An Effective Fault Diagnosis Technique for Wind Energy Conversion Systems Based on an Improved Particle Swarm Optimization

Abstract

The current paper proposes intelligent Fault Detection and Diagnosis (FDD) approaches, aimed to ensure the high-performance operation of Wind energy conversion (WEC) systems. First, an efficient feature selection algorithm based on particle swarm optimization (PSO) is proposed. The main idea behind the use of the PSO algorithm is to remove irrelevant features and extract only the most significant ones from raw data in order to improve the classification task using a neural networks classifier. Then, to overcome the problem of premature convergence and local sub-optimal areas when using the classical PSO optimization algorithm, an improved extension of the PSO algorithm is proposed. The basic idea behind this proposal is to use the Euclidean distance as a dissimilarity metric between observations in which a single observation is kept in case of redundancies. In addition, the proposed reduced PSO-NN (RPSO-NN) technique not only enhances the results in terms of accuracy but also provides a significant reduction in computation time and storage cost by reducing the size of the training dataset and removing irrelevant and redundant samples. The experimental results showed the robustness and high performance of the proposed diagnosis paradigms.

1. Introduction

Wind energy has received worldwide support for renewable energy development and has gained considerable attention over the past decade. To decrease carbon dioxide emissions, renewable energy has been commercialized and developed by most countries in the world [1]. Therefore, reducing the maintenance and operating costs of WEC systems is used in order to guarantee high reliability and effectiveness [2]. The main sources of failures, which include the power electronics interfaces, wind variability, and vibrations, are usually accompanied by unexpected faults; these must be detected and classified at an early stage to avert a system avalanche [2,3]. Thus, the development of fault detection and diagnosis (FDD) methods becomes mandatory to ensure the safe and uninterrupted operation of WEC systems with low maintenance costs. Recently, different FDD techniques for WEC systems have been developed to enhance their capacity and performance. Over the years, several FDD studies were conducted with various methodologies, advantages, and limitations [4,5]. Most of these studies prove the effectiveness of optimization algorithms [6,7]. In Ref. [8], a data-driven method combines fuzzy systems and neural networks to handle the nonlinear relationships between measurement and faults in order to diagnose a wind turbine benchmark. A deep auto-encoder (DAE) network using operational supervisory control and data acquisition (SCADA) data of wind turbines technique has been developed to detect and analyze fault in wind turbine systems [9]. This proposal can not only implement early warning of faulty components but also infer the physical location of a faulty component. A review of the fault diagnosis approaches based on machine learning techniques for wind power systems is reviewed in Ref. [10]. In Ref. [11], a hybrid approach is used for anomaly prediction in a wind turbine that uses a recurrent neural network approach for classification and XGBoost decision tree classifier for transparent outputs. In Ref. [12], a fault diagnosis technique that includes a multiscale convolutional neural network technique is developed to perform multiscale feature extraction and classification tasks. This proposal incorporates multiscale learning into the CNN model in order to improve the feature extraction ability and enhance the diagnosis efficacy. In Ref. [13], a fault diagnosis methodology based on the ensemble learning (EL) technique is developed. It aims to use ensemble techniques based on a neural network as the classifier and random subspace, boosting, and bagging as combination methods to diagnose WEC systems. In other studies, a diagnostic method based on stochastic subspace identification and multi-kernel support vector machine has been developed [14]. In this study, stochastic subspace identification is used for processing the collected signal of the wind turbine, and Gauss kernel SVM is used for the classification tasks. In Ref. [15], a hybrid approach based on the Hidden Markov model (HMM) and the principal component analysis (PCA) model to diagnose the WEC system is proposed. In this case, the PCA method is used for extracting and selecting the more significant features to be fed to the HMM classifier to perform the classification task. Due to their high effectiveness in nonlinear environments, ANN is used to distinguish sensor errors from faults originating from other turbine components. However, the main drawback of using the ANN technique presents the problem of overfitting, and it is suffering from high complexity time for training. To overcome these limitations, different advanced extensions of ANN such as backpropagation algorithms and multilayer perceptron networks (MLPN) were developed [16].

Recently, various FDD techniques based on optimization algorithms have been proposed [17,18]. Particle swarm optimization (PSO) is among the most powerful and intelligent optimization algorithms that demonstrate their effectiveness in several real-world applications such as features extraction, nonlinear process control, data clustering, and system optimization [17,19]. The main idea of the PSO technique is to use a swarm heuristic search algorithm to optimize a problem by iteratively trying to enhance a candidate solution against a given quality measure. Feature selection is addressed by selecting only relevant features by eliminating irrelevant and redundant features from raw data. The features selection using the PSO algorithm is started with all the original features (raw data). Then, the subset discovery step is done to generate the candidate subsets discovery procedure in order to search for the best subset of original features. Next, a subset evaluation step is done to measure the goodness of the generated feature subsets. Finally, the algorithm will stop according to the predetermined maximum number of iterations reached. The latter step includes whether the deletion of any feature makes a better subset. Therefore, we obtain new features less than the original features and more significant to achieve better classification performance than using all features.

Unfortunately, PSO suffers from some limitations, in particular for complex problems that become trapped in local sub-optimal areas due to premature convergence, which is caused by the lack of diversity. Therefore, the PSO algorithm suffers from early convergence, which leads to a low-quality solution. In addition, the PSO algorithm needs memory to update the velocity. To address these issues, we proposed a data reduction tool based on the use of the Euclidean distance (ED) algorithm as a dissimilarity metric between samples. The main idea behind the ED technique is to extract only a single sample in case of redundancy to construct the reduced data. ED techniques remove irrelevant and redundant samples from the original data. Therefore, the new reduced data are characterized by high diversity between samples. This, in turn, plays a pivotal role in overcoming the premature convergence problem and also provides memory in order to update the velocity. In Ref. [20], the authors indicate how the use of genetic algorithms as dimensional reduction can improve the classification task.

To address the above problems, we propose a neural networks-based fault diagnosis technique optimized by an improved PSO algorithm for WEC systems. Therefore, the contributions of this paper are two-fold: first, an efficient feature selection technique based on particle swarm optimization (PSO) is proposed. The basic objective behind the PSO algorithm is to remove irrelevant features and extract only the most significant ones from raw data in order to improve the classification task using a neural networks classifier. Second, an improved extension of the PSO model based on the use of Euclidean distance (ED) is presented. The main idea of using the ED technique is to avoid the problem of premature convergence and local sub-optimal areas when using the classical PSO optimization algorithm. In addition, the proposed reduced PSO-NN (RPSO-NN) algorithm aims to improve the results in terms of accuracy as well as in terms of complexity, time, and storage cost. The proposed algorithms will be accurate, robust, and capable of detecting and diagnosing the presence of incipient or drift WEC faults in addition to the severity of the fault. The proposed solutions will be effective in diagnosing the characteristics of WEC faults (focusing more on incipient or drift faults) under both normal and abnormal modes.

The rest of the paper is structured as follows. Section 1 introduces a detailed description of the developed methods. Section 2 presents a brief description of the used techniques. Section 3 describes the WEC system and data collection. In Section 4, the developed techniques are evaluated using the WEC system. Section 5 concludes the article.

2. System Description

In this proposal, special attention is paid to diagnosing various incipient faults during different modes of operation while taking into consideration WEC System uncertainties. A description of the variable speed wind turbine based on asynchronous machine is shown in Figure 1.

The model of the system contains two parts. One part consists of the model of the turbine and the squirrel cage induction machine (SCIG). This part of the system will be controlled through the stator side AC/DC Converter. The second part of the model consists of the grid side DC/AC converter sub-system. This configuration allows unlimited variable speed operation. The generated voltage is rectified and transformed into direct current and voltage whatever the rotation speed of the machine. The detailed description of the turbine is given in Ref. [15]. In the wind chain, the power converters topology is on two levels (Figure 2). Each converter consists of three arms. Each arm contains high and low IGBTs.

In this study, the diagnosis study concerns the IGBT11 for the generator converter and the IGBT21 for the grid converter. In this case, we considered three types of faults (short-circuit, open-circuit, and wear-out) and their description is provided in

Table 1. Main electrical faults in wind energy converters.

Fault Symbol	Description $\left({\mathit{IGBT}}_{11}\right)$	Fault Symbol	Description $\left({\mathit{IGBT}}_{21}\right)$
$S{C}_{11}$	Short-Circuit	$S{C}_{21}$	Short-Circuit
$O{C}_{11}$	Open-Circuit	$O{C}_{21}$	Open-Circuit
$W{O}_{11}$	Wear-Out	$W{O}_{21}$	Wear-Out

. The last fault is modeled by an internal resistance equal to two Ohms.

The behavior of some electrical and mechanical variables for different fault scenarios is presented in Figures 5–9 of Ref. [15].

In this study, 12 variables have been generated for modeling and fault classification as listed in

Table 2. Variables description.

Variables	Description
$C_m$	Mechanical torque (Nm)
$N_g$	Generator speed ($tr/m$)
$i_{sag}$	Generator current phase a (A)
$i_{sbg}$	Generator current phase b (A)
$I_{sd}$	Generator current along d-axis (A)
$I_{sq}$	Generator current along q-axis (A)
$V_{DC}$	Bus voltage (V )
$P_{Out}$	Output power (W)
$i_{sar}$	Grid current phase a (A)
$i_{sbr}$	Grid current phase b (A)
$I_{sd}$	Grid current along d-axis (A)
$I_{sq}$	Grid current along q-axis (A)

[13]. To demonstrate the effectiveness of the developed methodologies, real bearing vibration data are used as an example [15].

3. Diagnosis Using the Reduced PSO-Based NN Method

In the proposed method, the so-called reduced PSO (RPSO)-based NN Method, the diagnosis is achieved so that the RPSO is applied for feature extraction and NN is utilized for fault classification. PSO is a very effective global search method that relies on the intelligence of the movement of swarms. It searches for the best solution using a sum of samples including a swarm moving through the search space. However, the main limitation of the PSO algorithm is the weak local optimal search capability, which leads to problems of premature convergence and local sub-optimal areas. To overcome these problems, we propose novel PSO based on a data reduction tool using the Euclidean distance (ED) algorithm. Hence, the ED algorithm is used as a dissimilarity metric between samples such that only one sample is preserved in case of redundancy. This, hence, plays a key role in dealing with the problem of premature convergence and local sub-optimal areas and also enhances the optimal local search capability when using the classical PSO method for feature selection. Then, the selected features from the reduced data using the PSO algorithm will be fed to the neural networks classifier to perform the classification task. The proposed RPSO-NN not only improves the results in terms of classification metrics but also reduces the complexity, cost, and computation time. The main phases of the presented IRPSO-NN technique are summarized in the schematic diagram in Figure 3.

3.1. Euclidean Distance Dissimilarity Metric

The main objective behind dissimilarity metrics is to assess the differences between samples in a set of variables [21]. In this case we use the ED to calculate the similarity between data samples. Euclidean distance represents the shortest distance between two samples [21]. The dissimilarity matrix D that represents the measurement of dissimilarity between all pairs of the samples for a data matrix X with N samples and m process variables is defined as,

$$D=\left[\begin{array}{cccc}{p}_{11}& p_{12}& \dots & p_{1N}\\ p_{21}& p_{22}& ...& p_{2N}\\ .& .& .& .\\ p_{N1}& p_{N2}& \dots & p_{NN}\end{array}\right]$$

(1)

where $p_{ij}$ represents the Euclidean distance between the rows $X_i$ and $X_j$ of the data matrix X. So, $p_{ij}$ is computed as,

$$p_{ij}=\sqrt{\sum _{k=1}^m{\left(X(i,k)-X(j,k)\right)}^2}$$

(2)

Thus, the new data matrix $X_{ED}$ is presented as,

$$X_{ED}={\left[\begin{array}{cccc}{x}^{\prime }\left(1\right)& x^{\prime }\left(2\right)& \dots & x^{\prime }\left(N^{\prime }\right)\end{array}\right]}^T\in {\mathbf{R}}^{N^{\prime }\times m}$$

(3)

where $N^{\prime }$ is the size of the new training data matrix.

Then, using the ED metric, all redundancies in samples are left out from the original data matrix.

3.2. Particle Swarm Optimization (PSO)

Particle swarm optimization (PSO) is a self-adaptive population-based optimization method that is inspired by swarm social behaviors [22]. The main idea of the PSO algorithm is to use the global best solution of the swarm and the best experience in history to adjust the velocity and position of the particles [23]. PSO is equivalent to many evolutionary algorithms such as the genetic algorithm (GA), since it updates the generations in the research of the optimal solution [24]. In the PSO algorithm, a few parameter types require setting, while in the GA algorithm, different evolutionary operators such as crossover and mutation require settings [24]. In the PSO algorithm there are four important parameters: swarm size, max iteration, cognition coefficient, and social coefficient [25]. One of the advantages of the PSO algorithm is the fast convergence compared to the GA algorithm, which has a positive impact on the research for the best NN architectures and also on the time complexity of the algorithm. More details of the PSO algorithm are given in Refs. [22,25].

3.3. Artificial Neural Network

An ANN is an interconnected network of nodes, inspired by the simplification of neurons in a brain so that the computer can learn things and make decisions in a human way [26]. The ANN organism can simulate the interactive reply of the biological nervous system to real-world objects [26]. The ANN algorithm uses several different mechanisms for learning such as gradient descent-based methods like the back-propagation algorithm. The feed-forward algorithm makes the networks more applicable in different applications [27]. Neural network techniques have been widely used for FDD in various applications [28,29]. ANN consist of input, hidden, and output layer with connected nodes [29]. Weighted connections and neurons present the basic components of ANN [29].The input-output function and the weights are the main criteria to evaluate the performance of artificial NN (ANN) [29]. The main known network architectures are the feedforward and recurrent architectures. The main difference between the two types of architectures lies in the feedback between the network during the backpropagation step, which is present when using a recursive architecture by taking the correct prediction. In this paper, we adopt the Levenberg–Marquardt Backpropagation (LMBP) algorithm to train the multilayer artificial neural networks (ANN). The backpropagation (BP) algorithm has been widely used as a learning algorithm in feedforward multilayer neural networks. BP is a supervised learning technique based on the Gradient Descent (GD) method that attempts to minimize network error by descending the gradient of the error curve [30,31]. The main advantage of the BP algorithm is that it is the most efficient and easy to learn model for complex and multi-layered networks to improve the efficacy of neural networks (ANN). The Levenberg–Marquardt (LM) algorithm is a combination of gradient descent and Gauss–Newton algorithms, which aims to choose one of these combination methods to update the solution at each iteration [32]. The main objective behind using the LM algorithm is to find an optimal solution to gradient descent or Gauss–Newton algorithms [33]. Therefore, in this work, we use BP learning algorithm to train the ANN model to perform the classification task based on the Levenberg–Marquardt algorithm. The proposed ANN is constructed with 10 hidden layers and the number of hidden neurons in the hidden layer is equal to 50 (please refer to Figure 4).

4. Experimental Results and Comparative Study

In this work, seven operating modes, including one healthy and six separate faulty operating modes of the WEC system, are tested in order to make simulation data series (

Table 3. Construction of the database.

State	Process Status	Training	Testing
$C_0$	Healthy	2000	2000
$C_1$	Faulty-$S{C}_{11}$	2000	2000
$C_2$	Faulty-$S{C}_{21}$	2000	2000
$C_3$	Faulty-$W{O}_{11}$	2000	2000
$C_4$	Faulty-$W{O}_{21}$	2000	2000
$C_5$	Faulty-$O{C}_{11}$	2000	2000
$C_6$	Faulty-$O{C}_{21}$	2000	2000

). Each operating mode is adequately qualified over 2000 10-time-lagged observations during a second time period with 20 kHz as the sampling frequency.

Multi-Class (MC) Classification Results

For the multi-class (MC) classification phase, one healthy case (assigned to class C0) and six different faulty cases (assigned to C1 up to C6) are considered (Table 3). In this case, a comparison study between the developed methodologies and other existing methods such as NN, CFNN, FFNN, GRNN, RNN, SVM, and KNN is presented. This is done using a MATLAB computing environment with a PC equipped with Intel Core i5-8600K 3.4 GHz CPU (5 cores), with 8 GB of RAM. Firstly, the database in faulty mode is collected and labeled using WEC data. Then, the labeled data is used as inputs for the presented techniques. To evaluate the FDD efficacy of the proposed methods, a 10-fold cross-validation metric was adopted to compute the accuracy. For the NN, FFNN, GRNN, CFNN, and RNN, the number of hidden layers chosen is 10 and the number of hidden neurons in the hidden layer is equal to 50. The K and C parameters for SVM are chosen with the lowest RMSE value, the K value for KNN is equal to 3. For the PSO algorithm, the swarm size, max iteration, cognition coefficient, and social coefficient are equal to 20, 100, 2, and 2, respectively.

Table 4. Performances comparison of different multi-class techniques.

Method	Training		Testing
	Accuracy	CT(s)	Accuracy	CT(s)
PSO-NN	98.20	36.14	98.16	2.47
RPSO-NN	98.47	19.87	98.52	1.19
NN	93.19	45.11	93.70	3.53
FFNN	96.96	124.66	97.17	8.14
CFNN	97.02	123.87	97.17	8.36
GRNN	96.79	97.09	97.01	7.06
RNN	97.70	363.08	98.15	18.93
SVM	83.74	128.75	92.14	15.69.
KNN	77.99	6.04	88.30	0.91

presents the results in terms of accuracy and computation time of different techniques. Table 4 shows that the proposed techniques afford the best results in terms of accuracy. The classification results, given in Table 4, show that the proposed methods outperform other classical advanced techniques in terms of accuracy. Besides, the proposed PSO-NN ($98.20/98.16$) technique performed with better accuracy compared to the NN ($93.19/93.70$) method for the training and testing phases, respectively. In addition, the presented results demonstrate that the developed reduced PSO-NN (RPSO-NN) ($19.87/1.19$) provides an important reduction in terms of computation time compared to the PSO-NN ($36.14/2.47$) technique for both the training and testing phases. Thus, the proposed RPSO-NN not only reduced the computation time but also provided a little improvement in the classification accuracy. Additionally, KNN and SVM present a low accuracy and they are not able to differentiate between the different operating modes. The RNN classifier presents good results in terms of accuracy, but it suffers from a high computation time for both the training and testing phases.

According to these results, one can conclude that the proposed techniques based on features selection and data-size reduction steps present a good trade-off between the classification accuracy and the computation time compared to existing techniques. Therefore, the proposed techniques are characterized by a high classification accuracy and robustness when dealing with a large-scale dataset.

To further show the capacity of the developed techniques, the obtained results are given in

Table 5. Confusion matrix of PSO-NN.

Conf. Matrix	Predicted Class								Recall
True classes	C0	2000	0	0	0	0	0	0	100
	C1	0	1928	72	0	0	0	0	100
	C2	0	0	2000	0	0	0	0	100
	C3	0	0	0	2000	0	0	0	100
	C4	0	0	0	0	2000	0	0	100
	C5	0	0	0	0	0	1866	134	100
	C6	0	0	0	0	0	0	2000	100
Precision		100	96.40	100	100	100	93.30	100	98.16

and

Table 6. Confusion matrix of RPSO-NN.

Conf. Matrix	Predicted Class								Recall
True classes	C0	2000	0	0	0	0	0	0	100
	C1	0	1802	198	0	0	0	0	100
	C2	0	0	2000	0	0	0	0	100
	C3	0	0	0	2000	0	0	0	100
	C4	0	0	0	0	2000	0	0	100
	C5	0	0	0	0	0	2000	0	100
	C6	0	0	0	0	0	0	2000	100
Precision		100	90.10	100	100	100	100	100	98.52

using the confusion matrix (CM). CM represents the ways the model gets confused during the prediction. The x-axis and y-axis represent the instances in an actual class and the instances in a predicted class, respectively. Besides, CM presents the correctly classified observations and misclassified observations for healthy mode ($C_0$) and faulty modes ($C_1$ to $C_6$). Table 6, which presents the confusion matrix of the proposed RPSO-NN, shows that for the healthy case ($C_0$) the recall is $100\%$ and the precision is $100\%$ with $0\%$ of misclassification. In this case, 2000 samples are identified as healthy among 2000 samples. In addition, for the faulty case ($C_1$), 1802 samples are identified as healthy among 2000 using RPSO-NN. In this case, 198 samples are classified as $C_2$. Using the PSO-NN method, 1928 samples are identified and 72 samples are classified as $C_2$ during the faulty case ($C_1$). During the faulty case ($C_5$), PSO-NN identifies 1866 samples among 2000 and the rest are classified as $C_6$. It should be noted that during the faulty case ($C_1$), the proposed RPSO-NN method identifies a lower number of samples than those identified using the PSO-NN method. The presented results using two proposed methods showed the effectiveness of these techniques and also demonstrated the impact of PSO for features selection and ED for data-size reduction to afford the best trade-off between high diagnosis metrics and low computation time.

5. Conclusions

Electric power generation using WEC systems is now an active and growing field in academic and industrial research. WEC systems are among the most successful power technologies with the highest growth rate. Therefore, their proper functioning and safe handling is a top priority. In this paper, special attention was paid to the detection and diagnosis of various incipient faults of wind energy conversion (WEC) during different operating modes. Firstly, particle swarm optimization (PSO) was used as a features selection tool in order to select only the most pertinent samples from raw data. The selected features were introduced to a neural networks (NN) classifier as inputs to perform the classification task. Secondly, an enhanced extension of the PSO algorithm was done with the use of the Euclidean distance method to reduce the number of observations in the training data set. Therefore, the proposed reduced PSO (RPSO) aims to avoid the problem of premature convergence and local sub-optimal areas caused by using the classical PSO optimization algorithm. In addition, the proposed RPSO-NN once again overcomes the problem of computation time and storage cost.

Different case studies were investigated in order to illustrate the efficacy of the proposed approaches. The experimental results demonstrated the fast training time, high classification results, and robustness of the proposed method, which would exhibit desirable performance when dealing with a large-scale dataset. Additionally, the obtained results show low computation time and high diagnosis accuracy of the proposed approaches (an average accuracy greater than $98\%$) using WEC data.