
flexible industrial production systems. Predictive save high costs of unnecessary maintenance or in the
maintenance and condition monitoring are important opposite case, costs associated with breakdowns
techniques in order to reduce costs associated with However, automated fault detection is a challenging task,
unnecessary maintenance or premature breakdowns. In as the system can react to different faults in different
this paper, we propose techniques from supervised ways. Fault detection means the tracking o f unexpected
learning for automated malfunctioning detection. For that behavior o f relevant parameters and the discovery of
purpose, we train an artificial neural network on time patterns in data that might indicate malfunctioning.
series data representing the internal system behavior. We Machine learning techniques are a common “working
present experimental results from an industrial motor horse” for the task o f pattern recognition. Several
control system. We use a digital twin o f the electronic machine learning algorithms have been used for the task
component that models the relevant features o f the o f predictive maintenance [2, 3].
physical system. The obtained information can be used
during the runtime o f technical systems and installations
for a criticality analysis and the subsequent selection of
measures.

1. Introduction
The prevalence o f software-controlled functionalities in
industrial systems is continuously increasing. The high
degree o f connections between system parts leads to a
strong interconnection o f the system. This means that in
the worst case scenario, small faults o f system parts can
propagate to cause failures o f the entire system.
Condition monitoring as part o f predictive maintenance Fig. 1: Schematic of the proposed automated fault
serves to determine the condition o f the overall system, detection pipeline
system parts or components to notice subtle changes in
relevant parameters in order to predict when maintenance
or other measures should be performed to prevent Virtual prototyping (VP) has gained popularity as it
failures. [1]. allows for a validation o f a new design during the
development stage prior to building a physical prototype
[4, 5]. In addition, VP is also used for maintenance
Proceedings of the 6th All-Russian Scientific purposes [6]. In this scenario, possible faults or
Conference "Information Technologies for Intelligent degradation effects are simulated virtually to predict
Decision M aking Support", M ay 28-31, Ufa - when maintenance should be provided to the physical
Stavropol, Russia, 2018 system.

6th All-Russian Scientific Conference "Information Technologies for Intelligent Decision Making Support", Ufa-Stavropol, Russia, 2018
141
In this paper, we present a supervised machine learning and a very deep residual network and argue that with
algorithm for the task o f automated fault detection. A deep learning, heavy pre-processing or feature crafting is
schematic o f the simulation and the automated fault no longer necessary to achieve premium performance.
detection pipeline is presented in Fig.1. The general idea They demonstrate their findings on the UCR data set.
is to train an artificial neural network (ANN) with labeled
Several authors have applied methods o f time series
data that is obtained from a virtual system modeled
classification for the task o f condition monitoring and
according to a physical system and therefore being its
automated fault detection. Zhong et al. studied fault
digital twin. For the task o f condition monitoring,
diagnosis for a gearbox based on Support Vector
different labels indicate different system states. Some
Machines for condition monitoring [10]. The data was
system states contain faults and thus, the ANN learns to
obtained from an experimental test rig. Relevant features
recognize when faults are present. After the training
were extracted from the vibration signals o f the gearbox
stage, the ANN can accept data from the physical system
with the wavelet packet transform as well as using time-
in order to track potentially dangerous faults. Our main
statistical features. For an optimal set of features, the
conclusion is that the ANN is very successful in
authors reported an accuracy o f 100% on the test set.
separating faulty from faultless cases, since we achieve
Shulian et al. presented an artificial neural network as a
an accuracy o f 100% on the validation set after only
classifier for the task o f fault diagnosis o f a gearbox [11].
training for 4 epochs.
The data was obtained from a physical gearbox and faults
The main contributions o f this paper are: were classified according to their severity such as e.g.
‘gentle fault’ or ‘bad fault’. Aydin et al. presented a
• We demonstrate a methodology to generate an
fuzzy c-means algorithm that was used to distinguish
automated condition monitoring system from
between broken rotor bar faults and the healthy condition
VP.
o f an induction motor at four different operation speeds
• We show that techniques from supervised [12]. Recently, Yang et al. showed a time-efficient
machine learning serve as a valid method to clustering-k-nearest-neighbors algorithm for the purpose
distinguish between faulty and faultless cases o f fault detection in gas sensor arrays [13]. The authors
and thus are suitable for the task o f fault verified the performance o f the algorithm with a real gas
detection. sensor array experimental system with different kinds o f
faults. Campbell et al. examined the suitability o f
• We illustrate our methodology using an artificial neural networks to be used for condition
industry-relevant use-case. monitoring o f electric power transformers [14]. Sreejith
The remainder o f the paper is structured as follows: In et al. demonstrated an algorithm for the task o f fault
Section 2, we discuss previous publications on the diagnosis o f rolling element bearings using time-domain
classification o f time series data and the use o f supervised features and feedforward neural networks [15]. After the
machine learning algorithms for condition monitoring. In training stage, an accuracy o f 100% to distinguish
Section 3, we present our analysis approach for the different states o f the bearing was reported.
employed methodology and time series classification. In The references mentioned above were all evaluated with
Section 4, we describe our industrial use-case in more data obtained from physical experimental setups and
detail. Experimental results are provided in Section 5 and made no use of VP. In general, most evaluations were
finally, the paper ends with a conclusion in Section 6 . performed on measured data. Li et al. used both VP as
References are provided in Section 7. well as experimental studies for the task o f gear multi­
fault diagnosis [16]. The employed methods include the
wavelet transform technique, Autoregressive models and
2. Related Work Principal Component Analysis. With our work, we
The data that is used for the fault detection algorithm has complement the existing research for automated fault
the nature o f time series. For the tasks o f condition detection in industrial systems.
monitoring and fault detection, time series classification
has risen in popularity in the last decade. Time series
analysis is a well-studied field [7]. In particular, time 3. Analysis Approach
series classification has been studied extensively. Bagnall
et al. presented a comprehensive review o f different 3.1. Methodology
classification algorithms and their evaluation on publicly The automated fault detection framework was briefly
available data sets from the University o f California introduced in the Introduction and Fig. 1 and is explained
Riverside time series classification archive (UCR) [8]. here in more detail. An extended schematic o f the
They have found that 1-NN DTW and Rotation Forest automated fault detection pipeline is displayed in Fig. 2.
classifiers offer the best results in most cases. Despite
these findings, we have used a shallow ANN as a We differentiate between the Virtual Domain, the
classifier, because it was very successful for our use-case. Analysis Domain and the Physical Domain. The Virtual
In a recent publication, Wang et al. showed that deep Domain contains the virtual prototype and data obtained
learning can be used for time series classification [9]. The from it, while the Physical Domain contains the physical
authors implement both a fully convolutional network counterpart that is modelled by the VP. Data output from
An Artificial Neural Network for Automated Fault Detection
142
both the Virtual and the Physical Domains is analyzed in evaluation o f time series data produced by the physical
the Analysis Domain. The arrows indicate the directions system. Analogously to the different classes in the Virtual
o f information flow between the various stages within the Domain, a criticality analysis can be performed for the
whole system. The domains and their interconnections system state o f the physical system. Based on the results
will be explained in the following, starting with the o f the criticality analysis, the subsequent decision making
Virtual Domain (left part o f Fig.2). and safety measure stages are analogous to the Virtual
Domain as described above.
The System under Test (SuT) and test bench
configurations are defined for the stimulation o f the VP The employment o f VP for the simulation o f different
within the test bench. Upon stimulation, the VP produces faults has several benefits. First, running a simulation is
time series as output. Depending on the parameters usually much faster than tracking the faults physically
defined in the SuT and test bench configurations, the and for example, in case o f degradation effects, gaining a
resulting time series have different shapes. In half o f the sufficient amount o f data in the physical world can take
cases, a fault is injected during the simulation. The time very long. Additionally, with the employment o f VP, it is
series are then used as training data to train an artificial
neural network (ANN) in the Analysis Domain. The def p a a ( t s , n ) :

ANN learns patterns in the time series and determines f = n p .z e ro s ( n ) # in itia liz e s
class labels to group similar time series into the same d = le n ( t s ) / / n # slic e length
classes. The classes can either represent different motor fo r i in ra n g e ( 0 , n ) :

states or, for the task o f automated fault detection, s l i c e = t s [ i * d : ( i + 1 ) * d ]

different faults. Thus, after classification, a criticality f [ i ] = n p .m e a n s l i c e )

analysis is performed to determine critical classes that re tu rn f

indicate critical faults. The information obtained from the


criticality analysis is used for decision making to possible to simulate many different faults and observe
conclude whether and which (safety) measures should be their propagation through the system. Some faults are
invoked to prevent system failures. These measures are very rare, so the simulation o f all the possible fault
then applied to the VP. manifestations allows for their discovery and prediction
in the physical world before they actually appear for the
first time. In the best case scenario, the simulation of
faults makes it unnecessary to obtain a data base of
possible faults in the physical world in the first place. In
that case, unexpected behavior o f physical machinery
actually becomes ‘expected’ and certain measures can be
invoked depending on the fault.
In this work, we focus on the first five stages in the
Virtual and Analysis Domains. Our use-case is a
SystemC-based VP o f an industrial motor control. The
employed SystemC framework is described in Ref. [17].
The motor control model and its parameters are described
in Section 4 in more detail. The output o f the simulation
is the engine speed over time which is a sequence of
discrete time data, and can therefore be treated as a time
series. We focus on the easiest case o f fault detection
with only two classes: ‘Pass’ if no fault is injected and
‘Fail’ if a fault is injected. The resulting labeled engine
speed data is then fed as input in the ANN that then
distinguishes faulty and faultless classes. In the future,
we intend to implement the next steps as displayed in
Fig.2 such as the stages o f criticality analysis, decision
making and measures.

Fig. 2: Flowchart of the Virtual, Analysis and Physical


Domains showing information flow and 3.2. Framework for time series classification
interconnections between the domains. The data that is used for the task o f condition monitoring
has the nature o f time series. A time series is a sequence
o f observations that are arranged according to the time of
The proposed fault detection algorithm is evaluated on their outcome [18]. In this work, we stick to univariate
data obtained by the VP. The VP is a model o f a physical time series, which have a real numbered value at each
system and the proposed algorithm can be used for point in time. A time series may be interesting as a
condition monitoring in the Physical Domain (right part whole, if it is the result o f the observation o f a finite-time
o f Fig.2). The trained ANN can be used for the process. Another category o f time series are single series
6th All-Russian Scientific Conference "Information Technologies for Intelligent Decision Making Support", Ufa-Stavropol, Russia, 2018
143
o f ongoing or online data such as temperature curves or implemented in SystemC. A schematic o f the simulated
stock value charts. The questions that pose themselves for control process and the fault injection is displayed in Fig.
those different categories of time series are somewhat 4. The test bench orders a programmable logic controller
differing, and in this paper we focus on the first kind. (PLC) when and how fast to run the motor. Given this
Thus, in the remainder o f this paper, by ‘time series’ we information, the PLC transmits timed instructions to a
specifically mean a univariate whole time series. We transaction-level modelling queue, which provides them
further assume that for a given problem, the data points the motor control unit. The motor control unit now
are time series which all have the same length. determines the duty cycle for the motor, i.e. what
percentage o f a fixed time period the motor should be
Since a time series consists o f one numeric value per time
powered. It thus conveys the number o f ‘on’ and the
step, the dimension o f a space o f possible time series is
number of ‘off’ time steps within a period to a pulse
its length, which equals to the total duration multiplied by
width modulator (PWM). The PWM applies power to the
the temporal resolution. This number can easily become
motor for the number o f ‘on’ time steps and then deprives
very large and make pointwise comparisons between time
the motor o f current for the number o f ‘off steps. The
series very time-consuming. Furthermore, many
PWM repeats this process until a new signal is received
classification algorithms require or benefit from each
from motor control.
sample being a point in a relatively low dimensional
space. Thus, while it is in principle possible to treat a The output of the simulation is the motor's rotational
time series as an element o f a very high dimensional speed over time, with subsequent time steps o f s
space, for many algorithms that try to ‘understand’ the between t = 0s and t = 0.5 s. In Fig.5, an example for
nature o f time series, it is necessary to reduce this such a time series produced by specific engine
dimension drastically. This can be done by extracting parameters is shown. After 68ms, the motor is started and
certain features from time series. Piecewise Aggregate quickly overshoots to a speed o f 305rpm at 74ms.
Approximation (PAA) is a simple feature extraction Afterwards, the system oscillates until it levels out to a
method [19]. Given a time series t s as a n u m p y array, running speed o f 197rpm. At t = 2 67ms, a load o f 7Nm
we extract n features with the following Python 3 is attached to the motor. Subsequent to a short period of
implementation o f PAA: disturbance, the motor speed falls to its final plateau of
186rpm. For this example, we disabled the feedback to
the motor controller and therefore the closed-loop
The features are the means o f n disjoint slices that cover control.
the original time series. The calculated features are Fig.6 shows the same motor with identical parameters,
returned in an array f . but this time a fault is induced into the process: at
After the feature extraction step, the data is fed into a t = 2 08ms, a transient bit error inside the PWM is
shallow artificial neural network with one hidden layer. simulated. This causes the PWM to operate on the basis
The architecture o f the ANN is depicted in Fig. 3. The o f an overlong
ANN has the structure affine ^ ReLU ^ affine ^
softmax. The softmax layer turns the score from the last
affine layer into the probability . Eventually,
'Fail' is predicted if p (' Fail ') is larger than 50% and 'Pass'
is predicted if is below 50%.
Our current analysis stops at this stage after the
separation o f faulty from faultless time series. To follow
the flowchart presented in Fig.2, a criticality analysis
would be the next step and we intend to implement it in
the future and thereby extend our automated fault
detection framework. In the current scenario, we only
consider two classes o f ‘Pass’ or ‘Fail’ and do not
distinguish different kinds o f faults. For the task o f a
criticality analysis, such a simple ‘Pass’/ ’Fail’ threshold
does not suffice. Instead, it is necessary to sort different
faults in different fault classes. Different faults exhibit
different behavior and thus, affect the system differently.
The purpose o f the criticality analysis is to determine
which faults are the most critical ones and therefore,
which measures must be invoked as a reaction. Fig. 3: Architecture of the considered artificial neural
network.

4. Use-Case
The use-case for the proposed time series classification
approach is a motor control model. This model is
An Artificial Neural Network for Automated Fault Detection
144
That means that from 500 time series used for training,
100 contain faults and 400 do not contain faults.
To reduce the samples’ dimension of to a
more manageable number, we define and extract a
desired

Fig. 4: Schematic of the simulated motor and the fault


injection process

duty cycle. Aggregated over time, this means that it


transmits an excessive amount o f power to the motor.
After 200ms, at t = 4 0 7ms, the bit error is corrected and
the PWM is controlled with the correct value. The fault
injection process results in an immediate disturbance of Fig. 5: Motor speed depending on the time without a
the speed curve and eventually the speed being increased fault being inserted (‘Pass’).
by a value o f 12rpm when running with no load and by a
value o f 16rpm while the load is present. Additionally,
the oscillation process that occurs when the load is
attached changes slightly in form.

5. Experimental Results
The full data set consists o f 8000 time series similar to
those displayed in Fig.5 and Fig.6. The motor parameters
that are modified to get this number are:

• The supply voltage (UDC): steps o f 1.4 V


between 126 V and 154 V

• The motor’s flux constant between 0.5 Vs and


0.6 Vs, step size o f 0.005 Vs Fig. 6: Motor speed depending on the time with an
inserted fault (‘Fail’).
• An amount o f inertia to start the motor from 0.5
gm2 to 1.5 gm2 with steps o f 0.1 gm2 number o f features with the method o f PAA described
earlier. We use as the number o f features. Fig. 7
These parameter variations give a total o f 4000 time shows features 10 and 15 for 25 samples with an injected
series. For each parameter configuration, the simulation fault and for 119 samples without a fault. While for those
was run once without fault injection and once with an time series without fault injection, there is a clear linear
injected fault. For the cases where a fault was injected, it correlation between the two regarded features, the
was done at a random time. This onset time for fault positions o f the samples with an injected fault are more
injection was drawn from a uniform distribution in the scattered.
interval between 200ms and 300ms.
The previously described ANN architecture was
The structure o f the data set thus contains 4000 tuples of implemented using the publicly available software
time series with and without injected faults and is package PyTorch. The width o f the hidden layer was set
therefore very regular. This regularity could lead to to 40. For backpropagation, the Adam optimizer that is
problems o f the learned classification heuristics o f bad implemented in PyTorch as t o r c h . o p t i m . A d a m was
generalization to random and independent data. Similar used [20]. The learning rate was set to 0.005. In Fig. 8,
concerns hold for the data generation part, where the training loss is shown. The results on the validation
parameter spaces were discretized in regular grids. To set are flawless: after 4 training epochs, we reach an
break these symmetries, for each run, we only used a accuracy o f 100% on the validation set. The accuracy on
fraction o f the data (500 time series for each) both for our the validation set is displayed in Fig.9. The training time
training and validation sets. As in applications with real for 4 epochs takes less than one second, which can be
data there might be relatively few samples where a fault attributed to the heavily reduced size of the training set
occurs, we chose a fault percentage o f 20% for training. due to the feature extraction. We have thus shown that a

6th All-Russian Scientific Conference "Information Technologies for Intelligent Decision Making Support", Ufa-Stavropol, Russia, 2018
145
shallow ANN can learn meaningful patterns from
features extracted with PAA and distinguish between
faulty and faultless time series.
It is important to discuss the limitations o f our approach.
We considered the task o f automated fault detection in
the context o f supervised learning. The disadvantage of
this approach is that we have to rely on the availability of
labeled data. As a next step, it would be interesting to
tackle the automated fault detection problem with
techniques from unsupervised learning. This way, a
solution for cases when there is no labeled data could be
offered.
Fig. 9: Validation accuracy during the first training
epochs of the ANN
6. Conclusion
We have demonstrated a methodology supporting
prototype is a digital twin of a physical system, the
automated fault detection using virtual prototyping. We
trained network can be applied for condition monitoring
have shown that methods from supervised machine
o f the physical machine.
learning are suitable for the task o f automated fault
detection. Based on the results from the automated fault
detection, a criticality analysis can be performed to judge
which measures must be triggered. We have evaluated 8. Acknowledgements
our methods on data obtained from VP and, since the This work has partially been supported by the German
virtual Ministry o f Science and Education (BMBF) in the project
SAFE4I under grant 01IS17032C.

9. References
1. An introduction to predictive maintenance. / Mobley,
R. K. Butterworth-Heinemann, 2002.
2. Susto G.A., et al. Machine learning for predictive
maintenance: A multiple classifier approach. // IEEE
Trans. Ind. Informat. 2015. Vol 11 № 3 (2015)
3. Li H. et al. Improving rail network velocity: A
machine learning approach to predictive maintenance.
// Transp. Res. Part C Emerg. Technol. 2014. Vol 45.
4. Schaaf, J. C. et al. Systems Concept Development
Fig. 7: Distribution of features 10 and 15 for 144 with Virtual Prototyping // Proc. o f the 29th
samples conference on Winter simulation 1997. P. 941-947.
5. Oetjens J-H., et al. Safety evaluation o f automotive
electronics using virtual prototypes: State o f the art
and research challenges. // Proc. o f the 51st Annual
Design Automation Conference. ACM, 2014.
6 . De Sa, Gomes A., Zachmann G. Virtual reality as a
tool for verification o f assembly and maintenance
processes. // CG 1999. Vol. 23 № 3 P. 389-403.
7. Fu T.C. A review on time series data mining // Eng.
Appl. Artif. Intell. 2011. Vol. 24 P. 164-181
8. Bagnall, A. et al. The great time series classification
bake off: a review and experimental evaluation of
recent algorithmic advances. // Data Min. Knowl.
Discov. 2017. Vol. 31 № 3 P. 606-660.
Fig. 8: Training loss for the training of the ANN
9. Wang, Z., Yan W., Oates T. Time Series
Classification from Scratch with Deep Neural
Networks: A Strong Baseline. arXiv:1611.06455

An Artificial Neural Network for Automated Fault Detection


146
10. Zhong J., Zhixin Y., Wong S.F. Machine condition
monitoring and fault diagnosis based on support
vector machine.// Industrial Engineering and
Engineering Management (IEEM), 2010 .
11. Shulian Y. et al. Intelligent condition monitoring and
fault diagnosis o f a gearbox based on Artificial Neural
Network. // Proc o f the ICEMI'07, IEEE, 2007.
12. Aydin I.et.al. A simple and efficient method for fault
diagnosis using time series data mining. // Proc. to
IEMDC 2007, IEEE International
13. Yang J., Sun Z., Chen Y. Fault Detection Using the
Clustering-kNN Rule for Gas Sensor Arrays. //
Sensors (Basel) 2016. Vol. 12. P. 2069.
14. Campbell B., McDonald J.R. The use o f artificial
neural networks for condition monitoring o f electrical
power transformers // Neurocomputing 1998. Vol 23
№ 1-3 (1998) P.97-109.
15. Sreejith B., Verma A. K., Srividya A. Fault diagnosis
o f rolling element bearing using time-domain features
and neural networks // Proc.of the ICIIS’2008
16. Li, Z. et al. Virtual prototype and experimental
research on gear multi-fault diagnosis using wavelet-
autoregressive model and principal component
analysis method // MSPP 2011. Vol 25 № 7
17. Reiter S. et al. Fault injection ecosystem for assisted
safety validation o f automotive systems. // High Level
Design Validation and Test Workshop (HLDVT),
2016 IEEE International.
18. http://statistik.mathematik.uni-
wuerzburg.de/fileadmin/10040800/user_upload/time_
series/the_book/2012-August-01-times.pdf (дата
обращения: 31.03.2018).
19. Yaodong Z., Glass J. A piecewise aggregate
approximation lower-bound estimate for
posteriorgram-based dynamic time warping. // Proc.
o f the 12th Annual Conference o f the International
Speech Communication Association. 2011.
20. Kingma D. P., Ba J. Adam: A method for Stochastic
Optimization. https://arxiv.org/abs/1412.6980

6th All-Russian Scientific Conference "Information Technologies for Intelligent Decision Making Support", Ufa-Stavropol, Russia, 2018
147

You might also like