A high-speed rail catenary component fault detection method based on an SCA-GAnomaly model

Abstract

This invention discloses a method for detecting faults in high-speed railway catenary components based on the SCA-GANomaly model. The method includes: a data preprocessing stage: first, acquiring images of normal high-speed railway catenary components, and then expanding the dataset of normal high-speed railway catenary component samples using DCGAN data augmentation; a training stage: feeding the expanded dataset of normal high-speed railway catenary component samples into the SCA-GANomaly network for training, and obtaining the data distribution of normal catenary component images after training; a testing stage: acquiring catenary component images of test samples, then inputting them into the trained SCA-GANomaly network, obtaining the output image, comparing it with the test sample to obtain a difference score, and determining whether the difference score is greater than a threshold K: if it is greater, the catenary component image of the test sample is considered to have a fault; otherwise, the catenary component image of the test sample is considered normal. This invention can achieve more stable and accurate fault detection of high-speed railway catenary components.

Description

Technical Field

[0001] This invention relates to the field of high-speed railway catenary fault detection technology, specifically to a method for detecting faults in high-speed railway catenary components based on the SCA-GANomaly model. Background Technology

[0002] my country's high-speed rail operating mileage is increasing year by year, and its development is gradually transitioning from the design and construction phase to the operation and maintenance phase. The rapid development of high-speed rail and the gradual improvement in operation and maintenance quality have placed higher demands on the safe operation of the current traction power supply system equipment. The high-speed rail overhead contact line is a key power supply facility in the high-speed rail traction power supply system. Failure of any component of the contact line will directly affect the safe operation of high-speed trains. Therefore, fault detection of high-speed rail overhead contact line components is of paramount importance. Fault detection refers to the process of discovering faults during equipment operation and then implementing corresponding measures for maintenance. Fault detection has always been a key focus and challenge in the fields of computer vision and industrial inspection. Currently, fault detection of my country's high-speed rail overhead contact line is still mainly done manually. However, manual inspection has a long inspection cycle, a large workload, and inspectors are prone to visual fatigue, leading to missed or incorrect detections.

[0003] To alleviate the pressure of manual inspection, researchers have adopted intelligent algorithms to address the problem of fault detection in high-speed rail catenary systems. Initially, scholars proposed using traditional image processing techniques based on manually designed features to detect faults in catenary components, achieving promising results in both theory and engineering practice. However, the process of designing manual features using traditional image processing techniques is complex, making it difficult to guarantee that all manually designed features meet actual requirements. Furthermore, the features of different components often differ, necessitating the design of different features for each component, significantly increasing the time cost of traditional image processing techniques. In contrast, deep convolutional neural networks (DNNs) proposed in recent years do not require manual feature design; the network implicitly and automatically learns the inherent features of the image, and these features are often universal for different high-speed rail catenary components. This enhances the network's generalization ability and shortens the fault detection time for high-speed rail catenary components. However, deep convolutional neural networks require a large amount of data to achieve good results, but the fault sample data in high-speed rail catenary fault detection tasks is extremely scarce, making it difficult to meet the training requirements of deep convolutional neural networks. Therefore, how to achieve fault detection of high-speed rail catenary components under conditions of scarce fault data is a key challenge of deep learning and an important direction for future exploration of high-speed rail catenary fault detection.

[0004] Currently, GANomaly is one of the most effective methods in the field of fault detection. It directly utilizes normal sample data for training, learns the distribution of normal sample data, and determines the presence of faults based on the differences between input and output during testing. For the challenging problem of "few fault samples in high-speed rail catenary components and difficulty in predicting fault states in advance," GANomaly has significant advantages compared to typical image processing techniques and deep convolutional neural networks. However, GANomaly networks often suffer from inappropriate image reconstruction by the generator and unstable training, leading to unsatisfactory network training results and consequently affecting the accuracy of fault detection for high-speed rail catenary components.

[0005] Chinese patent document CN106340019A discloses a method for detecting the malfunction of the fixing hook of the overhead contact line in high-speed railways. Although this method can detect the condition of the fixing hook component of the overhead contact line in high-speed railways and provide objective, true and accurate detection and analysis results, its manual design process is complex and it is only effective for a single fault, with poor generalization ability.

[0006] Chinese patent document CN114202540A discloses an intelligent detection method for cotter pin defects in high-speed railway catenary based on deep learning. While this method achieves defect identification and detection in high-speed railway catenary cotter pins, it is highly dependent on the number of fault samples. In engineering practice, it is often difficult to obtain sufficient fault samples, which leads to insufficient training of the network model and consequently low fault detection accuracy.

[0007] Chinese patent document CN112184654A discloses a fault detection method for high-voltage line insulators based on GANomaly. While this method addresses the problem of limited insulator fault data, its network image fault reconstruction capability is too strong. During testing, the fault area is easily completely reconstructed, resulting in small differences in the abnormal scores between faulty and normal insulators, making it difficult to distinguish between normal and abnormal insulators during testing. Furthermore, the network training is unstable, prone to phenomena such as gradient explosion and mode collapse. Summary of the Invention

[0008] The purpose of this invention is to propose a fault detection method for high-speed railway catenary components based on the SCA-GANomaly model. This method does not require manual feature design, does not rely on fault samples, and has strong generalization ability for fault detection of different high-speed railway catenary components.

[0009] To achieve the above objectives, this application proposes a fault detection method for high-speed railway catenary components based on the SCA-GANomaly model, comprising:

[0010] In the data preprocessing stage: first, images of normal components of the high-speed rail catenary are acquired, and then the sample dataset of normal components of the high-speed rail catenary is expanded using DCGAN data augmentation techniques;

[0011] During the training phase: The expanded dataset of normal components of the high-speed rail catenary is fed into the SCA-GANomaly network for training. After training, the data distribution of images of normal components of the catenary is obtained.

[0012] During the testing phase: Images of the overhead contact line components of the test sample are acquired and then input into the trained SCA-GANomaly network. After obtaining the output image, it is compared with the test sample to obtain the difference score. It is determined whether the difference score is greater than the threshold K. If it is greater, the image of the overhead contact line components of the test sample is considered to have a fault; otherwise, the image of the overhead contact line components of the test sample is considered to be normal.

[0013] Furthermore, in the data preprocessing stage: multiple images of normal components of the high-speed rail catenary are collected and fed into the DCGAN network, enabling the DCGAN network to learn the normal data distribution and generate richer images of normal components of the high-speed rail catenary; the collected and generated images of normal components of the high-speed rail catenary are used to construct a new dataset for training the SCA-GANomaly network.

[0014] Furthermore, the SCA-GANomaly network includes a generator network G, an encoder network E, and a discriminator network.

[0015] Furthermore, the generator network G includes an encoder G. E (x) and decoder G D (z); where x represents a normal component image of the high-speed rail catenary in the input SCA-GANomaly network, and its dimension is C represents the number of image channels, H represents the height, and W represents the width; x passes through encoder G. E After downsampling (x), a latent vector z is generated, with dimension . d represents the vector length; decoder G D (z) Upsample the latent vector z to obtain the reconstructed image. Consistent with the x-dimensional dimension, all are The encoder G E The second layer feature map and decoder G of (x) D The second layer feature map of (z) is connected, and the encoder G is connected. E The penultimate layer feature map of (x) and decoder G D The penultimate layer feature map of (z) is connected.

[0016] Furthermore, a hybrid attention mechanism of CBAM and Attention is added during the upsampling, downsampling, and encoder-decoder connection processes;

[0017] The CBAM attention mechanism includes a channel attention module and a spatial attention module. The channel attention module utilizes the max pooling and average pooling layers of the shared network to aggregate the spatial information of the feature maps. The formula for obtaining channel attention is as follows:

[0018]

[0019] Where F is the input normal component feature map, M c Here, σ represents the output feature map of the channel attention function, MLP represents the multilayer perceptron, and AvgPool and MaxPool represent the average pooling operation and the max pooling operation, respectively. This represents element-wise multiplication;

[0020] Feature map M is obtained after passing through the channel attention module. c This is used as the input to the spatial attention module, and the formula for obtaining spatial attention is:

[0021]

[0022] Among them, f 5*5 This indicates that a 5x5 convolution kernel is used for convolution, [AvgPool(M c MaxPool(M) c This indicates that the two feature maps, after average pooling and max pooling, are concatenated.

[0023] The attention module first uses a Multi-Head Attention structure to obtain the feature information of feature map I, the expression of which is:

[0024] MultiHeadAttention(Q,K,V)=Concat(head1,...,head8)

[0025] Where Q, K, and V represent the Query, Key, and Value matrices, respectively, Concat represents the concatenation operation, and head i The attention module is represented by the following expression:

[0026] head i =Attention(QW i Q KWi K VW i V )

[0027] Among them, W i Q W i K W i V Let Q, K, and V represent the weight matrices respectively. Next, the feature map processed by the Multi HeadAttention structure and the input feature map I are added together and normalized to obtain the feature map O1. The formula for this step is as follows:

[0028] O1=Layer Normalization(I+Multi0HeadAttention(I))

[0029] Where Layer Normalization represents the layer normalization operation; then, the feature map O1 is input into the FeedForward layer, and layer normalization is performed to obtain the final output feature map O2 of the Attention module. The formula for this step is as follows:

[0030] O2=Layer Nomalization(O1+Feed Forward Network(O1))

[0031] The Feed Forward layer consists of two fully connected layers.

[0032] Furthermore, the discriminator network D takes as input images x of normal components of the high-speed rail overhead contact system and reconstructed images. The discriminator network D has two outputs: the output feature map of the Softmax layer and the output feature map of the layer before the Softmax layer.

[0033] Furthermore, the loss function of the SCA-GANomaly network consists of two parts: a generator loss and a discriminator loss. The generator loss includes adversarial loss, context loss, encoder loss, and bulldozer loss, defined as follows:

[0034] ① Counteracting losses:

[0035] ② Contextual loss:

[0036] ③ Encoder loss:

[0037] ④ Bulldozer losses:

[0038] Where f(x) represents the output obtained by inputting the image of a normal component of the high-speed rail catenary into the discriminator network D. Let E represent the output obtained by inputting the reconstructed image into the discriminator network D, and let E represent the expectation. Therefore, the generator loss of the SCA-GANomaly network is a weighted sum of the above four losses, defined as follows:

[0039] L gen =W adv L adv +W con L con +W enc L enc +W WGAN L WGAN

[0040] Among them, W con W adv W enc and W WGAN This is the weighting coefficient, the default value.

[0041] The loss of the SCA-GANomaly network discriminator is:

[0042]

[0043] Furthermore, the difference score is

[0044] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows:

[0045] (1) Before inputting normal samples into the SCA-GANomaly network, the normal sample dataset is first expanded through the DCGAN network so that the SCA-GANomaly network can fully learn the distribution of normal sample data, thereby improving the accuracy of fault detection.

[0046] (2) A selective connection structure is added to the GANomaly generator structure to improve the network's reconstruction capability. A hybrid attention mechanism is added to the selective connection structure and the image upsampling and downsampling process to make the network pay more attention to the high-speed rail catenary components and ignore information in the background area.

[0047] (3) Integrating EM Distance into GANomaly generative adversarial network and modifying the loss function of GANomaly can effectively improve the stability of GANomaly network and avoid problems such as gradient vanishing, gradient exploding and mode collapse.

[0048] (4) The present invention can achieve more stable and accurate fault detection of high-speed rail catenary components. Attached Figure Description

[0049] Figure 1 This is a diagram of the SCA-GANomaly network framework.

[0050] Figure 2 Flowcharts for the training and testing phases;

[0051] Figure 3 Here is a diagram of the generator framework;

[0052] Figure 4 This is a schematic diagram of the CBAM attention mechanism.

[0053] Figure 5 This is a diagram illustrating the principle of the Attention mechanism.

[0054] Figure 6 This is a distribution map of normal and faulty images. Specific implementation methods

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit the application; that is, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0056] Example 1

[0057] like Figure 1 As shown, a novel SCA-GANomaly (Selective Connection Attention GANomaly) network architecture is presented. This embodiment provides a fault detection method for high-speed railway catenary components based on the SCA-GANomaly model, which is divided into three stages: data preprocessing, training, and testing. Figure 2 As shown:

[0058] In the data preprocessing stage, images of normal components of the high-speed rail catenary are first acquired, and then the sample dataset of normal components of the high-speed rail catenary is expanded using DCGAN data augmentation techniques.

[0059] Specifically, multiple images of normal components of the high-speed rail catenary are first collected and fed into the DCGAN network, allowing the DCGAN network to learn the normal data distribution and generate richer images of normal components of the high-speed rail catenary. The collected and generated images of normal components of the high-speed rail catenary are then used to construct a new dataset for training the SCA-GANomaly network.

[0060] Using the DCGAN network to augment normal images of high-speed rail catenary effectively increases the sample size of normal images and generates realistic images, providing sufficient data for training the SCA-GANomaly network and improving model accuracy.

[0061] During the training phase, the expanded dataset of normal components of the high-speed rail catenary is fed into the SCA-GANomaly network for training. After training, the data distribution of images of normal components of the catenary is obtained, and the trained network is saved and used in the testing phase to realize fault detection of high-speed rail catenary components.

[0062] Specifically, after constructing a new dataset of normal components for the high-speed rail overhead contact system, it is fed into the SCA-GANomaly network for training. The SCA-GANomaly network comprises a generator network G, an encoder network E, and a discriminator network D; their structures are as follows:

[0063] The generator network G includes the encoder G. E (x) and decoder G D (z); where x represents a normal component image of the high-speed rail catenary in the input SCA-GANomaly network, and its dimension is C represents the number of image channels, H represents the height, and W represents the width; x passes through encoder G. E After downsampling (x), a latent vector z is generated, with dimension . d represents the vector length; decoder G D (z) Upsample the latent vector z to obtain the reconstructed image. Consistent with the x-dimensional dimension, all are To address the challenge of reconstructing complex high-speed railway overhead contact line components using a generator network G, this invention considers incorporating encoder G... E (x) and decoder G D (z) Skip-layer connections are added between corresponding feature layers. However, since shallow features contain too much original image information of high-speed rail catenary components, adding skip-layer connections makes the input and output of generator G almost equivalent, thus preventing the faulty image from being reconstructed into an image of a normal catenary component during the testing process; deep features contain too little original image information of high-speed rail catenary components, so adding skip-layer connections has little impact on the effect of generator network G. Therefore, a suitable selective connection scheme needs to be designed so that generator network G can reasonably reconstruct a normal image at the optimal cost. This invention selects encoder G... E The second layer feature map and decoder G of (x) D The second layer feature map of (z) is connected, and the encoder G is connected. E The penultimate layer feature map of (x) and decoder G DThe penultimate layer feature map of (z) is connected, and G is used appropriately. E (x) and G D Based on the feature information of (z), this scheme can effectively reconstruct images of normal components of the high-speed rail catenary.

[0064] In addition, to make the network focus more on the foreground part of the high-speed rail catenary component image and ignore the background area, a hybrid attention mechanism of CBAM and Attention was added in the selective connection, upsampling and downsampling process, which further improved the reconstruction capability of the generator network G.

[0065] The CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module utilizes the max pooling and average pooling layers of the shared network to aggregate the spatial information of the feature maps. The formula for obtaining channel attention is:

[0066]

[0067] Where F is the input normal component feature map, M e Here, σ represents the output feature map of the channel attention function, MLP represents the multilayer perceptron, and AvgPool and MaxPool represent the average pooling operation and the max pooling operation, respectively. This represents element-wise multiplication;

[0068] Feature map M is obtained after passing through the channel attention module. c This is used as the input to the spatial attention module, and the formula for obtaining spatial attention is:

[0069]

[0070] Among them, f 5*5 This indicates that a 5x5 convolution kernel is used for convolution, [AvgPool(M c MaxPool(M) c The expression )] indicates that the two feature maps after average pooling and max pooling are concatenated.

[0071] The attention module first uses a Multi-Head Attention structure to obtain the feature information of feature map I, the expression of which is:

[0072] MultiHeadAttention(Q,K,V)=Concat(head1,...,head8)

[0073] Where Q, K, and V represent the Query, Key, and Value matrices, respectively, Concat represents the concatenation operation, and head iThe attention module is represented by the following expression:

[0074] head i =Attention(QW i Q KW i K VW i V )

[0075] W i Q W i K W i V Let Q, K, and V represent the weight matrices, respectively. Next, the feature map processed by the Multi-Head Attention structure is added to the input feature map I and normalized to obtain the feature map O1. The formula for this step is as follows:

[0076] O1=Layer Normalization(I+Multi0HeadAttention(I))

[0077] Layer Normalization represents the layer normalization operation. Next, the feature map O1 is input into the FeedForward layer, and layer normalization is performed to obtain the final output feature map O2 of the Attention module. The formula for this step is as follows:

[0078] O2=Layer Nomalization(O1+Feed Forward Network(O1))

[0079] The Feed Forward layer consists of two fully connected layers.

[0080] The structure of the generator network G is as follows: Figure 3 As shown in the figure. The structures of the CBAM and Attention hybrid attention modules are respectively as follows: Figure 4 , Figure 5 As shown.

[0081] The encoder network E's role is to process the reconstructed image obtained from the generator network G. Compressed into another latent vector The structure of encoder network E and encoder G in the generator E (x) has a completely identical structure, and the latent vectors it generates are identical. The dimension is consistent with that of the potential vector z to ensure that the losses between them can be compared.

[0082] The discriminator network D receives images x of normal components of the high-speed rail overhead contact system and reconstructed images. The discriminator network D has two outputs: the output feature map of the Softmax layer and the output feature map of the layer before the Softmax layer.

[0083] The SCA-GANomaly loss function consists of two parts: the generator loss and the discriminator loss. To overcome the instability during the training process of traditional GANomaly networks, the GANomaly loss function is optimized by introducing EM Distance. The definition of EM Distance is as follows:

[0084]

[0085] Among them, P r P represents the distribution that the real data follows. g This represents the distribution that the generated data follows. П(P) r ,P g ) represents the marginal distribution P r and marginal distribution P g The set of all combined joint distributions γ(x,y). x and y represent images of normal components of the high-speed rail catenary and images reconstructed by the generated network, respectively. Using EM Distance to measure the smoothness of the x and y distributions provides a more effective measure of the distance between them. The above equation can be transformed into the following equation:

[0086]

[0087] Where k and K are both constants, and E represents the expectation. ||F|| L The slope ≤ K is called the Lipschitz continuity condition. The Lipschitz continuity condition restricts the slope of the function F, ensuring the smoothness of the function F, and is defined as follows:

[0088] |F(x)-F(y)|≤K|xy|

[0089] F(x) represents x satisfying the Lipschitz continuity condition. F(y) represents y satisfying the Lipschitz continuity condition. Therefore, we only need to find a function that satisfies the Lipschitz continuity condition as the loss function to achieve EM Distance.

[0090] Consider incorporating EM Distance into the GANomaly network to enhance its stability. The loss definitions for the generator and discriminator of the improved SCA-GANomaly network are as follows:

[0091] The loss function of the SCA-GANomaly generator consists of four parts: adversarial loss, context loss, encoder loss, and bulldozer loss. The first three losses aim to minimize the difference between the input high-speed rail catenary normal component image x and the reconstructed image. More similarly, the last bulldozer loss is primarily used to improve the stability of model training. Their loss functions are defined as follows:

[0092] ① Counteracting losses:

[0093] ② Contextual loss:

[0094] ③ Encoder loss:

[0095] ④ Bulldozer losses:

[0096] Where f(x) represents the output obtained by inputting the image of a normal component of the high-speed rail catenary into the discriminator network D. Let represent the output obtained by inputting the reconstructed image into the discriminator network D, and E represent the expectation. The loss function of the entire SCA-GANomaly generator is a weighted sum of the four losses mentioned above, defined as follows:

[0097] L gen =W adv L adv +W con L con +W enc L enc +W WGAN L WGAN

[0098] Among them, W con W adv W enc and W WGAN The weighting coefficient can take the value W. con =50,W adv =W enc =1,W WGAN =20.

[0099] The loss of the SCA-GANomaly discriminator is:

[0100]

[0101] Replacing the original discriminator loss with this loss function can effectively improve the training stability of the SCA-GANomaly network, while also enhancing the generation effect of images of normal components of the high-speed rail catenary.

[0102] After defining the SCA-GANomaly network structure and loss function, the SCA-GANomaly network was trained. This network ran on Ubuntu 20.04 using the PyTorch deep learning framework and an NVIDIA RTX 8000 graphics card, terminating training after 300 epochs. The model weights with the minimum loss during training were saved for experimental testing during the network testing phase.

[0103] During the testing phase, images of the overhead contact line components of the test sample are collected and then input into the trained SCA-GANomaly network. After obtaining the output image, it is compared with the test sample to obtain the difference score. It is determined whether the difference score is greater than the threshold K: if it is greater, the image of the overhead contact line components of the test sample is considered to have a fault; otherwise, the image of the overhead contact line components of the test sample is considered to be normal.

[0104] Specifically, the input for the testing phase consists of images of normal and faulty components of the high-speed rail overhead contact system. For each input image, it is fed into a pre-trained SCA-GANomaly network. During the testing phase, the network outputs an anomaly score, which is defined as the encoder loss during the training phase, i.e., the anomaly score.

[0105] Figure 6 To test the distribution of abnormal scores between normal and faulty components in the overhead contact system, 34 faulty images and 128 normal images were selected. It was found that the abnormal scores for normal components were concentrated between 1 and 3, while those for faulty components were concentrated between 3 and 6. There is a clear boundary between the distribution of normal and faulty images in the high-speed rail overhead contact system. A threshold of 3.12 was ultimately selected, resulting in a detection accuracy of 0.95 for normal images and 1.00 for faulty images.

[0106] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for fault detection of high-speed railway catenary components based on the SCA-GANomaly model, characterized in that, include: In the data preprocessing stage: first, images of normal components of the high-speed rail catenary are acquired, and then the sample dataset of normal components of the high-speed rail catenary is expanded using DCGAN data augmentation techniques; During the training phase: The expanded dataset of normal components of the high-speed rail catenary is fed into the SCA-GANomaly network for training. After training, the data distribution of images of normal components of the catenary is obtained. During the testing phase: images of the overhead contact line components of the test sample are collected and then input into the trained SCA-GANomaly network. After obtaining the output image, it is compared with the test sample to obtain the difference score. It is determined whether the difference score is greater than the threshold K: if it is greater, the image of the overhead contact line components of the test sample is considered to have a fault; otherwise, the image of the overhead contact line components of the test sample is considered to be normal. The SCA-GANomaly network includes a generator network G, an encoder network E, and a discriminator network; The generator network G includes encoders. and decoder ;in, This represents an image of a normal component of the high-speed rail overhead contact system from the input SCA-GANomaly network, with dimensions of [missing information]. C represents the number of image channels, H represents the height, and W represents the width; After encoder After downsampling, a latent vector is generated. Its dimensions are d represents the vector length; decoder Upsampling the latent vector z yields the reconstructed image. , and Dimensions are consistent, all are The encoder Second layer feature map and decoder The second layer feature map is connected, and the encoder is connected. The penultimate layer feature map and decoder The penultimate layer feature maps are connected; A hybrid attention mechanism combining CBAM and Attention is added during the upsampling, downsampling, and encoder-decoder connection processes. The CBAM attention mechanism includes a channel attention module and a spatial attention module; the channel attention module utilizes the max pooling layer and average pooling layer of the shared network to aggregate the spatial information of the feature maps, and the formula for obtaining channel attention is: Where F represents the input normal component feature map, This represents the output feature map of channel attention, where σ denotes the sigmoid function and MLP stands for Multilayer Perceptron. and These represent average pooling and max pooling operations, respectively. This represents element-wise multiplication; Feature maps are obtained after passing through the channel attention module. This is used as the input to the spatial attention module, and the formula for obtaining spatial attention is: in, This indicates that a 5x5 convolution kernel is used for convolution. This means concatenating two feature maps that have undergone average pooling and max pooling. The attention module first uses a Multi-Head Attention structure to obtain the feature information of feature map I, the expression of which is: Where Q, K, and V represent the Query, Key, and Value matrices, respectively, and Concat represents the concatenation operation. The attention module is represented by the following expression: in, These represent the weight matrices corresponding to Q, K, and V, respectively. Next, the feature map processed by the Multi HeadAttention structure and the input feature map I are added together and normalized to obtain the final feature map. The formula for this step is as follows: in Representation layer normalization operation; then feature map The feed forward layer is input, and layer normalization is performed to obtain the final output feature map of the attention module. The formula for this step is as follows: The Feed Forward layer consists of two fully connected layers.

2. The method for fault detection of high-speed railway catenary components based on the SCA-GANomaly model according to claim 1, characterized in that, In the data preprocessing stage: First, multiple images of normal components of the high-speed rail catenary are collected and fed into the DCGAN network, allowing the DCGAN network to learn the normal data distribution and generate richer images of normal components of the high-speed rail catenary; then, the collected and generated images of normal components of the high-speed rail catenary are used to construct a new dataset for... Train the SCA-GANomaly network.

3. The method for fault detection of high-speed railway catenary components based on the SCA-GANomaly model according to claim 1, characterized in that, The discriminator network D input is an image of a normal component of the high-speed rail overhead contact line. and reconstructed images The discriminator network D has two outputs: the output feature map of the Softmax layer and the output feature map of the layer before the Softmax layer.

4. The method for fault detection of high-speed railway catenary components based on the SCA-GANomaly model according to claim 1, characterized in that, The loss function of the SCA-GANomaly network consists of two parts: generator loss and discriminator loss. The generator loss includes adversarial loss, context loss, encoder loss, and bulldozer loss, defined as follows: Combat losses: Context loss: Encoder loss: Bulldozer losses: in, This represents the output obtained by inputting an image of a normal component of the high-speed rail overhead contact system into the discriminator network D. Let E represent the output obtained by inputting the reconstructed image into the discriminator network D, and let E represent the expectation. Therefore, the generator loss of the SCA-GANomaly network is a weighted sum of the above four losses, defined as follows: in, , , and These are the weighting coefficients; The loss of the SCA-GANomaly network discriminator is: 。 5. The method for fault detection of high-speed railway catenary components based on the SCA-GANomaly model according to claim 1, characterized in that, The difference score is .

Images