Traction rectifier open-circuit fault diagnosis method based on feature transfer learning

Abstract

The invention provides a traction rectifier open-circuit fault diagnosis method based on feature transfer learning, and belongs to the technical field of power electronic fault diagnosis. Comprising the steps of S1, data preparation; s2, dividing a test set and a training set; s3, training to obtain an IGBT open-circuit fault diagnosis model suitable for the high-power traction rectifier based on the training set data in combination with a weighted balance distribution adaptive WBDA algorithm and an RVFL algorithm; s4, performing hyper-parameter optimization on the training model in the S3 by using Bayesian optimization; s5, the optimal parameters obtained in the step S4 are used for training the step S3 again, a WBDA + RVFL classification model is obtained, test set data are used for verification, and a diagnosis result of the model in a test set is obtained. According to the method, the problems of low fault diagnosis precision and poor reliability caused by sample number and quality problems in open-circuit fault diagnosis of the high-power traction rectifier are solved.

Description

Technical Field

[0001] The invention provides a traction rectifier open circuit fault diagnosis method based on feature transfer learning, belonging to the technical field of power electronic fault diagnosis. Background Art

[0002] As an important part of the energy conversion of the electric traction drive system, the train traction rectifier is under high-intensity alternating thermal fatigue load for a long time, which accelerates the aging of the material and internal packaging structure, and is frequently subjected to electrothermal stress shock, resulting in a high proportion of its failures in the electric traction drive system. As the main body of realizing electric energy conversion, the IGBT module has a very high failure rate in the traction rectifier components. Therefore, it is particularly important to study the fault diagnosis of the traction rectifier IGBT module. The fault diagnosis objects of existing traction rectifiers are concentrated on the open circuit faults of the IGBT module, and the fault diagnosis methods used can be divided into three categories: model-based, signal-based, and knowledge-based. Among them, the model-based method, such as the "Open Circuit Fault Diagnosis Method of Single-Phase Three-Level Rectifier in High-Speed ​​Railway Traction System", compares the relationship between the residual feature quantity and the adaptive threshold after establishing a mathematical model to diagnose and locate the switch tube fault, but due to the load of the traction system, this type of method often has the problems of difficulty in establishing a mathematical model and low accuracy. Signal-based and knowledge-based methods, such as CN114624630B discloses a three-phase rectifier fault diagnosis method and system based on spectrum analysis, and CN114778969B discloses a rectifier fault diagnosis method based on RBF-Elman neural network. By extracting features from the rectifier voltage and current signals, fault diagnosis and positioning are performed by capturing feature changes or machine learning artificial neural networks, which can achieve high diagnostic accuracy. However, these two methods are highly dependent on data, especially the knowledge-based method, which requires the use of massive historical data during the training process. In actual applications, there are problems such as little real operating data and unbalanced fault data, which makes it difficult to establish a diagnostic model.

[0003] In summary, the existing technologies for traction rectifier open circuit faults have the following disadvantages:

[0004] 1) The problem of insufficient actual operating data of traction rectifiers is not considered;

[0005] 2) The problem of imbalanced fault data in actual operation is not considered. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, a traction rectifier open-circuit fault diagnosis method based on feature transfer learning provided by the present invention utilizes the idea of ​​feature transfer learning to reduce the distribution distance between samples, solves the feature transfer matrix of low-power rectifier simulation data and high-power rectifier operation data, and combines the random vector functional link network (RVFL) algorithm to obtain a fault diagnosis model suitable for high-power rectifiers, thereby solving the problems of small actual operation data of traction rectifiers and unbalanced fault data that may not be considered in the prior art.

[0007] The specific technical solution is:

[0008] A traction rectifier open circuit fault diagnosis method based on feature transfer learning comprises the following steps:

[0009] S1. Data preparation;

[0010] S1 specifically includes the following sub-steps:

[0011] S1-1. Build a low-power single-phase two-level pulse rectifier model, collect grid-side current data under normal conditions and after the IGBT modules T1, T2, T3, and T4 are open-circuited. The grid-side current data of each fundamental wave cycle (0.02s) is grouped together, and different fault type labels {y=1, 2, 3, 4, 5} are assigned to obtain the source domain database.

[0012] S1-2. Establish a high-power traction rectifier model, collect grid-side current data under normal conditions and after the IGBT modules T1, T2, T3, and T4 are open-circuited. The grid-side current data of each fundamental wave period (0.02s) is grouped together, and different fault type labels {y=1,2,3,4,5} are assigned to obtain the target domain database.

[0013] S1-3. Use resampling technology to ensure the consistency of the feature numbers of target domain samples and source domain samples, and normalize the source domain and target domain data.

[0014] S2, divide the test set and training set;

[0015] All source domain data are used for transfer learning model training; except for some labeled target domain data, the rest of the target domain data are used to test the performance of the diagnostic model;

[0016] The specific method of S2 is to simulate the real data situation and extract The sample data of each group is added to the training set. For each fault type of fault data, only k groups are extracted and added to the training set, which is expressed as

[0017]

[0018] The training set data is represented as

[0019] S3. Based on the training set data, combined with the weighted balanced distribution adaptation (WBDA) algorithm and the random vector functional link network (RVFL) algorithm, an IGBT open circuit fault diagnosis model suitable for high-power traction rectifiers is trained;

[0020] S3 specifically includes the following sub-steps:

[0021] S3-1, load training set data X and training set label Y;

[0022] S3-2, set the parameters of the WBDA algorithm and the RVFL algorithm, including the kernel bandwidth gamma, regularization parameter λ, kernel type kernel_type, number of iterations T, adapted dimension dim, balance factor μ of the WBDA algorithm, and the number of hidden layers Number of Hidden Neurons and activation function type of the RVFL algorithm;

[0023] S3-3, perform model initialization, initialize the centralization matrix H and the edge distribution distance matrix M0;

[0024] H=I-(1 / n)1 (1)

[0025]

[0026] In the formula, m represents the number of samples in the training set X, and n represents the number of features of each sample. is the identity matrix;

[0027] S3-4. Use the RVFL algorithm to perform preliminary classification on the target domain training data. The training input data of the RVFL algorithm is the labeled source domain and target domain data, and obtain the target domain data pseudo label y_pseudo;

[0028] In RVFL, by The initial input data X = [x1, x2, ..., x N ] is mapped to the hidden layer, where g(.) is a nonlinear activation function, ω j is the weight of the jth hidden node, b j is the bias; combining the direct nodes and the hidden nodes, the RVFL network is expressed as:

[0029]

[0030] Where J is the total number of hidden nodes, N is the total number of input nodes, and β j is the jth weight term; the training of the RVFL network is to minimize the squared output error (ε by optimizing the optimal hidden layer weights and output layer weights 2 ):

[0031]

[0032] Among them B t It consists of (J+N) weight values, where P is the total number of training samples, t is the target vector, d is the vector including the output node, and (p) is the sample index; the weights between the input layer and the hidden layer neurons are randomly sampled from a uniform distribution [-1, 1], and then the weights between the hidden layer and the output layer neurons are optimized by the least squares method;

[0033] Given a training sample X, the multi-output nodes of RVFL are f(X)=[f1(x),...,f E (x)], where E is the number of output nodes, then the general decision function of the RVFL classifier with multiple outputs is:

[0034]

[0035] S3-5. Use the target domain pseudo label y_pseudo and the source domain true label to initialize the conditional distribution distance matrix M c :

[0036]

[0037] S3-6. Set the balance factor μ and calculate the total distribution distance matrix M:

[0038] M=(1-μ)×M0+μ×M c 0≤μ≤1 (7)

[0039] S3-7. Consider minimizing the distribution distance between the source domain and the target domain data, and obtain the formula:

[0040]

[0041] Where tr(·) represents the trace of the matrix, and the A matrix is ​​the characteristic migration matrix, which is a Hermitian matrix;

[0042] S3-8. Considering the divergence of the data before and after the migration of the constraint feature, the formula is obtained:

[0043]

[0044] S3-9. Combining equations (8) and (9), adding a regularization term with a coefficient of λ>0, we get the final learning goal of transfer learning:

[0045]

[0046] S3-10. Solve equation (10) using the Lagrangian method. Its Lagrangian function is expressed as:

[0047]

[0048] Let its derivative have to

[0049]

[0050] Where φ is the Lagrange multiplier;

[0051] Before calculating the matrix A, select the appropriate kernel type to put the sample X into high-dimensional mapping, and obtain the mapping matrix K to replace X to solve the feature migration matrix A;

[0052]

[0053] The matrix A is solved using the eigs function in matlab;

[0054] S3-10. Find the sample feature matrix Z after feature migration, which contains the source domain data Z after feature migration S and target domain data Z T :

[0055]

[0056] S3-11. Using source domain data Z after feature migration S and target domain data Z T , combined with known labels, train the RVFL algorithm, classify the target domain training data, and obtain the target domain fault type diagnosis result y_tar_pseudo;

[0057] S3-12. Repeat steps 3-5 to 3-11 according to the number of iterations T to obtain the classification label results of the target domain data of the final training set, compare them with the true value, and obtain the model training accuracy.

[0058] S4. Use Bayesian optimization to optimize the hyperparameters of the training model of S3. The hyperparameter optimization objects include: the kernel bandwidth gamma, regularization parameter λ, adapted dimension dim, balance factor μ of the WBDA algorithm, and the number of hidden layers Number of Hidden Neurons of the RVFL algorithm;

[0059] S4-1. Define the parameter range that needs to be optimized, including gamma, λ, dim, μ and Number of Hidden Neurons;

[0060] S4-2, Bayesian Optimization:

[0061] Use the bayesopt function in MATLAB to perform Bayesian parameter optimization, the goal is to minimize the output of the K-fold cross validation function; the K-fold cross validation function takes K=3, and calculates the average accuracy and standard deviation of the training model in step 3; the optimization goal is to maximize the mean of the accuracy while controlling the size of the standard deviation, improving the accuracy of the training model while ensuring the stability of the training results;

[0062] S4-3. Obtaining optimal parameters:

[0063] Extract the optimal parameters from the Bayesian optimization results and save the values ​​of the optimal parameters.

[0064] S5. Retrain step S3 using the optimal parameters obtained in step S4 to obtain a WBDA+RVFL classification model, save the model parameters, and verify using the test set data to obtain the diagnostic results of the model in the test set.

[0065] Compared with the prior art, the beneficial effect of the present invention is that the present invention adds easily acquired small-power rectifier simulation data to the open-circuit fault diagnosis model training of high-power traction rectifiers with a small number of training samples and a normal data volume far greater than the fault data through the idea of ​​feature transfer learning. This solves the problems of low fault diagnosis accuracy and poor credibility of high-power traction rectifiers caused by sample quantity and quality issues, has reference value for the research on open-circuit fault diagnosis of traction rectifiers in actual operation, and contributes to the safe and stable operation of the electric traction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the present invention;

[0067] Figure 2 is the confusion matrix of the training set test results of the embodiment.

[0068] Figure 3 is the test set diagnosis result confusion matrix of the embodiment. DETAILED DESCRIPTION

[0069] The specific technical solution of the present invention is explained in conjunction with embodiments.

[0070] like Figure 1 As shown, the traction rectifier open circuit fault diagnosis method based on feature transfer learning includes the following steps:

[0071] S1. Data preparation;

[0072] S1-1. Establish a low-power single-phase two-level pulse rectifier model, collect the grid-side current data under normal conditions and after the IGBT module T1, T2, T3, and T4 are open-circuited. The sampling rate is 5e-5. The grid-side current data of each fundamental wave period (0.02s) is a group, and different fault type labels are assigned to obtain the source domain database. Label i Contains 1-5 corresponding to different fault types.

[0073] S1-2. In order to test the accuracy of the traction rectifier open circuit fault diagnosis method based on feature transfer learning, a high-power traction rectifier model is established to obtain data. The grid-side current data of normal conditions and after the IGBT modules T1, T2, T3, and T4 are open-circuited are collected. The sampling rate is 4e-5, and the grid-side current data of each fundamental wave period (0.02s) is a group. Different fault type labels {y=1,2,3,4,5} are assigned to obtain the target domain database.

[0074] S1-3. Resample the target domain data to maintain the consistency of the feature numbers of the source domain and target domain samples, and normalize the source domain and target domain data.

[0075] S2. Divide the test set and training set.

[0076] All source domain data is used for transfer learning model training, that is, To simulate the real situation, 500 sets of sample data (without labels) are extracted from the normal operation data of the target domain and added to the training set for feature transfer learning, which is expressed as In addition, 8 sets of labeled data are extracted from each fault type and added to the training set, which is expressed as The remaining target domain data are used to test the performance of the diagnostic model. The training set data is represented as in,

[0077] S3, based on training set data By combining the WBDA algorithm and the RVFL algorithm, an IGBT open-circuit fault diagnosis model suitable for high-power traction rectifiers is trained.

[0078] The training process is as shown in steps S3 and S4 in the previous part of the manual. The kernel type of the WBDA algorithm is selected as the rbf kernel, and the activation function of the RVFL algorithm is the hardlim function. After K-fold cross validation and Bayesian hyperparameter optimization, the optimal parameters of the WBDA algorithm and the RVFL algorithm are obtained: kernel bandwidth gamma = 1.1804, regularization parameter λ = 4.4638, number of iterations T = 15, adapted dimension dim = 30, balance factor μ = 0.53467, and the number of hidden layers of the RVFL algorithm Number of Hidden Neurons = 395. The WBDA+RVFL classification model is obtained by training, and the target domain training data is tested to obtain a training accuracy of 100%. The confusion matrix of the training set test results is as follows: Figure 2 shown.

[0079] S4. Use the test set data to verify the WBDA+RVFL classification model obtained in the previous step and obtain the diagnosis result of the model in the test set. The confusion matrix of the diagnosis result is as follows: Figure 3 As shown, the total number of test set samples is 10,000 and the test result accuracy is 96.95%.

Claims

1. A traction rectifier open circuit fault diagnosis method based on feature transfer learning, characterized in that: The following steps are involved: S1. Data preparation; S2, divide the test set and training set; All source domain data are used for transfer learning model training; except for some labeled target domain data, the rest of the target domain data are used to test the performance of the diagnostic model; S3. Based on the training set data, combined with the weighted balanced distribution adaptive WBDA algorithm and the random vector function chain network RVFL algorithm, an IGBT open circuit fault diagnosis model suitable for high-power traction rectifier is trained; S4, use Bayesian optimization to optimize the hyperparameters of the training model in S3; S5. Retrain step S3 using the optimal parameters obtained in step S4 to obtain a WBDA+RVFL classification model, and save the model parameters; verify using the test set data to obtain the diagnostic result of the model in the test set.

2. A traction rectifier open circuit fault diagnosis method based on feature transfer learning according to claim 1, characterized in that: S1 specifically includes the following sub-steps: S1-1. Build a low-power single-phase two-level pulse rectifier model, collect grid-side current data under normal conditions and after the IGBT modules T1, T2, T3, and T4 are open-circuited. The grid-side current data of each fundamental wave cycle (0.02s) is grouped together, and different fault type labels {y=1, 2, 3, 4, 5} are assigned to obtain the source domain database. S1-2. Establish a high-power traction rectifier model, collect grid-side current data under normal conditions and after the IGBT modules T1, T2, T3, and T4 are open-circuited. The grid-side current data of each fundamental wave period (0.02s) is grouped together, and different fault type labels {y=1,2,3,4,5} are assigned to obtain the target domain database. S1-3. Use resampling technology to ensure the consistency of the feature numbers of target domain samples and source domain samples, and normalize the source domain and target domain data.

3. The method for diagnosing open-circuit fault of a traction rectifier based on feature transfer learning according to claim 1, characterized in that: The specific method of S2 is to simulate the real data situation and extract The sample data of each group is added to the training set. For each fault type of fault data, only k groups are extracted and added to the training set, which is expressed as The training set data is represented as 4. The method for diagnosing open-circuit fault of a traction rectifier based on feature transfer learning according to claim 1, characterized in that: S3 specifically includes the following sub-steps: S3-1, load training set data X and training set label Y; S3-2, set the parameters of the WBDA algorithm and the RVFL algorithm, including the kernel bandwidth gamma, regularization parameter λ, kernel type kernel_type, number of iterations T, adapted dimension dim, balance factor μ of the WBDA algorithm, and the number of hidden layers Number of Hidden Neurons and activation function type of the RVFL algorithm; S3-3, perform model initialization, initialize the centralization matrix H and the edge distribution distance matrix M0; H=I-(1 / n)1(1) In the formula, m represents the number of samples in the training set X, and n represents the number of features of each sample. is the identity matrix; S3-4. Use the RVFL algorithm to perform preliminary classification on the target domain training data. The training input data of the RVFL algorithm is the labeled source domain and target domain data, and obtain the target domain data pseudo label y_pseudo; In RVFL, by The initial input data X = [x1, x2, ..., x N ] is mapped to the hidden layer, where g(.) is a nonlinear activation function, ω j is the weight of the jth hidden node, b j is the bias; combining the direct nodes and the hidden nodes, the RVFL network is expressed as: Where J is the total number of hidden nodes, N is the total number of input nodes, and β j is the jth weight term; the training of the RVFL network is to minimize the squared output error (ε by optimizing the optimal hidden layer weights and output layer weights 2 ): Among them B t It consists of (J+N) weight values, where P is the total number of training samples, t is the target vector, d is the vector including the output node, and (p) is the sample index; the weights between the input layer and the hidden layer neurons are randomly sampled from a uniform distribution [-1, 1], and then the weights between the hidden layer and the output layer neurons are optimized by the least squares method; Given a training sample X, the multi-output nodes of RVFL are f(X)=[f1(x),...,f E (x)], where E is the number of output nodes, then the general decision function of the RVFL classifier with multiple outputs is: S3-5. Use the target domain pseudo label y_pseudo and the source domain true label Initialize the conditional distribution distance matrix M c : S3-6. Set the balance factor μ and calculate the total distribution distance matrix M: M=(1-μ)×M0+μ×M c 0≤μ≤1 (7) S3-7. Consider minimizing the distribution distance between the source domain and the target domain data, and obtain the formula: Where tr(·) represents the trace of the matrix, and the A matrix is ​​the characteristic migration matrix, which is a Hermitian matrix; S3-8. Considering the divergence of the data before and after the migration of the constraint feature, the formula is obtained: S3-9. Combining equations (8) and (9), adding a regularization term with a coefficient of λ>0, we get the final learning goal of transfer learning: S3-10. Solve equation (10) using the Lagrangian method. Its Lagrangian function is expressed as: Let its derivative have to Where φ is the Lagrange multiplier; Before calculating the matrix A, select the appropriate kernel type to put the sample X into high-dimensional mapping, and obtain the mapping matrix K to replace X to solve the feature migration matrix A; The matrix A is solved using the eigs function in matlab; S3-10. Find the sample feature matrix Z after feature migration, which contains the source domain data Z after feature migration S and target domain data Z T : S3-11. Using source domain data Z after feature migration S and target domain data Z T , combined with known labels, train the RVFL algorithm, classify the target domain training data, and obtain the target domain fault type diagnosis result y_tar_pseudo; S3-12. Repeat steps 3-5 to 3-11 according to the number of iterations T to obtain the classification label results of the target domain data of the final training set, compare them with the true value, and obtain the model training accuracy.

5. The method for diagnosing open-circuit fault of a traction rectifier based on feature transfer learning according to claim 1, characterized in that: The objects of hyperparameter optimization in S4 are: kernel bandwidth gamma, regularization parameter λ, adapted dimension dim, balance factor μ of WBDA algorithm, and number of hidden layers Number of Hidden Neurons of RVFL algorithm; It includes the following sub-steps: S4-1. Define the parameter range that needs to be optimized, including gamma, λ, dim, μ and Number of Hidden Neurons; S4-2, Bayesian Optimization: Use the bayesopt function in MATLAB to perform Bayesian parameter optimization, the goal is to minimize the output of the K-fold cross validation function; the K-fold cross validation function takes K=3, and calculates the average accuracy and standard deviation of the training model in step 3; the optimization goal is to maximize the mean of the accuracy while controlling the size of the standard deviation, improving the accuracy of the training model while ensuring the stability of the training results; S4-3. Obtaining optimal parameters: Extract the optimal parameters from the Bayesian optimization results and save the values ​​of the optimal parameters.

Images