A method and system for open circuit fault diagnosis of a train traction converter

Abstract

The present application relates to the field of rail train traction system circuit fault diagnosis, and discloses a kind of open-circuit fault diagnosis method and system of train traction converter, the method comprises: build orthogonal auto-encoder to construct the association between sensor sampling data Association graph, construct orthogonal projection loss function to guide the encoder to fit out orthogonal mapping, improve the explainability of encoder;Build the association relationship graph optimization model based on nearest neighbor optimization strategy, delete the neighbor node far away to accurately describe the association between data, and prevent model overfitting;Association relationship graph is converted into association relationship matrix, constructs graph convolution network, uses association relationship matrix in network forward propagation, trains graph convolution network, and carries out fault diagnosis based on the network after training is completed. Supervised association relationship graph construction can be realized, and the fault diagnosis accuracy of traction motor is improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of traction system circuits in rail trains, and particularly to a method and system for diagnosing open-circuit faults in train traction converters. Background Technology

[0002] In complex systems, with advancements in technology and increasing societal demands, industrial instruments are gradually becoming more complex and larger-scale. The growing sophistication of various industrial systems has greatly promoted social progress and development. However, the various components of these instruments often operate for extended periods in complex and harsh environments, exposed to factors such as high temperature and humidity. If not maintained or replaced in a timely manner, malfunctions can easily occur, affecting the overall healthy and stable operation of the instrument. Therefore, fault diagnosis in complex systems has always been a major research focus.

[0003] In recent years, thanks to the increased computing speed of computers and the enhanced data acquisition capabilities of sensors, various neural network models have become a focus of attention. Considering the powerful feature extraction capabilities and convenient end-to-end characteristics of neural network models, researchers have begun to focus on combining neural networks with fault diagnosis problems. However, most common neural networks process regular Euclidean structured data such as image data and one-dimensional time series data. On the other hand, non-Euclidean structured graph data is widely present in industrial systems. In this type of data structure, the nodes are arranged in an unordered manner, and the number of neighbors of different nodes is usually different, which poses a challenge to neural network-based fault diagnosis models.

[0004] Currently, fault diagnosis methods based on graph convolutional networks can construct relationship graphs to record the correlations between sampled data, enabling the analysis and processing of non-Euclidean graph data. These methods exhibit good diagnostic performance and wide applicability. However, common methods for constructing relationship graphs rely on analyzing data similarity. Since the similarity between nodes is often measured using Euclidean distance or cosine similarity, and the feature vectors of each node are composed of different types of features with varying domains and distributions, and complex coupling relationships between features, simple Euclidean distance or cosine similarity cannot accurately reflect the similarity between nodes. Furthermore, traditional relationship graph construction methods operate under unsupervised learning, meaning parameters (such as the number of neighbors for each node) cannot be optimized in a supervised manner, directly impacting the accuracy of the relationship graph.

[0005] Therefore, there is an urgent need to provide a fault diagnosis method and system based on graph convolutional networks with high accuracy of relational graphs. Summary of the Invention

[0006] This invention provides a fault diagnosis method and system for train traction motors, which solves the problem that the accuracy of current common correlation graphs is difficult to guarantee, leading to difficulties in the practical application of fault diagnosis methods based on graph convolutional networks.

[0007] To solve the above-mentioned technical problems, the present invention proposes the following technical solution:

[0008] In a first aspect, this application provides a method for diagnosing faults in a train traction motor, comprising:

[0009] S1: Collect the original fault diagnosis dataset of the traction motor and determine the possible fault types of the traction motor based on the industrial field conditions of the traction motor.

[0010] S2: The dataset is preprocessed using the sliding window method, and the dataset is divided into a training set, a test set, and a validation set;

[0011] S3: Construct an orthogonalized autoencoder;

[0012] S4: Construct the orthogonal projection loss function of the autoencoder, and guide the encoder to fit an orthogonal mapping based on the loss function to construct the association graph between data within the dataset;

[0013] S5: Build an association graph optimization model based on the nearest neighbor optimization strategy;

[0014] S6: Transform the relationship graph optimization model into a relationship matrix, use the relationship matrix and the dataset as input, construct a graph convolutional neural network, use the relationship matrix in the network forward propagation formula, and use the training set to iteratively train the graph convolutional network until it converges.

[0015] S7: In the graph convolutional neural network that has been iteratively converged based on the test set and the validation set, if the output of the graph convolutional neural network meets the set requirements, it is considered to be successfully trained.

[0016] S8: Input the real-time operating data of the traction motor into the successfully trained graph convolutional neural network for fault diagnosis.

[0017] In a second aspect, the present invention also provides a fault diagnosis system for a train traction motor, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0018] The present invention has the following beneficial effects:

[0019] This invention provides a fault diagnosis method for train traction motors. It constructs an orthogonalized autoencoder to build a correlation graph between sensor-sampled data, and constructs an orthogonal projection loss function to guide the encoder in fitting an orthogonal mapping, improving the encoder's interpretability. A correlation graph optimization model based on a nearest neighbor optimization strategy is built to remove distant neighbor nodes to accurately describe the correlation between data and prevent overfitting. The correlation graph is transformed into a correlation matrix, and a graph convolutional network is constructed. The correlation matrix is ​​used during the network's forward propagation to train the graph convolutional network. The network output after training is used as the fault diagnosis prediction result. This invention can diagnose traction motor faults, including motor demagnetization faults, inter-turn short circuit faults, and bearing faults, improving the accuracy and versatility of the diagnostic method and effectively reducing the difficulty and cost of motor maintenance.

[0020] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0022] Figure 1 This is one of the flowcharts of a fault diagnosis method for a train traction motor according to a preferred embodiment of the present invention;

[0023] Figure 2 This is a second flowchart illustrating a preferred embodiment of the fault diagnosis method for a train traction motor of the present invention.

[0024] Figure 3 This is a comparison diagram of the fault diagnosis results of a preferred embodiment of the present invention and the prior art.

[0025] Figure 4 This is a comparison diagram showing the impact of the fault diagnosis method of the preferred embodiment of the present invention and different relationship graphs obtained by the prior art on the dataset in the graph convolutional network model. Detailed Implementation

[0026] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0028] It should be understood that the fault diagnosis method for train traction motors in this application can be applied to rail train scenarios; this is only an example and not a limitation.

[0029] Please see Figures 1-2 This application provides a fault diagnosis method for train traction motors, including:

[0030] S1: Collect the original fault diagnosis dataset of the traction motor and determine the possible fault types of the traction motor based on the industrial field conditions of the traction motor.

[0031] S2: The dataset is preprocessed based on the sliding window method, and the dataset is divided into training set, test set and validation set;

[0032] S3: Construct an orthogonalized autoencoder;

[0033] S4: Construct the orthogonal projection loss function of the autoencoder, and guide the encoder to fit the orthogonal mapping based on the loss function to construct the association graph between the data within the dataset;

[0034] S5: Build an association graph optimization model based on the nearest neighbor optimization strategy;

[0035] S6: Transform the graph optimization model into a relation matrix, use the relation matrix and dataset as input, construct a graph convolutional neural network, use the relation matrix in the network's forward propagation formula, and use the training set to iteratively train the graph convolutional network until it converges.

[0036] S7: In the graph convolutional neural network that verifies iterative convergence based on the test set and validation set, if the output of the graph convolutional neural network meets the set requirements, it is considered to be successfully trained.

[0037] S8: Input the real-time operating data of the traction motor into the successfully trained graph convolutional neural network for fault diagnosis.

[0038] In this implementation, there are four types of faults, including:

[0039] A. Motor demagnetization fault caused by long-term operation of the traction motor; the deviation between the motor current waveform under fault conditions and the normal condition gradually increases over time. B. Inter-turn short circuit fault caused by motor overheating or insulation layer peeling. C. Bearing failure caused by bearing wear or lack of lubrication. D. Motor in a normal, fault-free state.

[0040] The aforementioned fault diagnosis method for train traction substations involves constructing an orthogonalized autoencoder to build a correlation graph between sensor sampling data, and constructing an orthogonal projection loss function to guide the encoder in fitting an orthogonal mapping, thereby improving the encoder's interpretability. A correlation graph optimization model based on a nearest neighbor optimization strategy is built to remove distant neighbor nodes to accurately describe the correlation between data and prevent overfitting. The correlation graph is transformed into a correlation matrix, and a graph convolutional network is constructed. The correlation matrix is ​​used during the network's forward propagation to train the graph convolutional network. The output of the trained network is used as the fault diagnosis prediction result. This invention can diagnose traction motor faults, including motor demagnetization faults, inter-turn short circuit faults, and bearing faults, improving the accuracy and versatility of the diagnostic method and effectively reducing the difficulty and cost of motor maintenance.

[0041] It should be noted that the dSPACE hardware-in-the-loop simulation platform for fault diagnosis research and experimentation of rail train traction systems proposed in this application can simulate various fault types of traction motors and collect corresponding motor data. The platform consists of a test box, a PXI box, a cRIO chassis, a real-time simulation unit, and a host computer. The signal conditioning system, composed of the test box, PXI box, and cRIO box, realizes signal conversion between the device under test and the real-time simulator. The real-time simulation unit performs simulation operations and provides sensor signal sampling values ​​to external devices.

[0042] Specifically, the above steps are described in detail below:

[0043] The preprocessing of the dataset based on the sliding window method includes: segmenting the original fault diagnosis dataset into fragmented datasets X = {x1, x2, ..., xn} ∈ R by setting the window length d and the sliding step size s. n×d Where d = s, n represents the total number of data points, and R represents the set of all real numbers.

[0044] It is important to note that to prevent potential data leakage, d should be set to s to ensure that there is no overlap between data segments. Next, the dataset is randomly divided into training and training sets according to a certain ratio. test set and validation set Where n = n1 + n2 + n3. In the three datasets above, only the training set contains label information in its corresponding one-hot encoded form. c represents the number of fault categories.

[0045] In this embodiment, the forward propagation formula of the orthogonalized autoencoder structure includes:

[0046] X fea =FC2(FC1(X)) (1);

[0047]

[0048]

[0049] Where X is a fragmented dataset obtained through sliding window data preprocessing; X fea ∈R n×r r is the length of the node embedding vector, which can be set according to the actual dataset; FC1(·) and FC2(·) are fully connected layers with activation functions, used for feature extraction; MLP(·) is a neural network module composed of multiple fully connected layers;

[0050] The training objective of the orthogonalized autoencoder proposed in this invention is not to obtain the reconstructed feature matrix. Instead, it is the node embedding matrix Z emb It can obtain the predicted adjacency matrix A' through simple matrix multiplication.

[0051] A'=σ(Z emb Z emb T (4);

[0052] Among them, Z emb T For matrix Z emb The rank transformation is performed. Since the elements in the standard adjacency matrix should only contain 0 and 1, this method uses the sigmoid function as the activation function to normalize all output values. σ represents the sigmoid activation function.

[0053] In this embodiment, based on the structure of the orthogonalized autoencoder and its forward propagation formula, the loss function of the autoencoder consists of three parts and four sub-loss functions:

[0054] A. The deviation between the mapping relationship of the current autoencoder model and the mapping relationship of the orthogonal projection form.

[0055]

[0056]

[0057] Reconstruction loss of autoencoder model

[0058]

[0059] Loss function of association graph in supervised learning mode

[0060]

[0061] Among them, Y GC Labels for the association graph constructed from the training set

[0062]

[0063] The above formula shows that for two nodes belonging to the same part of the training set, if they have the same label, then their position in Y... GC The value of the corresponding element in Y is 1, and it is 0 if it has a different label; while the value of the other nodes in Y is 0. GC The value of the corresponding element in A' is the same as that in A'. This special design method is to ensure that the loss function is constructed by accumulating only the loss values ​​of nodes within the training set.

[0064] In this embodiment, the optimization objective during the backpropagation and iterative optimization process of the autoencoder is the sum of the products of each loss function and its corresponding weight coefficient, including:

[0065] L OAE =α1Loss1+α2Loss2+α3Loss mse +α4Loss GC (10);

[0066] Here, α1 to α4 represent the weight coefficients corresponding to the four loss functions. By setting the values ​​of each weight coefficient, the influence of each loss function on the iterative optimization process can be adjusted.

[0067] After completing iterative training, the trained autoencoder outputs a normalized dataset association matrix prediction value A'. This matrix is ​​then thresholded to obtain a standard adjacency matrix A consisting only of 0 and 1 values. OAE :

[0068]

[0069] Among them, J GC This is the preset threshold.

[0070] In this embodiment, the adjacency matrix A obtained by the orthogonalized autoencoder... OAEGiven a node, all its neighbors are selected, and only those neighbors that are close enough are retained, while other nodes that are far away are discarded.

[0071]

[0072] Wherein KNN(k,x) i Using the classic K-nearest neighbor algorithm, calculate the distance to node x. i The set of the k nearest nodes.

[0073] In this embodiment, the graph convolutional network takes the adjacency matrix A, obtained by sequentially passing it through an autoencoder and an association graph optimization model, as input to aggregate data and extract features from the fragmented dataset X, obtaining their feature vectors. Then, several fully connected layers and a normalized Softmax layer are connected to the graph convolutional network to obtain the diagnostic results for the motor fault type.

[0074] The forward propagation formula for graph convolutional networks is:

[0075] Z GCN =GCN l ...(GCN1(X,A)) (13);

[0076] Y GCN =Softmax(FC2(FC1(Z) GCN ))) (14);

[0077] Where GCN(·) is a graph convolutional layer, its forward propagation formula is:

[0078]

[0079] in, σ is the activation function, W∈R n×c It is a parameter matrix.

[0080] In this example, the training loss function of the graph convolutional network includes:

[0081]

[0082] Among them, considering that only the training set portion X train There is a corresponding label information Y train Therefore, when calculating the loss function, only the cross-entropy loss function of the training set portion is accumulated; Y train Y represents the fault type corresponding to each data point in the training set. GCN The result obtained by formula (14) represents the output result of the graph convolutional network for the motor fault type.

[0083] The present invention also provides a fault diagnosis system for a train traction motor, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the above embodiments.

[0084] The method of this invention uses a small dataset as the training set to train a graph convolutional network (GCN), and verifies its accuracy on a test set to judge the effectiveness of the GCN. To comprehensively verify the effectiveness of the method of this invention in traction motor fault diagnosis, it is compared with different existing technologies. Specifically, KNN-GCN and Radius-GCN combinations are used to verify the effectiveness of the graph construction method. Furthermore, a CNN model is selected as a representative of classic neural network models. Finally, a graph attention network model is used to replace the graph convolutional network in this method. In summary, the method of this invention is compared with four existing methods, and the results are as follows: Figure 3 As shown in the figure, the method used in this invention has the highest diagnostic accuracy.

[0085] Furthermore, a comparison diagram showing the impact of different relationship graphs obtained using the fault diagnosis method of the preferred embodiment of the present invention and existing technologies on the dataset in a graph convolutional network model is shown below. Figure 4 As shown. First, based on the orthogonalized autoencoder and the nearest neighbor-based graph optimization model proposed in this application, the adjacency matrix A1 within the dataset is obtained; then, the corresponding adjacency matrices A2 and A3 are obtained through two traditional graph construction methods based on KNN and Radius, respectively. The dataset X is then multiplied by the normalized adjacency matrix to obtain three dataset feature matrices for visualization.

[0086]

[0087] Among them, A nor This is the set of matrices obtained after normalizing three adjacency matrices. Then, the data dimensionality reduction and visualization method TSNE is used to display the original dataset X and the three dataset feature matrices used for visualization, resulting in... Figure 4 ,in, Figure 4 (a) is a visualization of the original dataset. Figure 4 (b) is the dataset Visual diagram, Figure 4 (c) is the dataset Visual diagram, Figure 4 (d) is the dataset A visual diagram. This diagram intuitively reflects the impact of the relationship graphs obtained by the three methods on the original dataset in the GCN model. The node colors in the diagram represent different fault types corresponding to the data. As can be seen from the diagram, the dataset feature matrix obtained by the method proposed in this invention... There is a clear clustering tendency; nodes of the same color in the graph tend to be closer to each other, which helps graph convolutional network models to perform data classification tasks well.

[0088] In summary, the fault diagnosis method and system for train traction transformers of this invention introduces an encoder-decoder network and uses a neural network to iteratively optimize the relationship graph, ensuring its accuracy. Then, a nearest neighbor-based graph optimization method is built to further optimize the relationship graph, while also preventing overfitting of the graph convolutional network. This invention proposes a supervised learning-based method for constructing data relationship graphs, improving the accuracy of traction motor fault diagnosis methods based on graph convolutional networks. It can diagnose traction motor faults including motor demagnetization faults, inter-turn short circuit faults, and bearing faults, improving the versatility of the diagnostic method and effectively reducing the difficulty and cost of motor maintenance.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of diagnosing a failure of a train traction motor, characterized by, include: S1: Collect the original fault diagnosis dataset of the traction motor and determine the possible fault types of the traction motor based on the industrial field conditions of the traction motor. S2: The dataset is preprocessed using the sliding window method, and the dataset is divided into a training set, a test set, and a validation set; S3: Construct an orthogonalized autoencoder; S4: Construct the orthogonal projection loss function of the autoencoder, and guide the encoder to fit an orthogonal mapping based on the loss function to construct a correlation graph between data within the dataset; S5: Build an association graph optimization model based on the nearest neighbor optimization strategy; S6: Transform the relationship graph optimization model into a relationship matrix, use the relationship matrix and the dataset as input, construct a graph convolutional neural network, use the relationship matrix in the network forward propagation formula, and use the training set to iteratively train the graph convolutional network until it converges. S7: In the graph convolutional neural network that has been iteratively converged based on the test set and the validation set, if the output of the graph convolutional neural network meets the set requirements, it is considered to be successfully trained. S8: Input the real-time operating data of the traction motor into the successfully trained graph convolutional neural network for fault diagnosis; The forward propagation formula of the orthogonalized autoencoder includes: （1）； （2）； （3）； wherein, is a fragmented dataset , is the length of the node embedding vector, and is a fully connected layer with an activation function, is a neural network module composed of multiple fully connected layers; The training objective of the orthogonalized autoencoder is to obtain a node embedding matrix ; The method further includes: based on the node embedding matrix satisfying the following relationship: （4）； in, For matrix The change of order, This represents the sigmoid activation function.

2. The fault diagnosis method for train traction motors according to claim 1, characterized in that, The preprocessing of the dataset based on the sliding window method includes: By setting the window length and sliding step size The original fault diagnosis dataset is segmented to obtain fragmented datasets. The following relationship must be satisfied: ； in, , Denotes the set of all real numbers. Indicates the total number of data sets; The dataset is randomly divided into training and training sets according to a certain proportion. test set and verification set ,in ; In the three datasets above, the training set data contains label information in the form of corresponding one-hot encoding. The following relationship must be satisfied: ； In the formula, c represents the number of fault categories.

3. The fault diagnosis method for train traction motors according to claim 1, characterized in that, The orthogonal projection loss function includes: A. The deviation between the mapping relationship of the current autoencoder model and the mapping relationship of the orthogonal projection form is expressed as follows: （5） （6） The reconstruction loss of the autoencoder model is expressed as follows: （7） The loss function for the association graph in supervised learning is as follows: （8） in, express, The labels of the association graph constructed from the training set satisfy the following relational expression; （9） For two nodes belonging to the same training set, if they have the same label, then in The value of the corresponding element in the middle is 1, and it is 0 if it has different labels, while the value of the other nodes is 0. The values ​​of corresponding elements and the adjacency matrix The same as in.

4. The fault diagnosis method for train traction motors according to claim 1, characterized in that, The step of guiding the encoder to fit an orthogonal mapping based on the loss function includes: The sum of the products of each loss function and its corresponding weight coefficient is used as the optimization objective. This objective enables the autoencoder to undergo backpropagation and iterative optimization to obtain an orthogonal mapping. The optimization objective satisfies the following relationship: （10）； in, ~ These represent the weight coefficients corresponding to the four loss functions.

5. The fault diagnosis method for train traction motors according to claim 1, characterized in that, Obtain the predicted adjacency matrix The method further includes: determining a standard adjacency matrix consisting only of 0 and 1 values ​​using a preset threshold and an adjacency matrix. It satisfies the following relationship: （11）； in, This is the preset threshold.

6. The fault diagnosis method for train traction motors according to claim 1, characterized in that, The construction of the association graph optimization model based on the nearest neighbor optimization strategy includes: For the adjacency matrix obtained by the orthogonal autoencoder Given a node in the graph, calculate its distance to its neighboring nodes. Then, sort the nodes in descending order and remove any neighboring nodes whose distance exceeds a set threshold to optimize the graph, satisfying the following relationship: （12）； in, This indicates that the K-nearest neighbor algorithm is used to calculate the distance between the nodes. closest A set of nodes.

7. The fault diagnosis method for train traction motors according to claim 1, characterized in that, S6 includes: The adjacency matrix obtained by sequentially passing through an autoencoder and an association graph optimization model As input to a graph convolutional network, the fragmented dataset Data aggregation and feature extraction are performed to obtain their feature vectors. Then, several fully connected layers and normalized Softmax layers are connected to the graph convolutional network to obtain the output results of the graph convolutional network for motor fault types. The training loss function of the graph convolutional network is as follows: （13）； Among them, considering only the training set portion There is corresponding tag information. Therefore, when calculating the loss function, only the cross-entropy loss function of the training set portion is accumulated; To determine the fault types corresponding to each data point in the training set, The result obtained by formula (14) represents the output result of the graph convolutional network for the motor fault type.

8. A fault diagnosis system for a train traction motor, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images