A fault diagnosis method and system for a train drive motor, gear box and axle box

Abstract

The application discloses a kind of train transmission motor, gear box and axle box fault diagnosis method and system, belong to train transmission system fault diagnosis technical field, method includes respectively designing special expert network for the key components of train transmission system;Introduce multi-head self-attention gate network, dynamically fuse the features output by special expert network;Adopt the joint loss function including cross-entropy loss, KL divergence regularization loss, load balancing loss, train special expert network and gate network;Based on teacher-student model architecture and knowledge distillation, realize new fault category adaptation and historical knowledge retention;The multi-component vibration signal to be diagnosed is input into the trained model, and the probability of the highest fault category is output as the diagnosis result through feature extraction, gate network fusion and classifier prediction.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for train transmission systems, and in particular to a fault diagnosis method and system for train transmission motors, gearboxes, and axle boxes. Background Technology

[0002] The train drive system is a core component of rail transit, and its reliability directly determines operational efficiency and passenger safety. Key components include gearboxes, traction motors, and left and right axle boxes. These components are prone to failure under long-term high-load and high-frequency vibration conditions, and the failure modes gradually emerge as the equipment ages and external conditions change, posing a severe challenge to fault diagnosis.

[0003] Traditional fault diagnosis methods often involve constructing a diagnostic network for a single component, which not only increases system complexity but also ignores the mutual influence of faults between components. Furthermore, cross-component comparative studies require re-selecting models, which is time-consuming and labor-intensive. Existing incremental learning methods rely on a large amount of new data to update the model, resulting in significant latency and difficulty in adapting to the gradual emergence of fault modes in real-world scenarios. Knowledge distillation techniques often rely on single-condition or single-module data, lacking a unified knowledge transfer mechanism for multi-module and multi-source signals, and are prone to "catastrophic forgetting," making it difficult to retain historical fault knowledge when adding new fault categories. Moreover, the strong heterogeneity of fault features under multiple conditions makes simple feature fusion strategies prone to redundancy, leading to a decrease in diagnostic accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a fault diagnosis method and system for train transmission motors, gearboxes, and axle boxes, which solves the problems of independent modeling of multiple components, incremental learning delay, catastrophic forgetting, and poor adaptability to multiple operating conditions in the existing fault diagnosis of train transmission systems, and achieves efficient fault diagnosis that balances the retention of historical knowledge and the ability to adapt to new faults.

[0005] To achieve the above objectives, the present invention provides a fault diagnosis method for train drive motors, gearboxes, and axle boxes, comprising the following steps:

[0006] S1. Construct dedicated expert networks for multiple components: Design dedicated expert networks for the key components of the train transmission system;

[0007] S2. Perform self-attention gating dynamic feature fusion: Introduce a multi-head self-attention gating network to perform sample-level dynamic fusion of features output by a dedicated expert network;

[0008] S3. Multi-objective joint training: Employing cross-entropy loss. KL divergence regularization loss Load balancing losses The joint loss function is used to train a dedicated expert network and a gated network;

[0009] S4. Construct an incremental learning and training mechanism: Based on the teacher-student model architecture and knowledge distillation, realize the adaptation of new fault categories and the retention of historical knowledge;

[0010] S5. Fault Diagnosis Output: Input the vibration signals of the multi-component components to be diagnosed into the trained model. Through feature extraction by a dedicated expert network, fusion by a gating network, and prediction by a classifier, the fault category with the highest probability is output as the diagnosis result.

[0011] Preferably, in S1, the key components include the gearbox, motor, left axle box, and right axle box, and the dedicated expert networks for each are as follows:

[0012] LHE-Net, a network of gearbox / motor experts, employs a multi-layer one-dimensional convolutional structure to extract high-dimensional discriminative features of vibration signals through convolutional blocks. Each convolutional block sequentially includes convolution operations, batch normalization (BN), ReLU activation function, and max pooling (MaxPool).

[0013] Axle box expert network RE-Net: Based on the residual block structure, it alleviates the degradation of deep networks through identity mapping and realizes the extraction of weak fault features of axle boxes. The residual block includes convolution operation, BN, dropout regularization and shortcut connection.

[0014] Preferably, in S1, the convolutional block parameters of the LHE-Net satisfy:

[0015] ;

[0016] kernel size is The number of output channels is One-dimensional convolution, setting the kernel size and number of output channels, padding parameters. Adaptively adjusts based on kernel size and input sequence length;

[0017] The l-th convolutional block is defined as the kernel size. Number of output channels The one-dimensional convolutional features are calculated as follows:

[0018] ;

[0019] in, For the weight of the l-th layer, connect the l-th layer. Layer Input channel and the l-th layer The first output channel Each weight; For the first Layer aisle, Input feature value at time t, For filling parameters, For the first Layer Channel bias term;

[0020] The overall feature mapping of the network is as follows:

[0021] ;

[0022] in, For the overall mapping of this network, For residual networks, Indicates input data, Indicates the number of convolutional layers; This indicates a combination of functions, where the output of the previous layer serves as the input of the next layer. For all trainable parameters, For global average pooling, For classifiers.

[0023] Preferably, in S1, the RE-Net includes several residual blocks, and the output of each residual block is... ,in For input features, The residual function is used; the complete forward mapping of RE-Net is:

[0024] ;

[0025] For left / right axle box signals, ; For input convolutional layers, For the first One residual module.

[0026] Preferably, in S2, the specific steps are as follows: the original features of the dedicated expert network are projected as query Q, key K, and value V; enhanced features are calculated by scaling dot product attention; the enhanced features are globally averaged and pooled to obtain contextual information; expert contribution weights are generated using MLP, satisfying the constraints that the weight sum is 1 and non-negativity; the multi-expert enhanced features are weighted and fused based on the contribution weights, and the weights are normalized using the Softmax function to obtain system-level fused features; the specific steps are as follows:

[0027] S21, Enhanced Bullish Attention: The h Each attention head projects expert features into query Q, key K, and value V:

[0028] ;

[0029] in, For learnable parameter matrix, For the original feature set from multiple experts; scaling dot product attention calculation:

[0030] ;

[0031] in, For single-head attention Feature dimensions; enhanced features are obtained by projecting the concatenated outputs of multiple attention heads. :

[0032] ;

[0033] in, For the number of attention heads, To output the projection matrix;

[0034] S22, Dynamic Weight Generation: For Enhanced Features Global average pooling is used to obtain contextual information, and expert contribution weights are generated through a two-layer MLP:

[0035] ;

[0036] in, For average pooling characteristics, This is the MLP weight matrix. Here, T is the bias term, a temperature parameter used to control the sharpness of the weight distribution, satisfying the following conditions: ;

[0037] S23, Feature-weighted fusion:

[0038] ;

[0039] in, For the first i The weight of each expert's contribution For the first i Enhanced features of individual experts.

[0040] Preferably, in S3, the joint loss function is:

[0041] ;

[0042] Wherein, λ1 and λ2 are hyperparameters, determined through cross-validation; Minimize classification error To prevent the gating weights from collapsing to a single expert, Experts are encouraged to use their expertise in a balanced manner; the specific calculation formula is as follows:

[0043] ;

[0044] ;

[0045] ;

[0046] in, It is the total number of samples. Calculate the gating weights of the nth sample KL divergence between u and a uniform distribution u The core function of symbols is to establish an asymmetric comparative relationship.

[0047] Preferably, step S4 specifically includes the following steps:

[0048] S41. Teacher Model Initialization: Based on pre-training with historical fault data, parameters are frozen as a historical knowledge repository, and these parameters are inherited from the parameters trained in the previous round of student model training; specifically: ;in, For the first Rotate student model parameters;

[0049] S42. Student Model Initialization: Expand the classifier to adapt to the new fault categories, and divide the classifier weights into historical fault weights. and new fault weights ,Keep Align with the teacher model output;

[0050] S43. Incremental training, employing knowledge distillation loss. and cross-entropy loss The student model is trained using a total loss function, which updates only the classifier weights and gating network parameters. The total loss function is: Where α is the weighting coefficient, 0≤α≤1; Used to retain historical fault knowledge Used to adapt to both old and new fault classifications.

[0051] Preferably, in S43, knowledge distillation loss The calculation method is as follows ,in, T For temperature parameters, For batch size, It is a category. This represents the number of historical fault categories. Softening the probability for the teacher model. The formula for calculating the softening probability of the student model is as follows:

[0052] ;

[0053] in, For the teacher model The first sample Logit for historical failures, The student model corresponds to the logit;

[0054] The cross-entropy loss The calculation method is as follows: ,in, This represents the total number of categories after the addition of faults. For the first The first sample The true label of the class, This is the logit output for the student model.

[0055] This invention also provides a fault diagnosis system for train drive motors, gearboxes, and axle boxes, comprising:

[0056] Signal acquisition module: Triaxial accelerometers deployed in the train transmission system gearbox, motor, left axle box, and right axle box are used to acquire vibration signals from each component; the input signals of the signal acquisition module must be compatible with the signal format of the bogie dataset; the sampling length and timing characteristics of the input signals must match the signal attributes of the bogie dataset;

[0057] Data preprocessing module: Performs sliding window truncation on the collected vibration signals to generate fixed-length input samples; divides the dataset into training set, test set and validation set according to the proportion;

[0058] Differentiated Hybrid Expert Diagnostic Module: Includes a dedicated expert network and a self-attention gating fusion submodule, used to achieve differentiated extraction and dynamic fusion of multi-component features;

[0059] Incremental learning training module: includes teacher model and student model. The teacher model stores historical fault knowledge, and the student model learns new fault categories and retains historical knowledge through extended classifier and knowledge distillation.

[0060] Diagnostic output module: Receives the fusion features from the differentiated hybrid expert diagnostic module, outputs the predicted probabilities of various faults through a classifier, and selects the category with the highest probability as the final diagnostic result.

[0061] Preferably, the training process of the incremental learning training module includes: a first stage, training the teacher model based on historical fault data and freezing all its parameters; a second stage, initializing the student model parameters and expanding the classifier to adapt to the newly added fault categories; and a third stage, training the total loss function and updating only the classifier weights and gating network parameters of the student model to complete the incremental learning.

[0062] The diagnostic output module also includes a dedicated visualization unit that outputs confusion matrices, precision, recall, and F1 scores in real time, as well as the probability distribution of fault category predictions, to help engineers locate the source of the fault.

[0063] Therefore, the present invention employs the above-mentioned fault diagnosis method and system for train drive motors, gearboxes, and axle boxes, which has the following beneficial effects:

[0064] 1) By integrating the fault characteristics and interactions of multiple components through a hybrid expert architecture and gating mechanism, the complexity of independent modeling is avoided, the training and management process is simplified, and the diagnostic efficiency is improved;

[0065] 2) Efficient incremental learning, without the need for a large amount of new data. Through classifier expansion and knowledge distillation, the model update latency is reduced, and it can quickly adapt to emerging new fault modes.

[0066] 3) No catastrophic forgetting; historical fault knowledge is solidified through distillation loss, maintaining the diagnostic accuracy of original faults after training with new faults.

[0067] 4) It has strong engineering practicality, and the total number of model parameters is reduced compared with existing deep learning methods, making it suitable for deployment in vehicles and edge devices.

[0068] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0069] Figure 1 This is a flowchart of an embodiment of the present invention;

[0070] Figure 2 This is a system framework diagram of an embodiment of the present invention;

[0071] Figure 3 This is a diagram illustrating the incremental learning process according to an embodiment of the present invention;

[0072] Figure 4 This is the confusion matrix of the present invention under different operating conditions;

[0073] Figure 5 This is a comparison chart of the accuracy, recall rate and F1 score of each fault category under different operating conditions in the embodiments of the present invention;

[0074] Figure 6 This is a comparison chart of the diagnostic accuracy of the embodiments of the present invention and existing mainstream methods. Detailed Implementation

[0075] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0076] 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. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0077] Example 1

[0078] The data used in this embodiment is based on the BJTU-RAO bogie dataset, collecting triaxial acceleration vibration signals from the gearbox (CH11-CH16), traction motor (CH1-CH6), left axle box (CH17-CH19), and right axle box (CH21-CH23), covering normal and different fault states. To verify the robustness of the model, Gaussian noise with a signal-to-noise ratio (SNR) of 20dB was added to the original signals to simulate environmental interference in actual operation. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set contained initial fault category samples, the validation set was used for hyperparameter tuning, and the test set was used to evaluate diagnostic performance. Long sequence signals were truncated using a sliding window to generate fixed-length input samples to ensure data integrity and continuity.

[0079] A fault diagnosis method for train drive motors, gearboxes, and axle boxes, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0080] S1. Construct dedicated expert networks for multiple components: Design dedicated expert networks for each key component of the train transmission system; the key components include the gearbox, motor, left axle box, and right axle box, and the dedicated expert networks for each are designed as follows:

[0081] LHE-Net, a network of experts for gearboxes and motors, employs a multi-layer one-dimensional convolutional structure to extract high-dimensional discriminative features of vibration signals through convolutional blocks. Each convolutional block sequentially includes convolution operations, batch normalization (BN), ReLU activation function, and max pooling (MaxPool).

[0082] The convolutional block parameters of LHE-Net satisfy:

[0083] ;

[0084] kernel size is The number of output channels is One-dimensional convolution is used, with the kernel size and number of output channels set. In this embodiment, the kernel size is 3-7, and the number of output channels is gradually increased to 256. The padding parameters are... Adaptively adjusts based on kernel size and input sequence length;

[0085] The l-th convolutional block is defined as the kernel size. Number of output channels The one-dimensional convolutional features are calculated as follows:

[0086] ;

[0087] in, The weights of the l-th layer (connecting the l-th layer) Layer Input channel and the l-th layer The first output channel (each weight) For the first Layer aisle, Input feature value at time t, For filling parameters, For the first Layer Channel bias term.

[0088] The overall feature mapping of the network is as follows:

[0089] ;

[0090] in, For the overall mapping of this network, For residual networks, Indicates input data, Indicates the number of convolutional layers; This indicates a combination of functions, where the output of the previous layer serves as the input of the next layer. For all trainable parameters, For global average pooling, For classifiers.

[0091] Axle box expert network RE-Net: Based on the residual block structure, it alleviates the degradation of deep networks through identity mapping and realizes the extraction of weak fault features of axle boxes. The residual block includes convolution operation, BN, dropout regularization and shortcut connection.

[0092] In this embodiment, RE-Net contains 3,000 residual blocks, and the output of each residual block is... ,in Input features; The residual function consists of two convolutional layers, batch normalization (BN), and dropout regularization. The complete forward mapping of RE-Net is:

[0093] ;

[0094] For left / right axle box signals, ; For input convolutional layers, For the first One residual module.

[0095] S2. Perform self-attention gating dynamic feature fusion: Introduce a multi-head self-attention gating network to perform sample-level dynamic fusion of features output by the dedicated expert network; specifically: project the original features of the dedicated expert network into query Q, key K, and value V, and calculate enhanced features through scaling dot product attention; globally average pool the enhanced features to obtain contextual information, generate expert contribution weights through MLP, and satisfy the constraints that the weight sum is 1 and non-negativity; perform weighted fusion of multi-expert enhanced features based on contribution weights, and use the Softmax function for weight normalization to obtain system-level fused features.

[0096] In this embodiment, the number of multi-head attention heads... The MLP hidden layer has a dimension of 128, a temperature parameter T=2, uses the Softmax function for weight normalization, a weight coefficient α=0.5, a batch size B=32, and an initial number of historical fault categories. The total number of fault categories after adding new fault categories Dynamic expansion. The specific steps are as follows:

[0097] S21, Enhanced Bullish Attention: The h Each attention focus projects expert features into query (Q), key (K), and value (V):

[0098] ;

[0099] in, For learnable parameter matrix, The original feature set is from multiple experts. Scaling dot product attention calculation:

[0100] ;

[0101] in, For single-head attention The feature dimensions are obtained by concatenating the outputs of multiple attention heads and then projecting them to obtain enhanced features. :

[0102] ;

[0103] in, For the number of attention heads, This is for outputting the projection matrix.

[0104] S22, Dynamic Weight Generation: For Enhanced Features Global average pooling is used to obtain contextual information, and expert contribution weights are generated through a two-layer MLP:

[0105] ;

[0106] in, For average pooling characteristics, The weight matrix of the MLP. Here, T is the bias term, a temperature parameter used to control the sharpness of the weight distribution, satisfying the following conditions: .

[0107] S23, Feature-weighted fusion:

[0108] ;

[0109] in, For the first i The weight of each expert's contribution For the first i Enhanced features of individual experts.

[0110] S3. Multi-objective joint training: Employing cross-entropy loss. KL divergence regularization loss Load balancing losses The joint loss function is used to train a dedicated expert network and a gated network;

[0111] The joint loss function is:

[0112] ;

[0113] Among them, λ1 and λ2 are hyperparameters used to balance the contributions of various losses. They are determined through cross-validation to ensure that various optimization objectives are promoted in a coordinated manner. The primary objective is to minimize classification error. To prevent gating weights from collapsing to a single expert, weight diversity should be encouraged. At the batch level, all experts are encouraged to be used evenly to avoid some experts being idle. The specific calculation formula is as follows:

[0114] ;

[0115] ;

[0116] ;

[0117] in, It is the total number of samples. Calculate the gating weights of the nth sample =[ , , The KL divergence between the uniform distribution u and the uniform distribution u. The core function of symbols is to establish an asymmetric comparative relationship.

[0118] S4. Construct an incremental learning and training mechanism: Based on the teacher-student model architecture and knowledge distillation, achieve adaptation to new fault categories and retention of historical knowledge; specifically including the following steps:

[0119] S41. Teacher Model Initialization: Based on pre-training with historical fault data, parameters are frozen as a historical knowledge repository, and these parameters are inherited from the parameters trained in the previous round of student model training; specifically: ;in, For the first Rotate student model parameters.

[0120] S42. Student Model Initialization: Expand the classifier to adapt to the new fault categories, and divide the classifier weights into historical fault weights. and new fault weights ,Keep Align with the teacher model output;

[0121] S43. Incremental training, the process is as follows: Figure 3 As shown, this employs a method that incorporates knowledge distillation loss. and cross-entropy loss The student model is trained using the total loss function, which updates only the classifier weights and gating network parameters. The total loss function is: Where α is the weighting coefficient, 0≤α≤1; Used to retain historical fault knowledge Used to adapt to both new and old fault classifications.

[0122] Knowledge distillation loss The calculation method is as follows ,in, T For temperature parameters, It is a category. For batch size, This represents the number of historical fault categories. Softening the probability for the teacher model. The formula for calculating the softening probability of the student model is as follows:

[0123] ;

[0124] in, For the teacher model The first sample Logit for historical failures, The logit corresponds to the student model.

[0125] Cross-entropy loss The calculation method is as follows: ,in, This represents the total number of categories after the addition of faults. For the first The first sample The true label of the class, This is the logit output for the student model.

[0126] S5. Fault Diagnosis Output: Input the vibration signals of the multi-component components to be diagnosed into the trained model. Through feature extraction by a dedicated expert network, fusion by a gating network, and prediction by a classifier, the fault category with the highest probability is output as the diagnosis result.

[0127] The fault diagnosis system used to implement the above method includes a signal acquisition module, a data preprocessing module, a differentiated hybrid expert diagnosis module, an incremental learning training module, and a diagnosis output module. The specific system architecture is as follows: Figure 2 As shown.

[0128] Signal acquisition module: Triaxial accelerometers deployed in the train transmission system gearbox, motor, left axle box, and right axle box are used to acquire vibration signals from each component; the input signals of the signal acquisition module must be compatible with the signal format of the bogie dataset; the sampling length and timing characteristics of the input signals must match the signal attributes of the bogie dataset;

[0129] In this embodiment, the deployment channels for the triaxial accelerometer are: motor (CH1-CH6), gearbox (CH11-CH16), left axle box (CH17-CH19), and right axle box (CH21-CH23). The acquisition channel definitions are consistent with the BJTU-RAO bogie dataset. The channel correspondence is as follows: motor signals correspond to channels CH1-CH6 (CH1-CH3 correspond to the drive end, CH4-CH6 correspond to the non-drive end), gearbox signals correspond to channels CH11-CH16 (CH11-CH13 correspond to the input shaft side, CH14-CH16 correspond to the output shaft side), left axle box signals correspond to channels CH17-CH19, and right axle box signals correspond to channels CH21-CH23. In terms of signal compatibility, the sampling length and timing characteristics of the input signals need to match the signal attributes of the BJTU-RAO bogie dataset to ensure that the model can be directly trained based on this dataset and adapted to signal diagnosis of similar acquisition specifications.

[0130] Data preprocessing module: The collected vibration signals are truncated using a sliding window to generate fixed-length input samples; the dataset is divided into training, testing, and validation sets according to the proportions; the dedicated process of the data preprocessing module in this embodiment is as follows: the sliding window length is 4096 sampling points, the step size is 2048 sampling points, and fixed-length samples are generated; the dataset is divided using stratified sampling to ensure that the sample ratio of each fault category in the training, validation, and testing sets is consistent, which is 7:1:2.

[0131] Differentiated Hybrid Expert Diagnostic Module: Includes a dedicated expert network and a self-attention gating fusion submodule, used to achieve differentiated extraction and dynamic fusion of multi-component features;

[0132] Incremental learning training module: includes teacher model and student model. The teacher model stores historical fault knowledge, and the student model learns new fault categories and retains historical knowledge through extended classifier and knowledge distillation.

[0133] The training process of the incremental learning training module includes: the first stage, training the teacher model based on historical fault data and freezing all its parameters; the second stage, initializing the student model parameters and expanding the classifier to adapt to the newly added fault categories; the third stage, training the total loss function and updating only the classifier weights and gating network parameters of the student model to complete the incremental learning.

[0134] The diagnostic output module receives fused features from the differentiated hybrid expert diagnostic module, outputs the predicted probabilities of various faults through a classifier, and selects the category with the highest probability as the final diagnostic result. The diagnostic output module also includes a dedicated visualization unit that outputs confusion matrices, precision, recall, and F1 scores in real time, as well as the predicted probability distribution of fault categories, facilitating engineers in locating the source of the fault.

[0135] This embodiment applies the above scheme to experiments on the BJTU-RAO dataset, and the results further verify the effectiveness of the proposed method. Particularly under low-frequency no-load conditions (20Hz / 0kN), this method exhibits superior classification performance, with an accuracy of 96.67%, as shown in Table 1. This outperforms the performance of traditional single-expert models, demonstrating stronger adaptability and robustness. Furthermore, when the model faces scenarios involving the introduction of new fault categories, by integrating a knowledge distillation mechanism, the system can not only quickly adapt and accurately identify the newly emerging fault types, but also maintain the high recognition accuracy of the original fault categories, without significant performance degradation. For example, the confusion matrix analysis results show that in operating condition 1 (20Hz / 0kN): the prediction for category 5 showed a confusion of 0.50-0.50, predicting it as both category 5 and category 6, while the other categories were correctly classified; in operating condition 2 (40Hz / 0kN): the prediction for category 5 was "0.84 (correct) + 0.16 (incorrectly predicted as category 6)," indicating a reduced level of confusion; in operating condition 3 (20Hz / 10kN): all categories achieved "all diagonal elements 1.0," indicating completely accurate classification. The diagonal elements clearly dominate; see details... Figure 4 This indicates that the classification results are highly reliable. Furthermore, the F1-score reached an extreme value of 1.0 in multiple categories, such as... Figure 5 As shown, in scenario 1 (20Hz / 0kN): the precision, recall, and F1 score for category 4 were all low (0.667, 0.500), while the metrics for the other categories were close to 1.0; in scenario 2 (40Hz / 0kN): the metrics for category 4 improved (0.833, 0.909), while the precision for category 8 slightly decreased (0.923); in scenario 3 (20Hz / 10kN): all three metrics for all categories remained at 1.0, indicating stable classification performance; this further confirms the model's excellent balance in handling multiple types of faults. These phenomena fully demonstrate that this method successfully alleviates the catastrophic forgetting problem in continuous learning. Furthermore, through horizontal comparison with current mainstream methods, this approach demonstrates advantages in multiple evaluation metrics, and its overall performance improvement is as follows: Figure 6 As shown, this highlights its potential value in engineering applications.

[0136] Table 1 Comparison of model accuracy under different working conditions

[0137]

[0138] Therefore, this invention adopts the above-mentioned fault diagnosis method and system for train drive motors, gearboxes, and axle boxes. By integrating the fault characteristics and interactions of multiple components through a hybrid expert architecture and gating mechanism, it avoids the complexity of independent modeling and simplifies the training and management process. It features efficient incremental learning without requiring a large amount of new data. Through classifier expansion and knowledge distillation, the model update latency is reduced, allowing for rapid adaptation to emerging new fault modes. It avoids catastrophic forgetting, as historical fault knowledge is solidified through distillation loss, maintaining the diagnostic accuracy of existing faults after training with new faults. It has strong adaptability to multiple operating conditions, maintaining high diagnostic accuracy under different speeds and load conditions. It has strong engineering practicality, with a smaller total number of model parameters compared to existing deep learning methods. It is suitable for onboard and edge device deployments, and its diagnostic accuracy is superior to mainstream methods such as OKD-MTDN and BDC.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fault diagnosis method for a train drive motor, gearbox, and axle box, characterized in that, Includes the following steps: S1. Construct dedicated expert networks for multiple components: Design dedicated expert networks for each key component of the train transmission system; the key components include the gearbox, motor, left axle box, and right axle box, and the dedicated expert networks for each are designed as follows: LHE-Net, a network of experts for gearboxes and motors, employs a multi-layer one-dimensional convolutional structure to extract high-dimensional discriminative features of vibration signals through convolutional blocks. Each convolutional block sequentially includes convolution operations, batch normalization (BN), ReLU activation function, and max pooling (MaxPool). Axle box expert network RE-Net: Based on the residual block structure, it alleviates the degradation of deep networks through identity mapping and realizes the extraction of weak fault features of axle boxes. The residual block includes convolution operation, BN, dropout regularization and shortcut connection. S2. Perform self-attention gating dynamic feature fusion: Introduce a multi-head self-attention gating network to perform sample-level dynamic fusion of the features output by the dedicated expert network; specifically: project the original features of the dedicated expert network into query Q, key K, and value V, and enhance the features by scaling dot product attention calculation; Context information is obtained by global average pooling of the enhanced features, and expert contribution weights are generated by MLP to satisfy the constraints that the weight sum is 1 and non-negativity. Based on the contribution weights, the multi-expert enhanced features are weighted and fused, and the weights are normalized using the Softmax function to obtain system-level fused features. The specific steps are as follows: S21, Enhanced Bullish Attention: The h Each attention head projects expert features into query Q, key K, and value V: ； in, For learnable parameter matrix, For the original feature set from multiple experts; scaling dot product attention calculation: ； in, For single-head attention Feature dimensions; enhanced features are obtained by projecting the concatenated outputs of multiple attention heads. : ； in, For the number of attention heads, To output the projection matrix; S22, Dynamic Weight Generation: For Enhanced Features Global average pooling is used to obtain contextual information, and expert contribution weights are generated through a two-layer MLP: ； in, For average pooling characteristics, The weight matrix of the MLP. For bias terms, T The temperature parameter is used to control the sharpness of the weight distribution, satisfying... ; S23, Feature-weighted fusion: ； in, For the first i The weight of each expert's contribution For the first i Enhanced features of individual experts; S3. Multi-objective joint training: Employing cross-entropy loss. KL divergence regularization loss Load balancing losses The joint loss function is used to train a dedicated expert network and a gated network; S4. Construct an incremental learning and training mechanism: Based on the teacher-student model architecture and knowledge distillation, realize the adaptation of new fault categories and the retention of historical knowledge; S5. Fault Diagnosis Output: Input the vibration signals of the multi-component components to be diagnosed into the trained model. Through feature extraction by a dedicated expert network, fusion by a gating network, and prediction by a classifier, the fault category with the highest probability is output as the diagnosis result.

2. The fault diagnosis method for train drive motor, gearbox, and axle box according to claim 1, characterized in that, In S1, the convolutional block parameters of the LHE-Net satisfy: ； kernel size is The number of output channels is One-dimensional convolution, setting the kernel size and number of output channels, padding parameters. Adaptively adjusts based on kernel size and input sequence length; The l-th convolutional block is defined as the kernel size. Number of output channels The one-dimensional convolutional features are calculated as follows: ； in, For the weight of the l-th layer, connect the l-th layer. Layer Input channel and the l-th layer The first output channel Each weight; For the first Layer aisle, Input feature value at time t, For filling parameters, For the first Layer Channel bias term; The overall feature mapping of the network is as follows: ； in, For the overall mapping of this network, For residual networks, Indicates input data, Indicates the number of convolutional layers; This indicates a combination of functions, where the output of the previous layer serves as the input of the next layer. For all trainable parameters, For global average pooling, For classifiers.

3. The fault diagnosis method for train drive motors, gearboxes, and axle boxes according to claim 2, characterized in that, In S1, the RE-Net contains several residual blocks, and the output of each residual block is... ,in For input features, The residual function is used; the complete forward mapping of RE-Net is: ； For left / right axle box signals, ; For input convolutional layers, For the first One residual module.

4. The fault diagnosis method for train drive motors, gearboxes, and axle boxes according to claim 3, characterized in that, In S3, the joint loss function is: ； Wherein, λ1 and λ2 are hyperparameters, determined through cross-validation; Minimize classification error To prevent the gating weights from collapsing to a single expert, Experts are encouraged to use their expertise in a balanced manner; the specific calculation formula is as follows: ； ； ； in, It is the total number of samples. Calculate the gating weights of the nth sample KL divergence between u and a uniform distribution u The core function of symbols is to establish an asymmetric comparative relationship.

5. The fault diagnosis method for train drive motor, gearbox, and axle box according to claim 4, characterized in that, S4 specifically includes the following steps: S41. Teacher Model Initialization: Based on pre-training with historical fault data, parameters are frozen as a historical knowledge repository, and these parameters are inherited from the parameters trained in the previous round of student model training; specifically: ;in, For the first Rotate student model parameters; S42. Student Model Initialization: Expand the classifier to adapt to the newly added fault categories. The classifier weights are divided into historical fault weights. and new fault weights ,Keep Align with the teacher model output; S43. Incremental training, employing knowledge distillation loss. and cross-entropy loss The student model is trained using a total loss function, which updates only the classifier weights and gating network parameters. The total loss function is: Where α is the weighting coefficient, 0≤α≤1; Used to retain historical fault knowledge Used to adapt to both old and new fault classifications.

6. The fault diagnosis method for train drive motor, gearbox, and axle box according to claim 5, characterized in that, In S43, knowledge distillation loss The calculation method is as follows ,in, T For temperature parameters, For batch size, It is a category. This represents the number of historical fault categories. Softening the probability for the teacher model. The formula for calculating the softening probability of the student model is as follows: ； in, For the teacher model The first sample Logit for historical failures, The student model corresponds to the logit; The cross-entropy loss The calculation method is as follows: ,in, This represents the total number of categories after the addition of faults. For the first The first sample The true label of the class, This is the logit output for the student model.

7. A fault diagnosis system for a train drive motor, gearbox, and axle box implementing the method as described in any one of claims 1-6, characterized in that, include: Signal acquisition module: Triaxial accelerometers deployed in the train transmission system gearbox, motor, left axle box, and right axle box are used to acquire vibration signals from each component; the input signals of the signal acquisition module must be compatible with the signal format of the bogie dataset; the sampling length and timing characteristics of the input signals must match the signal attributes of the bogie dataset; Data preprocessing module: Performs sliding window truncation on the collected vibration signals to generate fixed-length input samples; divides the dataset into training set, test set and validation set according to the proportion; Differentiated Hybrid Expert Diagnostic Module: Includes a dedicated expert network and a self-attention gating fusion submodule, used to achieve differentiated extraction and dynamic fusion of multi-component features; Incremental learning training module: includes teacher model and student model. The teacher model stores historical fault knowledge, and the student model learns new fault categories and retains historical knowledge through extended classifier and knowledge distillation. Diagnostic output module: Receives the fusion features from the differentiated hybrid expert diagnostic module, outputs the predicted probabilities of various faults through a classifier, and selects the category with the highest probability as the final diagnostic result.

8. The fault diagnosis system for train drive motors, gearboxes, and axle boxes according to claim 7, characterized in that, The training process of the incremental learning training module includes: the first stage, training the teacher model based on historical fault data and freezing all its parameters; the second stage, initializing the student model parameters and expanding the classifier to adapt to the newly added fault categories; and the third stage, training the total loss function and updating only the classifier weights and gating network parameters of the student model to complete the incremental learning. The diagnostic output module also includes a dedicated visualization unit that outputs confusion matrices, precision, recall, and F1 scores in real time, as well as the probability distribution of fault category predictions, to help engineers locate the source of the fault.

Images