Urban rail vehicle bottom abnormality detection method based on multi-scale mask feature self-encoder

Abstract

The application discloses a kind of urban railcar bottom abnormality detection methods based on multi-scale mask feature self-encoder, it is related to rail transit technical field, the method includes based on car bottom monitoring data construction training set, the training set includes large-scale unlabeled training set and small amount of labeled training set;Multi-scale mask feature self-encoder is constructed, the unlabeled training set is used to the multi-scale mask feature self-encoder is carried out self-supervised training of reconstruction image, obtains model parameter;The model parameter and corresponding network structure of the encoder part in the obtained multi-scale mask feature self-encoder are embedded as skeleton network in the network structure of downstream car bottom abnormality detection task, and the model parameter is supervised fine-tuning training using small amount of labeled training set, obtains car bottom abnormality identification model.The present application can effectively solve the problems of data labeling difficulty, low utilization rate of metadata information, uneven positive and negative samples and interference of optical imaging environment in the prior art.

Description

Technical Field

[0001] This invention relates to the field of train inspection technology, specifically to a method for detecting anomalies under urban rail vehicles based on a multi-scale mask feature autoencoder. Background Technology

[0002] As the mileage of urban rail trains increases and their operating speeds rise, the frequency of maintenance on the train bodies also increases to ensure safe operation. To maintain efficient and high-speed operation of urban rail trains, it is necessary to promptly detect potential safety risks under the train cars and increase the frequency of maintenance. Common inspection items for abnormalities under urban rail train cars include missing, damaged, or foreign object-laden parts of key components (axle box covers, shock absorber mounting devices, air springs).

[0003] Current research on defect and anomaly detection in key parts of urban rail train undercarriages mainly focuses on the following three aspects:

[0004] Firstly, regarding template matching, Zhang et al. used a contour-based template matching algorithm to detect loose or missing faults in train components. They compared images of specific locations with a standard template library and categorized samples based on differences. This method is highly flexible and performs well for images with high pixel similarity. Lu et al. used template matching based on the shape descriptor of the target's bounding rectangle and contour to detect defects in truncated door handles. Experimental results showed that shape descriptor matching outperformed grayscale template matching in terms of time and accuracy. However, template matching algorithms are susceptible to factors such as lighting and imaging angle, and fail significantly when the target is deformed, exhibiting poor robustness and making large-scale application in practical train inspection systems difficult.

[0005] Secondly, in the area of ​​traditional machine learning, Zhou et al., in detecting faults in brake shoe pins in trains, used Gradient Encoding Histogram (GEH) features and Support Vector Machine (SVM) algorithms to locate the component position and used SVM to determine the fault state, achieving an accuracy of 99.2% and an image processing speed of 5 fps / s. Qin et al. used SVM to automate the detection of missing handle faults in train corner cocks, achieving good real-time performance and high detection accuracy. However, both of these methods, which use traditional machine learning algorithms for feature extraction, require manual design based on specific scenarios, resulting in low efficiency in locating target objects. They also rely on traditional methods such as grayscale projection, contour detection, and sliding window exhaustive search, which are often only applicable to specific components with strong identifiable features, such as bolts and keyways, and their accuracy still falls short of real-world requirements.

[0006] Third, in terms of deep learning, Sun et al. used convolutional neural networks to coarsely and finely locate the key and axle bolt areas of the train side frame, and then trained a multi-classification model to identify four typical faults related to loss and loosening. This achieved high recognition ability and good robustness even under low-quality imaging conditions. Zhou et al. combined traditional methods with deep learning methods and applied them to train component fault detection. They located the bolt area using traditional methods, and then used a convolutional neural network combined with a stacked autoencoder (SAE) to identify the bolt state. Ultimately, they achieved high detection accuracy for fault images in the middle plate bolts. As can be seen, deep learning-based defect detection methods have stronger robustness and can cope with more complex on-site environments, imaging conditions, and weather conditions. They have better detection performance for subtle faults such as bolts, brake pads, and keyways, and can achieve high-precision positioning of the target under complex conditions. However, due to limitations in sample acquisition, especially in fault discrimination due to the lack of negative samples, the trained model is prone to overfitting, and the detection time is relatively longer.

[0007] In summary, while existing technologies can detect defects and anomalies in key components of rail trains through template matching, machine learning, and conventional deep learning, the accuracy, reliability, and stability of defect and anomaly detection remain unsatisfactory for large-scale industrial applications due to the complex and variable imaging environment, lens optical distortion, complex backgrounds, and other factors. These issues include unknown defects, limited accurate defect annotations, high cost of annotation samples, uneven distribution of positive and negative samples, irregularity of defects, and large range of variation in defect shape and size. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method for detecting anomalies under urban rail vehicles based on a multi-scale mask feature autoencoder.

[0009] The technical solution of this invention is: a method for detecting anomalies under urban rail vehicles based on a multi-scale mask feature autoencoder, comprising:

[0010] A training set is constructed based on vehicle undercarriage monitoring data. The training set includes a large-scale unlabeled training set and a small number of labeled training sets. The vehicle undercarriage monitoring data includes point cloud data and image data.

[0011] A multi-scale mask feature autoencoder is constructed, and the reconstructed image is trained by the multi-scale mask feature autoencoder using the unlabeled training set to obtain model parameters.

[0012] The model parameters and corresponding network structure of the encoder part in the multi-scale mask feature autoencoder are embedded as the backbone network into the network structure of the downstream vehicle underside anomaly detection task. The model parameters are then trained in a supervised manner using a small amount of labeled training set to obtain the vehicle underside anomaly recognition model.

[0013] Preferably, the training set constructed based on undercarriage monitoring data includes:

[0014] The point cloud data and the image data are preprocessed separately.

[0015] The preprocessed point cloud data and image data are correlated to obtain correlated data;

[0016] The point cloud data and image data in the associated data are weighted and fused to obtain a multimodal fusion matrix.

[0017] Preferably, the preprocessing of the point cloud data includes filtering, registration, and resampling, and the preprocessing of the image data includes bilateral filtering and histogram equalization.

[0018] Preferably, when training the multi-scale mask feature autoencoder to reconstruct images using the unlabeled training set, the reconstruction criterion is to minimize the scaling cosine error of the mask feature map and the reconstructed image.

[0019] Preferably, the self-supervised training of the reconstructed image using the unlabeled training set for the multi-scale mask feature autoencoder includes:

[0020] The model parameters of the multi-scale mask feature autoencoder are updated using backpropagation;

[0021] After each backpropagation, a momentum update mechanism is used to update the encoder model parameters in the multi-scale mask feature autoencoder.

[0022] Preferably, the encoder in the multi-scale mask feature autoencoder includes a multi-scale mask feature extraction module and a mask feature fusion module.

[0023] Preferably, the multi-scale mask feature extraction module includes a first feature extraction stage, a second feature extraction stage, and a third feature extraction stage;

[0024] The first feature extraction stage includes a patch embedding module and a Masked Convolutional Block module, used to obtain mask features at a first scale; the second feature extraction stage includes a patch embedding module and a Masked Convolutional Block module, used to obtain mask features at a second scale; the third feature extraction stage includes a patch embedding module and a Transformer module, used to obtain mask features at a third scale.

[0025] A downsampling layer is provided between the first feature extraction stage and the second feature extraction stage, and a downsampling layer is provided between the second feature extraction stage and the third feature extraction stage.

[0026] Preferably, the mask feature fusion module is used to downsample the first-scale mask feature and the second-scale mask feature respectively, and then fuse the downsampled first-scale mask feature, second-scale mask feature and third-scale mask feature to obtain a fused feature map.

[0027] Preferably, downsampling the first scale mask feature and the second scale mask feature respectively includes: downsampling the first scale mask feature with stride=2 and downsampling the second scale mask feature with stride=4.

[0028] Preferably, the loss function of the multi-scale mask feature autoencoder is:

[0029] L = L recon +L cos +λ1L feat +λ2L fusion

[0030]

[0031]

[0032]

[0033]

[0034] Among them, L recon L represents the reconstruction error. cos L represents the scaling cosine error. feat L represents the feature map vector error. fusion The feature map error is represented by λ1 and λ2, which are hyperparameters; ||.||1 represents the L1 norm, x i and Let S(A, B) represent the original image of the i-th small block and the reconstructed image of the τ-th small block output by the decoder, respectively; S(A, B) represents the cosine similarity, and F... i and Let z represent the mask feature map of the i-th small block and the mask feature map of the τ-th small block output by the decoder, respectively; z and Let ||f| and ||2| represent the vector representations of the fused feature map and the decoder output, respectively, where ||.||2 represents the L2 norm; E t and These represent the fused feature map and the fused feature map output by the decoder, respectively.

[0035] The beneficial effects of this invention are as follows: This invention provides a method for detecting anomalies under urban rail vehicles based on a multi-scale mask feature autoencoder. By collecting point cloud data and image data and performing multi-modal fusion, the idea of ​​a mask autoencoder is transferred to the image. Through multi-scale mask feature autoencoder, the high computational cost of high-resolution images is solved from the perspectives of loss function and model structure. At the same time, a mask feature reconstruction strategy with scaling cosine error as the reconstruction criterion is designed to achieve adaptive sample re-weighting and solve the problem of uneven positive and negative samples. Finally, the representation encoding of self-supervised learning is transferred to few-sample learning, making full use of the advantages of automatic feature or representation learning in unsupervised learning, making the model more reliable and robust, and reducing the dependence on labeled data. Applying the anomaly recognition model under urban rail vehicles provided by this invention to the field of anomaly detection of foreign objects under urban rail vehicles can effectively solve problems such as difficulty in data labeling, low utilization of metadata information, uneven positive and negative samples, and interference from the optical imaging environment. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0037] Figure 1 A flowchart of an urban rail vehicle undercarriage anomaly detection method based on multi-scale mask feature autoencoder provided in an embodiment of the present invention;

[0038] Figure 2 A framework diagram of the urban rail vehicle undercarriage anomaly detection method based on multi-scale mask feature autoencoder provided in an embodiment of the present invention;

[0039] Figure 3 A framework diagram of a multi-scale mask feature autoencoder provided in an embodiment of the present invention;

[0040] Figure 4This is a task phase diagram of the urban rail vehicle undercarriage anomaly detection method based on multi-scale mask feature autoencoder provided in an embodiment of the present invention. Detailed Implementation

[0041] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0042] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0043] Please see Figure 1 , 2 This invention provides a method for detecting anomalies under urban rail vehicles based on a multi-scale mask feature autoencoder, comprising:

[0044] Step 1: Construct a training set based on the vehicle undercarriage monitoring data. The training set includes a large-scale unlabeled training set and a small number of labeled training sets. The vehicle undercarriage monitoring data includes point cloud data and image data.

[0045] In this embodiment of the invention, constructing a training set based on vehicle undercarriage monitoring data includes: preprocessing the point cloud data and the image data respectively; associating the preprocessed point cloud data and image data to obtain associated data; and weighted fusing the point cloud data and image data in the associated data to obtain a multimodal fusion matrix.

[0046] Specifically, point cloud data is acquired through radar, and image data is acquired through a 2D linear array camera.

[0047] Specifically, the preprocessing of the point cloud data includes filtering, registration, and resampling. Specifically, a Gaussian filter is used to smooth the point cloud data to remove noise and enhance the signal. Simultaneously, the point cloud data is registered with a reference point cloud and downsampled to eliminate the influence of different positions and poses, while reducing the data volume. Specifically, the preprocessing of the image data includes bilateral filtering and histogram equalization. Specifically, a bilateral filter is used to denoise the input image to reduce noise and artifacts in the image, and then adaptive histogram equalization is used to enhance the image to improve contrast and robustness.

[0048] After preprocessing the point cloud data and image data, it is necessary to match the point cloud data and image data one by one through timestamp alignment and external calibration, and convert the camera pixel coordinates into real-world coordinates through internal calibration to facilitate subsequent weighted fusion.

[0049] To improve data utilization and quality, point cloud data and image data need to be fused. Specifically, different weighting coefficients are configured for point cloud data and image data, and weighted fusion is performed based on these weighting coefficients. The formula for weighted fusion is as follows:

[0050] f fusion (x i ) = w radar ·f radar (x i )+w camera ·f camera (x i (1)

[0051] Among them, f radar (x i f represents the feature vector of point cloud data. camera (x i f represents the feature vector of the image data. fusion (x i ) represents the multimodal fusion matrix after weighted fusion.

[0052] The above method enables multimodal fusion of point cloud and linear scan camera image data, improving the comprehensiveness and accuracy of the data.

[0053] Step 2: Construct a multi-scale mask feature autoencoder, and use the unlabeled training set to perform self-supervised training on the reconstructed image to obtain model parameters.

[0054] Please see Figure 3 In this embodiment of the invention, when training the multi-scale mask feature autoencoder to reconstruct images using the unlabeled training set, the reconstruction criterion is to minimize the scaling cosine error of the mask feature map and the reconstructed image.

[0055] In this embodiment of the invention, the self-supervised image reconstruction training of the multi-scale mask feature autoencoder using the unlabeled training set includes: updating the model parameters of the multi-scale mask feature autoencoder using backpropagation; and updating the model parameters of the encoder in the multi-scale mask feature autoencoder using a momentum update mechanism after each backpropagation.

[0056] To better understand the scaling cosine error, the basic definition of scaling cosine is explained in detail below. The cosine distance between two vectors u and v is defined as follows:

[0057]

[0058] in,<u,v> Let ||u||² represent the dot product of vectors u and v, and ||u||² represent the L2 norm of vector u; assume the input mixing matrix is ​​X∈R. W×H×C Where C is the number of channels, and the output is the restored matrix. The original matrix is ​​divided into N equal-sized blocks, each of size W×H×C, and a binary mask m is used. i This indicates whether the small block contains target information; if it does, then m i =1, otherwise m i =0. Based on the above explanation, it should be understood that if the parameters of the encoder and decoder in the multi-scale mask feature autoencoder are θ respectively... E and θ D Then the potential representation is z∈R K K << W × H × C, and its reconstruction criterion is to minimize the mask feature map F and the decoder output. This is achieved through the scaling cosine error between them, that is:

[0059]

[0060] Among them, F i Let i represent the mask feature map of the i-th small block. This represents the mask feature map output by the corresponding decoder.

[0061] Furthermore, during training, the latent representation is computed through forward propagation:

[0062]

[0063] and reconstructed feature maps:

[0064]

[0065] Finally, the gradient is calculated and the parameters are updated through backpropagation.

[0066] It should be noted that throughout the training process, the encoder network and the decoder network jointly learn the representation and reconstruction of the mask feature map.

[0067] In this embodiment of the invention, the encoder in the multi-scale mask feature autoencoder includes a multi-scale mask feature extraction module and a mask feature fusion module. The multi-scale mask feature extraction module includes a first feature extraction stage, a second feature extraction stage, and a third feature extraction stage.

[0068] The first feature extraction stage includes a patch embedding module and a Masked Convolutional Block module, used to obtain mask features at a first scale; the second feature extraction stage includes a patch embedding module and a Masked Convolutional Block module, used to obtain mask features at a second scale; the third feature extraction stage includes a patch embedding module and a Transformer module, used to obtain mask features at a third scale.

[0069] A downsampling layer is provided between the first feature extraction stage and the second feature extraction stage, and a downsampling layer is provided between the second feature extraction stage and the third feature extraction stage.

[0070] In this embodiment of the invention, the mask feature fusion module is used to downsample the first-scale mask feature and the second-scale mask feature respectively, and then fuse the downsampled first-scale mask feature, second-scale mask feature and third-scale mask feature to obtain a fused feature map.

[0071] A multi-scale masked feature autoencoder progressively abstracts the input multimodal feature vector into multi-scale token embeddings. It generates feature maps of different resolutions using a pyramid-like approach, then divides each feature map into blocks, each corresponding to a token. For the initial high-resolution token embeddings, convolutional blocks encode local content. These high-resolution embeddings extract local spatial features and transform them into lower-dimensional feature vectors as token representations. For the later low-resolution token embeddings, Transformer blocks fuse global contextual information. Furthermore, Transformer blocks encode the relationships between all tokens to obtain a more global feature representation. To avoid feature confusion, the mask is progressively upsampled to a larger resolution in the early convolutional stages. Simultaneously, masked convolutions are added in the early stages to ensure that the features processed in the convolutional blocks are completely separated into mask tokens and visible tokens, thus avoiding feature confusion between masked and visible regions in later stages.

[0072] Specifically, the multi-scale masked feature autoencoder combines convolutional neural networks and the Transformer architecture. First, the input low-resolution image undergoes a series of convolutional and masked convolutional operations to extract feature maps. Then, the feature maps of the masked regions are concatenated with those of the visible regions through upsampling to obtain a high-resolution feature map. In the later Transformer stage, the high-resolution feature map is processed, and a mask is used to achieve the autoencoding purpose.

[0073] The network structure uses a special scenario coding structure plus a proposed general network structure as the encoder or decoder.

[0074] The specific encoder part consists of 3 stages. Let H and W be the dimensions of the input matrix, and the output features of each stage are respectively...

[0075] The first two stages are hybrid convolutional modules that use Masked Convolutional Blocks to operate on features, and their structure is as follows: Figure 2 As shown in the bottom right corner (where the dilated convolution uses a 3×3 kernel), a convolution with a stride of 2 is performed between each stage to perform a downsampling operation.

[0076] The final stage is the general encoder network, which uses the Transformer fusion module. Its main function is to expand the receptive field and fuse features from all patches. The network part of the final stage can be replaced with a commonly used ResNet residual network or a fully convolutional network structure of FCN type such as U-Net, depending on different downstream tasks.

[0077] To improve training, the acquired first-scale mask features E1, second-scale mask features E2, and third-scale mask features E3 need to be fused. Specifically, E1 and E2 are downsampled to stride=2 and stride=4 respectively, and then added to E3 to perform multi-scale feature fusion. The fused result is then passed through a Linear Transformer to obtain the final feature E to be input to the Decoder. t :

[0078] E t =Linear(StrideConv(E1,4)+StrideConv(E2,2)+E3)

[0079] Where StrideConv(·, k) represents the convolution with stride = k, E t A fused feature map representing features at three scales.

[0080] In this embodiment of the invention, the loss function of the multi-scale mask feature autoencoder is:

[0081] L = L recon +L cos +λ1L feat +λ2L fusion

[0082]

[0083]

[0084]

[0085]

[0086] Among them, L recon L represents the reconstruction error. cos L represents the scaling cosine error. feat L represents the feature map vector error. fusion The feature map error is represented by λ1 and λ2, which are hyperparameters; ||.||1 represents the L1 norm, x i and Let A and B represent the original image of the i-th small block and the reconstructed image of the τ-th small block output by the decoder, respectively; S(A, B) represents the cosine similarity. F i and Let z represent the mask feature map of the i-th small block and the mask feature map of the τ-th small block output by the decoder, respectively; z and Let ||f| and ||2| represent the vector representations of the fused feature map and the decoder output, respectively, where ||.||2 represents the L2 norm; E t and These represent the fused feature map and the fused feature map output by the decoder, respectively.

[0087] It should be noted that the mask feature map in this embodiment includes a first-scale mask feature, a second-scale mask feature, and a third-scale mask feature.

[0088] Step 3: The model parameters and corresponding network structure of the encoder part of the obtained multi-scale mask feature autoencoder are embedded as the backbone network into the network structure of the downstream vehicle underside anomaly detection task. The model parameters are then fine-tuned in a supervised manner using a small amount of labeled training set to obtain the vehicle underside anomaly recognition model.

[0089] Specifically, the three mask feature vectors obtained in the previous step are further input into a multi-scale cascaded Transformer network (the backbone network can be arbitrarily selected, such as variants of ResNet, Mask R-CNN series networks, etc., as long as the matrix input of the three branches is handled well, and can be dynamically selected according to different downstream tasks). After obtaining the hidden representations through self-supervised training of the proposed autoencoder, these representations are used to solve various downstream tasks, making full use of the advantages of automatic learning of features or representations in unsupervised learning, making the model more general and robust, while also reducing the dependence on labeled data.

[0090] Specifically, the parameters of the encoder network are fixed, and the latent representation is used as the output. This output is then combined with a few-shot learning model, and the entire network is fine-tuned using labeled data to obtain the final prediction model. By leveraging the representational capabilities learned by the pre-trained model on large-scale unlabeled data, the model performance can be improved by fine-tuning labeled data. This method can usually achieve better results with less labeled data than training from scratch.

[0091] In summary, this invention presents an anomaly detection method for urban rail vehicle undercarriages based on a multi-scale mask feature autoencoder. By collecting point cloud data and image data and performing multi-modal fusion, the mask autoencoder concept is transferred to graphs. The multi-scale mask feature autoencoder addresses the high computational cost of high-resolution images faced by general graph autoencoders from the perspectives of image reconstruction, loss function, and model structure. Simultaneously, a mask feature reconstruction strategy using scaling cosine error as the reconstruction criterion is designed, achieving adaptive sample re-weighting and resolving the problem of uneven positive and negative sample distribution. Finally, the representation encoding of self-supervised learning is transferred to few-shot learning, fully utilizing the advantages of automatic feature or representation learning in unsupervised learning, making the model more robust and reducing dependence on labeled data. Applying the undercarriage anomaly recognition model provided by this invention to the field of anomaly detection of foreign objects under urban rail vehicles can effectively solve problems such as difficult data labeling, low utilization of metadata information, uneven positive and negative sample distribution, and interference from the optical imaging environment.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for detecting abnormalities of a metro vehicle bottom based on a multi-scale mask feature autoencoder, characterized in that, The method comprises the following steps: constructing a training set based on vehicle bottom monitoring data, wherein the training set comprises a large-scale unlabeled training set and a small amount of labeled training set, and the vehicle bottom monitoring data comprises point cloud data and image data; constructing a multi-scale mask feature autoencoder, performing self-supervised training of the multi-scale mask feature autoencoder on the unlabeled training set to obtain model parameters; embedding the model parameters and the corresponding network structure of the encoder part of the obtained multi-scale mask feature autoencoder into the network structure of a downstream vehicle bottom anomaly detection task as a skeleton network, and performing supervised fine-tuning training on the model parameters by using a small amount of labeled training set to obtain a vehicle bottom anomaly recognition model; constructing a training set based on vehicle bottom monitoring data comprises: respectively pre-processing the point cloud data and the image data; associating the pre-processed point cloud data and image data to obtain associated data; weighting and fusing the point cloud data and image data in the associated data to obtain a multi-modal fusion matrix; the pre-processing of the point cloud data comprises filtering, registration and resampling, and the pre-processing of the image data comprises bilateral filtering and histogram equalization; when performing self-supervised training of the multi-scale mask feature autoencoder on the unlabeled training set, the reconstruction criterion is to minimize the scaling cosine error of the mask feature map and the reconstructed image; the encoder in the multi-scale mask feature autoencoder comprises a multi-scale mask feature extraction module and a mask feature fusion module.

2. The metro vehicle bogie anomaly detection method based on multi-scale mask feature autoencoder according to claim 1, characterized in that, performing self-supervised training of the multi-scale mask feature autoencoder on the unlabeled training set comprises: updating the model parameters of the multi-scale mask feature autoencoder using back propagation; updating the model parameters of the encoder in the multi-scale mask feature autoencoder using a momentum update mechanism after each back propagation.

3. The metro vehicle bogie anomaly detection method based on multi-scale mask feature autoencoder according to claim 1, characterized in that, The multi-scale mask feature extraction module comprises a first feature extraction stage, a second feature extraction stage and a third feature extraction stage. The first feature extraction stage comprises a patch embeding module and a Masked Convolutional Block module for obtaining first scale mask features; the second feature extraction stage comprises a patch embeding module and a Masked Convolutional Block module for obtaining second scale mask features; and the third feature extraction stage comprises a patch embeding module and a Transformer module for obtaining third scale mask features. A down-sampling layer is arranged between the first feature extraction stage and the second feature extraction stage, and a down-sampling layer is arranged between the second feature extraction stage and the third feature extraction stage.

4. The metro vehicle bogie anomaly detection method based on multi-scale mask feature autoencoder according to claim 3, characterized in that, The mask feature fusion module is used for down-sampling the first scale mask features and the second scale mask features respectively, and fusing the down-sampled first scale mask features, second scale mask features and third scale mask features to obtain a fused feature map.

5. The metro vehicle bogie anomaly detection method based on multi-scale mask feature autoencoder according to claim 4, characterized in that, The downsampling on the first scale mask feature and the second scale mask feature respectively comprises: stride=2 downsampling on the first scale mask feature, and stride=4 downsampling on the second scale mask feature.

6. The method of claim 1, wherein the method further comprises: The loss function of the multi-scale mask feature self-encoder is: wherein denotes the reconstruction error, denotes the scaled cosine error, denotes the feature map vector error, denotes the feature map error, and are hyperparameters; denotes the L1 norm, and denote the original image and the decoder output of the the first small patch, respectively; denotes the cosine similarity, and denote the mask feature map and the decoder output of the the first small patch, respectively; and denote the vector representation of the fused feature map and the vector representation of the fused feature map output by the decoder, respectively, denotes the L2 norm; and denote the fused feature map and the fused feature map output by the decoder, respectively.

Images