You are on page 1of 5

Wind Turbine Fault Classification Using Support

Vector Machines with Fuzzy Logic


Colton Seegmiller Blake Chamberlain Jordan Miller
Engineering Department Engineering Department Engineering Department
2022 Intermountain Engineering, Technology and Computing (IETC) | 978-1-6654-8653-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/IETC54973.2022.9796919

Utah Valley University Utah Valley University Utah Valley University


Orem, UT, 84058, USA Orem, UT, 84058, USA Orem, UT, 84058, USA
10803013@uvu.edu 10737027@uvu.edu 10767354@uvu.edu

Mohammed A.S. Masoum Mohammad Shekaramiz


Engineering Department Engineering Department
Utah Valley University Utah Valley University
Orem, UT, 84058, USA Orem, UT 84058, USA
m.masoum@ieee.org mshekaramiz@uvu.edu

Abstract— Rapid and accurate identification of faults on wind include leading edge erosion, and cracks running along the
turbine blades is important to ensure the continued operation of surface of the blades. When visually inspecting wind turbine
wind power generation systems. This paper explores the blades, these types of faults can be clearly identified. However,
implementation of Support Vector Machines (SVM) combined it is not always feasible to perform manual visual inspections
with fuzzy logic for image recognition and fault classification of on every wind turbine system during normal operation. Because
wind turbine blades. We discuss the concept, ideas, and of this, maintenance must be scheduled on wind turbines, which
implementation of SVM for image recognition, and the final result can be costly, hard to implement, and complex to schedule [7].
is to implement these features into a system for detecting the
various cracks and damages on the blades of wind turbines using To solve this issue, it is beneficial to develop an
a scale model. The final system will be tested on a scale model of a autonomous system which detects faults with minimal manual
wind turbine blade. We will focus on what SVM is, what the intervention. Furthermore, an autonomous system minimizes
crossover between SVM and fuzzy may look like, and how it will the labor involved in manually identifying faults in wind
effectively be able to detect cracks in the blades of wind turbines. turbine blades. This provides a more cost-effective maintenance
solution while still accurately identifying faults present.
Keywords—Fuzzy, SVM, fault diagnosis, wind turbine blades
In this paper, the autonomous detection of faults in machine
I. INTRODUCTION learning image recognition technology is used. The
In wind turbine systems, continual monitoring and implementation involved uses a Support Vector Machines
maintenance is necessary to ensure continued operation of the (SVM) algorithm to identify faults in images of wind turbine
system. Due to the harsh operating conditions experienced by blades. Additionally, fuzzy logic is used to classify faults.
most wind turbines, it is common for faults and damage to be Using fuzzy logic helps to make the classification after the
incurred through normal operation [1]. Quick and accurate SVM algorithm more accurate. Instead of a binary
identification of potential faults is important to ensure that no classification provided by an SVM algorithm, fuzzy logic
lasting damage is caused to the wind turbine system, and allows for more accurate classification of faults.
maintenance of a wind turbine system incurs significant costs This paper is organized as follows. Section II presents an
through the operational lifespan of the wind turbine [2]. overview of SVM algorithms and their applications in image
Typically, faults are identified using sensor data such as recognition. Section III is a description of fuzzy logic and how
through a Supervisory Control and Data Acquisition (SCADA) it applies to classifying faults. Section IV discusses how the
system [3]-[4]. Detecting anomalies through this data, such as SVM and fuzzy logic system will be developed and
vibrations, can be useful for detecting faults on certain implemented. The conclusion and summaries are presented in
components in wind turbines. However, when identifying faults Section V.
present on the wind turbine blades, manual visual inspections II. SUPPORT VECTOR MACHINES (SVM)
are necessary. This is necessary because when a fault is present
on the wind turbine blades, the SCADA system data does not A. Modeling of SVM
show any anomalies as a result of the data [5]. Thus, the SVMs are supervised learning models used in machine
imbalances and faults in the wind turbine system are difficult to learning that are primarily employed for classification but can
detect through sensor data, and alternate methods of identifying also be used for regression and outlier detection.
the faults must be implemented. Applications of these models can be seen in face recognition
The nature of faults present on wind turbine blades means [8], text categorization [9], predicting protein-protein
that they can be detected through visual inspection [6]. For interaction sites [10], image classification [11], and for fault
example, some common types of faults on wind turbine blades

978-1-6654-8653-8/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on November 10,2022 at 19:54:19 UTC from IEEE Xplore. Restrictions apply.
detection in various devices such as wind turbine blades
[12][13][14].
SVM models use training techniques to develop an
algorithm that ultimately classifies the data with more accuracy.
The goal of the SVM is to create a hyperplane that splits the
data into a processible output of different classes. The SVM
starts by finding points from each of the classes that are closest
to the hyperplane; these points are called support vectors. The
distance between these points and the hyperplane is called the
margin. This distance is calculated and maximized as much as
possible to obtain the optimal hyperplane.

Fig. 2 An example of SVM mapping [13]

The nature of a dataset will determine what type of kernel


function to use. Some common kernel functions include the
linear kernel function, polynomial kernel function, Radial Basis
Function (RBF), sigmoid kernel function, Fourier kernel
function, B-spline kernel function, and the wavelet kernel
function [15]. The RBF kernel, specifically a Gaussian radial
basis function, is a commonly used kernel function known for
accurate results. This function can be very versatile for different
dimensions, and it is more concise since it uses fewer

, used to optimize an SVM [15]:


parameters. Below shows the equation for an RBF kernel
function
Fig. 1 Optimal hyperplane of SVM [12]

| |
, : exp 2
2
In Fig. 1, there is a sample of circles and a sample of
triangles. H2 and H1 are the lines that cross through the support
vector points for each of the samples. The margin between these
lines is equal and they are parallel to H, which is the optimal
hyperplane that was calculated through the SVM. The equation Equation 2 shows an expansion of the kernel . On the
of the hyperplanes is [12]: numerator is the square of the distance between two of the
feature vectors x and y, on the denominator is which
represents the variance. When training and testing an SVM,
: ⋅ 0 adequate width selection of the kernel for the best accuracy is
1: ⋅ 1 1
also necessary. The output of an SVM is a binary classification
between two or more classes. If a percent accuracy is wanted
2: ⋅ 1, instead, fuzzy inference can be implemented into an SVM
model for even higher accuracy.
III. FUZZY-SVM
where is the normal vector of the hyperplane , and
represents the segmentation threshold from the origin. Since all The crossover between the previously discussed topics in
of the samples are outside the boundaries of hyperplanes 1 SVM and fuzzy is an idea that can bring a lot of different
and 2, the constraints to the equation are satisfied [12]. options with the implementation of fuzzy logic. Fuzzy logic
refers to the space between 1 and 0 where a truth value can lie
B. Optimization of SVM [16]. This truth value can express variables of partial truth that
There are cases where a dataset can be non-linear, which lie in the “fuzzy” space between 1 and 0 where binary has no
requires the SVM to be optimized to separate the data more output case. Fuzzy has been applied to many different cases and
accurately. If the dataset is non-linear, a higher dimensional systems, and our project is no exception. In this case, however,
plane can be implemented to the equations using kernel the challenge will be to find the correct crossover between
functions. A kernel function is how an SVM does mapping fuzzy and SVM.
from a low space to a high dimension space [13]. Fig. 2 depicts The advantages for using an SVM system with fuzzy logic
how an SVM maps through a dataset in a high dimension integrated make the concept attractive for use in many different
feature space. applications, one of those being image recognition. While this

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on November 10,2022 at 19:54:19 UTC from IEEE Xplore. Restrictions apply.
idea is fairly new in the industry, the concept is one that is very
The first image categories discussed in [17] was outdoor
advantageous when working with an output that could benefit
from a more complex classification than a simple binary result. versus indoor images where the fuzzy SVM would classify the
The biggest advantage of fuzzy SVM lies in the fact that the given image into an outdoor or an indoor class with a
output will give the user a better idea of what the output entails. confidence level attached. The accuracy of this first
For example, the fuzzy output combined with the output of the classification set was reported to be 86% for the outdoor images
SVM can provide a higher confidence level than a simple 1 or and 84% for the indoor images.
0 that would come from the SVM alone. The second image categories discussed in [17] was natural
A. Introduction of fuzzy logic in SVM versus artificial landscapes. The fuzzy SVM took images from
The fuzzy SVM combines SVM and fuzzy to create a deeper the decided outdoor images classification from the previous
reading of the data. When the data set used is linear or semi step and from there classified each as either a natural setting or
linear, a linear kernel is desired to best classify the data and find an artificial one. The accuracy during this step for the natural
the optimal hyperplane for the data. In a nonlinear case, settings was 92.2%, whereas the accuracy level for the artificial
however, an RBF kernel is often used. The reason is that an settings was 93%. In this step the fuzzy SVM was even
RBF kernel better classifies the complex data points that often compared to the accuracy of a Learning Vector Quantization
come in nonlinear data, as discussed earlier in this paper. There (LVQ) classifying the same images. It was found that the fuzzy
are many different kernel types for SVM functionality, but the SVM was more accurate in its classifications.
RBF kernel is one of the most common options due to its The last/third image category discussed in [17] stemmed
success rate in classifying data, especially nonlinear data. One from the natural landscapes category, which was classified into
of the biggest advantages of the crossover between fuzzy and different categories such as Mountain, Sea, Forest, and others.
SVM is how it handles noisy data. Noisy data in our case would In the Sea subcategory the images were assigned a choppy sea
refer to data in the images that is useless, corrupted, or ruined versus calm sea classification. The calm sea not only received
by different factors. A normal SVM has trouble classifying this a positive classification, but also an “easy” feeling, whereas the
data since it is so difficult to detect, whereas a fuzzy SVM choppy sea received a negative classification accompanied with
system can handle this type of data due to the ability to classify a “nervous” feeling. Specifically, in this last classification step,
it in the fuzzy space between 1 and 0. Fuzzy SVM brings the we can see the power that fuzzy logic brings to the classification
ability of “producing unambiguously a generalized separating process by adding another layer of detail to the system that is
hyperplane with the uncertainty from the noisy samples which not possible through an SVM alone.
are neighboring around the decision boundary” [17].
IV. IMPLEMENTATION AND FUTURE RESEARCH
B. Previous applications of fuzzy SVM
A. Implementation of fuzzy SVM
There have been many previous applications of fuzzy SVM,
some being powerful image recognition [17] and even protein This system will be implemented in a scale model of a wind
binding sites [10]. In reference [17], an example of a fuzzy turbine. To do this: i) initial image processing is executed in
SVM applied to image processing is presented. In reference Python to reduce the noise present in the images while image
[17] the authors use the fuzzy SVM to process images from formatting is performed in MATLAB, ii) the SVM algorithm
three different categories. and the image recognition and classification as well as

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on November 10,2022 at 19:54:19 UTC from IEEE Xplore. Restrictions apply.
additional classification through fuzzy logic is performed in every hour, to ensure that faults are quickly and accurately
MATLAB, and iii) the resulting classification for images is detected. The final system will be tested using a real-world
outputted and checked for accuracy. Fig. 4 shows the block scale model of a wind turbine.
diagram of the complete classification process.
V. CONCLUSION
First, the images of wind turbine blades with and without
Autonomous detection of faults in wind turbine blades is
faults are obtained from the real-world wind turbine prototype
important to ensure continued operation of wind energy
system (Fig. 5). These images are processed in Python to reduce conversion systems. Utilizing SVM algorithms for image
the noise within the photos and to format the data for use in the recognition along with fuzzy reasoning for classification will
MATLAB script for the SVM algorithm. A set of wind turbine allow faults to be quickly and more accurately detected with no
images with faults is required to train the SVM algorithm. input necessary from a user. Fuzzy logic is used to make the
Once the images are formatted using Python, the MATLAB classification more precise, instead of a typical binary
script for the SVM algorithm and fuzzy logic is used. The SVM classification used in an SVM algorithm.
algorithm uses the training set to predict if faults are present in This paper introduced the implementation of SVM
the testing data set. Because a typical SVM algorithm returns a combined with fuzzy logic for image recognition and fault
binary classification, after the SVM algorithm is applied, fuzzy classification of wind turbine blades. The scale model of a wind
logic is used to further classify the faults on the images. turbine prototype at Utah Valley University was used for
imaging of faulty blades. Initial image processing is executed
in Python and the Fuzzy-SVM system is implemented and
Data tested in MATLAB. Using an SVM machine learning algorithm
Acquisition for detecting faults along with fuzzy classification helps to
more accurately identify faults present in wind turbine blades.
The future direction of this undergraduate research will be
Image on the implementation and testing of an autonomous drone-
Processing based system to have a reliable and effective fault detection
technology for wind energy conversion system.

Binary Fault
Classification
Based on SVM

Classification with No Fault (Final


Fuzzy Logic Classification)

Probability of
Fault (Final
Classification)

Fig. 4 The block diagram of the implemented process flow to be improved in


future work

B. Future research
Our ongoing research aims to have a fully autonomous
system which does not require inputs from the user to detect
faults. To do this, an autonomous drone will be developed. The
drone will be equipped with a camera to take photos of the wind
turbine systems and its blades. Then, the images will be
uploaded for image processing and used for the SVM algorithm
and fuzzy classification system to detect faults. A drone flight
path which takes photos from several angles is to be developed
and implemented.
Utilizing an autonomous drone will help to make the fault Fig. 5 Prototype of wind turbine with faults (shown in red)
(Photo taken at UVU on 3/5/22)
detection system fully autonomous. A maintenance schedule
can be developed for the drone, such as flying and taking photos

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on November 10,2022 at 19:54:19 UTC from IEEE Xplore. Restrictions apply.
ACKNOWLEDGMENTS [10] B. K. Sriwastava, S. Basu and U. Maulik, "Predicting protein-protein
interaction sites” with a novel membership based fuzzy SVM classifier,"
IEEE/ACM Transactions on Computational Biology and Bioinformatics,
This work is supported by the Office of the Commissioner vol. 12, no. 6, pp. 1394-1404, 1 Nov.-Dec. 2015, doi:
of Utah System of Higher Education (USHE)-Deep Technology 10.1109/TCBB.2015.2401018.
Initiative Grant 20210016UT. [11] H. B. Wang and J. H. Ma, "A classification method of multispectral
images which is based on fuzzy SVM," 2008 International Conference on
Computer Science and Software Engineering, 2008, pp. 815-818, doi:
REFERENCES 10.1109/CSSE.2008.784.
[12] S. Qiang, H. Lin and Z. Yu, "Online faults diagnosis of wind turbine
[1] J. Ribrant and L. Bertling, “Survey of failures in wind power systems blades based on support vector machines," 2016 3rd International
with focus on Swedish wind power plants during 1997-2005,” 2007 IEEE Conference on Systems and Informatics (ICSAI), 2016, pp. 247-250, doi:
Power Engineering Society General Meeting, 2007, pp. 1–8. 10.1109/ICSAI.2016.7810962.
[2] H. Yin, R. Jia, F. Ma, and D. Wang, “Wind turbine condition monitoring [13] J. H. Shin, Y. S. Lee and J. O. Kim, "Fault prediction of wind turbine by
based on scada data analysis,” 2018 IEEE 3rd Advanced using the SVM method," 2014 International Conference on Information
InformationTechnology, Electronic and Automation Control Conference Science, Electronics and Electrical Engineering, 2014, pp. 1923-1926,
(IAEAC), 2018, pp. 1101–1105. doi: 10.1109/InfoSEEE.2014.6946258.
[3] T. Matsui, K. Yamamoto, S. Sumi and N. Triruttanapiruk, "Detection of [14] W. Tuerxun, X. Chang, G. Hongyu, J. Zhijie and Z. Huajian, "Fault
lightning damage on wind turbine Blades Using the SCADA System," diagnosis of wind turbines based on a support vector machine optimized
IEEE Transactions on Power Delivery, vol. 36, no. 2, pp. 777-784, April by the sparrow search algorithm," IEEE Access, vol. 9, pp. 69307-69315,
2021, doi: 10.1109/TPWRD.2020.2992796. 2021, doi: 10.1109/ACCESS.2021.3075547.
[4] Y. Shi, Y. Liu, and X. Gao, “Study of wind turbine fault diagnosis and [15] W. An-na, Z. Yue, H. Yun-tao and L. I. Yun-lu, "A novel construction of
early warning based on scada data,” IEEE Access, vol. 9, pp. 124 600– SVM compound kernel function," 2010 International Conference on
124 615, 2021. Logistics Systems and Intelligent Management (ICLSIM), 2010, pp.
[5] Y. Yu, H. Cao, S. Liu, S. Yang, and R. Bai, “Image-based damage 1462-1465, doi: 10.1109/ICLSIM.2010.5461210.
recognition of wind turbine blades,” 2017 2nd International Conference [16] P. V. S. Reddy, "Generalized fuzzy logic with twofold fuzzy set: learning
on Advanced Robotics and Mechatronics (ICARM), 2017, pp. 161–166. through neural net and application to business intelligence," 2021
[6] F. Zhang, Y. Li, Z. Yang and L. Zhang, "Investigation of wind turbine International Conference on Fuzzy Theory and Its Applications
blade monitoring based on optical fiber Brillouin sensor," 2009 (iFUZZY), 2021, pp. 1-5, doi: 10.1109/iFUZZY53132.2021.9605090.
International Conference on Sustainable Power Generation and Supply, [17] J. Chen, D. Zhang and H. Li, "The research of SVM introjecting fuzzy
2009, pp. 1-4, doi: 10.1109/SUPERGEN.2009.5348095. theory in image affective recognition," 2008 International Seminar on
[7] L. Liu, Y. Zhou, B. Yan and J. Liu, "Large-scale maintenance scheduling Future BioMedical Information Engineering, 2008, pp. 71-74, doi:
of wind turbines," 2019 Prognostics and System Health Management 10.1109/FBIE.2008.69.
Conference (PHM-Qingdao), 2019, pp. 1-5, doi: 10.1109/PHM- [18] B.B. Gao, J.J. Wang and C.Y. Yang, “Coordinate descent fuzzy twin
Qingdao46334.2019.8943053. support vector machine for classification,” 2015 IEEE 14th International
[8] Zhao Lihong, Song Ying, Zhu Yushi, Zhang Cheng and Zheng Yi, "Face Conference on Machine Learning and Applications (ICMLA), 2015, pp.
recognition based on multi-class SVM," 2009 Chinese Control and 7-12, doi: 10.1109/ICMLA.201.
Decision Conference, 2009, pp. 5871-5873, doi:
10.1109/CCDC.2009.5195250.
[9] M. A. Kumar and M. Gopal, "An investigation on linear SVM and its
variants for text categorization," 2010 Second International Conference
on Machine Learning and Computing, 2010, pp. 27-31, doi:
10.1109/ICMLC.2010.64.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on November 10,2022 at 19:54:19 UTC from IEEE Xplore. Restrictions apply.

You might also like