Attention-enhanced dual-path network-based railway carriage axle box cover bolt fault detection method

Abstract

The invention discloses a passenger train axle box cover bolt fault detection method based on an attention-enhanced double-path network, and belongs to the technical field of passenger train fault detection. The method aims at solving the problem that axle box cover bolts are not easy to accurately recognize in a dense and complex background environment. Comprising the steps of constructing a dual-path neural network model, and obtaining axle box cover macroscopic geometric features by adopting a global feature path; using a local detail feature path to obtain an enhanced image microscopic feature; performing attention fusion by adopting a feature fusion module to obtain a fused feature image, and obtaining an axle box cover type, axle box cover direction and bolt state preliminary identification result; and in combination with the predefined mask matrix, transforming the predefined mask matrix to a to-be-identified image axle box cover space by adopting a spatial attitude matching algorithm to carry out spatial matching alignment, and determining a final state of the bolt through intersection operation of the predefined mask matrix and a bolt state preliminary identification result label output by the trained dual-path neural network model. The method is used for bolt fault detection.

Description

Technical Field

[0001] The invention relates to a railway passenger car axle box cover bolt fault detection method based on an attention enhanced dual-path network, and belongs to the technical field of railway passenger car fault detection. Background Art

[0002] The traditional railway passenger car fault detection method that uses manual map checking is time-consuming and labor-intensive, and is easily affected by subjective judgment. Especially when faced with a variety of axle box cover types and uneven bolt distribution, the missed detection and false detection rates are high.

[0003] Since the axle box cover bolts are located in the bogie area where the side parts of the bus are densely and complexly distributed, the bolt image targets are small and relatively blurred. At the same time, there are many types of axle box covers, which leads to the diversity of the shapes of the axle box cover bolt targets. The traditional Faster R-CNN has limitations in small target recognition in dense and complex environments. Summary of the invention

[0004] Aiming at the problem that axle box cover bolts are difficult to accurately identify in dense and complex background environments, the present invention provides a railway passenger car axle box cover bolt fault detection method based on an attention enhanced dual-path network.

[0005] A railway passenger car axle box cover bolt fault detection method based on an attention-enhanced dual-path network of the present invention comprises:

[0006] Construct a dual-path neural network model and axle box cover bolt image sample set;

[0007] The dual-path neural network model includes a global feature path, a local detail feature path and a feature fusion module. The dual-path neural network model is trained using axle box cover bolt image sample set, including:

[0008] The global feature path uses a lightweight convolution module to extract the overall structural layout features of the axle box cover bolt image samples, and then uses a large kernel convolution layer to obtain the macroscopic geometric features of the axle box cover from the overall structural layout features; the local detail feature path first extracts image microscopic features from the axle box cover bolt image samples, and then enhances the details through the channel space attention mechanism to obtain the enhanced image microscopic features;

[0009] The feature fusion module is used to fuse the macroscopic geometric features of the axle box cover and the microscopic features of the enhanced image to obtain a fused feature image.

[0010] The fully connected layer is used to identify the faults of the fused feature images, and the preliminary identification results of the axle box cover type, axle box cover direction and bolt status are output;

[0011] During the training process, the network parameters are adjusted by combining cross entropy loss and positioning loss to obtain a trained dual-path neural network model for axle box cover bolt fault detection.

[0012] In the fault detection task, the corresponding predefined mask matrix is ​​determined according to the axle box cover type and axle box cover direction of the image to be identified output by the trained dual-path neural network model. The predefined mask matrix is ​​transformed to the axle box cover space of the image to be identified by the spatial posture matching algorithm for spatial matching and alignment. The final state of the bolt is determined by the intersection operation of the predefined mask matrix and the preliminary identification result label of the bolt state output by the trained dual-path neural network model, thereby realizing bolt fault detection.

[0013] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, the axle box cover bolt image sample set covers all known axle box cover types and different bolt configurations;

[0014] The axle box cover bolt image samples include real collected images and enhanced images; the enhanced images are obtained by performing data enhancement on the real collected images.

[0015] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual-path network of the present invention, the channel space attention mechanism dynamically adjusts the channel and space weights of the image micro features, highlights the bolt state related features and suppresses irrelevant noise, so as to obtain enhanced image micro features.

[0016] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual-path network of the present invention, the channel spatial attention mechanism extracts the channel attention of image micro features through maximum pooling and average pooling operations, and then determines the channel weight of each channel through two fully connected layers; the spatial attention map of image micro features is obtained through convolution operation, and then the spatial weight is determined.

[0017] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, the predefined mask matrix is ​​a template generated by binary image processing method and accurate to pixel level, 0 indicates that there is an illegal area of ​​the bolt, and 1 indicates that there is a legal area of ​​the bolt.

[0018] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, affine transformation and SIFT feature matching method are used to transform the predefined mask matrix to the axle box cover space of the image to be identified for spatial matching and alignment.

[0019] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, the method for determining the final state of the bolt is:

[0020] Calculate the percentage of overlapping pixels between the bolt abnormal label in the preliminary identification result label of the bolt status and the legal area of ​​the bolt in the predefined mask matrix. If the percentage of overlapping pixels exceeds the preset alarm threshold, it is determined as a valid abnormal identification result; otherwise, it is determined as a false identification and filtered out; and the final status of the bolt is determined.

[0021] According to the railway passenger car axle box cover bolt fault detection method based on the attention enhanced dual-path network of the present invention, the preset alarm threshold is 0.55-0.8.

[0022] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, the macro-geometric features of the axle box cover include the shape and size of the axle box cover and the distribution pattern of the bolts thereon.

[0023] According to the railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network of the present invention, the data enhancement method includes rotation, flipping, brightness adjustment and occlusion simulation.

[0024] Beneficial effects of the present invention: The present invention utilizes the global feature path (GFP) and the local detail feature path (LDFP) to capture the macroscopic layout and microscopic features of the image respectively, and LDFP is embedded in the attention mechanism CBAM to improve recognition ability. A multi-task learning framework is adopted to simultaneously predict the type, state and position of the bolt, and the model performance is optimized in combination with the comprehensive loss function design. A post-processing algorithm is designed for the results of the model output to achieve spatial matching and alignment through mask definition, SIFT feature matching and affine transformation, filter unreasonable position false alarms, and ensure accurate detection. The system of the present invention can be directly deployed on edge computing devices, effectively solving the detection challenges of bolt diversity and complex backgrounds, significantly improving detection accuracy and efficiency, realizing real-time detection and immediate maintenance decisions, and providing scientific and technological guarantees for safe railway operation and maintenance.

[0025] The method of the invention combines deep learning with a customized dual-path network model and attention mechanism, realizes accurate identification of bolt status, greatly improves the automation level and accuracy of detection, solves the problems of low efficiency and insufficient precision of manual detection, and ensures the safety of railway transportation while reducing costs. Through targeted optimization, the accurate detection capability of the axle box cover bolt fault detection in the bogie area is effectively improved.

[0026] The dual-path neural network model structure of this implementation effectively integrates feature information of different scales, improves recognition capabilities, and solves the limitations of traditional single-path networks in recognizing small targets in complex environments. In multi-task learning and loss function optimization, shared representation learning is used to promote knowledge transfer between tasks and optimize resource utilization. At the same time, the customized design of the loss function combines cross-entropy loss and positioning loss, and by carefully balancing the weights of each loss, the system can simultaneously pursue a macroscopic understanding of global features and accurate capture of local details. The multi-task learning framework and comprehensive loss function design improve the overall performance and efficiency of the model, and enhance the generalization ability and recognition accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the training process of the dual-path neural network model;

[0028] Figure 2 is a schematic diagram of a process for determining the final state of a bolt using a predefined mask matrix;

[0029] Figure 3 It is a schematic diagram of the image processing process status of axle box cover bolt fault detection. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0032] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0033] Specific implementation method 1. Combination Figures 1 to 3 As shown, the present invention provides a railway passenger car axle box cover bolt fault detection method based on an attention-enhanced dual-path network, comprising:

[0034] Construct a dual-path neural network model and axle box cover bolt image sample set;

[0035] The dual-path neural network model includes a global feature path, a local detail feature path and a feature fusion module. The dual-path neural network model is trained using axle box cover bolt image sample set, including:

[0036] The global feature path uses a lightweight convolution module to extract the overall structural layout features of the axle box cover bolt image samples, and then uses a large kernel convolution layer to obtain the macroscopic geometric features of the axle box cover from the overall structural layout features; the local detail feature path first extracts image microscopic features from the axle box cover bolt image samples, and then enhances the details through the channel space attention mechanism to obtain the enhanced image microscopic features;

[0037] The feature fusion module is used to fuse the macroscopic geometric features of the axle box cover and the microscopic features of the enhanced image to obtain a fused feature image.

[0038] The fully connected layer is used to identify the faults of the fused feature images, and the preliminary identification results of the axle box cover type, axle box cover direction and bolt status are output;

[0039] During the training process, the network parameters are adjusted by combining cross entropy loss and positioning loss to obtain a trained dual-path neural network model for axle box cover bolt fault detection.

[0040] In the fault detection task, the corresponding predefined mask matrix is ​​determined according to the axle box cover type and axle box cover direction of the image to be identified output by the trained dual-path neural network model. The predefined mask matrix is ​​transformed to the axle box cover space of the image to be identified by the spatial posture matching algorithm for spatial matching and alignment. The final state of the bolt is determined by the intersection operation of the predefined mask matrix and the preliminary identification result label of the bolt state output by the trained dual-path neural network model, thereby realizing bolt fault detection.

[0041] Furthermore, the axle box cover bolt image sample set covers all known axle box cover types and different bolt configurations; the axle box cover types in the axle box cover bolt image sample set can be timely added with the axle box cover types corresponding to the newly added models, so that the sample types meet the diversity. Currently, it includes all axle box cover types of about 5 types of EMU vehicles and 5 types of ordinary passenger vehicles, and can be supplemented according to the newly added platform models in the future.

[0042] The axle box cover bolt image samples include real collected images and enhanced images; the enhanced images are obtained by performing data enhancement on the real collected images.

[0043] In this embodiment, the real collected image is captured by high-definition imaging technology.

[0044] In this embodiment, the channel space attention mechanism is used to dynamically adjust the channel and space weights of the image micro features, highlight the bolt state related features and suppress irrelevant noise, so as to obtain enhanced image micro features.

[0045] As an example, the channel spatial attention mechanism extracts the channel attention of image micro features through maximum pooling and average pooling operations, and then determines the channel weight of each channel through two fully connected layers; the spatial attention map of image micro features is obtained through convolution operations, and then the spatial weight is determined.

[0046] The dual-path neural network model constructed in this embodiment is designed to accurately identify the state of the axle box cover bolts. It integrates the global feature path (GFP) and the local detail feature path (LDFP) to effectively integrate the macro layout and micro features to improve the recognition ability. The specific description is as follows:

[0047] Global Feature Path (GFP): This path uses a lightweight convolutional module to capture the overall structural layout information of the axle box cover. Through the convolutional layer with a large receptive field, GFP can extract the macroscopic geometric features of the axle box cover, such as shape, size, and approximate bolt distribution pattern, to provide context for subsequent bolt status judgment.

[0048] Local Detail Feature Path (LDFP): In parallel with GFP, LDFP focuses on the microscopic feature extraction of bolts, using higher-resolution feature maps to capture the detailed morphological changes of bolts. To enhance the perception of key details, an advanced attention mechanism CBAM (Convolutional Block Attention Module) is embedded in this path to dynamically adjust the channel weights in the feature map, highlight the features related to the bolt state, and suppress irrelevant noise.

[0049] The core of the dual-path neural network model structure is to separate and optimize the visual features of different scales: on the one hand, the global path (GFP) uses a lightweight convolutional structure to focus on extracting the overall layout and structural information of the axle box cover, such as the outline, size ratio, and the approximate distribution pattern of the bolts, to build a macro background framework for subsequent analysis. On the other hand, the local path (LDFP) goes deep into the details, focusing on analyzing the individual characteristics of the bolts, and using high-resolution feature representation to capture subtle morphological differences. The combination of the two can accurately distinguish the bolt status even in cases with similar morphology but actual differences.

[0050] Attention optimization of LDFP: Since the local feature path needs to accurately capture the morphological changes of the bolts, the advanced CBAM (Convolutional Block Attention Module) attention mechanism is embedded. CBAM focuses on the key areas of the bolts by dynamically adjusting the channel and spatial weights of the feature map, while suppressing irrelevant background interference, significantly enhancing the recognition of sensitive features of the bolt state. It first extracts channel attention through maximum pooling and average pooling operations, then determines the weight of each channel through two fully connected layers, and then uses convolution operations to obtain the spatial attention map to emphasize important areas. This mechanism not only considers which features are important (channel attention), but also where these features are important (spatial attention), thereby highlighting the details of the bolts in complex images and improving the accuracy and robustness of recognition.

[0051] Feature fusion and comprehensive analysis: The outputs of GFP and LDFP are combined through a feature fusion layer, which adopts an attention-based fusion strategy to effectively integrate global structural information and local detail features, providing a comprehensive and rich feature basis for subsequent judgment.

[0052] Multi-task learning framework: Multi-task learning strategies are incorporated into model training to simultaneously predict the bolt type, state and position, as well as the identification of the axle box cover type and direction. Joint training is performed within the same network framework to promote knowledge transfer between tasks through shared representation learning, optimize resource utilization, effectively reduce the number of model parameters, enhance the generalization ability of the model, and improve learning efficiency and recognition accuracy.

[0053] Customized design of loss function: In order to adapt to multi-task requirements, this implementation adopts a comprehensive loss function, combining cross entropy loss to optimize classification tasks, and possible positioning loss to ensure coordinated optimization of various task objectives. Through multiple experiments with different parameters to balance the weights of each loss, it is possible to simultaneously pursue a macro understanding of global features and accurate capture of local details to achieve the best recognition effect.

[0054] Combination Figure 1 As shown in the figure, the dual-path neural network model is used to process images containing axle box covers to achieve accurate recognition of the bolt status on them. In the initialization stage, GFP initially uses a lightweight convolutional layer to extract features and reduce the computational burden. The subsequent large-core convolutional layer analyzes macro features such as shape, size, and bolt layout. At the same time, LDFP extracts high-resolution details and uses the CBAM module to dynamically emphasize bolt-related features and suppress irrelevant information. The enhanced features are combined in the fusion stage to integrate global and local perspectives. Finally, the fused features are used to judge the bolt status through a fully connected layer, and a multi-task loss function (category cross entropy and positioning loss) is used to balance and optimize global and local feature recognition to achieve accurate detection.

[0055] The following is an explanation of the post-processing algorithm for the output results of the dual-path neural network model:

[0056] This implementation method uses a post-processing algorithm to accurately identify the type, position and bolt distribution of the axle box cover, while ensuring the efficiency and accuracy of bolt status detection, greatly reducing the false alarm rate. The specific implementation steps are as follows:

[0057] The regional mask is adaptively generated based on the preliminary recognition results of the axle box cover type, axle box cover direction and bolt status output by the dual-path neural network model:

[0058] According to the type of axle box cover, the corresponding predefined mask matrix is ​​automatically called. The predefined mask matrix is ​​a template generated by binary image processing method and accurate to pixel level. 0 indicates that there are illegal areas with bolts, and 1 indicates that there are legal areas with bolts. Taking into account the diversity of axle box cover structures, the mask matrix specifically marks the legal areas where bolt holes are expected to exist (value is 1) and the inherent bolt-free vacancies or illegal areas of the axle box cover (value is 0) to distinguish normal structures from potential fault areas. In order to adapt to the different postures of the axle box cover in the image (rotation, horizontal flipping, etc.), a posture matching algorithm is introduced to ensure that the generated mask is consistent with the actual axle box cover direction according to the axle box cover direction information output by the model, so as to ensure the accuracy of subsequent analysis.

[0059] Affine transformation and SIFT feature matching methods are used to transform the predefined mask matrix to the axle box cover space of the image to be identified for spatial matching and alignment. SIFT feature matching and transformation calculation methods are used to achieve registration, and affine transformation (Affine Transformation) and SIFT (Scale-Invariant Feature Transform) feature matching methods are applied to ensure that it is aligned with the spatial layout of the identified axle box cover image.

[0060] Combination Figure 2 and Figure 3 As shown, the method to determine the final state of the bolt is:

[0061] Traverse all the abnormal bolt labels identified by the model, and use the location information of these labels to perform spatial intersection operations with the previously generated and matched masks.

[0062] The percentage of overlapping pixels between the bolt abnormal label in the preliminary identification result label of the bolt status and the legal area where the bolt exists in the predefined mask matrix (the part with a value of 1 in the mask) is calculated. If the percentage of overlapping pixels exceeds the preset alarm threshold, it is determined as a valid abnormal identification result, indicating that the bolt is indeed missing or abnormal; otherwise, it is determined as a misidentification and filtered out; and the final status of the bolt is determined.

[0063] The preset alarm threshold range is selected as 0.55-0.8.

[0064] As an example, the preset alarm threshold is 0.65.

[0065] The axle box cover bolt abnormal mask matching algorithm of this embodiment generates a predefined mask matrix by defining and binarizing the axle box cover mask, marking the areas where bolts are allowed and not allowed. Then, the SIFT feature matching and affine transformation methods are used to achieve spatial matching and alignment between the mask and the axle box cover image. Finally, by calculating the percentage of overlapping pixels between abnormal labels and legal areas, false alarms are filtered to ensure efficient and accurate bolt status detection. This post-processing algorithm significantly reduces the false alarm rate and improves the practicality and reliability of the detection system.

[0066] As examples, data augmentation methods include rotation, flipping, brightness adjustment, and occlusion simulation.

[0067] In this implementation, by collecting high-quality images of multiple types of axle box covers and the normal and abnormal states of the bolts on them, balanced sampling and diversified data enhancement are performed to build a sample data set to ensure that the model training covers all key scenarios. Specifically, it includes:

[0068] 1. Diversity collection: Build a large image dataset containing various types of axle box covers and their bolt states. Ensure that the dataset covers all known axle box cover types and their different bolt configurations, especially including various visual representations of some easily confused bolts in normal and abnormal states.

[0069] 2. Enhanced display: Use data enhancement techniques (such as rotation, flipping, brightness adjustment, occlusion simulation, etc.) to increase data diversity and improve the generalization ability of the model. Design targeted data enhancement strategies for the specific forms formed after the bolt is lost, such as simulating different lighting conditions, angle changes, color jitter, etc., to increase the generalization ability of the model. In particular, for a small number of samples with high background complexity, more samples with complex backgrounds are generated through a variety of image synthesis techniques, while ensuring that the boundaries between positive and negative samples are clearer. For example, image stretching is used to simulate the phenomenon of different image stretching ratios caused by different vehicle speeds, and color jitter, contrast and brightness adjustment are used to simulate the color differences of component imaging caused by different lighting and different vehicle conditions. Homography Transform is used to create new data with more complex scenes, which can effectively simulate the changes in perspective caused by differences in camera laying angles at different sites.

[0070] 3. Fine marking: Each sample is accurately marked to distinguish normal bolts, missing bolts, and features that may cause confusion.

[0071] 4. Sample balance: Since some positive and negative samples have similar shapes and the ratio of the number of axle box cover samples and axle box cover bolt samples is quite different, it is necessary to ensure that the samples in the training set are balanced and solve the class imbalance problem by oversampling minority class samples.

[0072] The implementation system of the method of the present invention is deployed to the edge computing device, which can realize real-time image input and bolt status output, and display the detection results in combination with the user interface to support immediate maintenance decisions. The dual-path neural network model has a continuous learning mechanism: a model update mechanism is established, which can perform online learning or fine-tuning based on new field data, continuously optimize model performance, and respond to new axle box cover types or bolt configuration changes.

[0073] The method of the present invention combines a deep learning model network with a post-processing algorithm, and can provide efficient and accurate detection results in practical applications, thereby improving the safety and efficiency of railway operation and maintenance.

[0074] Example:

[0075] Table 1 shows an optimization scheme for the specific implementation of the present invention:

[0076] Table 1 Optimization scheme

[0077]

[0078] Table 2 compares the detection results of different combinations of detection networks:

[0079] Table 2 Comparison of results of different improvement schemes for detection network

[0080]

[0081] The blank control detection model in the first row of Table 2 is the unoptimized Faster R-CNN+ResNet50+FPN.

[0082] It can be seen from Table 2 that the evaluation index value of the method of the present invention is much higher than the evaluation index value obtained by the comparison network structure, so the method of the present invention greatly improves the accuracy of fault identification.

[0083] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network, characterized in that: include: Construct a dual-path neural network model and axle box cover bolt image sample set; The dual-path neural network model includes a global feature path, a local detail feature path, and a feature fusion module; The dual-path neural network model is trained using the axle box cover bolt image sample set, including: The global feature path uses a lightweight convolution module to extract the overall structural layout features of the axle box cover bolt image samples, and then uses a large kernel convolution layer to obtain the macroscopic geometric features of the axle box cover from the overall structural layout features; the local detail feature path first extracts image microscopic features from the axle box cover bolt image samples, and then enhances the details through the channel space attention mechanism to obtain the enhanced image microscopic features; The feature fusion module is used to fuse the macroscopic geometric features of the axle box cover and the microscopic features of the enhanced image to obtain a fused feature image. The fully connected layer is used to identify the faults of the fused feature images, and the preliminary identification results of the axle box cover type, axle box cover direction and bolt status are output; During the training process, the network parameters are adjusted by combining cross entropy loss and positioning loss to obtain a trained dual-path neural network model for axle box cover bolt fault detection. In the fault detection task, the corresponding predefined mask matrix is ​​determined according to the axle box cover type and axle box cover direction of the image to be identified output by the trained dual-path neural network model. The predefined mask matrix is ​​transformed to the axle box cover space of the image to be identified by the spatial posture matching algorithm for spatial matching and alignment. The final state of the bolt is determined by the intersection operation of the predefined mask matrix and the preliminary identification result label of the bolt state output by the trained dual-path neural network model, thereby realizing bolt fault detection.

2. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 1 is characterized in that: The axle box cover bolt image sample set covers all known axle box cover types and different bolt configurations; The axle box cover bolt image samples include real collected images and enhanced images; the enhanced images are obtained by performing data enhancement on the real collected images.

3. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 1 is characterized in that: The channel-spatial attention mechanism dynamically adjusts the channel and spatial weights of the image micro-features, highlights the bolt state-related features and suppresses irrelevant noise, thereby obtaining enhanced image micro-features.

4. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 3 is characterized in that: The channel spatial attention mechanism extracts the channel attention of image micro features through maximum pooling and average pooling operations, and then determines the channel weight of each channel through two fully connected layers; the spatial attention map of image micro features is obtained through convolution operations, and then the spatial weight is determined.

5. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 4 is characterized in that: The predefined mask matrix is ​​a template generated by a binary image processing method and accurate to the pixel level, with 0 indicating that the bolt exists in an illegal area and 1 indicating that the bolt exists in a legal area.

6. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 5 is characterized in that: Affine transformation and SIFT feature matching method are used to transform the predefined mask matrix to the axle box cover space of the image to be identified for spatial matching and alignment.

7. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 6 is characterized in that: The method to determine the final state of the bolt is: Calculate the percentage of overlapping pixels between the bolt abnormal label in the preliminary identification result label of the bolt status and the legal area of ​​the bolt in the predefined mask matrix. If the percentage of overlapping pixels exceeds the preset alarm threshold, it is determined as a valid abnormal identification result; otherwise, it is determined as a false identification and filtered out; and the final status of the bolt is determined.

8. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 7 is characterized in that: The preset alarm threshold is 0.55-0.

8.

9. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 1 is characterized in that: The macro-geometric characteristics of the axle box cover include the shape, size and bolt distribution pattern of the axle box cover.

10. The railway passenger car axle box cover bolt fault detection method based on attention enhanced dual path network according to claim 2, characterized in that: Data augmentation methods include rotation, flipping, brightness adjustment, and occlusion simulation.

Images