A train vehicle fault image intelligent analysis method based on a lightweight deep learning technology

Abstract

The application discloses a kind of train vehicle fault image intelligent analysis methods based on lightweight deep learning technology, comprising: obtaining train image data, then the train fault component to be detected in image is labeled, then the overall data is divided into training set, test set two parts;Data enhancement operation is carried out, the hyperparameter required by algorithm is configured, and the training sample is input into the MobileDetectNet neural network model formed after improvement to carry out feature learning;The learned feature model is input to the new vehicle image to identify faults, and finally the area of the fault occurs is framed and the corresponding alarm information is output.The MobileDetectNet network intelligent identification model of the application has few parameters, occupies less memory and has high recognition rate, can improve the recognition rate of EMU fault detection under the original equipment CPU environment, reduce the work intensity of artificial, shorten the maintenance operation time, reduce the missed detection probability, so as to ensure the safe operation of EMU.

Description

Technical Field

[0001] This invention relates to the field of vehicle technical condition detection technology research and equipment development, specifically to a method for intelligent analysis of train vehicle fault images based on lightweight deep learning technology. Background Technology

[0002] Currently, to ensure the safe operation of high-speed trains, TEDS (Train Operation Fault Image Detection System) equipment is installed on major railway lines across the country. This system is an intelligent system integrating high-speed digital image acquisition, large-capacity image data storage, and pattern recognition technology. The system consists of equipment for image information acquisition, image information processing and transmission, and a monitoring platform at the train inspection center. Through a high-speed camera array, the system captures all visible information about the train's undercarriage and sides. After digital processing, this information is displayed on the monitoring platform at the inspection center. Personnel at the monitoring center then use the images displayed on the terminal to diagnose faults based on the appearance of observable components during train operation. Currently, the system has initially achieved automatic fault diagnosis on servers with GPU computing cards. However, it still suffers from low fault detection rate and low alarm accuracy on existing servers without GPUs. Furthermore, many components of high-speed trains still rely on manual inspection of diagrams to confirm faults. Given the high density of high-speed train traffic and the large number of vehicles passing through, the workload of monitoring center personnel is extremely heavy. During high-speed train operation, the effectiveness of fault confirmation heavily depends on the experience and work status of the monitoring center personnel. Therefore, there is an urgent need for an intelligent identification method and equipment for fault detection during high-speed train operation to reduce the workload of monitoring center personnel, improve fault detection effectiveness, and ensure the safe operation of high-speed trains.

[0003] Currently, most existing methods for detecting fault states in high-speed trains address the aforementioned issues by relying on image processing techniques. However, high-performing deep learning network models often depend on GPU acceleration. Therefore, deploying recognition models on devices without GPUs and achieving good results has become a key research area for lightweight networks. Several related patents have been found, with the following general content:

[0004] 1) A face recognition method based on an improved MobileNetV3, patent publication number CN110647840A

[0005] This invention provides a face recognition method based on an improved MobileNetV3, comprising the following steps: S1. Performing face detection and feature point detection on the image; S2. Performing face alignment using the detected face feature points; S3. Creating a training set from the aligned image and its labels; S4. Constructing the network structure, transforming the final pooling layer into a convolutional layer, and reducing the channel dilation factor. This invention only modifies the last layer of MobileNetV3, resulting in limited improvements and making the face detection task easy to implement.

[0006] 2) A real-time face detection method based on MobileNet V3, patent publication number CN110647817A

[0007] This invention designs a real-time face detector based on an SSD architecture. A lightweight feature enhancement module is proposed to improve the representation capability of shallow features in MobileNetV3, thereby improving the detection accuracy of extremely small faces. Finally, to improve the recall rate of extremely small faces, an anchor density strategy is used, ensuring that different types of anchors have the same density in the image. Although this invention improves upon MobileNetV3, the anchor density strategy increases computational complexity, which is detrimental to improving recognition efficiency.

[0008] 3) Intelligent inspection robot defect identification system and method for power transmission lines, patent publication number CN112836659A

[0009] This invention provides a defect identification system for intelligent inspection robots of power transmission lines. The system includes the SDLiteMobilenetV3 lightweight network, the TensorFlow framework, TensorRT technology, and the NVIDIA Jetson Nano embedded platform. The invention also utilizes this system to provide a defect identification method. The target detection model of this method uses the SDLiteMobilenetV3 lightweight network, is built and trained based on the TensorFlow framework, uses TensorRT technology for inference acceleration, and is deployed on the NVIDIA Jetson Nano embedded platform. Real-time inference video is pushed to the server via video stream, enabling real-time detection of targets on power transmission lines. This invention reduces latency by 10ms, improves computational speed, and achieves no loss of accuracy, reducing computational consumption while maintaining accuracy. This invention only applies the SDLiteMobilenetV3 lightweight network and does not improve upon it.

[0010] 4) A rolling bearing fault diagnosis method based on enhanced lightweight multi-scale CNN, Publication No. CN112577748A

[0011] This invention provides a rolling bearing fault diagnosis method based on a reinforced lightweight multi-scale CNN. First, a constructed CNN model is used to extract multi-dimensional features from the original vibration signal. Then, this feature is input into a lightweight multi-scale feature learning network to extract multi-scale features. The learned multi-scale features are selectively reinforced using a discriminative fault feature enhancement mechanism (DFRM) to strengthen fault features and weaken general features. The reinforced multi-scale features are then fused. This process is repeated several times, and the resulting high-level abstract feature map is input into a classifier for training. Finally, the trained CNN model is used to identify faults in the test samples. The model training of this invention is relatively complex, and the recognition network uses a simple CNN convolutional neural network. Although it is a lightweight network, the recognition performance needs further verification.

[0012] 5) A method and system for large-scale train displacement fault detection based on deep learning, publication number CN110264440A

[0013] This invention primarily utilizes object detection models in deep learning to locate components and semantic segmentation models to segment component contours, thereby obtaining the component's contour and positional information. Then, it uses relative positional information to determine whether the component has shifted. However, the invention only indicates the use of a deep learning network model, which encompasses a wide range of models; it does not specify the particular network structure used. Furthermore, this invention merely applies a deep learning network model and does not involve any improvements to the model itself.

[0014] 6) A method for augmenting train fault images based on deep learning, authorized announcement number CN112699952B. This invention only augments train fault images, that is, it artificially expands train fault images, and does not involve train fault recognition or deep learning model improvement.

[0015] With the rapid development of my country's high-speed rail infrastructure construction and the rapid increase in high-speed rail operating mileage in recent years, train safety inspection has become a popular research topic for many domestic research institutions and enterprises. At present, there are many types of high-speed trains in actual operation, and the number of vehicles passing through the main line is large. The types of components and faults to be inspected are also diverse. The main line refers to the actual operating line on site.

[0016] Currently, the Train Operation Fault Image Detection System (TEDS) possesses certain image algorithm capabilities for train fault identification. However, due to factors such as server hardware, its automatic identification capability still falls short of the demands of field applications. With the continuous increase in the operating mileage and train formations of high-speed trains, on-site personnel face heavy workloads and require high concentration, leading to fatigue. Simultaneously, the number of personnel at TEDS monitoring centers across various railway bureaus is gradually increasing, resulting in rising labor costs. Faced with these challenges, there is an urgent need to strengthen and optimize image intelligent recognition algorithms for fault images of key undercarriage components of high-speed trains through new technologies and methods. This would allow on-site applications to gradually transition from a system primarily based on human inspection with machine inspection as a supplement to one primarily based on machine inspection with human inspection as a supplement.

[0017] For intelligent image recognition algorithms of high-speed train faults, feature extraction and difference comparison methods are commonly used. However, due to the wide variety of high-speed train models, components, and fault modes, as well as the influence of external environmental changes and speed variations, these two methods still suffer from low fault detection rates and accuracy. Another method based on deep learning heavily relies on GPU computing power, placing high demands on server hardware. With the development of artificial intelligence technology, recognition algorithms are constantly iterating and updating, and the large number of fault sample images accumulated during the application of TEDS provides a solid foundation and guarantee for further deepening intelligent image recognition algorithms.

[0018] Not all equipment on a high-speed train contains GPUs, and most deep neural networks rely on the high computing power of GPUs for model inference. Therefore, considering algorithm compatibility, a lightweight network that does not rely on GPUs, has fewer network parameters, and is faster is used for high-speed train vehicle fault image detection. Deep network models based on lightweight network designs include the MobileNet series, ShuffleNet series, and XceptionNet. This invention uses MobileNet-v3 as the basic model structure and makes targeted improvements. The original CNN feature extraction layer is replaced with an EESP module to reduce the number of model parameters and improve the model's feature representation ability. Two bendeck layers in MobileNet-v3 are removed to further reduce network parameters. Additionally, the pooling layers and fully connected layers of the backbone network are replaced with coupled detection heads to predict the target's classification and location coordinates. Summary of the Invention

[0019] The purpose of this invention is to provide an intelligent analysis method for train vehicle fault images based on lightweight deep learning technology, in order to solve the problems existing in the prior art.

[0020] To achieve the above objectives, this invention provides an intelligent analysis method for train vehicle fault images based on lightweight deep learning technology. Specifically, the method includes the following steps:

[0021] Step 1: Acquire train image data, then label the faulty parts of the train to be detected in the image, and then divide the overall data into two parts: training set and test set.

[0022] Step 2: Perform data augmentation, configure the hyperparameters required by the algorithm, and input the training samples into the MobileDetectNet neural network model, which is an optimized and improved version of the MobileNet-v3 backbone feature extraction network, for feature learning.

[0023] Step 3: Perform fault identification on the feature model learned from the new vehicle image input, and finally outline the faulty area and output the corresponding alarm information.

[0024] Furthermore, the training of the improved MobileDetectNet neural network model involves collecting a large number of original images of EMU malfunctions, augmenting the sample data using image enhancement algorithms, and adding normal passing images as background to form an image dataset A. Then, the image dataset A is divided into a training dataset and a test dataset.

[0025] Furthermore, the training dataset comprises 9 / 10 of the entire image dataset, and the test dataset comprises 1 / 10 of the entire image dataset. The training dataset is labeled using the VOC dataset format, and then the EMU fault identification model is trained using the stochastic gradient descent (SGD) optimization algorithm. The initial learning rate is 0.001, and the training is iterated for 300 epochs. The learning rate gradually decreases as the number of iterations increases. The trained network model parameter file is tested using the test dataset, and the optimal network model parameter file is selected.

[0026] Furthermore, the specific process of the algorithm network for feature learning by inputting the improved MobileDetectNet neural network model with training samples includes: the improved MobileDetectNet neural network model includes an input layer, an EESP feature extraction network, a 9-layer bneck feature extraction network, and a prediction output layer; preprocessing the input image by scaling it to 224×224×3; constructing a feature extraction network, including an EESP network layer and a bneck network layer, wherein the EESP network uses point-wise convolution and dilated convolution spatial pyramid modules to split the standard convolution and introduces HFF to remove the grid effect of dilated convolution, which can maintain good performance while reducing parameters; reducing the 11-layer bneck network in MobileNet-v3 to a 9-layer bneck network to reduce network parameters for further image feature extraction; and finally, using a coupled detection head for target detection.

[0027] Furthermore, the EESP module of the improved MobileDetectNet neural network model is derived from the ESPNetV2 network model.

[0028] Furthermore, the hyperparameters required for the configuration algorithm include the learning rate, gradient descent method, number of iterations, ratio of training set to test set, and training set size.

[0029] Furthermore, the optimization method based on the MobileNet-v3 backbone feature extraction network includes: replacing the original CNN feature extraction layer with an EESP network module, wherein the EESP network module modifies the standard CNN convolution into a depthwise separable convolution, each layer has a different dilation rate, and the feature maps of different scales are connected using a concatenation operation to increase the dimension of the output features.

[0030] Furthermore, the improvement of the detection network for target detection using the coupled detection head includes: replacing the pooling layer and fully connected layer of the original backbone network with the coupled detection head and predicting the target classification and location coordinates; performing a 1×1 convolution operation on the further extracted image features, and finally the network obtains a 75-dimensional feature vector after performing image feature extraction.

[0031] Furthermore, the improvement of the detection network for target detection using coupled detection heads further includes: training a model on train fault images using the improved MobileDetectNet network; and using the trained network model to determine the faults of the EMU (Electric Multiple Unit).

[0032] Furthermore, the method of using the trained fault detection model to determine the faults of the EMU includes: using the trained network model to perform fault detection on the EMU passing images, with the input image normalized to 224×224×3; extracting features from the image through the EESP structure, and the extracted feature map is then processed by a subsequent 9-layer bneck network to extract deep image features; if a faulty component is present in the image, the detection model provides the fault type and location coordinates of the faulty component and issues an alarm.

[0033] The method of the present invention has the following advantages:

[0034] This invention uses the lightweight MobileNet-v3 network as its basic model structure and makes targeted improvements. The original CNN feature extraction layer is replaced with an EESP module, reducing the number of model parameters and improving its feature representation capability. Two bridge layers in MobileNet-v3 are removed to further reduce network parameters. Furthermore, the pooling and fully connected layers of the backbone network are replaced with coupled detection heads to predict the classification and location coordinates of targets. Compared to the original MobileNet-v3, the proposed MobileDetectNet intelligent recognition model has fewer parameters, lower memory consumption, a larger receptive field, and a higher recognition rate. Moreover, this network is compatible with both CPU and GPU devices and can be used as a target detection network to accurately locate faulty components in high-speed train images.

[0035] The application of this intelligent recognition model can improve the recognition rate of EMU fault detection in the CPU environment of the original equipment, greatly reduce the workload of manual labor, shorten the maintenance operation time, improve the efficiency of maintenance work, reduce the probability of missed detection, reduce labor costs and management costs, thereby ensuring the safe operation of EMUs. Attached Figure Description

[0036] Figure 1 The improved MobileDetectNet network structure is shown.

[0037] Figure 2 The EESP network structure is shown;

[0038] Figure 3 The flowchart shows the training process of high-speed train fault samples using the MobileDetectNet network model;

[0039] Figure 4 The flowchart shows the process of using a trained fault detection model to determine faults in high-speed trains. Detailed Implementation

[0040] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art should understand that the embodiments described below are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Automatic fault detection systems for high-speed trains, as a means of detecting train operation status, play a crucial role in ensuring safe train operation. Currently, high-speed train fault detection equipment varies widely in product and hardware configuration, and the compatibility of different detection algorithms is not ideal. This invention selects a lightweight network from deep learning object detection algorithms as the research object, using MobileNet-v3 as the base network. Based on MobileNet-v3, the feature extraction layer (CNN) is replaced with an EESP module. This EESP module uses point-wise convolution and dilated convolution spatial pyramid modules to split standard convolutions and introduces HFF to remove the grid effect of dilated convolutions, maintaining good performance while reducing parameters. Two bneck layers in MobileNet-v3 are removed to further reduce network parameters. Additionally, the pooling layers and fully connected layers of the MobileNet-v3 backbone are replaced with coupled detection heads to classify targets and regress their location coordinates. The improved MobileDetectNet network, compared to the original MobileNet-v3, has fewer parameters, less memory usage, a larger receptive field, and a higher recognition rate. Furthermore, this network is compatible with both CPU and GPU devices and can be used as a target detection network to accurately locate faulty components in high-speed train images. The receptive field refers to the range of image values ​​calculated by the network model in a single step.

[0042] The improved MobileDetectNet network structure is as follows: Figure 1 As shown, the input image is normalized to 224×224×3, and then the EESP structure is used to extract features from the image. Subsequently, a 9-layer bneck network is used to extract deep image features. The pooling layers and subsequent fully connected layers of the MobileNet-v3 network are replaced with convolutional layers to provide the extracted image features to the subsequent detection network for target classification and coordinate regression.

[0043] The improved MobileDetectNet network's EESP module is derived from the ESPNetV2 network model. The EESP module structure is as follows: Figure 2As shown, to learn multi-scale features of images more efficiently, this EESP module modifies the standard CNN convolution into a depthwise separable convolution, with each layer having a different dilation rate. Feature maps of different scales are concatenated using a concatenation operation, increasing the dimensionality of the output features and fusing multi-scale features without increasing the number of parameters. An HFF operation is added before concatenation to perform layer-by-layer summation of the outputs, eliminating the mesh texture introduced by the concatenation operation. Grouped point-wise convolutions in this EESP module reduce computational complexity. To prevent information loss due to downsampling, a shortcut connecting the input image is added. This shortcut uses multiple pooling operations to make its spatial size match the feature size of the module's output. Figure 1 Then, two convolutions are used to extract features and adjust the dimensions, and finally the feature map matrices are added together.

[0044] The improved MobileDetectNet network parameters are shown in Table 1, where bneck is the basic structure of the MobileNet-v3 network. EESP represents the EESP feature extraction module, exp size represents the dimension of the network after convolution, out represents the number of output channels, and SE represents whether channel attention mechanism is used. NL represents the type of activation function, including HS (h-swish) and RE (ReLU). NBN represents no BN (batch normalization) operation. s stands for stride, and the network uses the convolution stride operation for downsampling.

[0045] Table 1 MobileDetectNet Network Parameters

[0046]

[0047] As shown in Table 1, the improved MobileDetectNet network structure first performs EESP operation on the normalized 3-channel image, then uses a 9-layer bneck module to extract feature maps, and performs a 1×1 convolution operation on the extracted feature maps. Finally, the network obtains a 75-dimensional feature vector after performing feature extraction.

[0048] The last dimension of the improved MobileDetectNet network structure is 75 because the network was designed based on the number of categories in the VOC dataset. The VOC dataset has 20 categories. For each feature point in each feature layer, 3 prior boxes are set, so the number of channels in the prediction result is 3×25. The 25 includes the probability of each category in the 20 categories, 4 coordinate parameters, and 1 parameter indicating whether there is an object.

[0049] This invention provides a technical solution, which is an intelligent analysis method for train vehicle fault images based on lightweight deep learning technology. The specific steps of the overall operation include:

[0050] First, train image data is acquired. Then, the faulty parts of the train to be detected in the images are labeled. Finally, the overall data is divided into training set and test set.

[0051] Data augmentation is performed, and the hyperparameters required by the algorithm are configured. Then, the training samples are input into the improved MobileDetectNet model for feature learning. In deep learning, hyperparameters are parameters whose values ​​are set before the learning process begins. Typically, hyperparameters need to be optimized to select an optimal set for the learning machine, thereby improving learning performance and efficiency.

[0052] The newly passed vehicle image is input into the learned feature model for fault identification. Finally, the area where the fault occurred is outlined and the corresponding alarm information is output.

[0053] The model was trained by collecting a large number of original images of EMU (Electric Multiple Unit) malfunctions. Image augmentation algorithms were used to augment the sample data, and normal passing train images were added as background to form image dataset A. Image dataset A was then divided into training and testing datasets, with the training dataset comprising 9 / 10 of the total image dataset and the testing dataset comprising 1 / 10. The training dataset was labeled using the VOC (Video Object Classification) format. The EMU malfunction identification model was then trained using the SGD (Stochastic Gradient Descent) optimization algorithm with an initial learning rate of 0.001 and iterative training for 300 epochs, with the learning rate gradually decreasing as the number of iterations increased. The trained network model parameter file was tested using the test dataset to select the optimal network model parameter file.

[0054] The specific process of the algorithm network includes:

[0055] The improved MobileDetectNet neural network model includes an input layer, an EESP feature extraction network, a 9-layer bneck feature extraction network, and a prediction output layer. First, the input image is preprocessed, scaling it to 224×224×3. Then, a feature extraction network is constructed, including an EESP network layer and a bneck network layer. The EESP network uses point-wise convolution and dilated convolution spatial pyramid modules to split standard convolutions and introduces HFF to remove the grid effect of dilated convolutions, maintaining good performance while reducing parameters. The 11-layer bneck network in MobileNet-v3 is reduced to 9 layers, further reducing network parameters for further image feature extraction. Finally, a coupled detection head is used for target detection.

[0056] The parameter settings include the learning rate, gradient descent method, number of iterations, ratio of training set to test set, and composition of the training set.

[0057] The optimization method based on the MobileNet-v3 backbone feature extraction network is as follows:

[0058] The original CNN feature extraction layer was replaced with an EESP network module. To more efficiently learn multi-scale image features, this module modifies the standard CNN convolution into depthwise separable convolutions, with each layer having a different dilation rate. Feature maps of different scales are concatenated using a concatenation operation, increasing the dimensionality of the output features and fusing multi-scale features without increasing the number of parameters. An HFF operation is added before concatenation to perform layer-by-layer summation of the outputs, eliminating the mesh-like texture introduced by the concatenation operation. Grouped point-wise convolutions in this EESP network module reduce computational complexity. To prevent information loss due to downsampling, a shortcut connecting the input image is added. This shortcut uses multiple pooling operations to make its spatial size match the feature size of the module's output. Figure 1 Then, two convolutions are used to extract features and adjust the dimensions, and finally the feature map matrices are added together.

[0059] The optimization based on the MobileNet-v3 feature extraction network is as follows:

[0060] The 11-layer bneck feature extraction network in the MobileNet-v3 network was changed to a 9-layer bneck feature extraction network to further reduce the number of network parameters and the amount of computation.

[0061] The detection network was improved as follows:

[0062] The original backbone network's pooling and fully connected layers were replaced with coupled detection heads to predict the target's classification and location coordinates. The extracted feature maps underwent 1×1 convolution operations, resulting in a 75-dimensional feature vector after feature extraction. The final dimension of the improved MobileDetectNet network structure is 75 because it was designed based on the number of categories in the VOC dataset (20 categories). For each feature point in each feature layer, three prior boxes are set, resulting in a prediction channel count of 3×25. The 25 channels include the probability of each of the 20 categories, four coordinate parameters, and one parameter indicating whether an object exists.

[0063] The implementation process mainly consists of two parts: training a model on train fault images using the improved MobileDetectNet network; and using the trained network model to determine the faults of the high-speed train.

[0064] 1) The MobileDetectNet network model is used to train the model on the fault samples of the EMU. The flowchart of the training process is as follows: Figure 3 As shown.

[0065] 2) Use the trained fault detection model to diagnose faults in high-speed trains.

[0066] This step outlines the overall fault diagnosis process. First, a trained network model is used to detect faults in images of passing trains. The input image is normalized to 224×224×3. Then, the EESP structure is used to extract features from the image. The extracted feature maps are then processed by a 9-layer Bneck network to extract deeper image features. Finally, the extracted feature maps undergo 1×1 point convolution operations to adjust their dimensions. The final network output is a 75-dimensional feature vector, representing three prior boxes. Each prior box contains probabilities for 20 categories, four coordinate values, and one parameter indicating whether an object exists. If a faulty component is present in the image, the detection model identifies the fault type and its location coordinates, and issues an alarm. The entire flowchart is shown below. Figure 4 As shown.

[0067] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for intelligent analysis of train vehicle fault images based on lightweight deep learning technology, comprising: Step 1: Acquire train image data, then label the faulty parts of the train to be detected in the image, and then divide the overall data into two parts: training set and test set. Step 2: Perform data augmentation, configure the hyperparameters required by the algorithm, and input the training samples into the MobileDetectNet neural network model, which is an optimized and improved version of the MobileNet-v3 backbone feature extraction network, for feature learning. Step 3: Perform fault identification on the feature model learned from the new vehicle image input, and finally outline the area where the fault occurred and output the corresponding alarm information; The specific process of the algorithm network for feature learning by inputting training samples into the improved MobileDetectNet neural network model includes: the improved MobileDetectNet neural network model includes an input layer, an EESP feature extraction network, a 9-layer bneck feature extraction network, and a prediction output layer; preprocessing the input image and scaling the image to 224×224×3; A feature extraction network is constructed, including an EESP network layer and a bneck network layer. The EESP network uses point-wise convolution and dilated convolution spatial pyramid modules to split the standard convolution and introduces HFF to remove the grid effect of dilated convolution, which can maintain good performance while reducing parameters. The 11-layer bneck network in MobileNet-v3 is reduced to a 9-layer bneck network to reduce network parameters for further image feature extraction. Finally, a coupled detection head is used for target detection.

2. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 1, characterized in that, The improved MobileDetectNet neural network model is trained by collecting a large number of original images of EMU malfunctions, using image enhancement algorithms to augment the sample data, and adding normal passing images as background to form an image dataset A. Then, image dataset A is divided into a training dataset and a test dataset.

3. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 2, characterized in that, The training dataset comprises 9 / 10 of the entire image dataset, and the test dataset comprises 1 / 10. The training dataset is labeled using the VOC dataset format. Then, the high-speed train fault identification model is trained using the stochastic gradient descent (SGD) optimization algorithm with an initial learning rate of 0.

001. The training is iterated for 300 epochs, and the learning rate gradually decreases as the number of iterations increases. The trained network model parameter file is tested using the test dataset, and the optimal network model parameter file is selected.

4. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 1, characterized in that, The EESP module of the improved MobileDetectNet neural network model is derived from the ESPNetV2 network model.

5. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 1, characterized in that, The hyperparameters required for the configuration algorithm include the learning rate, gradient descent method, number of iterations, ratio of training set to test set, and training set size.

6. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 1, characterized in that, The optimization method based on the MobileNet-v3 backbone feature extraction network includes: replacing the original CNN feature extraction layer with an EESP network module. The EESP network module modifies the standard CNN convolution into a depthwise separable convolution, with each layer having a different dilation rate. Feature maps of different scales are connected using a concatenation operation to increase the dimension of the output features.

7. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 1, characterized in that, The improvement to the target detection network using the coupled detection head includes: replacing the pooling layer and fully connected layer of the backbone network with the coupled detection head and predicting the target's classification and location coordinates; performing a 1×1 convolution operation on the further extracted image features, and finally obtaining a 75-dimensional feature vector after the network performs image feature extraction.

8. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 7, characterized in that, The improvement of the detection network for target detection using coupled detection heads further includes: training a model on train fault images using the improved MobileDetectNet network; and using the trained network model to determine the faults of the EMU (Electric Multiple Unit).

9. The intelligent analysis method for train vehicle fault images based on lightweight deep learning technology according to claim 8, characterized in that, The method of using a trained fault detection model to determine the faults of the EMU includes: using a trained network model to detect faults in the EMU passing images, with the input image normalized to 224×224×3; extracting features from the image through an EESP structure, and then using a subsequent 9-layer bneck network to extract deep image features; if a faulty component is present in the image, the detection model provides the fault type and location coordinates of the faulty component and issues an alarm.

Images