Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept.

21-24, 2016, Jaipur, India

Predictive Maintenance for Wind Turbine Diagnostics


using Vibration Signal Analysis based on
Collaborative Recommendation Approach
Gopi Krishna Durbhaka Barani Selvaraj
Sathyabama University, Sathyabama University
Chennai, India Chennai, India
gopikrishna.durbhaka@gmail.com baraniselvaraj77@gmail.com

Abstract— The application of decision making knowledge The historical trend analysis of multiple data
based methods is to analyze the system and identify its in- parameters enables the analytic model to analyze the
depth diagnosis and fault behavior by simulating the performance and behavior pattern of the entire system.
expert knowledge from a similar domain. Rule based Alternatively, analyzing the behavior pattern of the different
systems with predefined conditions have been replaced parameters enables the system to suggest the replacement of
and or upgraded to expert knowledge based systems and the units and the respective transducers instead of any
further replaced / upgraded by applying machine learning emergency to happen which may lead to severe system break
techniques wherein, association rules, reasoning and downs and disruptions.
decision making processes have been considered similar to
This is to analyze different measures of similarity between the
expert knowledge in resolution of diagnostics during
predicted results of the models and used in the
critical scenarios.
The application of condition monitoring technique has
recommendations.
been widely applied to analyze the behavior pattern of the The proposed Condition Monitoring System (CMS)
system. In this paper, vibration signal analysis is performed to in this study enables to derive and analyze the failure pattern
study and extract the behavioral pattern of the bearings.
of a bearing on the gear shaft during operation. Moreover, it
Further, machine learning models such as k-Nearest Neighbor
(k-NN), Support Vector Machine (SVM), k-Means have been
further helps to classify the type of fault based on the CRA
applied to classify the type of fault. Further, Collaborative recommendation approach. It also provides an in depth insight
Recommendation Approach (CRA) has been applied here to of the type of fault that has occurred also, which is achieved
analyze the similarity of all the model results to suggest in by using few machine learning algorithms. The three types of
advance, the replacement and correction of the deteriorating the racing faults considered in this study for a bearing have
units and prevent severe system break downs and disruptions. been enlisted as inner, outer and ball.

Keywords— Vibration; bearings; Wind Turbine; behavior


II. LITERATURE REVIEW
pattern; Support Vector Machine; k-Means; k-Nearest
Neighbor; Machine learning is a field of study that provides the
system the ability to learn without explicitly being
I. INTRODUCTION programmed and acquire knowledge on its own. In the past,
for fault diagnosis, rule based engines were used with
The proceedings are the records of the conference. The predefined conditions. Recently, expert and knowledge based
productivity of the wind farms depends on the availability of systems have been overruled by applying machine learning
wind, which varies seasonally in a year. The downtime of a techniques, where in association rules, reasoning and decision
wind turbine leads to high expenses and reduce the making processes have been considered similar to expert
productivity. The logistical expenses for break down and knowledge for resolution of diagnostics during critical
disruptions is very high. Hence, predictive maintenance of a scenarios.
turbine plays a vital role during this time so as to avoid all
such expenses. Signal analysis based fault diagnosis are the most
prevalent in the literature. Signal processing is one of the
The working as well as fault behavior pattern of different classical technique which is widely adopted and applied to
parameters need to be monitored and analyzed to perform understand the performance of the system, in this case we are
health monitoring of the wind turbine. specific about the gearbox and generator of the wind turbines.

978-1-5090-2029-4/16/$31.00
Authorized @2016 IEEE
licensed use limited to: UNIVERSIDADE 1839on November 29,2022 at 10:25:45 UTC from IEEE Xplore. Restrictions apply.
DE SANTIAGO. Downloaded
2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

First a literature by Indeed, Yang [10] states about a A brief study of the widely applied techniques of Artificial
wavelet-based adaptive filter in order to extract the energy of Intelligence have been discussed below.
the generator power signal along with the fault related
Knowledge Based Systems – This system has only the
frequencies. In addition, the non-stationarity signal was
Knowledge Base (KB) explicitly programmed with a set of
treated by adjusting the filter bandwidth according to the
rules for defined conditions.
fluctuation of the wind speed. Both mechanical and electrical
abnormalities were assessed experimentally on a wind turbine Expert Systems – In artificial intelligence, an expert system is
test rig. a computer system that emulates the decision-making ability
of a human expert. Expert systems are designed to solve
Second a literature by Amirat, Choqueuse and Benbouzid
complex problems by reasoning about knowledge, like an
[2], where in fault diagnosis on generator has been done. The
expert, and not by following the procedure of a developer as is
primary focus of this literature was on mechanical failures that
the case in conventional programming.
lead to stator current amplitude modulation. In this paper,
Hilbert and Concordia transform have been applied on the A typical expert system consists of a knowledge
stator current data. base and an inference engine. The knowledge base stores facts
about the world. The inference engine applies logical rules to
Vibration signals were also widely used with classical
the knowledge base and deduces new knowledge. This process
signal processing tools in both time and frequency domain [4].
would iterate as each new fact in the knowledge base could
trigger additional rules in the inference engine.

FIGURE 1: ARCHITECTURE FLOW OF VIBRATION SIGNAL ANALYSIS SHOWCASING THE INSIGHTS OF THE BEARING FAULTS

III. CONDITION MONITORING OF BEARINGS B. Data Acquisition


Among the different types of the vibration sensors,
A. Vibration Signal Analysis commonly used for data acquisition are accelerometer,
The Vibration Signal Analysis has been evaluated here to displacement sensor, and velocity as well as torque sensors. In
evaluate the performance of non-stationary components. In a wind turbine vibration is widely applied to monitoring the
this case study it has been used for monitoring bearings and condition of bearings, blades, gearbox, shaft and tower.
deriving their failure pattern and categorizing the fault
Few of the common fault parameters of a bearing that
accordingly.
require condition monitoring have been enlisted below. The
Although, the data analysis performed in this study is following parameters shall help further to increase the
limited to a gear shaft the same approach can be applied on the accuracy by localizing a defect of the bearing and determining
gearbox bearings of the wind turbines also and predict the its root cause too.
faults accordingly. Significant Parameters of the bearing for fault detection are
ƒ Wear
ƒ Defects of bearing shells and rolling elements

1840on November 29,2022 at 10:25:45 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSIDADE DE SANTIAGO. Downloaded
2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

ƒ Crack information in shafts


ƒ Fatigue
ƒ Shaft Torque
In this study, it has been limited to predict the defect of the
bearing alone. Localizing the defect is in progress as the other
parameters also need to be acquired.

C. Signal Processing
1) Hilbert Transform
A vibration signal is a modulated signal (with FIGURE 2: FREQUENCY SPECTRUM PLOT OF HILBERT TRANSFORM
vibration, electrical signals, speed, torque etc.) and is difficult
to demodulate so as to extract the respective features and D. Data Evaluation – Results And Discussion
classify the faults accordingly. Hence, demodulation is
performed for the fault feature extraction. Hilbert Transform In this study, four different machine learning (ML) models
(HT) is applied here primarily applied here to demodulate the have been used here for evaluating the data.
signal for the fault pattern analysis of bearings. ƒ Euclidean Distance,
ƒ k-Means,
This is classical method wherein the instantaneous ƒ k-Nearest Neighbour (k-NN)
amplitude and phase of the signal are computed. ƒ Support Vector Machine (SVM)
In the below Table 1, the different types of faults have
In the field of signal processing, Hilbert transform shall
been labelled for the bearings considering two diameters 7mm
be can be computed [5] in a following steps:
and 14 mm along with the normal state.
A. First, calculate the Fourier transform of the given
signal TABLE 1. CLASSIFICATION OF FAULT LABELS
B. Second, reject the negative frequencies
C. Finally, calculate the inverse Fourier transform, and S No Fault Labels
the result will be a complex-valued signal where the 1 Inner
real and the imaginary parts form a Hilbert-transform
pair 2 Ball7
Envelope detection is applied here to extract the 3 Ball14
modulating signal from the amplitude modulated signal.
4 Outer14
Envelope Analysis shall be applied to investigate the faults of
a system that have an amplitude modulating effect on their 5 Outer7
characteristic frequencies of the system. System could be any 6 Normal
machinery, such as gearbox and generators of the turbine. Few of the vibration signal cases evaluated using the k-
2) Feature Extraction NN model have been shared below in the Table 2. With k-NN
Feature Extraction from the processed signal in frequency the computed accuracy of 87% (fine tuning with 6 labels).
spectra is performed by extracting the Peak Co-ordinates Among the 33 cases, 4 cases have been predicted incorrect; of
extracted from Hilbert Plot. The figure below shows the which only two cases (#6 & #22) have been are predicted
Hilbert plot with the Peak Co-ordinates extracted from the false, and other two cases (#29 & #32) are false due to the
Plot for a case#9 is: (108.398438, 0.314505). outlier conditions. This can be improved further fine tuning it
by availing with more data and also considering different
parameters.

TABLE 2. SUMMARY OF THE EVALUATED K-NN MODEL RESULTS

Case Peak_X Peak_Y Actual Predicted Compared


5 43.94531 0.052797 Ball7 Ball7 TRUE
6 105.4688 0.060729 Ball7 Outer7 FALSE
7 76.17188 0.058509 Ball7 Ball7 TRUE
8 76.17188 0.067902 Ball7 Ball7 TRUE
21 161.1328 0.146043 Inner14 Inner TRUE
22 87.89063 0.142216 Inner14 Outer7 FALSE
23 158.2031 0.165434 Inner14 Inner TRUE
24 155.2734 0.131393 Inner14 Inner TRUE

1841on November 29,2022 at 10:25:45 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSIDADE DE SANTIAGO. Downloaded
2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

29 131.8359 0.054599 Outer14 Outer7 FALSE


30 17.57813 0.056189 Outer14 Outer14 TRUE
31 17.57813 0.078937 Outer14 Outer14 TRUE
32 251.9531 0.049055 Outer14 Inner7 FALSE

The accuracy of the different models evaluated have been failure pattern of the bearings helps in diagnosis and locating
listed in the Table 3. the defect. A quick methodology to identify the state of
bearings is more needed for predictive maintenance. In this
TABLE 3. SUMMARY OF THE EVALUATED MODEL RESULTS paper, the scope has been limited to the prediction type of the
S No ML Model Tuner Accuracy bearing faults based on signal processing for feature extraction
1 Euclidean Distance - 81.8 % and applying few of the machine learning models along with
2 K-means k=6 81.8 % novel approach known as Collaborative Recommendation
3 k-NN k=6 87 % Approach. The effectiveness of this methodology has been
4 SVM Linear 78.8 %
showcased analyzing the faults of bearings using multiple
IV. COLLABORATIVE RECOMMENDATION APPROACH machine learning models. Finally, the CRA has been used to
recommend the fault with an accuracy of 93%.
Collaborative Recommendation Approach (CRA) is applied
here to analyze the similarity of all the above evaluated results VI. FUTURE ENHANCEMENTS
of the different models to predict the type of fault more The process of localizing the defect is in progress where in
precisely. This has been achieved by applying weightage to the additional parameters such as velocity, torque etc., also
each result of the model and predicting the fault accordingly. need to be acquired and considered for the data analysis.
The heuristics considered here for the CRA are based on Further, looking forward for support so as to test the same on
the confirmation of the user with the type of recommended a wind turbine with real time data both offline and online for
fault it actually is, at the end. Hence, the weights get adjusted recommending the replacement of the units and preventing
and tuned accordingly based on their accuracy. catastrophic failures by predictive maintenance.
For better understanding of the Collaborative
Recommendation Approach (CRA) model, consider the case REFERENCES
#32 (actual is outer racing fault) wherein each model has [1] F. N. Ford. 1985. Decision support systems and expert systems: a
predicted as follows show in the Table 4. comparison. Inf. Manage. 8, 1 (January 1985), 21-26.
[2] Amirat, Y., Choqueuse, V., & Benbouzid, M. E. H. (2010). Wind
TABLE 4 CRA RESULT ANALYSIS turbines condition monitoring and fault diagnosis using generator
current amplitude demodulation. Proceedings of IEEE Energy
S No ML Model Predicted Conference and Exhibition (EnergyCon), pp. 310-315.
1 Euclidean Distance Outer14 [3] I. Jaksch, “Fault diagnosis of three-phase induction motors using
2 K-means Outer14 envelope analysis,” in SDEMPED, Atlanta, USA, 2003, pp. 289–293.
3 k-NN Inner7
[4] Zijun Zhang; Verma, A.; Kusiak, A., "Fault Analysis and Condition
4 SVM Outer14 Monitoring of the Wind Turbine Gearbox," in Energy Conversion, IEEE
Transactions on , vol.27, no.2, pp.526-535, June 2012, doi:
The model holds few heuristics as inputs such as priority 10.1109/TEC.2012.2189887
where the input shall be 0 / 1. Given the models predicted [5] Andrew Kusiak, Zijun Zhang, Anoop Verma, “Prediction, operations,
results and based on the outcome frequency from the group of and condition monitoring in wind energy”.
models, the type of fault shall be recommended by CRA. In [6] Yi-Wen Liu (2012). “Hilbert Transform and Applications, Fourier
Transform Applications”, Dr Salih Salih (Ed.), ISBN: 978-953-51-0518-
the above case it is outer racing fault for the bearing with 3.
14mm diameter. [7] Jiang YH, Tang BP, Qin Y, Liu WY. Feature extraction method of wind
CRA further has additional heuristics and is in progress of turbine based on adaptive Morlet wavelet and SVD. Renewable Energy.
enhancement along with the additional parameter to localize 2011; 36(8): 2146-2153.
the defect appropriately. Moreover, option to bias the CRA [8] Abbasion S, Rafsanjani A, Farshidianfar A, Irani N. Rolling element
bearings multi-fault classification based on the wavelet denoising and
priority heuristic by choosing the output of the best-fit model support vector machine. Mechanical Systems and Signal Processing.
(because in industry few analysts have their preferences in 2007; 21(7): 2933-2945.
choosing models on which detail analysis and study would [9] Konar P, Chattopadhyay P. Bearing fualt detection of induction motor
have been performed in their datasets) or by analyzing the using wavelet and Support Vector Machines (SVMs). Applied Soft
given weights on its own, both have been provided as option Computing. 2011; 11(6): 4203-4211.
for the user to select with. [10] Bouthaina Abichou1 , Diana Flórez 2 , Moamar Sayed-Mouchaweh1 ,
Houari Toubakh1 , Bruno François2 and Nicolas Girard3 “Fault
Diagnosis Methods for Wind Turbines Health Monitoring: a Review”
V. CONCLUSION
[11] Krishna, D. Gopi. "Preventive maintenance of wind turbines using
Bearings are the one of the most essential parts in a rotating Remote Instrument Monitoring System." Power India Conference, 2012
machinery especially wind turbines. Hence, identifying the IEEE Fifth. IEEE, 2012.

1842on November 29,2022 at 10:25:45 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSIDADE DE SANTIAGO. Downloaded

You might also like