A methodological approach for detecting multiple faults in wind turbine blades based on vibration signals and machine learning

4.1 Experimental rig

The experimental test rig used in this study is based on a wind turbine blade and utilizes the Computer-Controlled Wind Energy Unit (EEEC) provided by Edibon Equipment, as shown in Figure 3(a). It comprises a laboratory-scaled aerogenerator with a rotor, generator, and computer-controlled axial fan. The rig allows air velocity control by adjusting the rotational speed and offers flexibility in blade configuration. The set-up includes a stainless steel tunnel with transparent windows to simulate natural wind conditions, with wind tunnel velocities ranging from 1.3 to 5.3 m/s. The aerogenerator has a diameter of 510 mm and generates 60 W of power with a maximum voltage of approximately 12 V and an operational charging current of 5 A.

Figure 3

Installation of the wind turbine system (a) EEEC wind turbine, (b) accelerometer attached to the turbine hub, and (c) DAQ of the DAQ model NI-USB-4431.

In the experimental design, as depicted in Figure 4(b) and (c), a piezoelectric accelerometer was employed as a transducer to capture vibration signals. This sensor is well-suited for detecting faults at high frequencies and is commonly used in condition monitoring. The specific accelerometer model utilized in the study is the PCB Piezotronics 352C65 uniaxial accelerometer; its specification is shown in Table 1. An adhesive mounting technique was employed to securely install the accelerometer on the nacelle near the wind turbine hub, enabling vibration data collection.

Figure 4

The simulated faults in the blades.

A cable connected the accelerometer to the data acquisition (DAQ) card. The NI USB 4431 DAQ card was utilized in the study, featuring five analog input channels, a sampling rate of 102.4 kS/s, and a resolution of 24 bits. The accelerometers and the DAQ devices interfaced to a Lenovo laptop equipped with Core i7 CPUs. The DAQ process was facilitated using LabVIEW software.

4.2 Experimental procedure

Initially, the wind turbine was healthy (without defects); the accelerometer was used to record the signals. These signals were captured using the listed requirements:

1. The sample length was established to maintain consistency, and the following factors were also considered. Statistical measures are more relevant when the number of samples is large enough. On the contrary, as the number of samples increases, so does the computation time. According to the Nyquist sampling theorem, the sampling frequency must be at least twice the maximum frequency to achieve balance [37]. Hence, the sampling rate was set at 1,000 Hz.

2. A minimum of 500 samples were collected for each state of the wind turbine blade, and vibration signals were recorded using LabVIEW 2020.

The turbine was operated at 240 rpm. The accelerometer is positioned vertically on top of the hub to monitor vibrations (y-axis), as illustrated in Figure 3(b). DAQ is used to gather vibration signals at a sampling rate of 1,000 Hz and a sample size of 500. This results in four rotations (240/60 ≈ 4 rotations per second). The following faults were simulated one at a time on the blade. In contrast, the remaining blades and components remained in good condition, and the relevant vibration signals were obtained. ‎

4.3 Intentionally adopted faults

In this research, we created models of wind turbine blades based on normal conditions and common blade fault states (cracks with different locations) to discuss the vibration signals generated by wind turbine blades in various states. The blades in this study were custom-designed by the manufacturer of genuine commercial wind turbines. The blades were made of FRP, measured 300 mm long, and were solid from the inside. Figure 4 shows the three simulated fault types in addition to the healthy case of the blades used in this investigation. This study defines three fault types such as F_a: Blade tip crack fault, F_b: mid-span crack fault, and F_c: crack ‎near the root fault.

5.1 Feature selection

Before pattern recognition, selecting features is crucial because it eliminates loud, redundant, or unnecessary features, significantly reducing the number of features. In most cases, it optimizes classification tasks and improves the performance of learning algorithms. In the diagnosis procedure, selecting the appropriate features or collecting features that reflect the device’s condition is important. It is believed that a good feature or set of features allows one to distinguish between normal and abnormal circumstances, enabling trend analysis while avoiding the impact of other device operating parameters [36]. In most cases, selecting features is considered a dimension-reduction problem. Techniques such as principal component analysis, multidimensional scale, factor analysis, projection search, and kernel Fisher discrimination analysis are used. However, these methods usually produce synthetic properties greater than the original set, so the reduced set properties are not of physical importance [37,40]. Fisher’s scores, ReliefF algorithms, Wilcoxon ranks, gains ratios, memetic characteristics selection, chi-squares, and IGs are used to select relevant characteristics and improve precision in the diagnosis of mechanical failures [41–43].

5.2 ReliefF algorithm

ReliefF is a supervised feature classification algorithm [44]. It is typically used in data preprocessing to select feature subsets. ReliefF is based on randomly generating instances, computing their nearest neighbors, and adjusting a feature weighting vector to provide greater weight to attributes that distinguish the instance from neighbors of other classes. It is an extension of the Relief approach (used for binary classification) that aims to evaluate the quality of the distinguishing factors of neighboring samples [45]. The ReliefF algorithm begins by selecting a random instance, followed by a search for the k nearest instances of the same class. This operation alters a weighting vector (W) that gives greater weight to traits that better differentiate between surrounding groups and is defined by [46]:

where W _f represents the weight of the feature f.

5.3 Chi-square (χ²)

The chi-square statistical optimal feature selection approach was used to improve prediction accuracy [47]. Feature ranking allows testing to determine whether a specific feature’s occurrence and a specific class’s occurrence are independent. Thus, when a feature is independent of the class, this is discarded [48]. It can be computed as follows:

In this equation, χ² represents the chi-square statistic, and the summation symbol (∑) signifies that the equation is computed for each category or cell in the dataset. u _j denotes the observed frequency in each category or cell, while u _j represents the expected frequency in each category or cell, assuming no association exists between the variables.

5.4 IG

IG is a crucial metric in machine learning for evaluating feature relevance in classification tasks [43–49]. It measures the reduction in entropy when a feature is known. The IG for a feature is calculated as the difference between the entropy of the original dataset and the weighted average of the entropies of the subsets created by splitting the dataset based on that feature. The equation for IG is

where IG(F) is the IG for feature F, H(D) is the entropy of the original dataset, |Dv| represents the number of instances in each subset, and H(Dv) is the entropy of each subset. Decision trees use IG to determine feature selection order, selecting features with higher IG as more relevant for accurate classification. Alternative metrics like gain ratio and Gini index address limitations of IG. Overall, IG quantifies the reduction in entropy and aids in identifying informative features for classification.

6.1 SVM

SVM is a supervised learning algorithm mainly used for classification and regression. Vapnik described the theoretical concepts of SVMs [50]. Due to its high precision and good generalizability, some researchers [51,52] have used SVM to classify mechanical failures in rotating machines, even if the sample is small. The formulation of the SVM is based on the principle of minimization of structural risk. For binary classification problems, the aim is to maximize the margin between the different planes. The maximum margin to separate hyperplanes (H ₁) can be used to classify the data sets into the classes to be considered. The equation of H ₁ can be written as follows.

where x is the point on the separator plane (H₁), and w is the vector on the plane. Normalization of the two-class w parameters can be represented as

and

By combining Eqs. (13) and (14), we obtain the following:

where ξ _i represents the slack parameter.

Due to better generalization capabilities, SVM is of great interest to academic and industrial societies as an algorithm for fault detection systems.

6.2 KNN

The KNN algorithm is a supervised learning approach used for classification and regression tasks. It is a nonparametric model that utilizes training datasets to classify new samples from test datasets based on nearest-neighbor criteria [53]. The algorithm searches for k samples in the training set that is closest to the new test sample. Classification is then based on the most prevalent classes among the nearest neighbors. Given a training set D(x, y) where x represents a sample and y its corresponding class, and a test sample z = (x′, y′), the algorithm calculates the distance between z and all training samples (x, y) in D to obtain a list of nearest neighbors [54]. The class assignment for y corresponding to the test sample x is determined by a majority vote of the neighboring classes.

The class assignment equation can be expressed as follows:

where v represents a class label, yᵢ denotes the neighboring class label, and I is an indicator function that returns 1 if the condition in parentheses is true. This equation allows for the determination of the class with the highest frequency among the neighboring samples.

In some cases, a weighted approach is used to account for the contribution of each neighboring sample based on its distance from the sample to be classified. The weight factor can be defined as the inverse square of the distance:

In this manner, the kNN algorithm can be defined as follows:

Another important consideration in the kNN algorithm is the choice of distance metric. The most common distance metric used is the Euclidean distance, but alternative metrics such as cosine similarity, Minkowski distance, correlation, and chi-square distance can also be employed [55].

In summary, the kNN algorithm utilizes nearest-neighbor principles to classify new samples based on their proximity to training samples. The class assignment is determined through majority voting, and a distance metric is used to measure the similarity between samples [56].

6.3 RF

RF is a machine-learning technique introduced by Breiman [57] that leverages an ensemble of decision trees. By combining the concepts of bagging and random feature selection, RF aims to address issues related to variance and overfitting. In this approach, each tree in the forest independently determines the class for a given sample, and the final class prediction is made through majority voting. The training data used to construct each tree are referred to as the “in-bag” data, while the remaining data constitute the “out-of-bag” observations (OOB) [58]. OOBt represents the OOB sample associated with tree t. The classification error of the forest, errForest, can be defined as follows:

where y i represents the true class label for the ith sample, and ȳ_i denotes the majority class predicted by the trees where the sample i is part of the OOBt. It is important to note that “Cart” in the equation may have been intended to represent a specific mathematical function, but its precise definition or context is not provided. Further elaboration or clarification is required to ensure an accurate understanding and interpretation of the equation.

6.4 Evaluating the machine learning model

Developing a machine learning model is a crucial skill for aspiring data scientists. However, the initial model is rarely the “best” model. Evaluating the quality of our machine learning model is crucial for improving its performance until it reaches its maximum potential. Evaluation metrics for classification problems compare the expected class label to the predicted class label or interpret the predicted probabilities for class labels. Classification problems are widespread and have numerous applications in the real world, such as identifying spam emails, targeting marketing, fraud detection, and determining whether a patient is at high risk of having a particular disease diagnosis. In this blog article, we examine various categorization evaluation metrics that can be applied to issues of this nature.

• Confusion matrix (CM)

  The CM is crucial in classification tasks, summarizing the predicted and observed values. It represents four outcome combinations and enables evaluation using precision, recall, accuracy, F1 score, and area under the curve (AUC)-receiver operating characteristics (ROC) metrics. Engineers and professionals in the wind industry rely on the accuracy and interpretability of the proposed model, which is visualized in a table-like format. The CM provides count values for accurate and inaccurate predictions, allowing for analysis and decision-making. Overall, the CM is a valuable asset for evaluating classification models in various fields, including the wind industry.

• Precision is the number of classified correct outputs or the exactness of the model. It is calculated using the following equation:

  (20) Precision = T p T p + F p .

• Recall: Recall is our model’s measurement to identify the real positive. The calculation is done using the following equation:

  (21) Recall = T p T p + F n .

• Accuracy: Accuracy is the percentage of production that is correctly predicted. Measure how many positive and negative observations were correctly classified. Calculations are made using the following equation:

  (22) Accuracy = T p + T n T p + F p + F n + T n .

• F 1 score: The F1 score is an average of accuracy and recall. They combine accuracy and memory into a single metric by calculating their harmony average. The formula is calculated using the following equation:

  (23) F 1 = 2 T p 2 T p + F p + F n .

• Specificity, S , also called the true negative rate, measures the proportion of negatives that are correctly identified, given by the following equation:

  (24) S = T n F p + F n ,

  where T _p represents actual positive values, T _n represents valid negative values, F _p represents false positive values, and F _n represents false negative values.

• ROC

The ROC based on the CM is used to evaluate the classification. The ROC curve extracts many indices to assess a classifier’s effectiveness. The region between the ROC curve and the negative diagonal is the AUC, with a value between 0.5 and 1 [59,60]. Because there is an area where the value of R and P is 1 for each cutoff point, AUC = 1 implies a perfect rating, while AUC = 0.5 shows that the classifier is faulty. The Wilcoxon rank test and AUC’s statistical characteristics are equal [61]. The Gini coefficient [62] is twice the area between the diagonal and the ROC curve, and the AUC is also closely connected. ‎

7.1 Time domain vibration signals analysis

In the present study, the state of the wind turbine blades included standard blades and blades with various faults. The wind turbine blades rotated at various wind speeds during health state detection. The blades rotated at a wind speed of (1.3–5.3) m/s, which is compatible with the Iraqi climb [2]. Recent research [16–25] has used vibration signals extensively due to their efficiency in forecasting difficulties. Using NI LabView signal processing software, accelerometer voltage values were collected and converted to time-domain acceleration signals. Figure 5 shows the reference vibration signals of the healthy blade and another blade with fault condition signals‎ taken when different cracks in the wind turbine blade at different wind speeds. They show the vibration signal plot (time vs amplitude) for a healthy condition blade, Fa (blade tip crack fault), Fb (blade mid-span crack fault), and Fc (blade root crack fault), respectively. This gives a basic idea about how the magnitude of the acquired vibration signal varies over time concerning simulated faults. It was observed that the vibration signals changed depending on wind speed, from Figure 5, V _wind = 1.3 m/s observed the vibration signals in the healthy case. Normally, a slight vibration would occur during the operation; therefore, the vibration measure. Therefore, the observed signals suggested that the wind turbine was operating normally. While observed, the signals changed and were affected by the type of ‎faults. It can be seen that the signs change ‎significantly compared to the rest of the cases‏. The increase in wind speed‏ ‏to reach 3.3 m/s observed the vibration signals for the blade with a crack near the tip more affected by another. It was repeated when the wind speed increased to reach 5.3 m/s. Time-dominant vibration signals provide the key indication of wind turbine faults.

Figure 5

Blade vibration signatures.

7.2 Result of the feature ranking

Eight-time domains are measured from the vibration signals of the wind turbine. They are input to machine learning algorithms for fault classification [34]. Three feature ranking algorithms were used to optimize the feature set, viz. the IG Chi-square (χ²) and ReliefF reduce the order of the set of features. Table 2 compares the results of the feature ranking algorithms and shows how the calculated features are ranked. Variance, standard deviation, and RMS are the best-ranked features based on IG, and variation is the most helpful feature. Although using Chi-square (χ²) to rank the characteristics, kurtosis, skewness, and RM were the most important. The kurtosis value in the time domain was the most important feature since it shows the average power of the measured signal and is given more weight than the other features.

Furthermore, Table 2 displays the ranked feature list generated by the ReliefF algorithm, which is utilized to demonstrate further the utility of the feature ranking method for fault classification and observed that the standard deviation and the variance and value become the most significant of the ten measured features. It is observed from Table 2 that the standard deviation features are more significant compared to other time domain features. Where it appears in each algorithm, the measure of the vibration signal’s actual energy or power content is weighted high compared to the other feature. The same observation is also found in Table 2. Using the RF-ranked feature set, two classifiers compare the classification accuracy with selected features. In the current study, the classification of features depends on the time domain from which the characteristics are calculated and the weight assigned by the ranking method.

7.3 Result of machine learning

This study focuses on the application of three widely used machine learning algorithms, namely SVM, KNN, and RF, to detect faults in wind turbine blades. Additionally, the integration of feature ranking techniques, specifically IG (IG), Chi-square (χ²), and ReliefF, is employed to identify the most effective methodology for integrating machine learning models. The objective is to identify the optimal combination of algorithms and feature ranking methods to enhance the accuracy and efficiency of fault detection in wind turbine blades.

The confusion matrices presented in this study illustrate the performance of three different types of machine learning algorithms, namely KNN, SVM, and RF, in classifying instances into four categories: Fa (crack at tip blade), Fb (crack at mid-span blade), Fc (crack near the blade root), and H (healthy state). The confusion matrices display the number of instances predicted for each category and compare them to the actual distribution of instances.

Table 3 presents the confusion matrices for the three machine learning algorithms (kNN, RF, SVM) when applying the IG feature ranking method. Each matrix shows the number of instances classified into different categories: Fa (crack at tip blade), Fb (crack at mid-span blade), Fc (crack near the blade root), and H (healthy state). The numbers in the matrix represent the count of instances correctly classified and misclassified for each category. For example, in the kNN CM, there are 963 instances of Fa correctly classified, 482 instances misclassified as Fb, 202 instances misclassified as Fc, and 453 instances misclassified as Healthy.

Table 3

Comparison of the evaluation matrices with other models for Info. gain

The Chi-square feature ranking method was employed as the second approach for fault selection in wind turbine blades. Table 4 presents the confusion matrices for each model utilizing the chi-square feature ranking. Within the confusion matrices, the diagonal elements represent correctly identified instances, while the off-diagonal elements indicate misclassifications. In this context, Fa corresponds to the root crack, Fb denotes the mid-span crack, and Fc represents the crack at the blade’s tip. The classification accuracy for the KNN model is 58.4097%, with 583.785 instances misclassified as Fb and 584.505 instances misclassified as Fc. Outperforming the kNN model, the SVM model achieves an accuracy of 79.8953% and correctly classifies 789.53 instances as Fb, yielding an F1 score of 0.798075. Similarly, the RF model demonstrates strong performance with an accuracy of 79.8456% and an F1 score of 0.798456.

Table 4

Comparison of the evaluation matrices with other models for the Chi-square

These findings emphasize the efficacy of the chi-square feature ranking method in accurately identifying faults in wind turbine blades. The kNN model exhibits moderate accuracy but encounters challenges in distinguishing between Fb and Fc instances, resulting in a significant number of misclassifications. On the other hand, the SVM and RF models showcase superior performance in correctly classifying Fb instances. Further investigations can delve into optimizing these models, refining feature selection techniques, and exploring ensemble methods to enhance the overall accuracy and robustness of the fault classification systems.

Table 5 displays the results of applying the ReliefF algorithm to three machine learning models: SVM, KNN, and RF. Table 5 presents confusion matrices, indicating the number of instances classified into categories representing blade conditions (H: healthy, Fa: crack at tip blade, Fb: crack at mid-span blade, and Fc: crack near blade root). SVM achieved accurate classification for Fa in 1,344 instances but misclassified 127 Fa instances as Fb, 108 as Fc, and 521 as H. kNN achieved perfect classification for the Healthy state (H) and accurate Fa classification for 2,083 instances. RF accurately identified Fa (1,997 instances) but misclassified some instances across other states. These matrices provide valuable insights into the performance of each algorithm, facilitating analysis of accuracy and misclassification patterns.

Table 5

Fault CM for ReliefF

The results obtained from the machine learning models for wind turbine blade fault detection demonstrate remarkable advancements in accuracy and precision. The application of the ReliefF feature ranking method yielded an impressive classification accuracy of 97%, surpassing the accuracy reported in the reference [61], which achieved a maximum of 94.94% accuracy [22]. Furthermore, the proposed hierarchical feature selection approach based on relative dependency exhibited a remarkable classification accuracy of 97.08% for gear fault diagnosis, as documented in the study of Manju et al. [63]. The utilization of multimodal deep support vectors with homologous features for gearbox malfunction diagnosis also showcased exceptional performance, with accuracy exceeding 97% using any of the four ranking features for the RF and kNN classifiers [63,64]. Comparatively, the precision achieved by these models ranged from 96.7 to 97%, outperforming the precision values reported in previous studies [63,64].

Table 6 provides a comprehensive overview of the precision values obtained by the kNN classifiers, revealing that the IG feature selection method resulted in the lowest precision, plunging as low as 54.0649%. Conversely, the ReliefF feature ranking method consistently yielded the highest precision and delivered superior results across all fault classifications. Table 6 further supports these findings, demonstrating that the IG algorithm yielded the lowest precision value, while the employment of ReliefF resulted in the highest precision and recall values. These results are congruent with the F score and AUC metrics, affirming the efficacy of the ReliefF algorithm for feature selection in identifying the most critical features and achieving exceptional precision in fault classification, as corroborated by prior research [65].

The exceptional precision and accuracy achieved by the machine learning models, particularly when employing the ReliefF feature ranking method, signify substantial progress in wind turbine blade fault detection. These findings significantly contribute to the advancement of fault diagnosis methodologies and bear vital implications for enhancing wind turbine maintenance strategies and minimizing operational downtime.