Hybrid strategy method for predicting cracking fault of fan blade by using SCADA (Supervisory Control and Data Acquisition) data

Abstract

The invention discloses a hybrid strategy method for predicting a cracking fault of a fan blade by using SCADA (Supervisory Control and Data Acquisition) data. The hybrid strategy method comprises the following steps: preprocessing marked SCADA data of the fan blade; constructing features highly related to the blade state in the SCADA data through mutual information screening; balancing an abnormal state data proportion based on a method of combining a support vector machine and a synthetic minority oversampling technology; and an improved Transform deep learning model and block design and feature separation input are integrated to realize accurate modeling and state prediction of long time series data. And normalizing a data set input into the prediction model, using oversampled data as model training data, and using real data of the same fan and different fans to carry out migration verification. Compared with the prior art, the method has the advantages that relevant features in the SCADA data can be extracted, and the problem of class imbalance can be effectively solved. The method can maintain high accuracy in fault prediction, has good migration generalization ability, and provides a low-cost, efficient and reliable solution for real-time prediction of fan blade crack faults.

Description

Technical Field

[0001] The invention relates to the field of data processing and equipment fault diagnosis, and in particular to a hybrid strategy method for predicting fan blade cracking faults using SCADA data. Background Art

[0002] As a key component of a fan, the reliability of fan blades has become increasingly prominent as the operating environment of the fan becomes more complex. When fan blades operate for a long time in various environments and under the action of superimposed stress, structural failures and performance degradation often occur. The failure of fan blades will cause the modal parameters of the blades to change, which will not only cause the fan to shut down and increase maintenance costs, but also cause huge operating accidents such as blade breakage. Among them, blade cracking is the most common and most serious type of failure of fan blades. Therefore, the diagnosis and prevention of fan blade cracking failures has become one of the key technologies to ensure the reliable operation of fans.

[0003] Since the SCADA system has been widely used in wind turbine condition monitoring, domestic and foreign experts and enterprises have carried out extensive and in-depth research and application by using SCADA data for wind turbine condition monitoring. SCADA data presents highly nonlinear, non-stationary and multidimensional characteristics. Therefore, the effective application of feature engineering, unbalanced data processing and extraction of SCADA data time series and spatial features in this field of research have brought many technical challenges to its use in predicting wind turbine blade cracking failures. In addition, the parameters of SCADA data lack standardized processing, which further increases the complexity of feature extraction. More importantly, SCADA data cannot reflect the status of wind turbine blades in real time, and complex data processing and analysis are required to draw meaningful conclusions. These challenges require the development of more advanced data processing and analysis methods to improve the accuracy and real-time performance of anomaly detection.

[0004] In view of these existing problems and current situations, the present invention considers the complex features and class imbalance characteristics of SCADA data, proposes a method combining mutual information feature selection to screen effective features; uses synthetic minority oversampling technology combined with support vector machine to perform class balance processing on abnormal data by focusing on boundary features; proposes a model algorithm that combines Transformer with block design and spatial information separation to train and classify and predict data, thereby improving the detection capability of the model. Summary of the invention

[0005] In response to the above technical challenges and research status, the present invention proposes a hybrid strategy prediction method for judging fan blade cracking faults using SCADA data. Targeted preprocessing and data screening are performed according to the characteristics of the data, and a new Transformer processing strategy is used to obtain high-precision prediction accuracy of fan blade cracking faults.

[0006] The technical solution adopted by the present invention is: a method for predicting crack failure of fan blades using SCADA data, which specifically includes the following steps:

[0007] S2: SCADA dataset preprocessing

[0008] In view of abnormal situations such as sensor errors and fan shutdown in data collection, invalid data is eliminated through data cleaning. Combined with the physical space similarity of SCADA data features, high similarity data is merged to extract low-dimensional features with new practical physical meanings, reduce data dimensions and size, optimize redundant information, and improve data training efficiency. Considering the different dimensions of data collected by sensors, each feature data is normalized to standardize it into relative quantities. Including:

[0009] S21: Select SCADA monitoring data, perform preliminary data cleaning, remove data in shutdown state and invalid data collected due to system failure, and retain normal data;

[0010] S22: Based on the physical similarity of features, merge features with the same or similar physical meanings to reduce redundant data and optimize feature expression;

[0011] S23: According to the characteristics of different dimensions of each sensor data, all SCADA data input into the model are normalized using formula (2) to convert the data into a normal distribution with a mean of 0 and a standard deviation of 1:

[0012]

[0013] In the formula, x ij * is the element in the i-th row and j-th column of the standardized SCADA matrix, x ij is the element in the i-th row and j-th column of the original data matrix D, μ j represents the mean of the jth feature in the matrix D; σ j represents the standard deviation of the jth feature in the matrix D. The standardized matrix is ​​D * =[x * T1 ,x * T2 ,…,x * TN ];

[0014] S24: Construct a hierarchical data set based on normalized relative quantities to provide a standardized input basis for data analysis and modeling;

[0015] S3: Constructing features highly correlated with blade status in SCADA data through mutual information screening

[0016] Based on the mutual information theory, the nonlinear correlation between the quantitative features and the blade state is calculated. By calculating the mutual information value between each feature and the wind turbine blade state information, the information sharing amount and mutual dependence between each variable and the state value are quantified, the nonlinear relationship between the variables is evaluated, and the mutual information threshold is set, so as to retain the features with significant mutual information values ​​​​with cracking faults, eliminate the features with mutual information values ​​​​lower than the set threshold, and construct a feature set with high sensitivity to crack faults. T1 ,x T2 ,…,x TN ], and its corresponding state label Y = [y1,y2,…,y T ]; The mutual information calculation method is shown in formula (3):

[0017]

[0018] In the formula, count(X i =x i ,Y=y) represents X i The value is x i And Y is the number of samples of y. The final mutual information vector MI = [MI(X1; Y), MI(X2, Y), ..., MI(X N ,Y)]; By setting the percentage threshold combined with the mutual information value, the characteristic variables within the threshold are retained;

[0019] S4: Balancing the proportion of abnormal state data based on support vector machine and synthetic minority oversampling technology

[0020] Establish boundary vectors based on optimal hyperplane nonlinear mapping, perform data augmentation on minority class abnormal data based on support vector machine, and establish a SCADA data oversampling method focusing on classification boundaries;

[0021] S41: First, the support vector machine projects the original data into a high-dimensional feature space through kernel function mapping, and searches for an optimal hyperplane in the feature space to maximize the interval between data of different categories, and searches for support vectors located at the classification boundary, where the support vectors are sample points located at the classification interval boundary. Then, for each boundary sample point, the minority class sample x * min Find k nearest neighbor samples NN k (x * min )={SV1,SV2,…SV k};

[0022] S42: Randomly select a nearest neighbor sample SV j , along x * min and SV jThe line segment between generates synthetic samples x syn =x * min +λ(SV j -x min ),λ~U(0,1), added to the original data. Repeat this process until the class imbalance reaches a relatively balanced state;

[0023] S5: Building a deep temporal model based on Transformer

[0024] The Transformer model is constructed based on block design, and the SCADA data is divided into input blocks with token attributes to extract deep and wide-area time series information. Each feature is separated independently and shares the same weight input model. Through the design of block input and shared weight, a model prediction process is established to optimize efficient independent features based on block input and shared weight. For the output multivariate SCADA data D′, the lookback window is K: (x1,…,x k ), x at each time step t t is a vector of dimension N, predicting whether the state at time step t is abnormal or normal;

[0025] S51: Each independent single feature sequence x i It is divided into blocks in the form of sliding windows. The length of each block is B and the sliding step is S. The total number of blocks that can be generated is M is the number of blocks,

[0026] S52: Transformer encoder completes the transformation of input signal to potential representation. Blocked Tokens are trained through linear projection Mapped to the Transformer latent space and applied additive positional encoding To maintain the time sequence, d is the embedding dimension after projection, and the representation after projection and position vector embedding As shown in the following formula (4):

[0027]

[0028] Then, the external category coding is introduced into the block representation by appending the category embedding at the end of the time series, as shown in the following formula (8):

[0029]

[0030] S53: The representation vector after embedding the category code and the position vector is taken as input, and the attention output is obtained by using the scaled dot product of the query matrix, key matrix and value matrix transformed in the multi-head attention mechanism;

[0031] The multi-head attention also includes a batch normalization layer and a feed-forward network with residual connections to generate the representation

[0032] S6: Classification prediction and target loss method

[0033] After the model training is completed, the new data generated by the system is input into the pre-trained model for state prediction. The data is pre-processed, the physical similarity features are merged, and the features after the mutual information screening are retained before input to identify abnormal states and obtain data detection results.

[0034] Finally, the classification result is obtained through the classification head. The classification head consists of a flattening layer and a linear layer, which are used to further process the extracted features. In this process, each feature first generates a corresponding prediction sequence, and then these prediction sequences are concatenated and fused to form a complete feature representation. Finally, the cross entropy loss function is used to measure the difference between the predicted probability distribution and the true distribution, and the ability of the model to judge blade cracking faults is optimized. The overall target loss of the cross entropy loss function is shown in the following equations (6) and (7):

[0035]

[0036] Where Loss represents the overall target loss, which measures the overall difference between the predicted and true categories; M represents the number of samples in the current batch, is the predicted value of the i-th sample. The cross entropy loss function CrossEntropy measures the cross entropy loss of the i-th sample, where y c is the label value of the true label vector in category c, is the value of the predicted probability vector output by softmax in category c. The final total loss value is obtained by averaging the cross entropy loss of all samples in the current batch, which is used for iterative model optimization and training.

[0037] As a preferred solution, the specific method for the support vector machine to find the classification boundary support vector in step S4 is:

[0038] Given a training matrix data sample and the class label y i ∈{-1,+1}, the support vector machine determines the optimal high-dimensional hyperplane by solving the optimization problem of the following formula (1). The optimal hyperplane is expressed as W·x+b=0, where x is the input vector and b is the bias term. The vectors distributed around the hyperplane are support vectors. This method solves the nonlinear regression problem in low-dimensional space by linear regression in the high-dimensional hyperplane.

[0039]

[0040] Where w is the weight vector, ξ i is the slack variable, and C is the penalty function. In order to solve the above optimization problem, the Lagrange multiplier α is introduced i Construct the Lagrangian function as shown in formula (2). By performing dual optimization on the Lagrangian function, its dual problem is obtained as shown in formula (3):

[0041]

[0042] Among them, K(x * i ,x * j ) is the kernel function, using the RBF kernel function: K(x * i ,x * j ) = exp(-γ||x * i -x * j || 2 , where γ is the RBF kernel parameter. SV is a sample point located on the boundary of the classification interval, satisfying w·x * i +b=±1, support vector satisfies α i > 0. According to the dual problem solving process, SV = {x * i |α i >0}, in geometric sense, SV={x * i |w·x * i +b=±1}.

[0043] As a preferred solution, the query matrix in step S53 is Key Matrix Sum Matrix The specific calculation method is in Scaling the dot product to get the attention output As shown in the following formula (4):

[0044]

[0045] In formula (4), the specific calculation steps of the attention mechanism are: first, through the weight matrix For the input vector Perform linear transformation to obtain the query matrix Key Matrix Sum Matrix Then, the attention score is calculated by dot product and scaling, and the score is converted into attention weight using the softmax function. Finally, the value matrix Perform weighted summation to generate attention output

[0046] The beneficial effects of the present invention are:

[0047] (1) Effectiveness of combining support vector machine and synthetic minority oversampling technology: By using oversampled data to train the pre-trained model, it shows significant advantages in dealing with the problem of SCADA data class imbalance. By oversampling minority class data samples, the model has a high level of performance.

[0048] (2) The combination of the Transformer model’s block approach and feature separation input significantly improves the accuracy and stability of wind turbine blade cracking fault prediction, especially when processing high-dimensional, nonlinear, and high-system complexity data, showing excellent performance.

[0049] (3) Effectiveness of feature selection method: Ablation experiments show that the mutual information feature selection used in this study has significant advantages in retaining key features and reducing dimensionality. Compared with other feature dimensionality reduction methods, it can more effectively improve the model performance in SCADA data and reduce the computing resources for training and inference.

[0050] (4) Reliability of migration generalization performance: In the verification of migration generalization performance of different types of wind turbines, the model still maintained a high level of performance, proving the strong adaptability and wide application potential of the invented method among different types of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the implementation scheme of the hybrid strategy prediction method for judging fan blade cracking faults using SCADA data in this application;

[0052] Figure 2 This is a data preprocessing flow chart of a hybrid strategy prediction method for judging fan blade cracking faults using SCADA data in this application;

[0053] Figure 3 A line graph of the feature mutual information value in the example of the hybrid strategy prediction method for judging the cracking fault of the wind turbine blade using SCADA data in this application;

[0054] Figure 4 A schematic diagram of the proportion of balanced abnormal state data based on support vector machine and synthetic minority oversampling technology for the hybrid strategy prediction method for judging fan blade cracking fault using SCADA data in this application;

[0055] Figure 5 This is a schematic diagram of the improved Transformer model in the hybrid strategy prediction method for judging wind turbine blade cracking failure using SCADA data in this application. DETAILED DESCRIPTION

[0056] To facilitate those skilled in the art to understand the specific implementation process of the present invention, the contents of the present invention are described in detail below with reference to the accompanying drawings.

[0057] The present invention proposes a hybrid strategy prediction method for judging fan blade cracking faults using SCADA data. The specific implementation process is as follows: Figure 1 As shown, the following steps are included:

[0058] S1: Wind turbine SCADA data acquisition and data set construction

[0059] Optionally, the SCADA system of the wind turbine collects data from several sensor points, including parameters such as temperature, pressure, current, yaw angle, torque, etc. at locations such as blades, hubs, nacelles, generators, converters, yaw systems, wind towers, bearings, etc., and collects SCADA data of 75 feature dimensions of the wind turbine. The collected SCADA data includes historical monitoring data and real-time monitoring data. Historical monitoring data is used for model training, and real-time monitoring data is used to determine the cracking of wind turbine blades. According to the fault records of cracked blades, the historical SCADA data of the week before the blade cracking is integrated. At the same time, the historical data of blades that operate normally is obtained. The data of every 10 minutes is set as a time window sample point, and the model is trained based on the data state label in the time window, so that the newly input unlabeled real-time monitoring data can be predicted and inferred to determine whether the wind turbine blade will crack in the next week when the real-time monitoring data in the time window is monitored. Based on the above method, a blade cracking fault data set containing two categories is constructed, and the data points are collected at the same sampling interval to form multiple multi-dimensional feature time series matrices with 10-minute data as samples. It can be expressed as formula (1):

[0060]

[0061] Where N is the number of features collected by SCADA data, T is the number of time series sampling data points collected, and x is the collected data point.

[0062] S2: SCADA dataset preprocessing

[0063] In the case of sensor errors, fan shutdown and other abnormal situations in data collection, invalid data is removed through data cleaning. Combined with the physical space similarity of SCADA data features, high similarity data is merged to extract low-dimensional features with new practical physical meanings, reduce data dimensions and size, optimize redundant information, and improve data training efficiency. Considering the different dimensions of data collected by sensors, each feature data is normalized to standardize it into relative quantities. The specific implementation process is shown in the attached figure. Figure 2 As shown, including:

[0064] S21: Select SCADA monitoring data, perform preliminary data cleaning, remove data in shutdown state and invalid data collected due to system failure, and retain normal data.

[0065] S22: Based on the physical similarity of features, merge features with the same or similar physical meanings to reduce redundant data and optimize feature expression.

[0066] Optional feature merging methods include: using the average blade angle to describe the average values ​​of multiple blade angles, using the average pitch motor current to describe the average current values ​​of multiple pitch motors, using the vector synthetic vibration value to describe the vibration values ​​of blades in different directions, using the converter temperature difference and pressure difference to describe the converter temperature and pressure difference, etc. Through feature merging based on physical similarity, 17 features with new physical meanings were obtained, and the features before merging were deleted, and the feature dimension was reduced from 75 to 43.

[0067] S23: According to the characteristics of different dimensions of each sensor data, all SCADA data input into the model are normalized using formula (2) to convert the data into a normal distribution with a mean of 0 and a standard deviation of 1:

[0068]

[0069] In the formula, x ij * is the element in the i-th row and j-th column of the standardized SCADA matrix, x ij is the element in the i-th row and j-th column of the original data matrix D, μ j represents the mean of the jth feature in the matrix D; σ j represents the standard deviation of the jth feature in the matrix D. The standardized matrix is ​​D * =[x * T1 ,x * T2 ,…,x * TN ].

[0070] S24: Construct a hierarchical data set based on normalized relative quantities to provide a standardized input basis for data analysis and modeling.

[0071] S3: Constructing features highly correlated with blade status in SCADA data through mutual information screening

[0072] Based on the mutual information theory, the nonlinear correlation between the quantitative features and the blade state is calculated. By calculating the mutual information value between each feature and the wind turbine blade state information, the information sharing amount and mutual dependence between each variable and the state value are quantified, the nonlinear relationship between the variables is evaluated, and the mutual information threshold is set, so as to retain the features with significant mutual information values ​​​​with cracking faults, eliminate the features with mutual information values ​​​​lower than the set threshold, and construct a feature set with high sensitivity to crack faults. T1 ,x T2 ,…,x TN ], and its corresponding state label Y = [y1,y2,…,y T ]; The mutual information calculation method is shown in formula (3):

[0073]

[0074] In the formula, count(X i =x i ,Y=y) represents X i The value is x i And Y is the number of samples of y. The final mutual information vector MI = [MI(X1; Y), MI(X2, Y), ..., MI(X N ,Y)].

[0075] Optionally, the number of features to be retained can be determined by setting a quantile threshold comparison experiment. In this example, 33 features are retained and the last 10 features with low mutual information values ​​are deleted. The mutual information values ​​between each feature and the status of the wind turbine blade are shown in the attached figure. Figure 3 shown.

[0076] S4: Balancing the proportion of abnormal state data based on support vector machine and synthetic minority oversampling technology

[0077] Establish a boundary vector based on the optimal hyperplane nonlinear mapping, perform data amplification on minority class abnormal data based on support vector machine, and establish a SCADA data oversampling method that focuses on classification boundaries. The schematic diagram of the method is shown in the attached figure. Figure 4 shown.

[0078] S41: First, the support vector machine projects the original data into a high-dimensional feature space through kernel function mapping, and searches for an optimal hyperplane in the feature space to maximize the interval between data of different categories, and searches for support vectors located at the classification boundary, where the support vectors are sample points located at the classification interval boundary. Then, for each boundary sample point, the minority class sample x *min Find k nearest neighbor samples NN k (x * min )={SV1,SV2,…SV k}.

[0079] The specific method of the support vector machine to find the classification boundary support vector is: given a training matrix data sample and the class label y i ∈{-1,+1}, the support vector machine determines the optimal high-dimensional hyperplane by solving the optimization problem of the following formula (4), and the optimal hyperplane is expressed as W·x+b=0, where x is the input vector and b is the bias term. The vectors distributed around the hyperplane are support vectors. This method solves the nonlinear regression problem in low-dimensional space by linear regression in the high-dimensional hyperplane.

[0080]

[0081] Where w is the weight vector, ξ i is the slack variable, and C is the penalty function. In order to solve the above optimization problem, the Lagrange multiplier α is introduced i Construct the Lagrangian function as shown in formula (5). By performing dual optimization on the Lagrangian function, we get its dual problem as shown in formula (6):

[0082]

[0083] Among them, K(x * i ,x * j ) is the kernel function, using the RBF kernel function: K(x * i ,x * j ) = exp(-γ||x * i -x * j || 2 , where γ is the RBF kernel parameter. SV is a sample point located on the boundary of the classification interval, satisfying w·x * i +b=±1, support vector satisfies α i > 0. According to the dual problem solving process, SV = {x * i |α i >0}, in geometric sense, SV={x * i |w·x * i +b=±1}.

[0084] S42: Randomly select a nearest neighbor sample SV j , along x * min and SV j The line segment between generates synthetic samples x syn =x * min +λ(SV j -x min ),λ~U(0,1), and add them to the original data. Repeat this process until the class imbalance reaches a relatively balanced state.

[0085] S5: Building a deep temporal model based on Transformer

[0086] The SCADA data is divided into input blocks with token attributes. Every 16 monitoring points are grouped into one block, and the step length between each block is set to 8. This operation extracts deep and wide-area timing information and reduces the computational complexity of the Transformer, reducing the number of input tokens from the original L to approximately L / S. This reduces the memory usage and computational complexity of the Transformer's attention graph from the original o(L 2 ) is reduced to o((L / S) 2 ). Each feature space information is independently separated and shares the same weight input model. Through the design of block input and shared weight, the temporal information of a single feature is accurately learned, and a model prediction process for optimizing efficient independent features based on block input and shared weight is established. The Transformer structure described in the present invention is an encoder structure of the Transformer, which includes a multi-head attention layer and a feedforward neural network, and both layers are calculated by residual connection. The multi-head attention layer associates the features of each time step in the sequence by calculating the self-attention of the input sequence, thereby realizing global information integration between different blocks. The feedforward neural network further processes the features output by the multi-head attention layer, so that the model can have a strong feature expression capability while capturing temporal dependencies. In addition, residual connections are used between the multi-head attention layer and the feedforward neural network layer, so that the model maintains the feature information of the original input during the layer-by-layer feature extraction process, which can prevent the gradient from disappearing or exploding, and improve the convergence and stability of the model.

[0087] For the output multivariate SCADA data D′, the lookback window is K:(x1,…,x k ), x at each time step t t is a vector with dimension N, which predicts whether the state of time step t is abnormal or normal. The structure of the improved Transformer model is shown in the attached figure. Figure 5 shown.

[0088] S51: Each independent single feature sequence x i It is divided into blocks in the form of sliding windows. The length of each block is B and the sliding step is S. The total number of blocks that can be generated is M is the number of blocks,

[0089] S52: Transformer encoder completes the transformation of input signal to potential representation. Blocked Tokens are trained through linear projection Mapped to the Transformer latent space and applied additive positional encoding To maintain the time sequence, d is the embedding dimension after projection, and the representation after projection and position vector embedding As shown in the following formula (7):

[0090]

[0091] Then, the external category coding is introduced into the block representation by appending the category embedding at the end of the time series, as shown in the following formula (8):

[0092]

[0093] S53: The representation vector after embedding the category code and the position vector is used as input, and the attention output is obtained by using the scaled dot product of the query matrix, key matrix and value matrix transformed in the multi-head attention mechanism; the query matrix Key Matrix Sum Matrix The specific calculation method is in Scaling the dot product to get the attention output As shown in the following formula (9):

[0094]

[0095] In formula (4), the specific calculation steps of the attention mechanism are: first, through the weight matrix For the input vector Perform linear transformation to obtain the query matrix Key Matrix Sum Matrix Then, the attention score is calculated by dot product and scaling, and the score is converted into attention weight using the softmax function. Finally, the value matrix Perform weighted summation to generate attention output

[0096] The multi-head attention also includes a batch normalization layer and a feed-forward network with residual connections to generate the representation

[0097] S6: Output feature concatenation for evaluation and prediction.

[0098] After the model training is completed, the new data generated by the system is input into the pre-trained model for state prediction. The data is pre-processed, the physical similarity features are merged, and the features after the mutual information screening are retained before input to identify abnormal states and obtain data detection results.

[0099] Finally, the classification result is obtained through the classification head. The classification head consists of a flattening layer and a linear layer, which is used to further process the extracted features. In this process, each feature first generates a corresponding prediction sequence, and then these prediction sequences are concatenated and fused to form a complete feature representation. Finally, the cross entropy loss function is used to measure the difference between the predicted probability distribution and the true distribution, and the ability of the model to judge blade cracking faults is optimized. The overall target loss of the cross entropy loss function is shown in the following equations (10) and (11):

[0100]

[0101] Where Loss represents the overall target loss, which measures the overall difference between the predicted and true categories; M represents the number of samples in the current batch, is the predicted value of the i-th sample. The cross entropy loss function CrossEntropy measures the cross entropy loss of the i-th sample, where y c is the label value of the true label vector in category c, is the value of the predicted probability vector output by softmax in category c. The final total loss value is obtained by averaging the cross entropy loss of all samples in the current batch, which is used for iterative model optimization and training.

[0102] Back propagation optimizes the model parameters and gradually improves the prediction accuracy. In the inference stage, the probability distribution of blade cracking failure is generated through the classification head to achieve fault prediction. The effectiveness of this application strategy is evaluated through four key indicators: accuracy, precision, recall and F1-score.

[0103] In model training, the hyperparameters of the model are set as follows: the number of Transformer encoder layers is 3; BatchSize is set to 32; the maximum epoch is set to 100, and an early stopping strategy is set. If the loss value does not decrease after more than 10 epochs, the training is stopped; the learning rate is set to 0.001. In order to comprehensively evaluate the effectiveness and advantages of the strategy of the present invention, a number of advanced models were selected for comparison, including four of the latest Transformer-based multidimensional time series classification methods: the original Transformer, Informer, Reformer, and Autoformer; MultiRocket, a time series classification method based on convolution kernel transformation; DLiner, a time series classification method based on linear classification; and the classic LSTM model. Through multi-dimensional comparisons, the data processing strategy and model used in the present invention are superior to the performance of other solutions in the detection of fan blade cracking faults.

[0104] The data of the same type of wind turbines were used for training and testing, and the comparison results with other methods are shown in Appendix 1.

[0105] Table 1 Comparison results of the present invention with other methods in the test of the same type of fans

[0106] Model Accuracy (%) Accuracy (%) Recall rate (%) F1(％) Method of the present invention 98.4 97.6 99.2 98.4 MultiRocket 91.0 98.0 86.0 91.6 LSTM 91.4 96.0 87.9 91.7 DLiner 70.0 62.7 73.6 67.7 Transformer 96.4 96.0 96.8 96.4 Informer 95.0 95.2 94.8 95.0 Reformer 94.8 94.0 95.5 94.7 Autoformer 94.8 97.6 92.4 94.9

[0107] The data of one type of wind turbines was used for training, and the data of another type of wind turbines was used for testing to evaluate the generalization ability of the strategy of the present invention. The comparison results are shown in Appendix 2.

[0108] Table 2 Comparison results of the present invention with other methods in cross-type fan testing

[0109] Model Accuracy (%) Accuracy (%) Recall rate (%) F1(％) Method of the present invention 93.9 93.6 94.1 93.8 MultiRocket 50.5 0.5 50.0 1.1 LSTM 52.9 100.0 51.2 67.8 DLiner 46.3 17.1 40.0 24.0 Transformer 52.6 7.5 70.0 13.5 Informer 55.0 61.0 54.0 57.3 Reformer 50.5 1.1 50.0 2.1 Autoformer 76.5 77.5 75.5 76.5

[0110] Those skilled in the art will appreciate that the above embodiments are intended to help readers understand the basic principles of the present invention, and should be understood that the protection scope of the present invention is not limited to the above specific embodiments. For those skilled in the art, the present invention may have various changes and variations. Any modifications and equivalent substitutions made within the basic principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A hybrid strategy method for predicting fan blade cracking failure using SCADA data, specifically comprising the following steps: S1: Wind turbine SCADA data acquisition and data set construction The SCADA system of the wind turbine collects data from several sensor points, including temperature, pressure, current, yaw angle, torque and other parameters at blades, hubs, nacelles, generators, converters, yaw systems, wind towers, bearings and other locations, to collect the SCADA data of the wind turbine. According to the fault records of cracked blades, the SCADA data of the week before the blades cracked are integrated. At the same time, the data of blades operating normally is obtained. The data of every 10 minutes of data is set as a time window sample point, and the model is trained based on the data state label in the time window, so that predictive reasoning can be performed on the newly input unlabeled data to determine whether the wind turbine blades will crack in the next week when the data in the time window is monitored. Based on the above method, a blade cracking failure data set containing two categories is constructed. The data points are collected at the same sampling interval to form multiple multi-dimensional feature time series matrices with 10-minute data as samples. It can be expressed as formula (1): Where N is the number of features collected by SCADA data, T is the number of time series sampling data points collected, and x is the collected data point; S2: SCADA dataset preprocessing In view of abnormal situations such as sensor errors and fan shutdown in data collection, invalid data is eliminated through data cleaning. Combined with the physical space similarity of SCADA data features, high similarity data is merged to extract low-dimensional features with new practical physical meanings, reduce data dimensions and size, optimize redundant information, and improve data training efficiency. Considering the different dimensions of data collected by sensors, each feature data is normalized to standardize it into relative quantities. Including: S21: Select SCADA monitoring data, perform preliminary data cleaning, remove data in shutdown state and invalid data collected due to system failure, and retain normal data; S22: Based on the physical similarity of features, merge features with the same or similar physical meanings to reduce redundant data and optimize feature expression; S23: According to the characteristics of different dimensions of each sensor data, all SCADA data input into the model are normalized using formula (2) to convert the data into a normal distribution with a mean of 0 and a standard deviation of 1: In the formula, x ij * is the element in the i-th row and j-th column of the standardized SCADA matrix, x ij is the element in the i-th row and j-th column of the original data matrix D, μ j represents the mean of the jth feature in the matrix D; σ j represents the standard deviation of the jth feature in the matrix D. The standardized matrix is ​​D * =[x * T1 ,x * T2 ,…,x * TN ]; S24: Construct a hierarchical data set based on normalized relative quantities to provide a standardized input basis for data analysis and modeling; S3: Constructing features highly correlated with blade status in SCADA data through mutual information screening Based on the mutual information theory, the nonlinear correlation between the quantitative features and the blade state is calculated. By calculating the mutual information value between each feature and the wind turbine blade state information, the information sharing amount and mutual dependence between each variable and the state value are quantified, the nonlinear relationship between the variables is evaluated, and the mutual information threshold is set, so as to retain the features with significant mutual information values ​​​​with cracking faults, eliminate the features with mutual information values ​​​​lower than the set threshold, and construct a feature set with high sensitivity to crack faults. T1 ,x T2 ,…,x TN ], and its corresponding state label Y = [y1,y2,…,y T ]; The mutual information calculation method is shown in formula (3): In the formula, count(X i =x i ,Y=y) means X i The value is x i And Y is the number of samples of y. The final mutual information vector MI = [MI(X1; Y), MI(X2, Y), ..., MI(X N ,Y)]; By setting the percentage threshold combined with the mutual information value, the characteristic variables within the threshold are retained; S4: Balancing the proportion of abnormal state data based on support vector machine and synthetic minority oversampling technology Establish boundary vectors based on optimal hyperplane nonlinear mapping, perform data augmentation on minority class abnormal data based on support vector machine, and establish a SCADA data oversampling method focusing on classification boundaries; S41: First, the support vector machine projects the original data into a high-dimensional feature space through kernel function mapping, and searches for an optimal hyperplane in the feature space to maximize the interval between data of different categories, and searches for support vectors located at the classification boundary, where the support vectors are sample points located at the classification interval boundary. Then, for each boundary sample point, the minority class sample x * min Find k nearest neighbor samples NN k (x * min )={SV1,SV2,…SV k }; S42: Randomly select a nearest neighbor sample SV j , along x * min and SV j The line segment between generates synthetic samples x syn =x * min +λ(SV j -x min ),λ~U(0,1), added to the original data. Repeat this process until the class imbalance reaches a relatively balanced state; S5: Building a deep temporal model based on Transformer The Transformer model is constructed based on block design, and the SCADA data is divided into input blocks with token attributes to extract deep and wide-area time series information. Each feature is separated independently and shares the same weight input model. Through the design of block input and shared weight, a model prediction process is established to optimize efficient independent features based on block input and shared weight. For the output multivariate SCADA data D′, the lookback window is K: (x1,…,x k ), x at each time step t t is a vector of dimension N, predicting whether the state at time step t is abnormal or normal; S51: Each independent single feature sequence x i It is divided into blocks in the form of sliding windows. The length of each block is B and the sliding step is S. The total number of blocks that can be generated is M is the number of blocks, S52: Transformer encoder completes the transformation of input signal to potential representation. Blocked Tokens are trained through linear projection Mapped to the Transformer latent space and applied additive positional encoding To maintain the time sequence, d is the embedding dimension after projection, and the representation after projection and position vector embedding As shown in the following formula (4): Then, the external category coding is introduced into the block representation by appending the category embedding at the end of the time series, as shown in the following formula (8): S53: The representation vector after embedding the category code and the position vector is taken as input, and the attention output is obtained by using the scaled dot product of the query matrix, key matrix and value matrix transformed in the multi-head attention mechanism; The multi-head attention also includes a batch normalization layer and a feed-forward network with residual connections to generate the representation S6: Classification prediction and target loss method After the model training is completed, the new data generated by the system is input into the pre-trained model for state prediction. The data is pre-processed, the physical similarity features are merged, and the features after the mutual information screening are retained before input to identify abnormal states and obtain data detection results. Finally, the classification result is obtained through the classification head. The classification head consists of a flattening layer and a linear layer, which are used to further process the extracted features. In this process, each feature first generates a corresponding prediction sequence, and then these prediction sequences are concatenated and fused to form a complete feature representation. Finally, the cross entropy loss function is used to measure the difference between the predicted probability distribution and the true distribution, and the ability of the model to judge blade cracking faults is optimized. The overall target loss of the cross entropy loss function is shown in the following equations (6) and (7): Where Loss represents the overall target loss, which measures the overall difference between the predicted and true categories; M represents the number of samples in the current batch, is the predicted value of the i-th sample. The cross entropy loss function CrossEntropy measures the cross entropy loss of the i-th sample, where y c is the label value of the true label vector in category c, is the value of the predicted probability vector output by softmax in category c. The final total loss value is obtained by averaging the cross entropy loss of all samples in the current batch, which is used for iterative model optimization and training.

2. A hybrid strategy method for predicting fan blade cracking failure using SCADA data as claimed in claim 1, characterized in that: The specific method for the support vector machine to find the classification boundary support vector in the above step S4 is: Given a training matrix data sample and the class label y i ∈{-1,+1}, the support vector machine determines the optimal high-dimensional hyperplane by solving the optimization problem of the following formula (1). The optimal hyperplane is expressed as W·x+b=0, where x is the input vector and b is the bias term. The vectors distributed around the hyperplane are support vectors. This method solves the nonlinear regression problem in low-dimensional space by linear regression in the high-dimensional hyperplane. Where w is the weight vector, ξ i is the slack variable, and C is the penalty function. In order to solve the above optimization problem, the Lagrange multiplier α is introduced i Construct the Lagrangian function as shown in formula (2). By performing dual optimization on the Lagrangian function, we get its dual problem as shown in formula (3): Among them, K(x * i ,x * j ) is the kernel function, using the RBF kernel function: K(x * i ,x * j ) = exp(-γ||x * i -x * j || 2 , where γ is the RBF kernel parameter. SV is a sample point located on the boundary of the classification interval, satisfying w·x * i +b=±1, support vector satisfies α i > 0. According to the dual problem solving process, SV = {x * i |α i >0}, in geometric sense, SV={x * i |w·x * i +b=±1}.

3. A hybrid strategy method for predicting fan blade cracking failure using SCADA data as claimed in claim 1, characterized in that: In the above step S53, the query matrix Key Matrix Sum Matrix The specific calculation method is in Scaling the dot product to get the attention output As shown in the following formula (4): In formula (4), the specific calculation steps of the attention mechanism are: first, through the weight matrix W Qh , W Kh , For the input vector Perform linear transformation to obtain the query matrix Key Matrix Sum Matrix Then, the attention score is calculated by dot product and scaling, and the score is converted into attention weight using the softmax function. Finally, the value matrix Perform weighted summation to generate attention output

Images