Fault Diagnosis Method, System and Medium for Train Traction System under Multiple Service Conditions

Abstract

The present invention relates to the technical field of fault diagnosis of rail train traction systems, and discloses a fault diagnosis method, system and medium for train traction systems under multiple service conditions; this method trains a pre-constructed fault diagnosis model framework to obtain a target multi-channel domain adaptive graph convolutional network model, and performs fault diagnosis based on the target multi-channel domain adaptive graph convolutional network model. In this way, based on the fault diagnosis model framework, not only can data features be extracted, but also the data correlation is emphasized, which can significantly reduce the number of required fault samples to a certain extent, solve the problems of the diversity of service conditions and the scarcity of fault samples, provide a feasible approach for fault diagnosis of traction systems under small samples, and can timely diagnose the minor faults of rail train traction systems.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of rail train traction systems, and particularly to a fault diagnosis method, system and medium for train traction systems under multiple service conditions. Background Art

[0002] With the long-term operation of rail trains, some faults are likely to occur in the key components of their traction systems as the service time increases. If not repaired in time, it is very likely to reduce the working performance of the train and even endanger the safety of passengers. Therefore, the design of fault diagnosis systems for rail trains has increasingly become the focus of research. Among them, the electric traction system (referred to as the traction system for short) is the main power source of rail trains and the core subsystem for the stable operation of rail trains. In some EMUs, the traction system is mainly composed of high-voltage electrical appliances (traction transformers, pre-charge circuits, etc.), traction converters (rectifiers, intermediate DC links, inverters), traction motors and traction control units in cascade. At the same time, the traction system is one of the subsystems with the highest frequency of faults. Each component of the traction system is prone to aging, damage, etc. as the operation time increases, ultimately resulting in the inability of the traction system to work properly. Due to the fault-oriented safety criterion of rail trains, the train is in a safety protection state for a long time, reducing the working efficiency. If a fault occurs in the train, it must be stopped for repair. Therefore, the fault data collection time is short, resulting in few fault data samples. At the same time, during the actual operation of rail trains, the traction system is mostly in normal service conditions. Even when a fault occurs, only a logical judgment of whether a fault has occurred is made. This is because of the lack of real-time and effective fault diagnosis strategies, making it difficult to obtain fault type labels. All these lead to the problem of small samples in the fault diagnosis of rail trains.

[0003] At the same time, the multiple service conditions under which the traction system operates are the main reason for the limited generalization performance of the diagnosis model. The working states of the traction system include the starting and running stages. When working in the starting stage, the train is still in the stopped state. When in the running stage, the traction system experiences different service conditions such as acceleration, coasting, conventional braking deceleration, deceleration or stopping under fault conditions. Due to the time-varying nature of the service conditions of the traction system operation and the complexity of the operating environment, the correlation of the observed variables changes with the change of the service conditions, causing the parameter relationship of the original correlation model to change, resulting in complex fault diagnosis and poor generalization performance of the traction system under small samples.

[0004] Currently, traditional modeling methods can establish mechanism models between the devices of the traction system, but they also ignore the correlation between strongly coupled devices. Therefore, it is difficult to diagnose minor faults. In addition, due to the existence of certain structural characteristics determined by the electrical connection method in the correlation between system devices, it is impossible to establish an integrated model by simply superimposing between devices. It can be seen that in the prior art, it is difficult to diagnose minor faults of the rail train traction system in a timely manner. Summary of the Invention

[0005] The present invention provides a fault diagnosis method, system and medium for a train traction system under multiple service conditions, so as to solve the problem in the prior art that it is difficult to diagnose minor faults of the rail train traction system in a timely manner.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] In the first aspect, the present invention provides a fault diagnosis method for a train traction system under multiple service conditions, including:

[0008] S1: Collect the original three-phase current data of the traction system under different service conditions and different fault states, and add fault type labels to the original three-phase current data;

[0009] S2: Preprocess the original three-phase current data to obtain a fault diagnosis data set, and divide the fault diagnosis data set into a source domain data set and a target domain data set according to the service condition;

[0010] S3: Input the source domain data set and the target domain data set into a pre-constructed fault diagnosis model framework, use the overall loss function as the network training index, and train based on the backpropagation algorithm to obtain a target channel domain adaptive graph convolutional network model;

[0011] S4: Input the three-phase current data of the traction system to be diagnosed into the target channel domain adaptive graph convolutional network model for fault diagnosis.

[0012] In the second aspect, the present application provides a fault diagnosis system for a train traction system under multiple service conditions, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect are implemented.

[0013] In the third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method steps described in the first aspect are implemented.

[0014] Advantageous Effects:

[0015] The fault diagnosis method for the train traction system under multiple service conditions provided by the present invention trains a target multi-channel domain adaptive graph convolutional network model from a pre-constructed fault diagnosis model framework, and performs fault diagnosis based on the target multi-channel domain adaptive graph convolutional network model. In this way, based on the fault diagnosis model framework, not only can data features be extracted, but also the data correlation is emphasized, which can significantly reduce the required number of fault samples to a certain extent, solve the problems of the diversity of service conditions and the scarcity of fault samples, provide a feasible approach for the fault diagnosis of the traction system under small samples, and can timely diagnose the minor faults of the rail train traction system.

[0016] In a preferred example, for the fault diagnosis method for the train traction system under multiple service conditions provided by the present invention, in the designed graph generation layer, two graph construction schemes are proposed. One is dynamic graph construction based on a multi-layer perceptron, and the other is fixed graph construction based on the K-nearest neighbor algorithm. The two graph construction methods perform data graph construction from different perspectives, increase the channels for node information propagation, and thus enhance the diversity of graph topology information. The graph construction process not only does not rely on mathematical modeling and prior knowledge, but also can adapt to the characteristics of the changes in the associated attributes of nodes, edges, etc. of the graph network model brought about by different service conditions and system characteristics, which can ensure the accuracy of the diagnosis result.

[0017] In a preferred example, for the fault diagnosis method for the train traction system under multiple service conditions provided by the present invention, in the designed graph information fusion layer, a multi-graph fusion method is adopted to reduce the parallel calculation amount. At the same time, the fused graph information includes the specific features and shared features of two graphs, and the consistency constraint and difference constraint are used to ensure the consistency and difference of the above feature representations, so as to improve the diversity of the node information propagated on the graph topology structure.

[0018] In a preferred example, for the fault diagnosis method for the train traction system under multiple service conditions provided by the present invention, in the process of multi-graph fusion, an adaptive update attention mechanism is adopted to learn the weights of different output feature representations, so as to judge which of the two graph construction methods is more representative, and thus select the most reasonable topological graph to describe the data correlation, thereby improving the diagnosis accuracy. Description of the Drawings

[0019] Figure 1 It is a flowchart of the fault diagnosis method for the train traction system under multiple service conditions of the preferred embodiment of the present invention;

[0020] Figure 2It is the three-phase current data measurement waveform of the traction converter of the rail train in the preferred embodiment of the present invention. Among them, (a) represents the three-phase current measurement waveform of the traction system converter when the motor speed is 1000 r / min; (b) represents the three-phase current measurement waveform of the traction system converter when the motor speed is 1800 r / min;

[0021] Figure 3 It is the traction system fault diagnosis model framework based on the multi-channel domain adaptive graph convolutional network in the preferred embodiment of the present invention;

[0022] Figure 4 It is the two-dimensional convolutional network structure in the preferred embodiment of the present invention;

[0023] Figure 5 It is the schematic diagram of the topological graph in the preferred embodiment of the present invention;

[0024] Figure 6 It is the graph convolutional network structure in the preferred embodiment of the present invention;

[0025] Figure 7 It is the fault diagnosis accuracy rate under different migration service conditions in the preferred embodiment of the present invention;

[0026] Figure 8 It is the schematic diagram of visualizing the features of the last layer of the feature extractor by t-SNE in the preferred embodiment of the present invention. Detailed implementation manners

[0027] The technical solutions of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without any creative work belong to the scope of protection of the present invention.

[0028] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, the terms such as "a" or "one" do not indicate a quantity limitation, but indicate that there is at least one. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0029] Please refer to Figure 1, an embodiment of the present application provides a fault diagnosis method for a train traction system under multiple service conditions, including:

[0030] S1: Collect the original three-phase current data of the traction system under different service conditions and different fault states, and add fault type labels to the original three-phase current data;

[0031] S2: Preprocess the original three-phase current data to obtain a fault diagnosis data set, and divide the fault diagnosis data set into a source domain data set and a target domain data set according to the service condition;

[0032] S3: Input the source domain data set and the target domain data set into a pre-constructed fault diagnosis model framework, use the overall loss function as the network training index, and train based on the backpropagation algorithm to obtain a target channel domain adaptive graph convolutional network model;

[0033] S4: Input the three-phase current data of the traction system to be diagnosed into the target channel domain adaptive graph convolutional network model for fault diagnosis.

[0034] In this embodiment, when collecting the original three-phase current data of the traction system under different service conditions, data such as motor speed and load torque under different conditions can be collected. Among them, data collection can be realized through high-precision sensors. In this embodiment, the measurement waveforms of the original three-phase current data of the traction system collected by high-precision sensors under different service conditions and different fault states are as Figure 2 shown.

[0035] In an example, after inputting the source domain data set and the target domain data set into a pre-constructed fault diagnosis model framework, using the overall loss function as the network training index, and training based on the backpropagation algorithm to obtain a target channel domain adaptive graph convolutional network model, the target domain test data set can also be input into the target channel domain adaptive graph convolutional network model for testing. When the accuracy rate of the test result meets the preset threshold, it is considered that the correctness of the target channel domain adaptive graph convolutional network model is relatively high.

[0036] For the above-mentioned fault diagnosis method for a train traction system under multiple service conditions, a target multi-channel domain adaptive graph convolutional network model is obtained by training a pre-constructed fault diagnosis model framework, and fault diagnosis is performed based on the target multi-channel domain adaptive graph convolutional network model. In this way, based on the fault diagnosis model framework, not only can data features be extracted, but also data correlation is emphasized. To a certain extent, the required number of fault samples can be significantly reduced, the problem of the diversity of service conditions and the scarcity of fault samples can be solved, a feasible approach is provided for the fault diagnosis of the traction system under small samples, and the minor faults of the rail train traction system can be diagnosed in time.

[0037] Optionally, the above S2 can be refined through the following steps:

[0038] First, collect the original three-phase current data x of the traction system using a high-precision sensor 0 , with a dimension of t×k, where t is the sampling time length and k is the number of sensors.

[0039] Perform normalization processing on each column of the original data x 0 (each column corresponds to the sampling data of a single sensor). The purpose is to limit the preprocessed data within the range of 0-1, thereby improving the computational efficiency of the model fitting process. The description formula for data normalization is:

[0040]

[0041] In the formula, x 0 represents the original three-phase current data, are the data of the i-th row and j-th column of the original three-phase current data x 0 and the normalized data x 1 respectively, where corresponds to all the data in the j-th column of x 0 , x 1 .

[0042] Based on the sliding window mechanism, segment the above normalized data x 1 . Let the window length be w and the sliding step be l. Then the number n of the segmented fault diagnosis data sets is:

[0043]

[0044] In the formula, [.] is the rounding symbol. The three-dimensional fault diagnosis data set obtained using the sliding window mechanism has a data size of n×w×k.

[0045] According to whether the service condition is known, divide the normalized data x 1 . Among them, the fault diagnosis data with known service conditions is divided into source domain data x s , and the unknown service conditions are divided into target domain data x t . At the same time, according to the use of the data in model training and verification, the source domain data is randomly divided into a training set and a test set in a ratio of 7:3. Similarly, the target domain data is randomly divided into a training set and a test set

[0046]

[0047] Before the above S3, the method further includes:

[0048] Construct a fault diagnosis model framework based on a multi-channel domain adaptive graph convolutional network, as Figure 3 shown. The fault diagnosis model framework includes an original data feature mapping layer, a graph generation layer, a graph information fusion layer, a domain adaptive module, and a label classifier. Among them, the graph generation layer is used to generate two graph topologies.

[0049] Next, the steps for constructing a fault diagnosis model framework based on a multi-channel domain adaptive graph convolutional network will be described in detail. It should be noted that the following method of constructing the framework is one implementation manner of this application, and the formula it satisfies is only one manifestation form. Here, it is only for example and not limited.

[0050] Specifically, to reduce the oscillation degree of model convergence and improve the training speed, mini-batch gradient descent is used for model training. Therefore, the input to the diagnostic model is mini-batch data where b is the number of batches for each training, and p = {s, t} indicates that the input data comes from the source domain s or the target domain t.

[0051] The first layer of the fault diagnosis model framework is the feature mapping layer. In this layer, the mini-batch data is input into a two-dimensional convolutional network to obtain a feature mapping matrix x p . Let m be the data feature dimension after feature mapping. Among them, the network structure is as Figure 4 shown, and the mapping process can be expressed by the following descriptive formula:

[0052]

[0053] In the formula, CNN(.) represents the feature mapping operation through the convolutional network, m is the data feature dimension after feature mapping, is the mini-batch data, b is the number of batches for each training, and R is the set of real numbers.

[0054] The second layer of the fault diagnosis model framework is the graph generation layer. The proposed layer can be used to generate two graphs. The graph here refers to the topological graph G = (V, E), and its topological graph is as Figure 5 shown, where Figure 5 each node v of represents a single data set, the node feature is the feature of the corresponding data, and if there is an edge connection e between nodes ij it means there is a correlation between data sets i and j, and the adjacent matrix A is used to quantify the adjacent relationship between nodes.

[0055] Among the two generated graphs, one is a dynamic graph construction based on a multi-layer perceptron, obtaining an adjacent matrix where b can also be regarded as the number of graph nodes; the other is a fixed graph construction based on the K-nearest neighbor algorithm, obtaining an adjacent matrix The two composition methods perform data composition from different perspectives, which can provide more channels for the node information transmission of the graph convolutional network in the third layer, thereby enhancing the diversity of graph topological information. Among them, the structure of the graph convolutional network is as Figure 6 shown. The detailed process of obtaining two adjacency matrices using the two composition methods is as follows:

[0056] 1) Dynamic composition: Input the feature mapping matrix x p into the double-layer convolutional network to obtain the feature output x M . Perform matrix multiplication between it and the transpose matrix, and after regularization, obtain the adjacency matrix A. Finally, according to the max-k sorting mechanism, set the first k largest elements in each row of A to 1, and set the other row elements to 0, so as to obtain the sparse adjacency matrix thus reducing the computational burden. This composition process can be represented by the following descriptive formula:

[0057]

[0058] where MLP(.) represents the feature mapping operation through a multi-layer perceptron; normalize(.) represents the regularization operation.

[0059] 2) Fixed composition: The row vectors of the feature mapping matrix x p ∈R b×m can be regarded as the feature vectors of the graph nodes. Thus, according to the node feature vectors, calculate the Euclidean distance of each node to obtain the Euclidean distance matrix E∈R b×b , where E ij represents the Euclidean distance between nodes i and j. To reduce the computational burden, according to the min-k sorting mechanism, select the first k minimum values in each row of the Euclidean distance matrix E, and set the corresponding row elements to 1, and the rest to 0, to obtain the sparse adjacency matrix to represent the k nearest neighbor nodes most similar to the target node.

[0060]

[0061] In this way, in the designed graph generation layer, two graph construction schemes are proposed. One is the dynamic composition based on a multi-layer perceptron, and the other is the fixed composition based on the K-nearest neighbor algorithm. The two composition methods perform data composition from different perspectives, increasing the channels for node information propagation, thereby enhancing the diversity of graph topological information. The composition process not only does not depend on mathematical modeling and prior knowledge, but also can adapt to the characteristics of the changes in the associated attributes of nodes, edges, etc. of the graph network model caused by different service conditions and system characteristics, which can ensure the accuracy of the diagnosis results.

[0062] Based on the node information propagation mechanism of the graph convolutional network, input the relevant data x and the corresponding adjacency matrix A. The forward propagation formula of the graph convolutional network is defined as:

[0063]

[0064] where c is the number of layers of the graph convolutional network, σ is the activation function, I is the identity matrix, is the degree matrix of:

[0065] Based on the above forward propagation formula of the graph convolutional network, the obtained feature mapping matrix x p and two types of graph adjacency matrices and are respectively imported into two specific graph convolutional network modules Z 1 (x, A), Z 2 (x, A) to extract two specific feature representations and The description formula is:

[0066]

[0067] And input them simultaneously into the common graph convolutional network Z 3 (x, A) with shared parameters to obtain the common feature representation of the two graphs. The description formula is:

[0068]

[0069]

[0070] Furthermore, use the attention mechanism to automatically learn the attention weights α m , α k , α c ∈R d×1 of the above three different feature representations (two specific feature representations and one common feature representation), and perform weighted summation according to the attention weights to adaptively fuse them to obtain the final feature representation Z p . The description formula is:

[0071]

[0072]

[0073] In this way, during the multi-graph fusion process, an adaptive update attention mechanism is adopted to learn the weights of different output feature representations, so as to judge which of the two graph construction methods is more representative, and thus select the most reasonable topological graph to describe the data correlation, thereby improving the diagnostic accuracy.

[0074] Among them, the attention weight α m , α k , α c The obtaining steps are as follows: Taking the i-th group of data (which can also be regarded as the features of node i) as an example, it corresponds to the i-th row of the feature representation Z M and is expressed as First, perform a non-linear transformation on to obtain The corresponding weight value:

[0075]

[0076] In the formula, W ∈ R k×1 and B ∈ R are the weight parameter and bias parameter to be learned respectively. Similarly, the weight value corresponding to node i in the feature representation can be obtained Finally, use the softmax function to regularize the weights of the above three feature representations to obtain the final weight value:

[0077]

[0078] Similarly, and Among them, the larger α is, the higher the importance of the corresponding feature representation. For all b nodes, Finally, use the weighted summation method to obtain the final feature representation Z p :

[0079]

[0080] Among them, the specific feature representation and the common feature representation satisfy the preset consistency constraint conditions and difference constraint conditions to ensure the above three feature representations and have consistency and difference. The two constraints can be obtained from the following description formulas:

[0081] 1) Consistency constraint: First, use the L2 norm to regularize the two feature representations output by the common graph convolutional network as and Then, use these two regularization matrices to obtain the similarity between nodes and

[0082]

[0083]

[0084] Consistency indicates that there should be consistency between two similarity matrices, resulting in the following similarity constraint L s :

[0085]

[0086] wherein, is the square of the Frobenius norm. Let M be an arbitrary matrix, and its description formula is as follows:

[0087]

[0088] 2) Difference constraint: In order to prompt the common graph convolutional network and the specific graph convolutional network to extract different features of the input data, the subspace orthogonality constraint L d is defined to enhance this difference, and the description formula is:

[0089]

[0090] Set the label classifier for outputting the fault diagnosis result. Specifically, input the obtained feature representation Z p into the fully connected layer. After obtaining the outputs of multiple neurons, use the softmax function to output the normalized classification result. Since the data categories in the target domain are unknown, when discussing the performance of the label classifier, only the source domain data is considered. Take the cross-entropy loss function of the true label and the predicted label as the measurement standard for the performance of the source domain classifier:

[0091]

[0092] wherein, the feature representation of the source domain data corresponds to the label are the feature spaces corresponding to the input data features and labels respectively. represents the prediction result of the label classifier. L(.,.) represents the cross-entropy loss, and E(.) represents the mathematical expectation.

[0093] In this way, in the graph information fusion layer designed in this embodiment, the multi-graph fusion method is adopted to reduce the parallel computation amount. At the same time, the fused graph information includes the specific features and shared features of the two graphs, and the consistency constraint and difference constraint are used to ensure the consistency and difference of the above feature representations, so as to improve the diversity of the node information propagated on the graph topology.

[0094] In addition, a domain adaptation module is set up to reduce the distribution difference between the source domain and the target domain. Specifically, a game mechanism between the domain adaptation module and the feature extractor can be used to extract transfer features. The feature extractor aims to extract domain-invariant features to "deceive" the domain adaptation module, while the domain adaptation module is trained to distinguish whether the features extracted by the feature extractor come from the source domain or the target domain. Through such adversarial training, when the min-max game between the domain adaptation module and the feature extractor reaches equilibrium, the game ends. At this time, domain-invariant features can be learned, that is, the features of different domains are difficult to be distinguished by the domain adaptation module, indicating that the data distribution difference between the source domain and the target domain has been greatly reduced, thus achieving the domain transfer effect. Using the domain adaptation module enables the diagnostic model to learn domain transfer knowledge, so as to quickly adapt to new tasks, and is suitable for solving the problem of inconsistent data distribution caused by multiple service conditions of the rail train traction system. Therefore, the proposed model maintains high generalization performance in solving small-sample problems.

[0095] Here, the binary cross-entropy loss function is used as the domain consistency loss:

[0096]

[0097] where and respectively represent the features extracted by the feature extractor from the i-th group of data in the source domain and the j-th group of data in the target domain. D(.) takes values of 0 (indicating the target domain) or 1 (indicating the source domain), E represents the mathematical expectation, D s represents the source domain distribution, and D t represents the target domain distribution.

[0098] It should be noted that after building the above-mentioned traction system fault diagnosis model framework, model training is required, that is, inputting the source domain training data set and the target domain training data set into the built multi-channel domain adaptation graph convolutional network framework, using the overall loss function as the network training index, and updating the network parameters based on the backpropagation algorithm. The formula for the overall loss function L is as follows:

[0099] L = L c + λL DA + βL s + γL d ;

[0100] In the formula, λ, β, and γ are all weight coefficients, λ, β, γ ∈ (0, 1), L c is the cross-entropy loss function between the true label and the predicted label, L DA is the domain consistency loss function, L s is the consistency constraint, and L d is the difference constraint.

[0101] Next, taking a certain experiment as an example, the fault diagnosis of the train traction system under multiple service conditions provided by the present application will be illustrated by examples. First, the three-phase current data of the rectifier of the traction system is obtained from the semi-physical platform of the rail train developed by a certain experiment. Among them, there are two types of motor faults: permanent magnet demagnetization and inter-turn short circuit, and the motor speeds of 1000 r / min, 1200 r / min, 1500 r / min, and 1800 r / min are used as service conditions. After multiple repeated experiments, the diagnostic effect diagram of the method proposed by the present invention is obtained as shown in Figure 7 shown, and the migration effect diagram under different service conditions is as shown in Figure 8 shown, where (a) is the motor data from 0->1, (b) is the motor data from 0->2, (c) is the motor data from 0->3, (d) is the motor data from 1->0, (e) is the motor data from 1->2, (f) is the motor data from 1->3, (g) is the motor data from 2->0, (h) is the motor data from 2->1, (i) is the motor data from 2->3, (j) is the motor data from 3->0, (k) is the motor data from 3->1, and (n) is the motor data from 3->2. Figure 7 Among them, the numbers 0, 1, 2, and 3 respectively represent the motor speeds of 1000 r / min, 1200 r / min, 1500 r / min, and 1800 r / min for the service conditions; 0->1 means that the service condition migrates from 1000 r / min to 1200 r / min, and so on. According to Figure 7 it can be seen that even in the case where the migration span of the service conditions between 1000 r / min and 1800 r / min is relatively large, the diagnostic accuracy rate of the method proposed by the present invention can reach more than 97.5%; Figure 8 Visualize the features of the last layer of the feature extractor through the t-SNE method. For example, in the service condition where the speed migrates from 1000 r / min to 1800 r / min, the data features from different domains but belonging to the same fault type can be well aligned, while the data features from the same domain but not belonging to the same fault type are divided into different intervals, and the boundaries between the intervals are obvious, indicating that the fault diagnosis method proposed by the present invention has good service condition migration effect and generalization performance in unknown service conditions.

[0102] The embodiment of the present application also provides a fault diagnosis system for a train traction system under multiple service conditions, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented. The fault diagnosis system for the train traction system under multiple service conditions can implement each embodiment of the above fault diagnosis method for the train traction system under multiple service conditions and can achieve the same beneficial effects, which will not be elaborated here.

[0103] The embodiments of the present application also provide a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the above-mentioned method are implemented. The readable storage medium can implement the various embodiments of the above-mentioned method and achieve the same beneficial effects, and will not be elaborated here.

[0104] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A fault diagnosis method for a train traction system under multiple service conditions, characterized in that, it includes: S1: Collect the original three-phase current data of the traction system under different service conditions and different fault states, and add fault type labels to the original three-phase current data; S2: Preprocess the original three-phase current data to obtain a fault diagnosis data set, and divide the fault diagnosis data set into a source domain data set and a target domain data set according to the service condition; S3: Input the source domain data set and the target domain data set into a pre-constructed fault diagnosis model framework, use the overall loss function as the network training index, and train based on the backpropagation algorithm to obtain a target channel domain adaptive graph convolutional network model; S4: Input the three-phase current data of the traction system to be diagnosed into the target channel domain adaptive graph convolutional network model for fault diagnosis.

2. The fault diagnosis method for a train traction system under multiple service conditions according to claim 1, characterized in that, the S2 specifically includes: Normalize the original three-phase current data using the following formula: where x 0 represents the original three-phase current data, which are the original three-phase current data x 0 and the normalized data x 1 at the i-th row and j-th column, where corresponds to all the data in the j-th column of x 0 and x 1 ; Based on the sliding window mechanism, for the normalized data x 1 is segmented. Let the window length be w and the sliding step be l. Then the number n of the segmented fault diagnosis data sets satisfies the following relational expression: where [.] is the rounding symbol and t is the length of the normalized data x 1 ; The fault diagnosis data set obtained using the sliding window mechanism is a three-dimensional fault diagnosis data set, and the data size is n×w×k; Divide the fault diagnosis data under known service conditions into source domain data x s , and divide the unknown service conditions into target domain data x t .

3. The fault diagnosis method for a train traction system under multiple service conditions according to claim 1, characterized in that, Before the S3, the method further includes: Construct a fault diagnosis model framework based on a multi-channel domain adaptive graph convolutional network. The fault diagnosis model framework includes an original data feature mapping layer, a graph generation layer, a graph information fusion layer, a domain adaptive module, and a label classifier. Among them, the graph generation layer is used to generate two graph topologies.

4. The fault diagnosis method for a train traction system under multiple service conditions according to claim 3, characterized in that, The construction of the fault diagnosis model framework based on the multi-channel domain adaptive graph convolutional network includes: Input small-batch data into a two-dimensional convolutional network to obtain a feature map matrix x p , let m be the data feature dimension after feature mapping, and the mapping process satisfies the following relational expression: Wherein, CNN(.) represents the feature mapping operation through a convolutional network, m is the data feature dimension after feature mapping, is a mini-batch of data, b is the number of batches for each training, and R is the set of real numbers; Input the feature mapping matrix x p into the multi-layer perceptron to obtain the feature output x M . Perform matrix multiplication with its transpose matrix, and after regularization, obtain the adjacency matrix A. According to the max-k sorting mechanism, set the first k largest elements in each row of A to 1, and set the other row elements to 0, so as to obtain the first sparse adjacency matrix Regarding the row vectors of the feature mapping matrix x p ∈R b×m as the feature vectors of graph nodes, calculate the Euclidean distance of each node based on the feature vectors of graph nodes to obtain the Euclidean distance matrix E ∈ R b×b , according to the min-k sorting mechanism, select the first k minimum values in each row of the Euclidean distance matrix E, and set the row elements at the corresponding positions to 1, and the rest to 0, to obtain the second sparse adjacency matrix Based on the forward propagation formula of the graph convolutional network, the obtained feature map matrix x p , the first sparse adjacency matrix and the second sparse adjacency matrix are respectively imported into two specific graph convolutional network modules Z 1 (x, A), Z 2 (x, A) to extract two specific feature representations and The two specific feature representations and are simultaneously input into the common graph convolutional network Z with shared parameters 3 (x, A) to obtain the common feature representations of the two graphs, and the attention weights α and are automatically learned using the attention mechanism m , α k , α c ∈R b×b , and weighted summation is performed according to the attention weights to obtain the final feature representation Z p ; wherein, the representation of the specific feature and the representation of the common feature satisfy a preset consistency constraint condition and a difference constraint condition; The obtained final feature representation Z p is input into the fully connected layer of the network. After obtaining the outputs of multiple neurons, the softmax function is used to output the normalized classification results. The cross-entropy loss function of the true label and the predicted label is used as the measurement criterion for the performance of the source domain classifier as follows: In the formula, the source domain data feature representation The corresponding label is They are the feature spaces corresponding to the input data feature and the label respectively, indicating the prediction result of the label classifier, L(.,.) represents the cross-entropy loss, and E(.) represents the mathematical expectation; Here, the binary cross-entropy loss function is used as the domain consistency loss as follows: In the formula, and respectively represent the features extracted by the feature extractor from the $i$-th group of data in the source domain and the $j$-th group of data in the target domain. $D(.)$ takes values of 0 or 1, $E$ represents the mathematical expectation, $D$ s represents the source domain distribution, and $D$ t represents the target domain distribution.

5. The fault diagnosis method for a train traction system under multiple service conditions according to claim 4, characterized in that, The automatic learning using the attention mechanism and attention weights α m , α k , α c ∈ R b×b , including: Select the i-th group of data, which corresponds to the feature representation Z M The i-th row of For Perform a non-linear transformation to obtain The corresponding weight value Is as follows: where W is the weight parameter to be learned, B is the bias parameter to be learned, W ∈ R k×1 , B ∈ R; Among them, the feature representation of node i The corresponding weight value Regularize the weights of the above three feature representations using the softmax function to obtain the final weight value: Then: where the larger α is, the higher the importance of the corresponding feature representation is. For all b nodes, 6. The fault diagnosis method for a train traction system under multiple service conditions according to claim 1, characterized in that, The overall loss function satisfies the following relationship: L = L c + λL DA + βL s + γL d ; where λ, β, and γ are all weight coefficients, λ, β, γ ∈ (0, 1), and L c is the cross-entropy loss function between the true label and the predicted label, and L DA is the domain consistency loss function, and L s is the consistency constraint, and L d is the difference constraint.

7. A fault diagnosis system for a train traction system under multiple service conditions, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the steps of any one of the methods described in claims 1 to 6 above.

8. A computer-readable storage medium, on which a computer program is stored, characterized in that, when the program is executed by the processor, it implements the method steps of any one of claims 1-6.

Images