Wind turbine generator variable pitch system fault diagnosis method for data balance

Abstract

The invention discloses a wind turbine generator variable pitch system fault diagnosis method for data balance. The method specifically comprises the following steps: acquiring operation data and fault data of a wind turbine generator through a monitoring and data acquisition system equipped on a wind power plant; based on the obtained data, screening sensitive features related to the fault of the variable pitch system, reducing the dimension of a data sample and reducing the input of redundant information; according to the quantitative relation, dynamic balance factors are distributed for various types of samples, the dynamic balance factors are introduced into a multi-stage diagnosis model Dualformer based on double Transformers, and a DBFDformer data balance fault diagnosis model is constructed; and training a DBFDform model by using the screened feature data, and completing dynamic updating of balance factors and optimization of model hyper-parameters through a snow ablation optimization algorithm. The wind turbine generator variable pitch system fault diagnosis method can effectively improve the diagnosis precision and performance of the wind turbine generator variable pitch system fault under the unbalanced data, has excellent generalization ability and robustness, provides a new thought for the fault diagnosis of the wind turbine generator variable pitch system under the unbalanced data sample, and is helpful for improving the utilization rate of wind energy.

Description

Technical Field

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

[0002] As a non-polluting, large-scale and widely distributed renewable clean energy, wind energy is increasingly accounting for the power supply of countries around the world. However, wind turbines, as carriers of power generation, have been operating in harsh environments for a long time. Coupled with the randomness of wind speed and direction and the uncertainty of load, wind turbines have a high frequency of failure.

[0003] As an important actuator for wind turbine control and protection, the pitch system controls the rotor speed and power output by adjusting the pitch angle of the wind turbine blades, which plays a very important role in the safe, stable and efficient operation of the wind turbine. Due to the strong coupling between the sub-components of the pitch system and the influence of the harsh operating environment, coupled with the load disturbance caused by the aerodynamic effect, the frequent operation of the pitch switch and the load inertia of the large-mass impeller, it has become one of the components with a higher failure rate in the wind turbine. According to a recent survey of wind power projects, the failure rate of the pitch system is as high as 21.29%, causing the downtime of wind turbines to account for as high as 23.32%. If the failure of the pitch system is not discovered in time and effective measures are taken to deal with it, a chain reaction will be triggered, causing the system to shut down. In severe cases, catastrophic accidents such as wind turbine loss of control and blade sweeping of the tower may occur. In addition, once a wind turbine fails, it is generally located in a place where few people go, and the installation position of the unit pitch system is high from the ground. There are many unfavorable factors such as inconvenient transportation and lack of auxiliary maintenance equipment, which make it impossible for maintenance personnel to repair wind turbines in time, resulting in great power generation losses in wind farms. Therefore, the fault diagnosis research on wind turbine pitch control system has important academic value and engineering practical significance.

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

[0005] Since SCADA data has the characteristics of high-dimensional variable correlation, dynamic variability and time correlation, and there are many characteristic parameters that are not related to the fault. As a result, when an emergency fault occurs in the wind turbine pitch system during operation, it is difficult to quickly and accurately locate the fault and diagnose the fault type. Therefore, screening out sensitive fault features and reducing the redundancy between input data are important prerequisites for improving the accuracy of pitch system fault diagnosis. In addition, it can effectively improve the diagnostic efficiency of the model to adapt to the daily operation and maintenance of wind farms. Liu et al. extracted the operating characteristics of the wind turbine pitch system based on SCADA data, and conducted statistical analysis on the relevant characteristic data, revealing the operating control logic of the pitch system, analyzing its fault generation mechanism, and providing a theoretical basis for the construction of fault diagnosis methods. Li et al. analyzed the main reasons for missed and false alarms of pitch system faults, and used the Relief method to mine features to improve the accuracy of fault diagnosis. On this basis, Wen et al. proposed a hybrid fault diagnosis method based on ReliefF-PCA and deep neural network (DNN), using the ReliefF method to screen fault features and reduce data dimensions. Secondly, the PCA algorithm is used to further reduce the data dimension, and finally the fault is diagnosed through a deep neural network. This method solves the redundancy between data and improves the accuracy of fault diagnosis.

[0006] The accuracy and performance of the above fault diagnosis methods rely on a key assumption that the number of training samples of different categories is balanced. However, in the actual operation of wind turbines, the probability of occurrence of different types of faults varies greatly, resulting in an imbalance in fault training data under different conditions. In addition, for safety reasons, wind turbines with faults are not allowed to operate continuously for a long time, resulting in a small amount of fault data in the variable pitch system and an extremely unbalanced ratio with the number of healthy data. In the training of models based on deep learning, the gradient update of the loss function tends to favor the majority class, and unbalanced data samples will reduce the learning attention to the minority class data. Therefore, the trained model lacks attention and sensitivity to the minority fault class, making it difficult to accurately diagnose faults and meet the needs of daily operation and maintenance of wind farms. In order to solve the problem of unbalanced training data, resampling and reweighting methods are usually used. The resampling method achieves the purpose of balancing training samples of different categories by enhancing minority class data or reducing majority class data. Jiang et al. used an improved synthetic resampling technique to detect and locate the main faults of wind turbines under unbalanced data, effectively alleviating the problem of high misdiagnosis rate of minority classes caused by unbalanced data sets. The reweighting method assigns corresponding weights according to the difference in the number of training samples of different categories, increases the attention to the minority class to achieve a balance between samples during model training. He et al. proposed a spatiotemporal multiscale neural network (STMNN) method for wind turbine fault diagnosis based on unbalanced SCADA data, in which the feature module is used to extract the multi-scale spatiotemporal features of the data, and the cross entropy loss is replaced by the focal loss as the loss function, providing an end-to-end fault diagnosis solution in the case of unbalanced SCADA data.

[0007] However, inappropriate data resampling will generate low-quality or invalid sample data, which can easily lead to overfitting of the training model and make it difficult to achieve good results in testing and practical applications. In addition, most of the currently adopted reweighting strategies weaken the weight of the majority class and the model's attention to it to ensure the balance of model training. In this case, the effectiveness of sample recognition in most categories is limited by overfitting and insufficient attention, and the inappropriate weight assignment mechanism will affect the generalization ability of the model and reduce the accuracy of the fault diagnosis results. Therefore, the existing methods are insufficient for wind turbine fault diagnosis with unbalanced data, and the research on variable pitch system faults with complex internal structures is even more seriously insufficient. It is urgent to explore more advanced fault diagnosis models to improve the diagnostic performance of variable pitch system faults under unbalanced data. Summary of the invention

[0008] The present invention provides a wind turbine pitch system fault diagnosis method for data balance, which effectively improves the wind turbine pitch system fault diagnosis accuracy and performance under unbalanced data.

[0009] The technical solution adopted by the present invention is:

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

[0011] Step 1: Data acquisition: obtain the operation data and fault data of wind turbines through the SCADA system equipped in the wind farm;

[0012] Step 2: Data preprocessing, remove features that are not related to the pitch system; remove invalid features with constant feature values, calculate the standard deviation of the remaining features, and remove invalid features with a standard deviation of 0; remove abnormal data with wind speeds less than the cut-in wind speed or greater than the cut-out wind speed, and use the persistence method to fill in missing values; replace data with active power exceeding the rated power with the rated power value, and replace data with negative active power with zero values; standardize the remaining feature data to eliminate the impact of dimensions between different features;

[0013] Step 3: Feature screening: The ReliefF algorithm is used to screen out sensitive features related to the pitch system fault, reduce the data sample dimension and reduce the input of redundant information. The number of sample sampling is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold of the feature weight.

[0014] Step 4: Construction and training of DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each category of samples according to the quantitative relationship, which is used to balance the attention of each category of samples during model training to reduce the influence of majority class samples. Secondly, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformer 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 melting optimization algorithm (SAO). Finally, the DBFDformer model is trained and the variable pitch system fault is diagnosed.

[0015] Further:

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

[0017] 1) According to the quantitative relationship between healthy samples and faulty samples, the sample initial balance factor α is set; according to the quantitative relationship between samples of different fault categories, the fault initial balance factor β is set. The calculation of each initial balance factor is shown in formulas (18)-(19):

[0018]

[0019] Among them, H is the total number of sample categories, N max Indicates the maximum number of category samples, Nh represents the number of samples in the hth category;

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

[0021]

[0022] Among them, N represents the total number of samples, y i Indicates the true category of the i-th sample, y i ∈{0,1},y i ′ is the model predicted probability of the i-th sample, y i ′∈[0,1];

[0023] 3) Use the initial fault balance factor β to construct the second-stage loss function L′ f ' l , as shown in formula (22), is used for training the fault type diagnosis model;

[0024]

[0025] Among them, β h Represents the balance factor of the h-th category samples.

[0026] 4) According to the loss value of each iterative training of the first and second stage diagnosis models, the SAO algorithm is used to dynamically update the balance factor to ensure that 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 diagnosis accuracy.

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

[0028] The health and fault diagnosis module is mainly used to identify the health and fault states of the variable pitch system, that is, to diagnose whether there is a fault: first, an adaptive maximum pooling layer is used to reduce the dimension of the input feature data and merge the key feature information of adjacent regions to reduce the complexity of the model and improve the computational efficiency; secondly, 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 state of the variable pitch system; finally, the health and fault diagnosis of the variable pitch system is completed through the fully connected layer and the Argmax layer;

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

[0030] The fault type diagnosis module is mainly used to quickly and accurately identify the specific fault type of the variable pitch system; the module first uses the Conv1D layer to obtain local feature information, capture the detailed features of the input data, and map the information to a high-dimensional space; then the BN&PReLU layer is used to control the transmission of the model training gradient to avoid the explosion and disappearance of the gradient and accelerate the convergence speed of the model training; secondly, the 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 specific fault type diagnosed is output through the fully connected layer and the Argmax layer to complete the second stage of fault diagnosis.

[0031] The present invention effectively improves the diagnostic accuracy and performance of wind turbine pitch system faults under unbalanced data. It has excellent generalization ability and robustness, provides a new idea for fault diagnosis of wind turbine pitch system under unbalanced data samples, and helps to improve the utilization rate of wind energy. The details are as follows:

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

[0033] (2) A sample dynamic balance factor is designed to balance the attention paid to samples of each category during model training to reduce the influence of majority class samples. The SAO optimization algorithm is introduced to dynamically update the balance factors of each category according to the training results to ensure that 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 diagnosis accuracy.

[0034] (3) A multi-level fault diagnosis model Dualformer based on dual Transformer is proposed, and a diagnostic control module is introduced to divide fault diagnosis into two stages: health and fault and fault type. This effectively shortens the diagnosis time of the fault diagnosis model and improves the diagnostic efficiency and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0036] Figure 2 The three-dimensional confusion matrix of the present invention. (a) Confusion matrix of fan #1; (b) Confusion matrix of fan #2;

[0037] Figure 3 It is a flow chart of the present invention. DETAILED DESCRIPTION

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

[0039] Reference Figure 3 A method for diagnosing a fault in a wind turbine pitch system for data balancing comprises the following steps:

[0040] Step 1: Data acquisition: obtain the operation data and fault data of wind turbines through the SCADA system equipped in the wind farm;

[0041] Step 2: Data preprocessing, remove features that are not related to the pitch system, including the annual / monthly / daily availability / power generation of wind turbines and features of other system components; remove invalid features with constant feature values, calculate the standard deviation of the remaining features, and remove invalid features with a standard deviation of 0; remove abnormal data with wind speeds less than the cut-in wind speed or greater than the cut-out wind speed, and fill in missing values ​​using the persistence method; replace data with active power exceeding the rated power with the rated power value, and replace data with negative active power with zero values; standardize the remaining feature data to eliminate the impact of dimensions between different features;

[0042] Step 3: Feature screening: The ReliefF algorithm is used to screen out sensitive features related to the pitch system fault, reduce the data sample dimension and reduce the input of redundant information. The number of sample sampling is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold of the feature weight.

[0043] Step 4: Construction and training of DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each category of samples according to the quantitative relationship, which is used to balance the attention paid to each category of samples during model training to reduce the influence of majority class samples. Secondly, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformer 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 melting optimization SAO algorithm. Finally, the DBFDformer model is trained and the variable pitch system fault is diagnosed.

[0044] The following details:

[0045] (1) ReliefF feature extraction algorithm

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

[0047] First, randomly select samples R from the training sample set i , from R i Find k nearest neighbor samples H from the same sample, and i Find k nearest neighbor samples M(C)(C≠Class(R i )). Then calculate R i Distance D from H H , R i The distance D from M(C) M , as shown in formulas (1)-(2). If D H Value less than D M , indicating that this feature is helpful in distinguishing the nearest neighbors of the same and different categories, then increase its weight value; if D H Value greater than D M , indicating that the feature has a negative impact on the nearest neighbors of the same and different classes, then reduce its weight value. The larger the weight of the feature, the stronger the classification ability, and vice versa. The weight is updated by continuous iteration m times to complete the feature selection. The weight calculation is as shown in formula (3).

[0048]

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

[0050] When the feature f is continuous:

[0051]

[0052] When feature f is discrete:

[0053]

[0054] (2) Snow Melt Optimization Algorithm

[0055] Snow Ablation Optimization (SAO) algorithm is a physics-based meta-heuristic algorithm that simulates the sublimation and melting process of snow in nature to achieve a trade-off between the development and exploration of the solution space. It uses a dual population mechanism for development and exploration to avoid premature convergence. The initial stage, exploration stage, and development stage of the optimization process are as follows:

[0056] 1) Initialization phase

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

[0058]

[0059] 2) Exploration phase

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

[0061]

[0062] By using dynamic and uniform step sizes, some potential areas in the search space can be explored, which well reflects the diffusion of steam in the search space. The position 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 a number of elite sets in the group, θ is a random number in [0,1], G(t) refers to the current optimal solution, represents the centroid position of the entire group, Z2(t) and Z3(t) represent the second best individual and the third best individual in the current group, respectively. c (t) represents the centroid position of the individuals whose fitness values ​​rank in the top 50%, and N1 represents the number of leaders, that is, N1 is equal to half of the size of the entire group.

[0067] 3) Development phase

[0068] When snow is converted into liquid water through melting behavior, the search agent is encouraged to develop high-quality solutions in the current optimal solution. According to the snowmelt model based on the degree-day method proposed by Zhou et al., the position update formula of the SAO algorithm is obtained:

[0069]

[0070] Where M is the snowmelt rate, t max represents the termination criterion, and δ is a random number in [-1,1].

[0071] (3) Transformer

[0072] Transformer consists of two parts: encoder and decoder, which mainly include multi-head self-attention (MHSA) module, feedforward neural network (FNN), residual connection and inter-layer norm. Among them, the MHSA module is the core component of Transformer. By calculating and allocating attention weights, it enhances the neural network's attention to key information and weakens the impact of redundant information on the target task. It has excellent global feature extraction capabilities and makes the interaction between encoder and decoder possible.

[0073] First, the original data is generated into query matrix Q, key matrix K and 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] Among them, χ∈R N×C represents the input, ω Q ,ω K ,ω V ∈R C×C are the weights of the corresponding linear mappings, denote the query matrix, key matrix, and value matrix in each header, respectively. are the weights of the corresponding transformations, H is the number of heads of the attention mechanism, N is the time dimension, C is the feature dimension of χ, and C H is the feature dimension of each head, C = H × C H .

[0077] Then, Q h and K h Perform a dot product to obtain global information, and then scale and normalize the result through the scaling factor and SoftMax function to obtain the attention weight, and finally multiply it by V h To obtain a weighted feature map, the specific expression is as shown in formula (16).

[0078]

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

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

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

[0082] (4) Dynamic balance factor

[0083] In view of the problem of unbalanced data samples in actual fault diagnosis, the usual solution is to assign a balancing factor to each category based on the quantitative relationship to enhance the model's recognition ability for minority categories. However, this method ignores the fact that samples of different categories have different differentiabilities, and a fixed balancing factor easily causes overfitting, which reduces the generalization ability of the model and thus affects the accuracy of fault diagnosis. To address this problem, the present invention designs a dynamic balancing factor and applies it to the cross entropy loss function to balance the attention paid to samples of each category during model training. The specific application process is as follows:

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

[0085]

[0086] Among them, H is the total number of sample categories, N max Indicates the maximum number of category samples, N h Represents the number of samples in the hth category.

[0087] 2) Based on the Binary Cross-Entropy Loss (BCELoss) function, as shown in Formula (20), the first-stage loss function L′ is constructed using the sample initial balance factor α fl , as shown in formula (21), is used for the training of the health and fault diagnosis model. If α>0, the loss of easy-to-classify samples can be reduced, so that the model training can focus more on 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 , represents the true category of the i-th sample, y i ∈{0,1},y i ′ is the model predicted probability of the i-th sample, y i ′∈[0,1].

[0090] 3) Use the initial fault balance factor β to construct the second-stage loss function L′ f ' l , as shown in formula (22), is used to train the fault type diagnosis model. By assigning corresponding weights to each fault category through the balancing factor, the model's learning attention and recognition sensitivity to minority category samples are increased, effectively balancing the influence of each fault category sample on model training, thereby improving the model's ability to recognize specific fault types.

[0091]

[0092] Among them, β h Represents the balance factor of the h-th category samples.

[0093] According to the loss value of each iterative training of the first and second stage diagnosis models, the SAO algorithm is used to dynamically update the balance factor to ensure that 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 diagnosis accuracy.

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

[0095] In order to further improve the accuracy and timeliness of fault diagnosis, this paper proposes a Dualformer multi-level fault diagnosis model based on the Transformer network. Figure 2 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 mainly used to identify the health and fault states of the variable pitch system, that is, to diagnose whether there is a fault. First, an adaptive maximum pooling layer is used to reduce the dimension of the input feature data and merge the key feature information of adjacent regions to reduce the complexity of the model and improve computational efficiency; secondly, 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 state of the variable pitch system; finally, the health and fault diagnosis of the variable pitch system is completed through the fully connected layer and the Argmax layer. Since this module is only used to diagnose whether there is a fault in the variable pitch system, it belongs to binary fault identification, so the network structure design of this module is simple and low in complexity, which can quickly and accurately identify faults and effectively improve the diagnostic efficiency of the model.

[0097] The diagnosis control module is mainly used to control the execution of different fault diagnosis modules, that is, whether to perform the diagnosis of a specific fault type. The diagnosis 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 diagnosis determines whether to perform the second stage diagnosis. If the diagnosis result of the first stage is healthy, it is directly output as the model result. If it is a fault, it enters the second stage, and the fault type diagnosis module identifies the specific type of the fault and outputs the diagnosed specific fault type as the model result. Since the fault type diagnosis module of the second stage is a multi-classification fault identification, compared with the health and fault diagnosis module of the first stage, its model is more complex, and the required parameters and operation time are more. Therefore, the diagnosis control module can realize the efficient connection of the first and second fault diagnosis stages, so that the model only performs high-precision diagnosis on the data that has been identified as a fault state, which can significantly improve the accuracy and runtime efficiency of the model diagnosis. In addition, due to the small amount of fault data and the imbalance in the ratio with healthy data, the number of high-precision diagnosis of fault types is extremely limited, and the operation efficiency of the model will be further improved.

[0098] The fault type diagnosis module is mainly used to quickly and accurately identify the specific fault type of the variable pitch system. The module first uses the Conv1D layer to obtain local feature information, capture the detailed features of the input data, and map the information to a high-dimensional space; then the BN&PReLU layer is used to control the transfer of the model training gradient to avoid the explosion and disappearance of the gradient and accelerate the convergence of the model training; secondly, the 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 specific fault type diagnosed is output through the fully connected layer and the Argmax layer to complete the second stage of fault diagnosis. Since this module only performs more accurate fault type diagnosis on fault data, it avoids the impact of the imbalance between fault data and healthy data on the model, and can effectively improve the diagnostic accuracy of the model.

[0099] (6) Experimental verification

[0100] 1) Evaluation indicators

[0101] In order to more objectively evaluate the diagnostic performance of the method of the present invention, F1-score and G-mean are used as evaluation indicators of the experimental results, which are defined as follows:

[0102]

[0103] TP N and FP N Represents the number of true positive and false positive samples, TN N and FN N Represents the number of true negative and false negative samples respectively. F1-score and G-mean are comprehensive evaluation indicators of 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 imbalanced categories. The higher the value of both, the better the classification effect of the model.

[0104] 2) Data processing

[0105] The present invention takes two Shanghai Electric W2000-111 wind turbines with a rated power of 2000KW as an example, numbered #1 and #2 respectively. The 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 contains 82 features and 52,572 data.

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

[0107]

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

[0109] Table 1 Description of wind turbine pitch system status and sample quantity

[0110]

[0111] 3) Wind turbine pitch system fault diagnosis results

[0112] Table 2 Evaluation of fault diagnosis results of wind turbine pitch system

[0113]

[0114] It can be seen from the F1-score and G-mean in the table that the diagnosis results of the pitch system health status of the method of the present invention are higher than 99.04%, and the diagnosis results of each type of fault are also higher than 91.94%, among which the diagnosis result of fault 2 is 100%. In addition, the three-dimensional confusion matrix diagram of the diagnosis results ( Figure 2 ) It can be observed more intuitively that the diagnostic accuracy of each state category reaches more than 89.47%, and most of them exceed 97.00%. The analysis shows that the present invention can effectively avoid the impact of the imbalance of fault and health data on the model, and shows good diagnostic ability and high diagnostic accuracy for the diagnosis of pitch system faults.

[0115] Since the present invention divides the fault diagnosis of the wind turbine pitch system into two stages: health and fault diagnosis and fault type diagnosis, in order to describe the diagnostic details of the method in more detail, the diagnostic results of the two stages are analyzed and studied respectively. The diagnostic results of the health and fault diagnosis stage are shown in Table 4. It can be observed from Table 4 that the recognition results of the proposed method for the health state are all above 98.80%, and the diagnosis results for the fault state are all above 97.73%. For a large number of healthy data samples, the rapid and accurate identification of the health and fault states of the pitch system is preferentially achieved, which greatly shortens the overall diagnostic time of the model and improves the diagnostic efficiency of the model. At the same time, it lays the foundation for the accurate diagnosis of the fault type diagnosis in the second stage, and ensures the overall diagnostic accuracy of the method of the present invention for the wind turbine pitch system fault.

[0116] Table 3 Evaluation of wind turbine pitch system health and fault diagnosis stage diagnosis results

[0117]

[0118] The diagnosis of the fault type in the second stage is determined according to the diagnosis results of the first stage. Its main function is to perform secondary diagnosis on the data diagnosed as faults in the first stage to avoid the influence of the majority of sample data and accurately diagnose the specific type of the fault. Table 4 shows the diagnosis results of the fault types of the wind turbine variable pitch system in the second stage. It can be seen from the table that the F1-score and G-mean of faults 1, 2 and 6 all reach 100%, while the reason why faults 4 and 5 in wind turbine #1 do not reach 100% is that there are two fault diagnosis errors, namely, fault 4 is misdiagnosed as fault 5 and fault 5 is misdiagnosed as fault 4. Similarly, the reason why faults 3, 4 and 5 in wind turbine #2 do not reach 100% is that there are two fault diagnosis errors, namely, fault 4 is misdiagnosed as fault 5 and fault 3 is misdiagnosed as fault 4. The analysis of the experimental results shows that the method of the present invention shows strong diagnostic performance and generalization ability in the diagnosis of various types of faults.

[0119] Table 4 Wind turbine pitch system fault types and the evaluation of the whole stage diagnosis results

[0120]

Claims

1. A wind turbine variable pitch system fault diagnosis method for data balancing, characterized in that: The specific steps include: Step 1: Data acquisition: obtain the operation data and fault data of wind turbines through the SCADA system equipped in the wind farm; Step 2: Data preprocessing, remove features that are not related to the pitch system; remove invalid features with constant feature values, calculate the standard deviation of the remaining features, and remove invalid features with a standard deviation of 0; remove abnormal data with wind speeds less than the cut-in wind speed or greater than the cut-out wind speed, and use the persistence method to fill in missing values; replace data with active power exceeding the rated power with the rated power value, and replace data with negative active power with zero values; standardize the remaining feature data to eliminate the impact of dimensions between different features; Step 3: Feature screening: The ReliefF algorithm is used to screen out sensitive features related to the pitch system fault, reduce the data sample dimension and reduce the input of redundant information. The number of sample sampling is 100, the number of nearest neighbor samples is 5, and the average value is used as the threshold of the feature weight. Step 4: Construction and training of DBFDformer data balancing fault diagnosis model. First, an initial dynamic balancing factor is assigned to each category of samples according to the quantitative relationship, which is used to balance the attention of each category of samples during model training to reduce the influence of majority class samples. Secondly, the dynamic balancing factor is introduced into the Dualformer multi-level fault diagnosis model based on dual Transformer 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 melting optimization algorithm (SAO). Finally, the DBFDformer model is trained and the variable pitch system fault is diagnosed.

2. A wind turbine pitch system fault diagnosis method for data balancing according to claim 1, characterized in that: The setting process of the dynamic balance factor DBF in step 4 is: 1) According to the quantitative relationship between healthy samples and faulty samples, the sample initial balance factor α is set; according to the quantitative relationship between samples of different fault categories, the fault initial balance factor β is set. The calculation of each initial balance factor is shown in formulas (18)-(19): Among them, H is the total number of sample categories, N max Indicates the maximum number of category samples, N h represents the number of samples in the hth category; 2) Based on the Binary Cross-Entropy Loss, BCELoss function, as shown in formula (20), the loss function L′ of the first stage is constructed using the sample initial balance factor α fl , as in formula (21), is used for training the health and fault diagnosis model; Among them, N represents the total number of samples, y i Indicates the true category of the i-th sample, y i ∈{0,1},y i ′ is the model predicted probability of the i-th sample, y i ′∈[0,1]; 3) Use the initial fault balance factor β to construct the second-stage loss function L′ f ' l , as shown in formula (22), is used for training the fault type diagnosis model; Among them, β h Represents the balance factor of the h-th category sample. 4) According to the loss value of each iterative training of the first and second stage diagnosis models, the SAO algorithm is used to dynamically update the balance factor to ensure that 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 diagnosis accuracy.

3. A wind turbine pitch system fault diagnosis method for data balancing according to claim 1, characterized in that: The DBFDformer data balance fault diagnosis model mainly consists of three modules: health and fault diagnosis, diagnosis control and fault type diagnosis; The health and fault diagnosis module is mainly used to identify the health and fault states of the variable pitch system, that is, to diagnose whether there is a fault: first, an adaptive maximum pooling layer is used to reduce the dimension of the input feature data and merge the key feature information of adjacent regions to reduce the complexity of the model and improve the computational efficiency; secondly, 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 state of the variable pitch system; finally, the health and fault diagnosis of the variable pitch system is completed through the fully connected layer and the Argmax layer; The diagnostic control module is mainly used to control the execution of different fault diagnosis modules, that is, whether to perform a diagnosis of a specific fault type; the diagnostic control module is divided into two stages: the first stage, that is, health and fault diagnosis, and the second stage, that is, fault type diagnosis. The diagnosis result of the first stage diagnosis determines whether to perform the second stage diagnosis. If the diagnosis result of the first stage is healthy, it is directly output as a model result. If it is a fault, it enters the second stage, and the fault type diagnosis module identifies the specific type of the fault; The fault type diagnosis module is mainly used to quickly and accurately identify the specific fault type of the variable pitch system; the module first uses the Conv1D layer to obtain local feature information, capture the detailed features of the input data, and map the information to a high-dimensional space; then the BN&PReLU layer is used to control the transmission of the model training gradient to avoid the explosion and disappearance of the gradient and accelerate the convergence speed of the model training; secondly, the 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 specific fault type diagnosed is output through the fully connected layer and the Argmax layer to complete the second stage of fault diagnosis.

Images