A Fault Diagnosis Method for Wind Turbine Pitch System for Data Balancing

Abstract

A fault diagnosis method for wind turbine pitch systems using data balancing is disclosed. Specifically, it involves: acquiring wind turbine operating data and fault data through a monitoring and data acquisition system installed in the wind farm; filtering sensitive features related to pitch system faults based on the acquired data to reduce the dimensionality of data samples and minimize redundant information input; assigning dynamic balancing factors to each category of samples according to quantitative relationships and introducing them into a Dualformer-based multi-level diagnostic model to construct a DBFDformer data balancing fault diagnosis model; training the DBFDformer model using the filtered feature data, and using a snow ablation optimization algorithm to dynamically update the balancing factors and optimize the model hyperparameters. This invention effectively improves the diagnostic accuracy and performance of wind turbine pitch systems under unbalanced data, exhibiting excellent generalization ability and robustness. It provides a new approach for fault diagnosis of wind turbine pitch systems under unbalanced data samples, contributing to improved wind energy utilization.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault diagnosis technology, and relates to a fault diagnosis method for wind turbine pitch system for data balancing. Background Technology

[0002] Wind energy, as a pollution-free, abundant, and widely distributed renewable clean energy source, is increasingly accounting for a larger share of electricity supply in countries around the world. However, wind turbines, as the carriers of power generation, operate in harsh environments for extended periods. Coupled with the randomness of wind speed and direction and the uncertainty of load, wind turbines experience a relatively high frequency of failures.

[0003] The pitch system, as a crucial actuator for the control and protection of wind turbine units, regulates the rotor speed and power output by adjusting the pitch angle of the turbine blades, playing a vital role in the safe, stable, and efficient operation of wind turbine units. Due to the strong coupling between the sub-components of the pitch system, the harsh operating environment, load disturbances caused by aerodynamic effects, frequent operation of the pitch switch, and the load inertia of the large-mass rotor, it is one of the components with a high failure rate in wind turbine units. According to a recent survey of wind power projects, the pitch system failure rate is as high as 21.29%, resulting in 23.32% of wind turbine downtime. Failure to promptly detect and effectively address pitch system failures can trigger a chain reaction, leading to system shutdown, and in severe cases, catastrophic accidents such as wind turbine runaway and blade sweeping. Furthermore, once a wind turbine fails, its location is generally in a remote area, and the pitch system is installed at a high altitude. The lack of convenient transportation and auxiliary maintenance equipment, among other unfavorable factors, prevents maintenance personnel from repairing wind turbines in a timely manner, resulting in significant power generation losses for wind farms. Therefore, research on fault diagnosis of wind turbine pitch systems has important academic value and practical engineering significance.

[0004] In recent years, with the continuous improvement of computer hardware computing power and the rapid development of artificial intelligence technology, deep learning and data-driven methods have gradually become the mainstream approach in wind turbine fault diagnosis research. Early on, Leahy et al. applied classification techniques to identify faulty and fault-free operating states of wind turbines in southeast Ireland based on Supervisory Control and Data Acquisition (SCADA) data. Later, Xiang et al. combined convolutional neural networks (CNNs) with long short-term memory (LSTM) models to construct a cascaded deep learning fault diagnosis model, enabling anomaly detection of key components in wind turbines and achieving the capture of early fault characteristics and identification of different fault types. Acho et al. proposed an iterative learning algorithm for sensor signal fusion and pitch system actuator fault detection. Building on this, Qin et al. fused multiple time-series signals and improved the LSTM network using a multi-channel attention mechanism, proposing an MCA-LSTM wind turbine pitch system fault diagnosis model based on a multi-channel attention mechanism, which can effectively detect faults in pitch bearings and hubs. For floating wind turbines, Cho et al. proposed a fault diagnosis method for the variable pitch system of floating wind turbines. This method uses a Kalman filter to estimate the blade pitch angle and valve core position of the variable pitch system, and an artificial neural network model is used to diagnose the fault type. Furthermore, Kong et al. fused SCADA spatiotemporal data, used a CNN to extract spatial features from the data, and then fused the temporal information from the spatial features using a gated recurrent unit (GRU). Finally, they used an exponentially weighted moving average (EWMA) control chart to identify the health status of the wind turbine. The effectiveness of the model was verified through actual wind farm testing.

[0005] Because SCADA data exhibits high-dimensional variable correlation, dynamic variability, and temporal correlation, and contains numerous feature parameters unrelated to faults, it is difficult to quickly and accurately locate and diagnose faults in wind turbine pitch systems during emergency operation. Therefore, selecting sensitive fault features and reducing redundancy among input data are crucial prerequisites for improving the accuracy of pitch system fault diagnosis. Furthermore, this can effectively improve the diagnostic efficiency of the model, adapting to the daily operation and maintenance of wind farms. Liu et al. extracted operating features of wind turbine pitch systems based on SCADA data and conducted statistical analysis on relevant feature data, revealing the operating control logic of the pitch system and analyzing its fault generation mechanism, providing a theoretical basis for the construction of fault diagnosis methods. Li et al. analyzed the main causes of false negatives and false negatives in pitch system faults, using the Relief method to mine features and improve the accuracy of fault diagnosis. Building on this, Wen et al. proposed a hybrid fault diagnosis method based on ReliefF-PCA and deep neural networks (DNN), using the ReliefF method to filter fault features and reduce data dimensionality. Secondly, the PCA algorithm is used to further reduce the dimensionality of the data, and finally, a deep neural network is used to diagnose the fault. This method solves the redundancy between data and improves the accuracy of fault diagnosis.

[0006] The accuracy and performance of the aforementioned fault diagnosis methods rely on a key assumption: that the number of training samples of different classes is balanced. However, in actual wind turbine operation, the probability of different types of faults varies significantly, leading to an imbalance in fault training data under different conditions. Furthermore, for safety reasons, wind turbines that have experienced faults are not allowed to operate continuously for extended periods, resulting in a small amount of fault data in the pitch system and an extremely unbalanced ratio between the fault data and healthy data. In deep learning-based model training, the gradient update of the loss function often favors the majority class; imbalanced data samples reduce the learning attention to the minority class. Therefore, the trained model lacks attention and sensitivity to the minority fault classes, making it difficult to accurately diagnose faults and meet the daily operation and maintenance needs of wind farms. To address the problem of imbalanced training data, resampling and reweighting methods are commonly used. Resampling methods balance the training samples of different classes by enhancing minority class data or reducing majority class data. Jiang et al. used an improved synthetic resampling technique to detect and locate major faults in wind turbines under imbalanced data, effectively alleviating the problem of high false diagnosis rates for the minority class caused by imbalanced datasets. The reweighting method assigns weights based on the difference in the number of training samples of different classes, increasing attention to the minority class to achieve balance among samples during model training. He et al. proposed a spatiotemporal multi-scale neural network (STMNN) method for wind turbine fault diagnosis based on imbalanced SCADA data. The feature module is used to extract multi-scale spatiotemporal features of the data, and the focus loss replaces the cross-entropy loss as the loss function, providing an end-to-end fault diagnosis solution in the case of imbalanced SCADA data.

[0007] However, inappropriate data resampling will generate low-quality or invalid sample data, easily leading to overfitting of the trained model and making it difficult to achieve good results in testing and practical applications. Furthermore, most current reweighting strategies weaken the weights of the majority class, reducing the model's attention to it to ensure balanced model training. In this case, the effectiveness of sample identification for the majority class is limited by overfitting and insufficient attention, and inappropriate weighting mechanisms will affect the model's generalization ability, reducing the accuracy of fault diagnosis results. Therefore, existing methods are insufficient for fault diagnosis of wind turbines using unbalanced data, especially for pitch system faults with complex internal structures. There is an urgent need to explore more advanced fault diagnosis models to improve the diagnostic performance of pitch system faults under unbalanced data. Summary of the Invention

[0008] This invention provides a fault diagnosis method for wind turbine pitch system for data balancing, which effectively improves the diagnostic accuracy and performance of wind turbine pitch system faults under unbalanced data.

[0009] The technical solution adopted in this invention is as follows:

[0010] A fault diagnosis method for wind turbine pitch system for data balancing specifically includes the following steps:

[0011] Step 1: Data acquisition. Obtain the operating data and fault data of the wind turbines through the SCADA system equipped in the wind farm.

[0012] Step 2: Data preprocessing, removing features irrelevant to the pitch system; removing invalid features with constant values, calculating the standard deviation of the remaining features, and removing invalid features with a standard deviation of 0; removing outlier data where wind speed is less than the cut-in wind speed or greater than the cut-out wind speed, and filling missing values ​​using the persistence method; replacing data where active power exceeds rated power with rated power values, and replacing data with negative active power with zero values; standardizing the remaining feature data to eliminate the influence of different dimensions between features.

[0013] Step 3: Feature screening. Sensitive features related to pitch system faults are screened using the ReliefF algorithm to reduce the dimensionality of data samples and reduce the input of redundant information. The number of sample samplings is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold for feature weights.

[0014] Step 4: Construction and training of the DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each class of samples according to the quantitative relationship. This is used to balance the attention given to each class of samples during model training, so as to reduce the influence of the majority class samples. Second, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformers to construct the DBFDformer data balancing fault diagnosis model. Then, the dynamic update of the balancing factor and the optimization of the model hyperparameters are completed through the Snow Ablation Optimization Algorithm (SAO). Finally, the DBFDformer model is trained and the pitch system fault diagnosis is performed.

[0015] Further:

[0016] The process of setting the dynamic equilibrium factor DBF in step 4 is as follows:

[0017] 1) Based on the quantitative relationship between healthy samples and faulty samples, set the initial balance factor α for the samples, and based on the quantitative relationship between samples of different fault categories, set the initial balance factor β for the faults. The calculation of each initial balance factor is shown in formulas (18)-(19):

[0018]

[0019] Where H is the total number of sample categories, and N is... max N represents the number of samples in the most frequent category.h This represents the number of samples in the h-th category;

[0020] 2) Based on the Binary Cross-Entropy Loss (BCELoss) function, as shown in Equation (20), the loss function L′ for the first stage is constructed using the initial balance factor α of the samples. fl As shown in formula (21), it is used for training the health and fault diagnosis model;

[0021]

[0022] Where N represents the total number of samples, y i Let y represent the true class of the i-th sample. i ∈{0,1},y i ' is the model prediction probability of the i-th sample, y i ′∈[0,1];

[0023] 3) Construct the loss function L′ for the second stage using the initial fault balancing factor β. f ′ l As shown in formula (22), it is used for training the fault type diagnosis model;

[0024]

[0025] Where, β h This represents the balance factor for the h-th class sample.

[0026] 4) Based on the loss value of each iteration of the first and second stage diagnostic model training, the SAO algorithm is used to dynamically update the balance factor to ensure that the samples of each category reach the best balance state during model training, thereby enhancing the generalization ability of the fault diagnosis model and improving the diagnostic accuracy.

[0027] The DBFDformer data balancing fault diagnosis model mainly consists of three modules: health and fault diagnosis, diagnostic control, and fault type diagnosis.

[0028] The health and fault diagnosis module is mainly used to identify the health and fault states of the pitch system, i.e., to diagnose whether a fault exists. First, an adaptive max pooling layer is used to reduce the dimensionality of the input feature data and merge key feature information in adjacent regions to reduce model complexity and improve computational efficiency. Second, the Transformer encoder performs global learning on the input features and uses the internal multi-head self-attention mechanism to capture the dependency between each feature and the pitch system state. Finally, the health and fault diagnosis of the pitch system is completed through a fully connected layer and an Argmax layer.

[0029] The diagnostic control module is mainly used to control the execution of different fault diagnosis modules, i.e. whether to perform specific fault type diagnosis. The diagnostic control module is divided into two stages: the first stage, namely health and fault diagnosis, and the second stage, namely fault type diagnosis. The diagnosis result of the first stage determines whether to perform the second stage diagnosis. If the diagnosis result of the first stage is health, it is directly used as the model result output. If it is fault, it enters the second stage, where the fault type diagnosis module identifies the specific type of fault.

[0030] The fault type diagnosis module is mainly used to quickly and accurately identify the specific fault type of the pitch system. This module first uses a Conv1D layer to acquire local feature information, capture the detailed features of the input data, and map the information to a high-dimensional space. Then, it uses BN & PReLU layers to control the propagation of gradients during model training, avoiding gradient explosion and vanishing and accelerating the convergence speed of model training. Next, a Transformer encoder performs global learning on the extracted feature information to obtain the nonlinear correlation between each feature and different types of faults. Finally, a fully connected layer and an Argmax layer output the specific fault type for diagnosis, completing the second stage of fault diagnosis.

[0031] This invention effectively improves the diagnostic accuracy and performance of wind turbine pitch system faults under unbalanced data. It exhibits excellent generalization ability and robustness, providing a new approach to fault diagnosis of wind turbine pitch systems under unbalanced data samples, and contributing to improved wind energy utilization. Specifically:

[0032] (1) The RelifF feature extraction algorithm was introduced to screen key feature data related to pitch system faults. While reducing the dimensionality of the data, redundant information between feature data was reduced, which effectively improved the training efficiency and diagnostic accuracy of the fault diagnosis model.

[0033] (2) A dynamic balance factor for samples was designed to balance the attention given to each class of samples during model training, so as to reduce the influence of the majority class of samples. The SAO optimization algorithm was introduced to dynamically update the balance factor of each class according to the training results, so as to ensure that each class of samples reaches the best balance state during model training, thereby enhancing the generalization ability of the fault diagnosis model and improving the diagnostic accuracy.

[0034] (3) A multi-level fault diagnosis model based on dual Transformers, Dualformer, was proposed. A diagnosis control module was introduced to divide the fault diagnosis into two stages: health and fault and fault type. This effectively shortened the fault diagnosis model diagnosis time and improved the model's diagnosis efficiency and reliability. Attached Figure Description

[0035] Figure 1This is a schematic diagram of the Dualformer multi-level fault diagnosis model structure of the present invention;

[0036] Figure 2 This is the three-dimensional confusion matrix of the present invention. (a) Confusion matrix of wind turbine #1; (b) Confusion matrix of wind turbine #2;

[0037] Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and experiments.

[0039] Reference Figure 3 A fault diagnosis method for wind turbine pitch system for data balancing specifically includes the following steps:

[0040] Step 1: Data acquisition. Obtain the operating data and fault data of the wind turbines through the SCADA system equipped in the wind farm.

[0041] Step 2: Data preprocessing. Features irrelevant to the pitch system are removed, including annual / monthly / day availability / power generation of the wind turbine and features of other system components. Invalid features with constant values ​​are removed; the standard deviation of the remaining features is calculated, and invalid features with a standard deviation of 0 are removed. Outlier data where wind speed is less than the cut-in wind speed or greater than the cut-out wind speed are removed, and missing values ​​are filled using the persistence method. Data with active power exceeding the rated power is replaced with the rated power value, and data with negative active power is replaced with zero. The remaining feature data is standardized to eliminate the influence of different dimensions.

[0042] Step 3: Feature screening. Sensitive features related to pitch system faults are screened using the ReliefF algorithm to reduce the dimensionality of data samples and reduce the input of redundant information. The number of sample samplings is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold for feature weights.

[0043] Step 4: Construction and training of the DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each class of samples according to the quantitative relationship. This is used to balance the attention given to each class of samples during model training, so as to reduce the influence of the majority class samples. Second, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformers to construct the DBFDformer data balancing fault diagnosis model. Then, the dynamic update of the balancing factor and the optimization of the model hyperparameters are completed by using the snow ablation optimization SAO algorithm. Finally, the DBFDformer model is trained and the pitch system fault diagnosis is performed.

[0044] The following is a detailed explanation:

[0045] (1) ReliefF feature extraction algorithm

[0046] The ReliefF algorithm is a classic multi-class filtering feature selection method, extended from the Relief algorithm. Its core idea is to calculate feature weights based on the feature's ability to distinguish between nearby samples. The specific calculation process is as follows:

[0047] First, randomly select sample R from the training sample set. i From R i Find the k nearest neighbor samples H from samples of the same type, and from R... i Find the k nearest neighbor samples M(C) among samples of different classes (C≠Class(R)). i Then calculate R respectively. i Distance D from H H R i The distance D between M(C) and M(C) M As shown in formulas (1)-(2). If D H Value less than D M If this indicates that the feature is helpful in distinguishing the nearest neighbors of the same and different classes, then its weight value should be increased; if D H Value greater than D M If the feature has a negative impact on distinguishing the nearest neighbors of the same class from the opposite class, its weight value is reduced. The larger the weight of the feature, the stronger the classification ability, and vice versa. By iterating m times to update its weight, the feature selection is completed. The weight is calculated as shown in formula (3).

[0048]

[0049] In the formula, f l Let f be the l-th feature of the sample, and diff(f,R1,R2) be the normalized distance between samples R1 and R2 on the f-th feature. j (j=1,2,...,k) is related to R i The j-th sample among the k nearest neighbors of the same type, M j (C)(j=1,2,…,k) is related to R i The j-th sample among the k nearest neighbors of different classes, W i (f l Let P(C) be the weight of the l-th feature f in the i-th samples, and let P(C) be the proportion of samples belonging to class C in the training samples; P(Class(R)) i )) for R i The ratio of samples of the same type to the total sample. The specific calculation of diff(f,R1,R2) is as shown in formulas (4)-(5).

[0050] When feature f is continuous:

[0051]

[0052] When feature f is discrete:

[0053]

[0054] (2) Snow Melting Optimization Algorithm

[0055] The Snow Ablation Optimization (SAO) algorithm is a physics-based metaheuristic algorithm that simulates the sublimation and melting process of snow in nature to achieve a trade-off between exploration and development of the solution space. It employs a dual-population mechanism for both exploration and development to avoid premature convergence. Its optimization process consists of the following initial, exploration, and development phases:

[0056] 1) Initialization phase

[0057] A random N×Dim matrix particle swarm is generated, as shown in formula (6), where N represents the size of the swarm and Dim represents the dimension of the solution space.

[0058]

[0059] 2) Exploration Phase

[0060] Brownian motion is used to simulate the irregular motion during the conversion of snow or liquid water into steam. The relevant mathematical formula is as follows:

[0061]

[0062] By utilizing dynamic and uniform step sizes, some potential areas in the search space can be explored, effectively reflecting the diffusion of steam within the search space. The location calculation formula during the exploration process is as follows:

[0063]

[0064] Elite(t)∈[G(t),Z2(t),Z3(t),Z c (t)](10)

[0065]

[0066] Among them, Z i (t) represents the i-th individual in the t-th iteration, BM i (t) represents Brownian motion, Elite(t) is a random individual selected from several elite sets in the swarm, θ is a random number in [0,1], and G(t) refers to the current optimal solution. Z represents the centroid position of the entire population, Z2(t) and Z3(t) represent the second and third best individuals in the current population, respectively. c (t) represents the centroid position of the individuals with fitness values ​​in the top 50%, and N1 represents the number of leaders, i.e., N1 is equal to half the size of the entire population.

[0067] 3) Development Phase

[0068] When snow melts into liquid water, the search agent is encouraged to develop high-quality solutions within the current optimal solution. Based on the snow melting model based on the degree-day method proposed by Zhou et al., the position update formula for the SAO algorithm is derived as follows:

[0069]

[0070] Where M is the snow melting rate, t max The termination criterion is δ, which is a random number in the range [-1, 1].

[0071] (3) Transformer

[0072] The Transformer consists of an encoder and a decoder, and mainly includes a multi-head self-attention (MHSA) module, a feedforward neural network (FNN), residual connections, and inter-layer norms. Among them, the MHSA module, as the core component of the Transformer, enhances the neural network's focus on key information and reduces the impact of redundant information on the target task by calculating and allocating attention weights. It has excellent global feature extraction capabilities and enables the interaction between the encoder and decoder.

[0073] First, the raw data is generated into a query matrix Q, a key matrix K, and a value matrix V through different learnable linear projections, as shown in formulas (14)-(15).

[0074] Q = χω Q K=χω K V=χω V (14)

[0075]

[0076] Where, χ∈R N×C Indicates the input, ω Q ,ω K ,ω V ∈R C×C These are the weights of the corresponding linear mappings. These represent the query matrix, key matrix, and value matrix in each header, respectively. Here, H represents the weights of the corresponding transformations, N represents the number of heads in the attention mechanism, C represents the time dimension, and C represents the feature dimension of χ. H It is the feature dimension of each head, C = H × C H .

[0077] Subsequently, regarding Q h and K h Perform a dot product to obtain global information, and then scale and normalize the result using a scaling factor and the SoftMax function to obtain the attention weights. Finally, multiply by V. h To obtain the weighted feature map, as shown in formula (16).

[0078]

[0079] in, It is the result of the attention calculation for the h-th head.

[0080] Finally, the attention in all heads is calculated and the final result is obtained through cascading and linear mapping, as shown in formula (17).

[0081] χ′=Concat(A1,A2,...,A H )ω χ (17)

[0082] (4) Dynamic equilibrium factor

[0083] To address the challenge of imbalanced data samples in practical fault diagnosis, a common solution is to assign a balancing factor to each category based on quantitative relationships to enhance the model's ability to identify the minority category. However, this method ignores the fact that different categories of samples have different differentiability, and a fixed balancing factor can easily lead to overfitting, reducing the model's generalization ability and thus affecting fault diagnosis accuracy. To address this problem, this invention designs a dynamic balancing factor and applies it to the cross-entropy loss function to balance the attention given to each category of samples during model training. The specific application process is as follows:

[0084] 1) Based on the quantitative relationship between healthy samples and faulty samples, set the initial balance factor α for the samples. Based on the quantitative relationship between samples of different fault categories, set the initial balance factor β for the faults. The calculation of each initial balance factor is shown in formulas (18)-(19).

[0085]

[0086] Where H is the total number of sample categories, and N is... max N represents the number of samples in the most frequent category. h This represents the number of samples in the h-th category.

[0087] 2) Based on the binary cross-entropy loss (BCELoss) function, as shown in formula (20), the loss function L′ for the first stage is constructed using the initial balance factor α of the samples. fl As shown in formula (21), it is used for training the health and fault diagnosis model. If α > 0, it can reduce the loss of easily classified samples, so that the model training can focus more on the fewer fault category samples, effectively solving the problem of difficulty in identifying minority class samples.

[0088]

[0089] Where N represents the total number of samples y i y represents the true class of the i-th sample. i ∈{0,1},y i ' is the model prediction probability of the i-th sample, y i ′∈[0,1].

[0090] 3) Construct the loss function L′ for the second stage using the initial fault balancing factor β. f ′ l As shown in formula (22), it is used for training the fault type diagnosis model. By assigning corresponding weights to each fault category through a balancing factor, the model's learning attention and recognition sensitivity to samples of a few categories are increased, effectively balancing the influence of each fault category sample on the model training, thereby improving the model's ability to identify specific fault types.

[0091]

[0092] Where, β h This represents the balance factor for the h-th class sample.

[0093] Based on the loss values ​​of each iteration of the first and second stage diagnostic model training, the SAO algorithm is used to dynamically update the balance factor to ensure that the samples of each category reach the optimal balance state during model training, thereby enhancing the generalization ability of the fault diagnosis model and improving the diagnostic accuracy.

[0094] (5) Multi-level fault diagnosis model Dualformer

[0095] To further improve the accuracy and timeliness of fault diagnosis, this invention proposes a Dualformer multi-level fault diagnosis model based on Transformer networks, the structure of which is as follows: Figure 2 As shown, the model mainly consists of three modules: health and fault diagnosis, diagnostic control, and fault type diagnosis.

[0096] The health and fault diagnosis module is primarily used to identify the health and fault states of the pitch system, i.e., to diagnose the presence of faults. First, an adaptive max-pooling layer is used to reduce the dimensionality of the input feature data and merge key feature information from adjacent regions, thereby reducing model complexity and improving computational efficiency. Second, a Transformer encoder performs global learning on the input features, utilizing an internal multi-head self-attention mechanism to capture the dependencies between each feature and the pitch system state. Finally, fully connected layers and Argmax layers complete the health and fault diagnosis of the pitch system. Since this module only diagnoses whether a fault exists in the pitch system, it falls under binary fault identification. Therefore, the network structure of this module is simple and has low complexity, enabling fast and accurate fault identification and effectively improving the model's diagnostic efficiency.

[0097] The diagnostic control module primarily controls the execution of different fault diagnosis modules, specifically whether to perform diagnoses for specific fault types. This module is divided into two stages: a first stage (health and fault diagnosis) and a second stage (fault type diagnosis). The decision to proceed to the second stage is based on the results of the first stage. If the first stage result is "healthy," it is directly output as the model result; otherwise, it proceeds to the second stage, where the fault type diagnosis module identifies the specific fault type and outputs it as the model result. Because the second-stage fault type diagnosis module involves multi-class fault identification, its model complexity is higher than that of the first-stage health and fault diagnosis module, requiring more parameters and runtime. Therefore, the diagnostic control module enables efficient connection between the first and second fault diagnosis stages, allowing the model to perform high-precision diagnoses only on data already identified as faulty, significantly improving the model's diagnostic accuracy and runtime efficiency. Furthermore, due to the limited number of fault data points and the imbalance between fault and health data, the number of high-precision fault type diagnoses is extremely limited, further improving the model's operational efficiency.

[0098] The fault type diagnosis module is primarily used to quickly and accurately identify the specific fault type of the pitch system. This module first uses a Conv1D layer to acquire local feature information, capturing the detailed features of the input data and mapping the information to a high-dimensional space. Then, it utilizes BN & PReLU layers to control the propagation of gradients during model training, avoiding gradient explosion and vanishing and accelerating model convergence. Next, a Transformer encoder performs global learning on the extracted feature information to obtain the nonlinear correlation between each feature and different fault types. Finally, a fully connected layer and an Argmax layer output the specific fault type for diagnosis, completing the second stage of fault diagnosis. Because this module only performs more accurate fault type diagnosis on fault data, it avoids the impact of imbalance between fault and healthy data on the model, effectively improving the model's diagnostic accuracy.

[0099] (6) Experimental verification

[0100] 1) Evaluation indicators

[0101] To more objectively evaluate the diagnostic performance of the method of this invention, F1-score and G-mean are used as evaluation indicators of the experimental results, and their definitions are as follows:

[0102]

[0103] Among them TP N and FP N TN represents the number of true positive and false positive samples, respectively. N and FN N These represent the number of true negatives and false negatives, respectively. F1-score and G-mean are comprehensive evaluation metrics for Precision, Recall, and Specificity, respectively. F1-score can more comprehensively reflect the accuracy of model classification, while G-mean is used to evaluate the classification performance of the model under class imbalance. For both, the higher the value, the better the model's classification effect.

[0104] 2) Data processing

[0105] This invention takes two Shanghai Electric W2000-111 wind turbine generator sets with a rated power of 2000KW as examples, numbered #1 and #2 respectively. SCADA data collection time is from 2021-01-01 00:00:00 to 2022-01-01 00:00:00, with a sampling interval of 10 minutes. Each wind turbine generator set contains a total of 82 features and 52,572 data points.

[0106] Since the pitch system fault samples of the wind farm are very limited, the unbalance rates of the health and fault samples of the two wind turbines were calculated according to formula (28) as 48.64 and 48.04, respectively, which are extremely unbalanced. There are a total of 7 wind turbine states involved, and the number of samples and descriptions of each state are shown in Table 1.

[0107]

[0108] Among them, I R N represents the imbalance rate between healthy and faulty samples. H and N F These represent the number of healthy samples and the number of faulty samples, respectively.

[0109] Table 1. Status Description and Sample Size of Wind Turbine Pitch System

[0110]

[0111] 3) Fault diagnosis results of wind turbine pitch system

[0112] Table 2 Evaluation of Fault Diagnosis Results for Wind Turbine Pitch System

[0113]

[0114] As can be seen from the F1-score and G-mean in the table, the diagnostic results of the method of this invention for the health status of the pitch system are all above 99.04%, and the diagnostic results for each type of fault are also above 91.94%, with the diagnostic result for fault 2 even reaching 100%. Furthermore, the three-dimensional confusion matrix diagram of the diagnostic results (…) Figure 2 As can be observed more intuitively, the diagnostic accuracy for each state category reaches over 89.47%, with most exceeding 97.00%. Analysis shows that this invention effectively avoids the impact of the imbalance between fault and health data on the model, demonstrating good diagnostic capability and high diagnostic accuracy for pitch system faults.

[0115] Since this invention divides the fault diagnosis of wind turbine pitch systems into two stages—health and fault diagnosis and fault type diagnosis—the diagnostic results of each stage are analyzed to describe the diagnostic details of the method in more detail. The diagnostic results of the health and fault diagnosis stage are shown in Table 4. As can be observed from Table 4, the proposed method achieves a health status identification rate of over 98.80% and a fault status diagnosis rate of over 97.73%. For a large number of health data samples, this method prioritizes the rapid and accurate identification of the health and fault status of the pitch system, greatly shortening the overall diagnostic time of the model and improving its diagnostic efficiency. Simultaneously, it lays the foundation for accurate diagnosis in the second stage of fault type diagnosis, ensuring the overall diagnostic accuracy of the wind turbine pitch system faults using the method of this invention.

[0116] Table 3 Evaluation of Diagnostic Results in the Health and Fault Diagnosis Phase of Wind Turbine Pitch System

[0117]

[0118] The second-stage fault type diagnosis is based on the results of the first stage. Its main function is to perform a secondary diagnosis on the data identified as faults in the first stage, avoiding the influence of majority sample data and accurately diagnosing the specific type of fault. Table 4 shows the diagnostic results of the second-stage wind turbine pitch system fault types. According to the table, the F1-score and G-mean of faults 1, 2, and 6 all reached 100%. However, in wind turbine #1, faults 4 and 5 did not reach 100% because of two fault diagnosis errors: fault 4 was misdiagnosed as fault 5, and fault 5 was misdiagnosed as fault 4. Similarly, in wind turbine #2, faults 3, 4, and 5 did not reach 100% because of two fault diagnosis errors: fault 4 was misdiagnosed as fault 5, and fault 3 was misdiagnosed as fault 4. Analysis of the experimental results shows that the method of this invention exhibits strong diagnostic performance and generalization ability in diagnosing various types of faults.

[0119] Table 4 Evaluation of Overall Stage Diagnostic Results for Fault Types in Wind Turbine Pitch System

[0120]

Claims

1. A fault diagnosis method for a wind turbine pitch system used for data balancing, characterized in that, Specifically, the following steps are included: Step 1: Data acquisition. Obtain the operating data and fault data of the wind turbines through the SCADA system equipped in the wind farm. Step 2: Data preprocessing, removing features irrelevant to the pitch system; removing invalid features with constant values, calculating the standard deviation of the remaining features, and removing invalid features with a standard deviation of 0; removing outlier data where wind speed is less than the cut-in wind speed or greater than the cut-out wind speed, and filling missing values ​​using the persistence method; replacing data where active power exceeds rated power with rated power values, and replacing data with negative active power with zero values; standardizing the remaining feature data to eliminate the influence of different dimensions between features. Step 3: Feature screening. Sensitive features related to pitch system faults are screened using the ReliefF algorithm to reduce the dimensionality of data samples and reduce the input of redundant information. The number of sample samplings is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold for feature weights. Step 4: Construction and training of the DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each class of samples according to the quantitative relationship. This is used to balance the attention given to each class of samples during model training, so as to reduce the influence of the majority class samples. Second, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformers to construct the DBFDformer data balancing fault diagnosis model. Then, the dynamic update of the balancing factor and the optimization of the model hyperparameters are completed through the Snow Ablation Optimization Algorithm (SAO). Finally, the DBFDformer model is trained and the pitch system fault diagnosis is performed.

2. The method for fault diagnosis of wind turbine pitch system for data balancing according to claim 1, characterized in that, The process of setting the dynamic balance factor DBF in step 4 is as follows: 1) Set the initial sample balance factor based on the quantitative relationship between healthy and faulty samples. Based on the quantitative relationship between samples of different fault categories, an initial fault balance factor is set. The initial equilibrium factors are calculated as shown in formulas (18)-(19): (18) (19) in, The total number of sample categories. This represents the number of samples in the most frequent category. Indicates the first Number of samples in each category; 2) Based on the Binary Cross-Entropy Loss (BCELoss) function, formula (20) is used, along with the initial balance factor of the samples. Constructing the loss function for the first stage Formula (21) is used for training the health and fault diagnosis model; (20) (21) in, Represents the total number of samples. Indicates the first The true category of each sample. , It is the first The model predicts the probability for each sample. ; 3) Use the initial fault balance factor Constructing the loss function for the second stage Formula (22) is used for training the fault type diagnosis model; (22) in, Indicates the first Balance factor for each class of samples; 4) Based on the loss value of each iteration of the first and second stage diagnostic model training, the balance factor is dynamically updated using the snow ablation optimization algorithm to ensure that the samples of each category reach the best balance state during model training.

3. The method for fault diagnosis of wind turbine pitch system for data balancing according to claim 1, characterized in that, The DBFDformer data balancing fault diagnosis model consists of three modules: health and fault diagnosis, diagnostic control, and fault type diagnosis. The health and fault diagnosis module is used to identify the health and fault states of the pitch system, i.e., to diagnose whether a fault exists. First, an adaptive max pooling layer is used to reduce the dimensionality of the input feature data and merge key feature information from adjacent regions to reduce model complexity and improve computational efficiency. Second, the Transformer encoder performs global learning on the input features and uses the internal multi-head self-attention mechanism to capture the dependency between each feature and the pitch system state. Finally, the health and fault diagnosis of the pitch system is completed through a fully connected layer and an Argmax layer. The diagnostic control module is used to control the execution of different fault diagnosis modules, i.e. whether to perform specific fault type diagnosis. The diagnostic control module is divided into two stages: the first stage, namely health and fault diagnosis, and the second stage, namely fault type diagnosis. The diagnosis result of the first stage determines whether to perform the second stage diagnosis. If the diagnosis result of the first stage is health, it is directly used as the model result output. If it is fault, it enters the second stage, where the fault type diagnosis module identifies the specific type of fault. The fault type diagnosis module is used to quickly and accurately identify the specific fault type of the pitch system. This module first uses a Conv1D layer to acquire local feature information, capture the detailed features of the input data, and map the information to a high-dimensional space. Then, it uses BN & PReLU layers to control the propagation of gradients during model training, avoiding gradient explosion and vanishing and accelerating the convergence speed of model training. Next, a Transformer encoder performs global learning on the extracted feature information to obtain the nonlinear correlation between each feature and different types of faults. Finally, the fully connected layer and Argmax layer output the specific fault type for diagnosis, completing the second stage of fault diagnosis.

Images