Intelligent diagnosis system for rail train axle box bearing fault and working method thereof

Abstract

This invention discloses an intelligent diagnostic system and its working method for axle box bearing faults in railway trains. The system includes a data preprocessing module, a model inference module, a human-computer interaction module, and a backend service module. The data preprocessing module receives and segments vibration signals, performs continuous wavelet transform to generate time-frequency images, and the model inference module incorporates a pre-trained TBAFP-GroupConv-KD model for feature extraction and classification inference. The human-computer interaction module provides an interface for data uploading and result display. The backend service module is deployed on a cloud server or edge computing platform to schedule data processing and model inference. This invention proposes an axle box bearing fault diagnosis method based on top-down and bottom-up attention feature pyramids and grouped convolution, combined with a knowledge distillation strategy, to improve the model's fault feature extraction capability, reduce the number of parameters, and maintain high diagnostic accuracy even in noisy environments, demonstrating good robustness and engineering application value.

Description

Technical Field

[0001] This invention relates to the field of rail transit equipment and intelligent operation and maintenance technology, and in particular to an intelligent diagnostic system for axle box bearing failures in rail trains and its working method. Background Technology

[0002] Axle box bearings are critical components in railway trains, and their performance directly affects the safety and stability of the train. However, due to the complex operating conditions of high speed, heavy load, and continuous vibration during train operation, axle box bearings are highly susceptible to damage. Failure to detect and address these issues promptly can lead to serious accidents. Therefore, researching efficient and accurate intelligent diagnostic methods is of great significance for ensuring the reliable operation of railway transportation systems.

[0003] Current bearing fault diagnosis technologies have mainly gone through three stages: traditional signal processing and machine learning methods, deep learning methods, and multi-scale feature fusion methods. Traditional methods rely on FFT, wavelet transform, empirical mode decomposition, and other techniques to extract features, and use classifiers such as SVM and ELM, but they lack robustness in noisy environments. Deep learning methods such as CNN and LSTM can automatically extract features, reducing manual intervention, and some researchers have improved performance by introducing ResNet, CapsNet, and attention mechanisms, but they have limitations in fixed-scale feature extraction. To address these issues, multi-scale feature fusion methods have emerged, such as Feature Pyramid Network (FPN), Path Aggregation Network (PAN), and multi-branch convolutional and attention fusion methods. While improving model performance, these methods inevitably increase computational complexity, limiting real-time applications.

[0004] However, current multi-scale feature fusion methods still have the following shortcomings: (1) Cross-scale feature interaction is mainly limited to adjacent layers, lacking dynamic fusion between non-adjacent layers; (2) Multi-scale feature fusion often brings large computational overhead, resulting in limited real-time performance; (3) Existing methods are not robust enough in strong noise environments. Therefore, it is necessary to propose an intelligent diagnostic method and system that can achieve cross-scale dynamic feature fusion while maintaining lightweight and high robustness. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent diagnostic system for axle box bearing failures in railway trains, addressing the technical deficiencies existing in the prior art.

[0006] Another objective of this invention is to provide a working method for an intelligent diagnostic system for axle box bearing failures in railway trains.

[0007] The technical solution adopted to achieve the purpose of this invention is: A working method for an intelligent diagnostic system for axle box bearing faults in railway trains includes the following steps: Step 1, Data Acquisition and Preprocessing: The raw vibration acceleration signal of the bearing is acquired by an acceleration sensor installed on the axle box of the railcar; the vibration acceleration signal is segmented according to a fixed window length and overlap rate, and continuous wavelet transform is performed on each segment to obtain a two-dimensional time-frequency image; Step 2, Model Construction and Training: A TBAFP-GroupConv-KD model is constructed using a ResNet18 network as the feature extraction backbone, integrating the TBAFP module, GroupConv module, and knowledge distillation strategy. The ResNet18 network processes the two-dimensional time-frequency image obtained in Step 1 to obtain feature maps, which are then transmitted to the TBAFP module. The TBAFP module fuses and enhances the received feature maps, outputting multi-scale features. The GroupConv module processes the multi-scale features output by the TBAFP module by grouping the feature maps with convolutional kernels and performing independent convolution operations within each group, outputting the final feature map. And output the category through the classifier. The probability output is used as the prediction result of the student model of the knowledge distillation strategy; the knowledge distillation strategy adopts a pre-trained teacher model, and the output distribution of the softmax function is softened by introducing a temperature parameter T as a soft label and then processed by the total loss function. By balancing soft and hard labels, the training of student models is guided, resulting in a fully trained TBAFP-GroupConv-KD model. Step 3, Fault Diagnosis: After preprocessing the bearing vibration acceleration signal to be diagnosed in Step 1, the fault type of the bearing is output using the TBAFP-GroupConv-KD model trained in Step 2.

[0008] In the above technical solution, the TBAFP module achieves feature fusion and enhancement through the following mechanisms: a bidirectional skip connection mechanism to establish a cross-scale feature transfer path; a local attention mechanism-global cross attention mechanism to calculate feature correlation in the local window and the global scope; and a cross-layer feature reweighting mechanism to adaptively adjust the contribution weights of features at different scales to the final output.

[0009] In the above technical solution, the local attention mechanism of the TBAFP module is based on a local window mask. Constrain the scope of attention calculation and output local attention. The local attention The calculation formula is:

[0010] In the formula, and These are query vectors and key vectors from different feature maps. Let be the dimension of the key vector. It is a value vector; The local window mask The calculation formula is:

[0011] In the formula, The index of the current image patch. The index of the candidate image patch. The total number of image patches, This is a local window radius parameter used to constrain the range of image patches that can be observed during attention calculations.

[0012] In the above technical solution, the cross-attention mechanism outputs features The The calculation formula is:

[0013] In the formula, and These are query vectors and key vectors from different feature maps. For value vectors, is the dimension of the key vector.

[0014] In the above technical solution, the group convolution operation of the GroupConv module is specifically as follows: Input feature map With convolution kernel Divided along the channel dimension Each group undergoes independent convolution operations, and the feature maps output by each group are combined. The final feature map is obtained by concatenating along the channel dimension. The first Feature maps of group output The calculation formula is:

[0015] In the formula, and The first Group input features and convolution kernels, This indicates a convolution operation.

[0016] In the above technical solution, the knowledge distillation strategy employs a pre-trained teacher model, and obtains the category by introducing a temperature parameter T to soften the output distribution of the Softmax function. Softening probability The category Softening probability The calculation formula is as follows:

[0017] In the formula, For the model, give the first The raw scores (logit) calculated for each category. For the model, iterate through all categories in the denominator. The raw score (logit), This refers to the temperature parameter.

[0018] In the above technical solution, the total loss function of the knowledge distillation strategy Loss due to knowledge distillation and cross-entropy loss The weighted composition, the total loss function The calculation formula is as follows:

[0019] In the formula, Learnable weighting factors to balance the two losses; The knowledge distillation loss function The calculation formula is:

[0020] In the formula, For temperature parameters, KL The distillation loss function is based on the Kullback-Leibler divergence. Output for the student model. Output for the teacher model; The cross-entropy loss function The calculation formula is:

[0021] In the formula, For the real category, The result of applying the Softmax function to the raw output of the student model is the result of applying the Softmax function to the true class. The predicted probability.

[0022] In another aspect of the present invention, a system for implementing the intelligent diagnosis method for axle box bearing faults in rail trains includes: a data preprocessing module, a model inference module, a human-computer interaction module, and a backend service module; the data preprocessing module is used to perform data acquisition and preprocessing operations; the model inference module has a pre-trained TBAFP-GroupConv-KD model built in it, which is used to perform feature extraction and fault classification inference; the human-computer interaction module provides an interface for receiving user input and displaying diagnostic results; the backend service module is deployed on a cloud server or edge computing platform, and is used to coordinate data processing and task scheduling among the data preprocessing module, the model inference module, and the human-computer interaction module.

[0023] In the above technical solution, the intelligent diagnostic system for axle box bearing faults of rail trains selectively incorporates a secondary verification module; the secondary verification module calls the large language model API to perform credibility verification and logical consistency analysis on the preliminary diagnostic results output by the model inference module, and generates a final diagnostic report containing explanatory notes.

[0024] In the above technical solution, the human-computer interaction module is built based on the Streamlit framework, the backend service module is deployed on Alibaba Cloud ECS server, and the operating system of the backend service module is Linux.

[0025] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention proposes a fault diagnosis method for axle box bearings based on top-down and bottom-up attention feature pyramids (TBAFP) and group convolution (GroupConv), and combines it with the knowledge distillation (KD) strategy to improve the model's fault feature extraction capability, reduce the number of parameters, and maintain high diagnostic accuracy in strong noise environments, thus possessing good robustness and engineering application value.

[0026] 2. This invention achieves cross-layer and cross-scale dynamic feature fusion through the TBAFP module, overcoming the limitation that traditional FPN and PAN can only interact between adjacent layers.

[0027] 3. This invention reduces model complexity and the number of parameters by using GroupConv, while maintaining feature expressiveness.

[0028] 4. This invention introduces teacher model knowledge through the KD training framework to enhance the generalization ability of student models. Attached Figure Description

[0029] Figure 1 The diagram shown is a schematic representation of the overall architecture of the TBAFP-GroupConv-KD model of this invention.

[0030] Figure 2 The diagram shows the network architecture of ResNet18 of this invention.

[0031] Figure 3 The image shown is an example of a time-frequency image obtained after CWT transformation.

[0032] Figure 4 The image shown is a t-SNE visualization of the output features of each module in the model under -4dB noise.

[0033] Figure 5 The diagram shown is a schematic diagram of the system structure of the present invention.

[0034] Figure 6 The figure shown is a schematic diagram of the system interface of the present invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0036] Example 1 Reference Figure 1 A smart diagnostic method for axle box bearing failure in rail trains includes the following steps: Step 1: Acquire the original vibration acceleration signal of the bearing by an acceleration sensor installed on the axle box of the railcar; divide the vibration signal into segments according to a fixed window and overlap rate, and perform continuous wavelet transform (CWT) on each segment to obtain a two-dimensional time-frequency image.

[0037] Specifically, the original vibration signal is processed by non-overlapping slicing, with each segment consisting of 2048 points. Then, a continuous wavelet transform (CWT) is performed using Morlet wavelets, with a scale range of 1–128, to extract feature information at multiple frequencies. The transformed time-frequency coefficients are saved as a PNG image in heatmap form, i.e., a two-dimensional time-frequency image, such as... Figure 3 As shown, the differences in energy distribution under different fault types are clearly visible. 2500 sample two-dimensional time-frequency images were generated for each fault type, totaling 20000 samples. The resulting two-dimensional time-frequency image dataset was divided into training, validation, and test sets in a 7:1:2 ratio, and five-fold cross-validation was used to ensure the robustness of the experimental results.

[0038] Step 2: Train the TBAFP-GroupConv-KD model using the two-dimensional time-frequency images in the training set from Step 1, and test and validate the trained TBAFP-GroupConv-KD model using the test set and validation set. The TBAFP-GroupConv-KD model is built on the ResNet18 network and integrates a top-down and bottom-up bidirectional attention feature pyramid (TBAFP) module, a group convolution (GroupConv) module, and a knowledge distillation strategy (KD). Figure 2 The ResNet network processes the two-dimensional time-frequency image to obtain a 7×7×512 feature map, which is then transmitted to the TBAFP module. The TBAFP module fuses and enhances the feature map output by the feature extraction backbone network through cross-layer skip connections, local-global cross-attention, and feature reweighting mechanisms, outputting multi-scale features. The bidirectional skip connections enable non-adjacent cross-layer interactions between high-level and low-level features; the local-global cross-attention mechanism preserves low-level details and high-level semantic features in the skip connection path; and the cross-layer feature reweighting mechanism adaptively adjusts the contribution of features at different scales to the prediction results. The GroupConv module groups the multi-scale features output by the TBAFP module with the feature map and convolutional kernels, performing independent convolution operations within each group to reduce model computational complexity and improve feature diversity, outputting the final features. Finally, a classifier outputs the classification. As the student model for the knowledge distillation (KD) strategy; the knowledge distillation (KD) strategy uses ResNet50 as the teacher model, employs the temperature parameter T to adjust the output distribution of the Softmax function as soft labels, and uses the total loss function Balance soft / hard labels to improve the performance of student models; The cross-attention mechanism output features of the TBAFP module The output features The calculation formula is:

[0039] In the formula, and These are query vectors and key vectors from different feature maps. For value vectors, is the dimension of the key vector.

[0040] The local attention mechanism of the TBAFP module is based on local window masks. Constrain the scope of attention calculation and output local attention. The local attention The calculation formula is:

[0041] In the formula, and These are query vectors and key vectors from different feature maps. Let be the dimension of the key vector. For local window mask, It is a value vector; The local window mask The calculation formula is:

[0042] In the formula, The index of the current image patch. The index of the candidate image patch. The total number of image patches, The local window radius parameter is used to constrain the range of image patches that can be observed during attention calculations for the current image patch.

[0043] The group convolution operation of the GroupConv module specifically involves: converting the input feature map... With convolution kernel Divided along the channel dimension Groups, each group performs convolution operations independently, and the outputs of each group are... The feature maps are concatenated along the channel dimension to output the final feature map. The final feature map Defined as:

[0044] In the formula, For the first The output feature map group obtained after performing convolution operation on each group For the first Grouped input features, For the first Each group corresponds to a convolutional kernel. The number of groups. The number of groups. Feature maps of group output The calculation formula is:

[0045] In the formula, and The first Group input features and convolution kernels, This indicates a convolution operation.

[0046] The KD strategy uses the temperature parameter T to adjust the output distribution of the Softmax function to obtain the category. Softening probability The calculation formula is as follows:

[0047] In the formula, For the model, give the first The raw scores (logit) calculated for each category. For the model, iterate through all categories in the denominator. The raw score (logit), This refers to the temperature parameter.

[0048] The total loss function The calculation formula is:

[0049] In the formula, For the knowledge distillation loss function, Let cross-entropy be the loss function. A learnable balance factor; The knowledge distillation loss function The calculation formula is:

[0050] In the formula, For temperature parameters, KL The distillation loss function is based on the Kullback-Leibler divergence. Output for the student model (TBAFP-GroupConv) Output for the teacher model (ResNet50); The cross-entropy loss function The calculation formula is:

[0051] In the formula, For the real category, The result of applying the Softmax function to the raw output of the student model is the result of applying the Softmax function to the true class. The predicted probability.

[0052] To evaluate the effect of temperature parameter T on the output distribution of distillation Softmax, a systematic experiment was conducted, and the results are shown in Table 1. When T=3, the model converged in the 74th epoch with an accuracy of 0.9945, demonstrating the best classification performance and convergence speed. Therefore, T=3 was fixed in subsequent training.

[0053] Table 1. Impact of different T values ​​on model classification performance

[0054] In this embodiment, to verify the generalization ability of the TBAFP-GroupConv model, a publicly available bearing dataset provided by the University of Paderborn, Germany, was selected for experiments. The rolling bearing, model 6203, underwent accelerated life testing on a test bench, encompassing various damage types and levels. Specifically, the fault types included: normal, inner ring single-point fault (IR), outer ring single-point fault (OR), and combined inner and outer ring multi-point fault (IR+OR); the damage levels were divided into three grades: 2 mm (Grade 1), 4.5 mm (Grade 2), and 13.5 mm (Grade 3). This resulted in eight different operating condition categories. The experimental conditions were set as follows: rotational speed 1500 rpm, torque 0.7 Nm, radial load 1000 N, and sampling frequency 64 kHz.

[0055] Parameter configuration: During model training, the Adam optimizer (β1=0.9, β2=0.999) is used, with an initial learning rate of 0.001, which decays by a factor of 0.1 every 10 epochs. A learning balance factor is also used. α Initialize to 0.5. Dropout is set to 0.3 in the fully connected layer to enhance generalization, while L2 regularization (weight decay of 1e) is introduced. 4) An early stopping mechanism (patience value of 10 epochs) is used to avoid overfitting. The remaining hyperparameter configurations are shown in Table 2.

[0056] Table 2 Hyperparameter Configuration

[0057] Step 3: Input the time-frequency image obtained in Step 1 into the TBAFP-GroupConv-KD deep learning model trained in Step 2 for feature extraction and fault identification, and output the diagnostic results of the health status and fault type of the axle box bearing.

[0058] Comparative Example The baseline model ResNet18 is denoted as Baseline. The baseline model ResNet18 with the TBAFP module is denoted as Baseline+TBAFP. The baseline model ResNet18 with the TBAFP module and the GroupConv module is denoted as Baseline+TBAFP+GroupConv. For comparison, the baseline model of this invention with the TBAFP module, the GroupConv module and the knowledge distillation strategy (KD) is denoted as Baseline+TBAFP+GroupConv+KD.

[0059] Example of effect Feature visualization results: such as Figure 4The feature visualization results further reveal the improvement effect of each module on feature distribution. Under the baseline, the feature distributions of different categories severely overlap, indicating that the model's ability to distinguish between different categories is limited under strong noise interference. After introducing the TBAFP mechanism, the intra-class clustering effect of Baseline+TBAFP is significantly enhanced; after adding the GroupConv module, the feature distribution of Baseline+TBAFP+GroupConv is more compact, and the inter-class boundaries are clearer. Finally, combined with the model distillation mechanism, each category of Baseline+TBAFP+GroupConv+KD forms independent and well-defined clusters in the t-SNE space, fully demonstrating that the present invention, Baseline+TBAFP+GroupConv+KD, has excellent feature extraction capabilities and robustness in complex noise environments.

[0060] Accuracy Comparison: To further verify the diagnostic performance of the model under different noise environments, this embodiment compares the performance of Baseline, Baseline+TBAFP, Baseline+TBAFP+GroupConv, and Baseline+TBAFP+GroupConv+KD. Experiments were conducted under signal-to-noise ratio (SNR) conditions of 0 dB, –4 dB, and –8 dB, respectively. Evaluation metrics included Matthews correlation coefficient (MCC), accuracy (Acc), precision (Pre), F1 score (F1), number of parameters (Para), and significance p-value based on MCC. As shown in Table 3, under noise-free conditions (SNR=0), the Baseline+TBAFP+GroupConv+KD model achieved optimal performance across all metrics, with an accuracy of 99.45% and an MCC of 0.9937, significantly outperforming the Baseline model. Under moderate noise conditions (SNR=–4), Baseline+TBAFP+GroupConv+KD maintained high robustness, with accuracy increasing to 96.48% and an MCC of 0.9599. Under high noise conditions (SNR = –8), the Baseline+TBAFP+GroupConv+KD still achieved an accuracy of 90.46% and an MCC value of 0.8907, demonstrating that the present invention can maintain excellent diagnostic performance in extreme noise environments. The gradual introduction of each module brought significant performance improvements, and their p-values ​​all met the statistical significance (p < 0.05). 0.05).

[0061] Table 3 Comparison of Accuracy of Different Models

[0062] Example 2 In this embodiment, to verify the portability and practical application effect of the intelligent diagnostic method, the present invention constructs and deploys a complete intelligent diagnostic system for axle box bearing faults in railway trains, referring to... Figure 5 The intelligent diagnostic system for axle box bearing faults in rail trains includes: The system comprises a data preprocessing module, a model inference module, a human-computer interaction module, and a backend service module. The data preprocessing module receives and segments vibration signals, performing continuous wavelet transform to generate time-frequency images. The model inference module uses a pre-trained TBAFP-GroupConv-KD model for feature extraction and classification inference. The human-computer interaction module provides an interface for data uploading and result display. The backend service module is deployed on a cloud server or edge computing platform to schedule data processing and model inference. The human-computer interaction module is based on the Streamlit framework, and the backend service module is deployed on an Alibaba Cloud Linux server, supporting concurrent inference and remote calls. The intelligent diagnostic system for axle box bearing faults in railway trains selectively incorporates a secondary verification module. This module calls a large language model API to verify the credibility of the diagnostic results output by the model inference module and generates an interpretive diagnostic report. The construction and application process of the intelligent diagnostic system for axle box bearing faults in railway trains is as follows: 1. Environment Setup This system is deployed on an Alibaba Cloud ECS server running Ubuntu and equipped with an Nvidia RTX 4070 GPU to ensure high-concurrency computing capabilities and stable inference performance. In terms of software, deep learning frameworks such as Python 3.8 and PyTorch 1.13, along with their dependencies, are installed and configured, providing a reliable platform for model execution and inference.

[0063] 2. Service Deployment After training, the TBAFP-GroupConv-KD model weight file (.pth format) is loaded into the PyTorch model in the model inference module. The front-end uses the Streamlit framework to develop the web application interface, including a data upload window and a result display module; the back-end implements the inference logic and task scheduling, and is responsible for receiving user input, performing signal preprocessing in the data preprocessing module (such as calling Continuous Wavelet Transform (CWT) if necessary), and the back-end service module runs the TBAFP-GroupConv-KD model to complete fault classification and returns the results to the front-end.

[0064] 3. Application Process The human-computer interaction module enables users to access the deployed system interface (such as...) through a browser. Figure 6As shown, the access address is: http: / / 7e53fa29.r7.cpolar.top. During use, users can drag and drop to upload vibration data files (.csv format) or time-frequency images (.png format), and input operating parameters such as rotational speed n, torque N, and sampling frequency fs. The system backend will automatically call the preprocessing module to generate model input, and the inference module will output the corresponding fault category. Furthermore, the system can call the OpenAI GPT API to generate a natural language explanation report for the diagnostic results, and the diagnostic conclusion and interpretable text will be displayed on the user interface.

[0065] 4. Engineering Application Effects Verification using real vehicle data shows that the system of this invention can effectively identify various operating conditions, including normal operation, outer ring fault, inner ring fault, rolling element fault, and different degrees of damage, with an overall diagnostic accuracy of 95.16%. Experimental results demonstrate that the system of this invention, while ensuring high-precision diagnosis, possesses good real-time performance and stability, meeting the requirements for high reliability and interpretability in rail transit engineering applications.

[0066] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A working method for an intelligent diagnostic system for axle box bearing faults in railway trains, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: The raw vibration acceleration signal of the bearing is acquired by an acceleration sensor installed on the axle box of the railcar; the vibration acceleration signal is segmented according to a fixed window length and overlap rate, and continuous wavelet transform is performed on each segment to obtain a two-dimensional time-frequency image; Step 2, Model Construction and Training: A TBAFP-GroupConv-KD model is constructed using a ResNet18 network as the feature extraction backbone, integrating the TBAFP module, GroupConv module, and knowledge distillation strategy. The ResNet18 network processes the two-dimensional time-frequency image obtained in Step 1 to obtain feature maps, which are then transmitted to the TBAFP module. The TBAFP module fuses and enhances the received feature maps, outputting multi-scale features. The GroupConv module processes the multi-scale features output by the TBAFP module by grouping the feature maps with convolutional kernels and performing independent convolution operations within each group, outputting the final feature map. And output the category through the classifier. The probability output is used as the prediction result of the student model of the knowledge distillation strategy; the knowledge distillation strategy adopts a pre-trained teacher model, and the output distribution of the softmax function is softened by introducing a temperature parameter T as a soft label and then processed by the total loss function. By balancing soft and hard labels, the training of student models is guided, resulting in a fully trained TBAFP-GroupConv-KD model. Step 3, Fault Diagnosis: After preprocessing the bearing vibration acceleration signal to be diagnosed in Step 1, the fault type of the bearing is output using the TBAFP-GroupConv-KD model trained in Step 2.

2. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 1, characterized in that, The TBAFP module achieves feature fusion and enhancement through the following mechanisms: a bidirectional skip connection mechanism to establish cross-scale feature transfer paths; and a local attention mechanism-global cross attention mechanism to calculate feature correlations within a local window and globally. A cross-layer feature reweighting mechanism is used to adaptively adjust the contribution weights of features at different scales to the final output.

3. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 2, characterized in that, The local attention mechanism of the TBAFP module is based on local window masks. Constrain the scope of attention calculation and output local attention. The local attention The calculation formula is: In the formula, and These are query vectors and key vectors from different feature maps. Let be the dimension of the key vector. It is a value vector; The local window mask The calculation formula is: In the formula, This is the index of the current image patch. The index of the candidate image patch. The total number of image patches. This is a local window radius parameter used to constrain the range of image patches that can be observed during attention calculations.

4. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 2, characterized in that, The cross-attention mechanism outputs features The The calculation formula is: In the formula, and These are query vectors and key vectors from different feature maps. For value vectors, is the dimension of the key vector.

5. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 1, characterized in that, The group convolution operation of the GroupConv module is specifically as follows: Input feature map With convolution kernel Divided along the channel dimension Each group undergoes independent convolution operations, and the feature maps output by each group are combined. The final feature map is obtained by concatenating along the channel dimension. The first Feature maps of group output The calculation formula is: In the formula, and The first Group input features and convolution kernels, This indicates a convolution operation.

6. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 1, characterized in that, The knowledge distillation strategy employs a pre-trained teacher model, which softens the output distribution of the Softmax function by introducing a temperature parameter T to obtain the category. Softening probability The category Softening probability The calculation formula is as follows: In the formula, For the model, give the first The raw scores (logit) calculated for each category. For the model, iterate through all categories in the denominator. The raw score (logit), This refers to the temperature parameter.

7. The working method of the intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 1, characterized in that, The total loss function of the knowledge distillation strategy Loss due to knowledge distillation and cross-entropy loss The weighted composition, the total loss function The calculation formula is as follows: In the formula, Learnable weighting factors to balance the two losses; The knowledge distillation loss function The calculation formula is: In the formula, For temperature parameters, KL The distillation loss function is based on the Kullback-Leibler divergence. Output for the student model. Output for the teacher model; The cross-entropy loss function The calculation formula is: In the formula, For the real category, The result of applying the Softmax function to the raw output of the student model is the result of applying the Softmax function to the true class. The predicted probability.

8. A system for implementing the intelligent diagnosis method for axle box bearing faults in rail trains according to any one of claims 1-7, characterized in that, include: The system comprises a data preprocessing module, a model inference module, a human-computer interaction module, and a backend service module. The data preprocessing module performs the data acquisition and preprocessing operations described in step 1 of claim 1. The model inference module incorporates a pre-trained TBAFP-GroupConv-KD model for feature extraction and fault classification inference. The human-computer interaction module provides an interface for receiving user input and displaying diagnostic results. The backend service module is deployed on a cloud server or edge computing platform to coordinate data processing and task scheduling among the data preprocessing module, the model inference module, and the human-computer interaction module.

9. The intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 8, characterized in that, The intelligent diagnostic system for axle box bearing faults in rail trains selectively incorporates a secondary verification module. The secondary verification module calls the large language model API to perform credibility verification and logical consistency analysis on the preliminary diagnostic results output by the model inference module, and generates a final diagnostic report containing explanatory notes.

10. The intelligent diagnostic system for axle box bearing faults in rail trains as described in claim 8, characterized in that, The human-computer interaction module is built on the Streamlit framework, and the backend service module is deployed on Alibaba Cloud ECS server. The operating system of the backend service module is Linux.

Images