Multi-target real-time perception method for track operation vehicle and coupler state detection system

Abstract

The present application discloses a track operation vehicle multi-target real-time sensing method and a coupler state detection system, belonging to the field of rail transit technology. The method comprises: collecting coupler and track images by multiple cameras; using the multi-task architecture of a shared backbone network to synchronously perform target detection, key point detection and semantic segmentation, and adopting feature orthogonal regularization to eliminate multi-task interference; comparing the recognition results with operation instructions and vehicle speed in real time, and realizing abnormality judgment and graded early warning through a Boolean logic matrix; based on a meta-learning and self-supervised joint framework, first pre-training the feature extractor by contrast learning, and then adapting to small sample defect detection by meta-training; constructing a space-time topology graph, and using a graph attention network combined with a speed and distance global reasoning to predict the influence of defects. The present application realizes high-precision and low-delay sensing on an embedded platform, solves the problem of scarce defect samples, and improves the safety and intelligence level of continuous coupling operation.

Description

Technical Field

[0001] This invention belongs to the field of rail transit technology, specifically a multi-target real-time perception method for rail work vehicles and a coupler status detection system. Background Technology

[0002] With the continuous improvement of mechanization in railway track maintenance, track work vehicles, as key carriers for railway inspection, maintenance, and rescue operations, have seen their operational safety and efficiency become core factors restricting the intelligent development of the railway operation and maintenance system. Against this backdrop, achieving real-time perception and intelligent early warning of the track work vehicle's surrounding environment, coupler connection status, and track structure health is a critical technical challenge that urgently needs to be addressed.

[0003] In the field of safety monitoring for railcars, existing technologies have begun to incorporate intelligent methods. For example, patent CN120932213A discloses a method and system for monitoring the safety of railway freight car couplers based on HRNet. This scheme achieves early warning by setting markers on the coupler, extracting features using a high-resolution network (HRNet), and calculating the angle between the centerline of the coupler tongue and the coupler body. However, this method is a single-task detection model, which not only heavily relies on manually set physical markers (which are easily detached or covered by oil in harsh environments), but also suffers from the high computational complexity of the HRNet architecture, making it impossible to achieve real-time concurrent processing of multiple targets such as coupler connections and track defects on onboard edge devices with limited computing power.

[0004] In track structure health monitoring, the traditional model of regular inspections by large inspection vehicles combined with manual inspections suffers from inherent drawbacks such as high cost, low frequency, and susceptibility to missed inspections. To improve the level of intelligence, patent CN122262803A proposes a fault prediction method for rail transit equipment based on graph neural networks. This method uses graph attention networks to model the topological relationships between equipment to predict faults. However, this approach is essentially a time-series prediction based on the system's endogenous historical data and cannot handle the instantaneous causal physical impact of sudden external environmental changes (such as track defects) on internal equipment (such as couplers). Furthermore, its model training heavily relies on a large amount of labeled data of past faults. In actual operation, rail cracks and missing fasteners are typical low-frequency events. This purely supervised learning method based on a large sample size suffers from severely insufficient generalization ability when faced with scenarios where defect samples are extremely scarce.

[0005] More critically, existing vehicle-mounted monitoring systems face a bottleneck in full video backhaul at the architectural level. In the weak network environment along railway lines (especially in tunnels and mountainous areas), uploading massive amounts of video data can lead to severe network congestion and high latency, and centralized cloud processing cannot guarantee millisecond-level timeliness for abnormal event warnings.

[0006] In summary, existing technologies cannot achieve synchronous, high-precision sensing of multiple targets on edge devices with low computing power and weak networks without artificial physical markers; they lack the ability to rapidly adapt models to a very small number of defect samples; and they have not established a cross-domain physical causal reasoning model between external disturbances (track defects) and internal responses (coupler attitude). Overcoming these bottlenecks under extremely limited hardware and data conditions has become a key research focus for those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-target real-time perception method and a coupler status detection system for rail vehicles, so as to solve the problems mentioned in the background art of the prior art that high-precision and low-latency perception cannot be achieved in low-bandwidth and weak network environments, and that it is difficult to build a detection model with strong generalization ability under the condition of scarce defect samples.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-target real-time perception method for rail-mounted work vehicles includes the following steps: Step S1: Use multiple cameras deployed on the rail work vehicle to collect image data of the coupler area and the track structure area in real time; Step S2: The acquired image data is input into a lightweight multi-task neural network architecture for synchronous processing. The lightweight multi-task neural network architecture includes an object detection branch, a key point detection branch, and a semantic segmentation branch. The three branches share a feature extraction backbone network. Specifically, the object detection branch outputs the target bounding box coordinates of the coupler and track structure targets, the key point detection branch outputs the spatial position and state information of the key components of the coupler, and the semantic segmentation branch outputs the pixel-level mask of the coupler gap area and the compression and stretching amount of the buffer device. Step S3: The identification result is compared with the operation process instructions and vehicle speed information in real time. When the detected actual state of the coupler is inconsistent with the expected state corresponding to the current operation instruction, an alarm is triggered. Step S4: Detect track structure defects based on the joint framework of meta-learning and self-supervised learning. In this step, a feature extractor is pre-trained on unlabeled normal track images using self-supervised learning. The parameters of the pre-trained feature extractor are used as the initialization of the meta-learning algorithm and meta-trained, so that the model can detect defects under the condition of a small number of labeled samples in the target domain. Step S5: Construct a topological relationship graph of the track structure and coupler components, and use a graph neural network for global reasoning. The topological relationship graph contains spatial topological edges and temporal correlation edges. The graph neural network combines vehicle speed and odometer information through a message passing mechanism to infer the degree of impact of local defects on the overall structure.

[0009] According to the above technical solution, in the lightweight multi-task neural network architecture, the three branches share the multi-scale spatial feature map extracted by the backbone network. Before the shared feature map is input into each branch, it is orthogonalized by the attention mask layer of the specific task. The joint loss function includes a feature orthogonal regularization term. This regularization term is constructed by calculating the inner product of the feature matrices generated by different branches after sharing the feature map, which forces the feature matrices of different tasks to remain orthogonal in order to eliminate gradient interference between branches.

[0010] According to the above technical solution, real-time comparison is specifically achieved by constructing a multi-dimensional Boolean logic state matrix: a mapping relationship between the set of operation instructions and the expected combination of coupler states is established in advance, and the real-time detected coupler state vector is XORed with the expected state vector found according to the current operation instruction. An alarm is triggered when any state bit does not match.

[0011] According to the above technical solution, in the joint framework of meta-learning and self-supervised learning, the contrastive learning framework is first used to construct positive sample pairs on unlabeled normal track images to pre-train a general feature extractor, enabling the feature extractor to learn robust feature representations that are insensitive to changes in illumination and viewpoint. Then, the parameters of the pre-trained feature extractor are used as the initialization parameters of the MAML algorithm, and meta-training is performed on a meta-task composed of normal samples from multiple source domains, enabling the model to quickly adapt to new tasks. Finally, fine-tuning is performed under the condition that there are only a few labeled defect samples in the target domain to achieve small sample defect detection.

[0012] According to the above technical solution, in the detection of track structure defects, a variational autoencoder is also used to train on normal track images to construct the feature manifold of the normal pattern. For the image to be tested, the abnormal score is obtained by calculating its reconstruction error or its Mahalanobis distance in the latent space. When the abnormal score exceeds a preset threshold, it is judged as a defect. In addition, sampling is performed outward along the center of the normal distribution in the latent space to generate reference points that deviate from the normal pattern. After being reconstructed by the decoder, these reference points are used as expanded negative samples to participate in meta-learning training to enhance the model's sensitivity to abnormal patterns.

[0013] According to the above technical solution, in the topological relationship graph, the spatial topological edges represent the front-to-back adjacency and left-to-right symmetry relationships between components along the mileage direction, and the temporal correlation edges represent the state changes of the same component at different time points, thus associating the historical state features of the component with the current state features; the graph neural network adopts a graph attention network architecture, and its node features are composed of the local defect features output by visual detection and the historical state features, and the edge features are composed of the physical connection relationship encoding and the temporal correlation relationship encoding.

[0014] According to the above technical solution, in the message transmission process of the graph neural network, when a node is a coupler component node and its neighboring node is a track defect node, the delay time from the defect location to the coupler location is calculated based on the train running speed, and dynamic weights are assigned to the temporal correlation edges to realize the spatiotemporal correlation prediction based on train speed and mileage markers, and increase the attention weight to the coupler status at the defect location in advance.

[0015] According to the above technical solution, the global reasoning also calculates the risk score of each node based on the output of the graph neural network. The risk score is obtained by weighted summation of the severity of each defect in the node's neighborhood and the detection confidence. When the risk score exceeds the preset threshold, a work order recommendation containing the defect type, location and suggested maintenance measures is generated.

[0016] According to the above technical solution, in real-time comparison, a graded warning is also implemented based on the relative distance between the two vehicles. The closer the distance, the higher the warning level. The distance measurement adopts a monocular depth estimation method. The network is trained by minimizing the relative gradient between the predicted depth map and the real depth map to ensure the accuracy of close-range measurement.

[0017] The coupler condition detection system includes: The image acquisition module is used to acquire image data of the coupler area and track structure area during the operation of the rail work vehicle; The edge computing module is used to run a lightweight multi-task neural network inference engine to process image data in real time and simultaneously perform object detection, key point detection and semantic segmentation tasks. The edge computing module also includes a model quantization unit and a model pruning unit to compress the model size and improve inference speed. The coupler status recognition module is used to identify the coupler connection status based on the output of the edge computing module and preset status judgment conditions, and to perform real-time comparison with the operation instructions and hierarchical early warning. The track defect detection module includes a meta-learning unit, a self-supervised learning unit, and an anomaly generation unit, which are used to execute a joint framework of meta-learning and self-supervised learning to realize track structure defect detection under the condition of scarce defect samples. The graph neural network inference module is used to construct a topological relationship graph and perform global graph neural network inference, outputting defect correlation analysis results and risk scores; The communication module is used for data interaction with the ground terminal and supports two modes: event-triggered upload and periodic health reporting. The auxiliary display module is used to provide drivers with a visual aid image that overlays the detection results with annotations.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention integrates object detection, keypoint detection, and semantic segmentation into a shared backbone network using a lightweight multi-task neural network architecture. It also introduces feature orthogonal regularization to eliminate gradient interference, achieving millisecond-level inference latency with a model parameter size not exceeding 8MB. A multidimensional Boolean logic state matrix is ​​constructed to map operation instructions to the expected coupler state, combined with monocular depth estimation for hierarchical early warning, enabling automated anomaly detection in coupled operations. Addressing the scarcity of track defect samples, a joint framework of meta-learning and self-supervised learning is employed. First, contrastive learning is used to pre-train environment-invariant features from unlabeled images, followed by meta-training using MAML initialization. This approach enables the model to adapt quickly to single-digit labeled samples. It utilizes a variational autoencoder to construct a normal manifold, performs anomaly detection on defect-free samples based on reconstruction error and Mahalanobis distance, and enhances the robustness of decision boundaries under small sample sizes by generating abnormal reference points through latent spatial extrapolation sampling. A graph attention network containing spatial topological edges and temporally correlated edges is constructed, integrating dynamic weights of vehicle speed and odometer markers to achieve spatiotemporal correlation prediction of track defects and coupler status, upgrading post-event alarms to proactive early warnings. Combined with model quantization pruning, it balances high precision and real-time performance on an embedded platform, significantly improving the coupling safety and intelligence level of the track maintenance vehicle. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multi-target real-time perception method for track maintenance vehicles according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0021] Example 1

[0022] like Figure 1 As shown, this embodiment provides a multi-target real-time perception method for a track maintenance vehicle, including the following steps: Step 1: Multiple high-definition cameras deployed on the front and sides of the rail work vehicle are used to collect real-time image data of the coupler area and the track structure area. The high-definition cameras support starlight-level night vision. When the ambient illuminance is below 0.1 lux, they automatically switch to infrared supplementary lighting mode. The infrared supplementary lighting wavelength is 850nm. Infrared light in this band mainly produces diffuse reflection and volume scattering on rusted and oxidized metal layers and surfaces covered with heavy oil stains, which significantly suppresses specular reflection on metal wear surfaces, thereby improving the imaging contrast of key features of the coupler.

[0023] Step two involves inputting the acquired image data into a lightweight multi-task neural network architecture for synchronous processing. This architecture includes an object detection branch, a key point detection branch, and a semantic segmentation branch. The object detection branch uses a one-stage object detection algorithm to output the coordinates of the target bounding boxes for the coupler, hook tongue, hook pin, lifting rod, buffer, and the relative positions of the ends of the two vehicles in real time. The key point detection branch outputs the opening and closing status of the hook tongue, the locking position of the locking iron, the insertion depth of the hook pin, the reset angle of the lifting rod, and the offset of the coupler centerline. The semantic segmentation branch outputs pixel-level masks of the coupler gap area and the compression and stretching of the buffer device.

[0024] Step 3: The recognition results output in Step 2 are compared in real time with the operation process instructions, vehicle speed information and traction braking status. When high-risk events such as unreliable coupler connection, unlocked coupler tongue, abnormal coupler pin posture or excessive coupling impact are detected, a voice alarm is triggered and the event triggering upload process is initiated.

[0025] Step four involves detecting track structure defects based on a joint framework of meta-learning and self-supervised learning. The meta-learning module uses the MAML algorithm (Model-Agnostic Meta-Learning, or MAML for short) to pre-train a feature extractor on normal samples from multiple source scenes, enabling the model to quickly adapt to new scenes. The self-supervised learning module uses a contrastive learning mechanism to mine environment-invariant features from a large number of unlabeled normal track images and constructs a normal feature manifold through an anomaly generation module for subsequent anomaly measurement.

[0026] Step 5: Construct a topological relationship graph of the track structure and coupler components, and use a graph neural network for global reasoning. The topological relationship graph includes spatial topological edges and temporal correlation edges. Spatial topological edges represent the front-to-back adjacency and left-to-right symmetry relationships between components along the mileage direction, while temporal correlation edges represent the state changes of the same component at different time points. The graph neural network combines vehicle speed and mileage information through a message passing mechanism to infer the potential impact of local defects on the overall structure.

[0027] Step six: Execute the periodic health reporting process, sending equipment status summary information and current identification result summary to the ground terminal every 30 seconds; equipment status includes CPU utilization, memory usage, temperature and algorithm running status; identification result summary includes signal status, coupler connection status and defect detection results.

[0028] As a preferred implementation, in step two, the backbone network of the lightweight multi-task neural network architecture adopts a lightweight convolutional neural network with approximately 2.5MB of parameters. After quantization pruning, the total number of parameters of the overall multi-task system is strictly controlled within 8MB, and the inference latency does not exceed 30ms. In the object detection branch, a feature pyramid network is used to achieve multi-scale feature fusion, and the feature pyramid has 3 levels. In the key point detection branch, a heatmap regression method is used to output the coordinates of 17 key points. In the semantic segmentation branch, a lightweight semantic segmentation head is used, and the output resolution is 4 times that of the backbone network.

[0029] The specific method for real-time comparison in step three is as follows: establish an operation status mapping table to map the preset set of operation instructions to the expected status of the corresponding coupler; when the detected actual status of the coupler is inconsistent with the expected status of the corresponding operation instruction in the mapping table, it is judged as an abnormal event; the set of operation instructions includes coupling instructions, uncoupling instructions, traction instructions and braking instructions, and each operation instruction corresponds to a unique combination of expected coupler statuses.

[0030] The specific training process of the meta-learning module in step four includes: the first stage is to perform meta-training on multiple source domain datasets, each source domain containing normal orbit image samples; during the meta-training process, the gradient descent strategy is used to update the initialization parameters of the feature extractor so that the parameters can obtain the minimum classification loss on the validation set after K steps of gradient update; the second stage is to fine-tune on a small number of labeled samples in the target domain, with the number of fine-tuning samples being 1 to 5.

[0031] The self-supervised learning module uses the SimCLR (Simple Framework for Contrastive Learning of Visual Representations) framework for contrastive learning. It constructs positive sample pairs of the same sample after different data augmentations and optimizes them using the InfoNCE loss function. This enables the feature extractor to learn robust feature representations that are insensitive to changes in environment such as lighting and viewing angle. The temperature parameter is set to 0.1, the batch size is set to 32, and the positive sample pairs are constructed using two data augmentation methods: random cropping and color jittering. The feature projection head uses a two-layer fully connected network with an output dimension of 128 for the first layer and 64 for the second layer.

[0032] In a preferred embodiment of the present invention, the anomaly generation module in step four adopts a variational autoencoder (VAE) architecture. The encoder maps the input image to the mean and variance of the latent space, and the decoder reconstructs the image from the sampling points in the latent space. After training on a large number of normal track images, this module constructs a feature manifold of the normal pattern. For the image to be tested, the anomaly score is obtained by calculating its reconstruction error or its Mahalanobis distance in the latent space. When the score exceeds a preset threshold, it is judged as a defect. In order to help the model learn a more robust decision boundary, sampling is performed outward along the center of the normal distribution in the latent space to generate reference points that deviate from the normal pattern. After being reconstructed by the decoder, these reference points are used as expanded negative samples to participate in the training of meta-learning, thereby enhancing the model's sensitivity to abnormal patterns.

[0033] In a preferred embodiment of the present invention, the graph neural network in step five adopts a graph attention network architecture. Its node features include local defect features and historical state features of visual detection output, and edge features include physical connection relationships and temporal correlation relationships between components. The graph attention network iteratively updates node features through a standard multi-head attention message passing mechanism, and realizes spatiotemporal correlation prediction based on vehicle speed and mileage markers by adding a cross-domain penalty coefficient.

[0034] This embodiment also provides a coupler status detection system, including: The image acquisition module is used to acquire image data of the coupler area and track structure area during the operation of the rail work vehicle. The module contains at least 4 high-definition cameras, which are deployed at the A end, B end, side and coupler position of the rail work vehicle. Each camera uses a 1 / 2.8-inch CMOS sensor with 2 million effective pixels, a frame rate of 25fps and an IP65 protection rating. The edge computing module is used to process the image data output by the image acquisition module in real time. This module includes an edge computing host with an 8-core ARM architecture processor with a main frequency of 2.5GHz, 8GB DDR4 memory, and 1TB NVMe SSD storage. The edge computing host runs a lightweight multi-task neural network inference engine to achieve simultaneous execution of object detection, key point detection, and semantic segmentation. The coupler status recognition module is used to identify the coupler connection status based on the output of the edge computing module. Specifically, this module is used to: determine the opening and closing status of the coupler tongue (an open state is determined when the coupler tongue rotation angle exceeds 15 degrees, and a locked state is determined when the coupler tongue rotation angle is less than 5 degrees); determine the locking position of the locking iron (a properly locked state is determined when the longitudinal displacement of the locking iron is less than 2mm); determine the insertion depth of the coupler pin (a reliable insertion is determined when the insertion depth is more than 90% of the total length of the coupler pin); determine the alignment of the coupler centerline (a qualified alignment is determined when the lateral offset is less than 10mm and the height difference is less than 15mm); and determine the compression and stretching of the buffer device (an abnormality is determined when the compression or stretching exceeds 80% of the rated stroke of the buffer device). The track defect detection module is used to detect track structural defects. This module includes a meta-learning unit, a self-supervised learning unit, and an anomaly generation unit. The meta-learning unit uses a prototype network for few-sample classification, the self-supervised learning unit uses a contrastive learning method to train the feature extractor, and the anomaly generation unit uses a variational autoencoder to construct a normal feature manifold for anomaly measurement. The graph neural network reasoning module is used to model and reason about the topological relationships of the track structure and coupler components. This module constructs a directed graph structure containing rail nodes, fastener nodes, sleeper nodes, ballast nodes and coupler component nodes, and infers the impact of local defects on the overall structural safety through message passing mechanism combined with spatiotemporal correlation. The communication module is used for data interaction with the ground terminal. This module includes a dedicated railway network interface and a 5G / 4G module interface, and operates in two modes: event-triggered mode and periodic health reporting mode. In event-triggered mode, when a high-risk event is detected, structured event metadata is uploaded immediately. The event metadata includes event type, level, timestamp, mileage marker, confidence level, and keyframe screenshot. In periodic health reporting mode, a device status and identification result summary are sent every 30 seconds. The auxiliary display module is used to provide visual assistance information to the driver. The module includes a high-definition display screen in the cab, which uses a 15.6-inch LCD panel with a resolution of 1920×1080. It supports split-screen display of 360-degree panoramic overhead view and close-up view of key blind spots. The screen overlays the coupler connection status label, defect marking box, distance measurement value and warning information. The storage module is used to store historical data; this module uses a 1TB NVMe SSD, supports at least 72 hours of continuous video loop recording, and supports USB 3.0 and Gigabit Ethernet high-speed dump interfaces; in playback mode, the video overlay displays the warning box, defect mark, coupler connection status, signal recognition results and distance information at that time.

[0035] Furthermore, the camera in the image acquisition module also includes an adaptive exposure control unit and an image enhancement unit; the adaptive exposure control unit dynamically adjusts the exposure parameters according to the ambient illumination, with an exposure time range of 1ms to 100ms and an ISO range of 100 to 3200; the image enhancement unit uses the CLAHE algorithm (Contrast Limited Adaptive Histogram Equalization, CLAHE) for adaptive histogram equalization and integrates a deep learning-based dehazing network to handle image degradation in rainy and foggy weather.

[0036] Furthermore, the edge computing module also includes a model quantization unit and a model pruning unit; the model quantization unit quantizes the network weights from 32-bit floating-point to 8-bit integers with a quantization error of less than 1%; the model pruning unit uses a structured pruning method to remove convolutional kernels with 16 channels, with a pruning rate of 30%; the quantized model inference speed is increased to 2.5 times that of the original, and the model size is reduced to one-quarter of the original.

[0037] Furthermore, the coupler status recognition module also includes a dynamic early warning unit for the coupling process, which is used to implement graded early warnings based on the relative distance between the two vehicles. When the relative distance is between 20m and 10m, the first-level voice prompt is output: Approaching the coupling vehicle, please pay attention to alignment. When the relative distance is between 10m and 3m, the second-level voice prompt is output: Coupler approaching, prepare for low-speed coupling. When the relative distance is less than 3m, the third-level voice prompt is output: Confirm coupler status, strictly control speed. The distance measurement adopts a monocular depth estimation method, and the error of close-range measurement is no more than 3%.

[0038] Furthermore, the meta-learning unit in the track defect detection module adopts the MAML algorithm, with an inner loop learning rate of 0.01, an outer loop learning rate of 0.001, 1000 meta-training rounds, and 5 support set samples and 15 query set samples in each round. The support set and query set of the prototype network are sampled from a small number of labeled samples in the target domain, with a feature dimension of 512, and a cosine similarity classifier is used as the classifier.

[0039] Furthermore, the graph neural network inference module also includes a risk propagation assessment unit; the risk propagation assessment unit calculates the risk score for each node based on the output of the graph neural network.

[0040] Furthermore, the event-triggered upload in the communication module uses the HTTPS protocol for secure transmission. The uploaded data packet size does not exceed 200KB and includes structured event metadata in JSON format and keyframe screenshots in JPEG format. The keyframe screenshots have a resolution of 640×480 and a JPEG compression quality of 70%. Periodic health reporting uses the MQTT protocol (Message Queuing Telemetry Transport), with a message period of 30 seconds and the message content encoded in a compact binary format.

[0041] Furthermore, the auxiliary display module also includes a blind spot stitching unit, which is used to stitch together images captured by multiple cameras into a 360-degree panoramic overhead view. The blind spot stitching unit adopts an image registration algorithm based on feature points, with a registration accuracy of less than 1 pixel. The refresh rate of the stitched panoramic image is 15fps, and the viewing angle coverage is 360 degrees horizontally and 180 degrees vertically.

[0042] Furthermore, the system also includes a post-event analysis module, which analyzes the stored video data and generates an operation report after the operation is completed. The post-event analysis module automatically marks every alarm event, coupler connection status change, signal change, and distance control anomaly event on the video timeline and supports jumping to key frames. The automatically generated operation report includes driver operation compliance score, route and operation anomaly statistics, and communication quality records.

[0043] Example 2

[0044] This embodiment provides a specific implementation of a multi-target real-time perception method for track maintenance vehicles. It acquires visual information about the coupler area and track structure area using a multimodal perception device deployed on the track maintenance vehicle. Simultaneous detection of multiple targets is achieved through a lightweight multi-task neural network on an edge computing device. Furthermore, it combines a meta-learning and self-supervised learning framework to address the challenge of track structure defect detection under conditions of scarce defect samples. The following details each step.

[0045] In step one, the image acquisition module consists of multiple high-definition cameras deployed on the front and sides of the track maintenance vehicle. Specifically, it includes four high-definition cameras: one at end A, one at end B, one on the side of the vehicle, and one at the coupler position. The deployment positions of each camera are precisely designed to ensure coverage of the coupler coupling area and blind spots around the vehicle. The end A and end B cameras are installed at the front and rear ends of the track maintenance vehicle, respectively, to acquire visual information about the relative positions of the two ends. The side camera is installed in the middle of the side of the vehicle to monitor the track structure and the side environment of the vehicle. The coupler position camera is installed near the coupler to acquire detailed images of various components of the coupler. All of these cameras use 1 / 2.8-inch CMOS sensors with 2 million effective pixels, a frame rate of 25fps, and an IP65 protection rating, enabling them to adapt to the harsh operating environment of railway sites.

[0046] In low-light environments such as nighttime or tunnels, the camera supports starlight-level night vision. When the ambient light level is below 0.1 lux, the camera automatically switches to infrared illumination mode with an infrared wavelength of 850nm. When this wavelength of infrared light illuminates rusted, oxidized metal layers and surfaces covered with heavy oil, diffuse reflection and volume scattering mainly occur, effectively suppressing specular reflections on worn metal surfaces. This significantly improves the imaging contrast of key features of the coupler while avoiding glare interference for drivers and passengers. The camera's adaptive exposure control unit dynamically adjusts exposure parameters according to ambient light, with an exposure time range of 1ms to 100ms and an ISO range of 100 to 3200, ensuring clear images are acquired even when lighting conditions change drastically. The image enhancement unit uses the CLAHE algorithm for adaptive histogram equalization and integrates a deep learning-based dehazing network to process image degradation under rainy and foggy weather. The contrast limit parameter of the CLAHE algorithm is set to 1.5, and the grid block size is set to 8×8 pixels. The dehazing network adopts an encoder-decoder architecture, with the encoder containing 4 convolutional layers and the decoder containing 4 transposed convolutional layers. The number of feature channels are 16, 32, 64, and 128, respectively.

[0047] In step two, the acquired image data is input into a lightweight multi-task neural network architecture for synchronous processing. This architecture is deployed on an edge computing host, which uses an 8-core ARM processor with a clock speed of 2.5GHz, 8GB of DDR4 memory, and 1TB of NVMe SSD storage. The lightweight multi-task neural network architecture includes three parallel detection branches: an object detection branch, a keypoint detection branch, and a semantic segmentation branch. These three branches share a feature extraction backbone network, enabling efficient reuse of inference computation.

[0048] The backbone network employs a lightweight convolutional neural network with approximately 2.5MB of parameters. After quantization pruning, the overall multi-task system's parameter count is strictly controlled to within 8MB, with an inference latency of no more than 30ms. In practical applications, the backbone network adopts the MobileNetV3 architecture, comprising five stages. Each stage includes several depthwise separable convolutional blocks and an SE attention module. The first stage has 16 output channels, the second stage has 24, the third stage has 40, the fourth stage has 80, and the fifth stage has 160. The backbone network extracts multi-scale spatial feature maps from the third, fourth, and fifth stages, projects them layer by layer to a unified shared channel, replacing the traditional global average pooling one-dimensional vector output. This preserves absolute spatial location information for subsequent feature pyramid-based branches, outputting to the three detection branches.

[0049] The target detection branch employs a one-stage target detection algorithm to achieve real-time detection of coupler-related targets and track structure targets. In its implementation, an improved YOLOv5 architecture is used. The detection head contains three output layers corresponding to large, medium, and small targets, respectively. The feature pyramid network has three levels for multi-scale feature fusion. Features are extracted from the third, fourth, and fifth stages of the backbone network at each level, and fusion is achieved through upsampling and feature concatenation operations. The target detection branch can output target categories including coupler, coupler tongue, coupler pin, lifting rod, buffer, and the relative positions of the ends of the two cars. The output information for each target includes the coordinates of the top-left corner of the target bounding box, the width and height of the target bounding box, and the target category confidence score.

[0050] The key point detection branch outputs the spatial position and status information of each key component of the coupler. It uses a heatmap regression method to output heatmaps of 17 key points, each corresponding to a Gaussian heatmap with a standard deviation of 4 pixels for the Gaussian kernel. The output of this branch includes three parts: heatmap, offset, and regression value. The heatmap is used to predict the existence probability of each key point, the offset is used to compensate for coordinate errors caused by downsampling, and the regression value is used to regress the precise coordinates of the key points. Key points that can be output include the hook tongue opening / closing state point, the lock iron locking position point, the hook pin insertion depth point, the hook rod reset angle point, and the coupler centerline offset point.

[0051] The semantic segmentation branch outputs pixel-level masks of the coupler gap region and the compression / stretching of the buffer device. It employs a lightweight semantic segmentation head, with an output resolution four times that of the backbone network. The semantic segmentation head includes a feature extraction upsampling path and a segmentation output layer. The upsampling path contains three upsampling convolutional blocks, each containing one transposed convolutional layer and two 3×3 convolutional layers. The segmentation output layer uses 1×1 convolutions to compress the number of feature channels to the number of target categories, and then upsamples to the original input resolution using bilinear interpolation. The segmentable regions include the coupler gap region, buffer device region, rail region, fastener region, and sleeper region.

[0052] The joint optimization objective function of this lightweight multi-task neural network architecture is built upon a feature decoupling mechanism. The shared feature maps output by the backbone network are orthogonalized by task-specific attention mask layers before being input to each branch, thus eliminating gradient interference between branches. The joint loss function of the overall network is defined as:

[0053]

[0054] in, Indicates the overall joint loss; , , These are the corresponding loss functions for object detection, keypoint detection, and semantic segmentation branches, respectively. , , The dynamic weight adjustment factor employs an uncertainty-weighted strategy, and its value is... , The first adaptively learned model during training The homoscedasticity noise parameters for each task, Initialize to ; , where is the static weight coefficient of the feature orthogonal regularization term, typically taking a value of 0.01; The orthogonal regularization term is specifically represented as:

[0055]

[0056] in and These are the corresponding feature matrices generated by the object detection branch and the keypoint detection branch after sharing the feature map, respectively. Using the Frobenius norm, it ensures that under extreme computing power constraints, a single feature map can maximize information reuse without semantic confusion by forcing the feature matrices of different tasks to remain orthogonal in high-dimensional space.

[0057] To further reduce the computational complexity and storage overhead of the model, the edge computing module also includes a model quantization unit and a model pruning unit. The model quantization unit quantizes the network weights from 32-bit floating-point to 8-bit integers with a quantization error of less than 1%. The model pruning unit uses a structured pruning method to remove convolutional kernels with 16 channels, achieving a pruning rate of 30%. After quantization and pruning, the model inference speed is increased to 2.5 times the original speed, and the model size is reduced to one-quarter of the original size. The measured inference latency of the quantized overall system model on the edge computing host is 12ms, supporting real-time processing at up to 83fps, fully meeting the real-time computing requirements of a 25fps acquisition frame rate in low-speed operation scenarios of rail vehicles.

[0058] In step three, the recognition results output from step two are compared in real time with the work process instructions, vehicle speed information, and traction and braking status. Specifically, a work status mapping table is established, mapping a preset set of work instructions to the corresponding expected coupler states. This comparison process essentially constructs a multidimensional Boolean logic state matrix: let the real-time work instruction vector of the railcar be... The real-time detected coupler state vector is Each element is a Boolean value (1 represents normal / in place, 0 represents abnormal / not in place); the system internally stores an ideal state mapping matrix. The real-time comparison process follows the decision equation:

[0059]

[0060] When the alarm flag is set When this happens, the system triggers an event. For example, for a hang-up command ( If visual detection indicates that the hook tongue is not locked (i.e., the hook tongue rotation angle is not detected), ),lead to XOR operation A logic high level is generated, and the system triggers the third-level voice alarm with zero delay and captures 15 frames before and after the current timestamp, which are then encapsulated as structured event metadata and stored in a buffer queue.

[0061] High-risk events that trigger the voice alarm include: unreliable coupler connection, manifested as a lateral offset of the coupler centerline exceeding 10mm or a height difference exceeding 15mm; coupler tongue not locking, manifested as a rotation angle exceeding 15 degrees; abnormal coupler pin posture, manifested as a coupler pin insertion depth less than 90% of the total coupler pin length; and excessive coupling impact, manifested as a compression or extension of the buffer device exceeding 80% of the buffer device's rated stroke. The voice alarm uses preset voice prompts, played through a speaker in the driver's cab.

[0062] The coupler status recognition module also includes a dynamic early warning unit for the coupling process, used to implement graded warnings based on the relative distance between the two cars: when the relative distance is between 20m and 10m, the first-level voice prompt is output: "Approaching coupling vehicle, please pay attention to alignment"; when the relative distance is between 10m and 3m, the second-level voice prompt is output: "Coupler approaching, prepare for low-speed coupling"; when the relative distance is less than 3m, the third-level voice prompt is output: "Confirm coupler status, strictly control speed." Distance measurement uses a monocular depth estimation method, with a near-distance measurement error of no more than 3%. The monocular depth estimation network is pre-trained on a dedicated railway track coupler dual-target positioning dataset, using scale-invariant logarithmic error as the loss function. High accuracy in near-distance measurement is ensured by minimizing the relative gradient between the predicted depth map and the true depth map. The network employs a lightweight encoder-decoder architecture, with a MobileNetV2 encoder and a 4-layer upsampled convolutional network decoder, with an output resolution four times that of the backbone network.

[0063] In step four, track structure defects are detected based on a joint framework of meta-learning and self-supervised learning, aiming to solve the problem of insufficient generalization ability of the detection model caused by the scarcity of defect samples and class imbalance. The specific process of the joint framework is as follows: First, a general feature extractor is pre-trained on a large number of unlabeled normal track images using the SimCLR framework in the self-supervised learning module. The SimCLR framework constructs positive sample pairs of the same sample after different data augmentations and uses the InfoNCE loss function (InformationNoiseContrastiveEstimation, or InfoNCE for short) for contrastive learning, enabling the feature extractor to learn robust feature representations that are insensitive to changes in environment such as illumination and viewing angle. The temperature parameter is set to 0.1, the batch size is set to 32, and the positive sample pairs are constructed using two data augmentation methods: random cropping and color jittering. The feature projection head uses a two-layer fully connected network, with the first layer having an output dimension of 128 and the second layer having an output dimension of 64. Then, the parameters of this pre-trained feature extractor are used as the initialization of MAML, and meta-training is performed on a meta-task consisting of normal samples from multiple source domains, enabling the model to quickly adapt to new tasks. Finally, fine-tuning is performed when there are only 1 to 5 labeled defect samples in the target domain to achieve small sample defect detection.

[0064] The meta-learning module employs a hybrid architecture nested with MAML and a prototype network. The MAML algorithm has an inner loop learning rate of 0.01 and an outer loop learning rate of 0.001, with 1000 meta-training epochs. Each epoch samples 5 support set samples and 15 query set samples. The feature extractor uses a ResNet18 architecture, outputting 512-dimensional features. In the second stage of fine-tuning, an Adam optimizer with a learning rate of 0.001 is used for 100 epochs. The prototype network, as a metric-based classification head, is embedded within the meta-learning unit's inner loop. It first extracts sample features from each class in the support set using the feature extractor, then calculates the prototype feature vector for each class (i.e., the mean of all sample features for that class), and finally classifies by calculating the cosine similarity between the query set sample features and the prototypes of each class. The classifier's temperature parameter is set to 0.1.

[0065] The self-supervised learning module employs the SimCLR framework for contrastive learning, using the InfoNCE loss function. By constructing positive sample pairs of the same sample after different data augmentations, the feature extractor learns robust feature representations insensitive to environmental changes such as illumination and viewing angle. The anomaly generation module uses a variational autoencoder architecture. The encoder contains four convolutional layers and two fully connected layers. The number of channels in the convolutional layers are 32, 64, 128, and 256, respectively. The fully connected layers output the mean and variance of the latent space, with the latent space dimension set to 32. The decoder contains two fully connected layers and four transposed convolutional layers, with the number of channels symmetrical to the encoder. The last layer uses a sigmoid activation function to obtain the reconstructed image. After training on a large number of normal track images, this variational autoencoder constructs a low-dimensional manifold of the normal pattern, whose latent space follows a multivariate Gaussian distribution with parameters of the mean and standard deviation of the normal data. For the test image, an anomaly score is obtained by calculating its reconstruction error or its Mahalanobis distance in the latent space. When the score exceeds a preset threshold, it is judged as a defect.

[0066] To help the model learn more robust decision boundaries, sampling can be performed outward from the center of the normal distribution in the latent space to generate reference points that deviate from the normal pattern. These reference points, after being reconstructed by the decoder, serve as augmented negative samples in the training of meta-learning, enhancing the model's sensitivity to abnormal patterns. This sampling process follows the equations below:

[0067]

[0068] in, For the generated reference point latent feature vector, and Let be the mean and standard deviation vectors of the normal data distribution in the latent space, respectively. To obtain from a uniform distribution Random direction vectors sampled in the middle, For Hadama product, The offset coefficient and This is to ensure that the sampling points are far from the normal area. These reconstructed images These are not real defects, but they provide diverse abnormal reference patterns to improve the model's generalization ability under conditions with very few real defect samples. The loss function of the variational autoencoder includes reconstruction loss and KL divergence loss, with a weight ratio of 1:0.01.

[0069] In step five, a topological relationship graph of the track structure and coupler components is constructed, and a graph neural network is used for global reasoning. The topological relationship graph includes spatial topological edges and temporal correlation edges. Spatial topological edges represent the front-to-back adjacency and left-to-right symmetry relationships between components along the mileage direction, while temporal correlation edges represent the state changes of the same component at different time points. By associating the historical state features of components with their current state features, temporal information is modeled.

[0070] The graph neural network employs a graph attention network architecture, comprising four graph attention layers, each with a hidden dimension of 256 and eight attention heads. The feature vector of each node is composed of 512-dimensional visual features and 128-dimensional historical state features, resulting in a total dimension of 640. Edge features are composed of 32-dimensional positional codes representing physical connections and 32-dimensional temporal codes representing temporal relationships, resulting in a total dimension of 64. This graph network constructs a heterogeneous spatiotemporal graph, with the node set containing environmental trigger nodes (e.g., rail wear and ballast settlement) and response receptor nodes (e.g., coupler alignment offset). The message passing function expression for the graph attention network is:

[0071]

[0072] in, and They represent the first Layer and first Layer nodes eigenvectors, Represents a node The neighborhood set, Represents a node With nodes Attention weights between them The weight matrix is ​​a learnable matrix. ReLU is used as the activation function. In the message passing computation, a cross-domain penalty coefficient is added to the attention weights. When the node is a coupler component node and its neighboring node is a track defect node directly below it, this coefficient calculates the impact delay time based on the train speed and assigns dynamic weights to the temporal edges, enabling the system to make spatiotemporal correlation predictions based on train speed and odometer markers: if the preceding frame is at the odometer marker... If severe track bed settlement is detected at a certain location, the system will increase the weight of attention given to the compression of the coupler buffer when passing through that location, thus improving the system from delayed alarm to proactive prediction.

[0073] The graph neural network inference module also includes a risk propagation assessment unit, which calculates the risk score for each node based on the output of the graph neural network, as shown in the following formula:

[0074]

[0075] in, Represents a node Risk score, Represents a node With nodes Edge weights between them Represents a node The severity of the corresponding defect, Represents a node The detection confidence level of the corresponding defect. When the risk score exceeds the preset threshold of 0.7, a maintenance work order recommendation is generated and uploaded to the ground terminal.

[0076] In step six, the periodic health reporting process is executed, sending equipment status summary information and current identification result summary to the ground terminal every 30 seconds. Equipment status includes CPU utilization, memory usage, temperature, and algorithm running status. CPU utilization is calculated by reading the ` / proc / stat` file, memory usage by reading the ` / proc / meminfo` file, and temperature by reading the ` / sys / class / thermal / thermal_zone0 / temp` file. Algorithm running status includes the running status flags of each module. The identification result summary includes signal status, coupler connection status, and defect detection results. Signal status includes signal color and display confidence level; coupler connection status includes the values ​​and results of each judgment indicator; and defect detection results include defect type, quantity, location, and confidence level.

[0077] This embodiment provides a coupler status detection system for implementing the above-mentioned method. It includes an image acquisition module, an edge computing module, a coupler status recognition module, a track defect detection module, a graph neural network inference module, a communication module, an auxiliary display module, and a storage module. The functions and hardware configurations of each module have been described in Embodiment 1 and will not be repeated here. The system also includes a post-processing analysis module for analyzing the stored video data and generating a work report after the operation is completed. It automatically marks every alarm event, coupler connection status change, signal change, and distance control anomaly event on the video timeline, supports jumping to keyframes, and the automatically generated work report includes driver operation compliance scores, track and operation anomaly statistics, and communication quality records.

[0078] To verify the technical effects of the present invention, specific test examples are provided below.

[0079] In the low-light environment coupler status detection test, the ambient light level was set to 0.05 lux, which is lower than the 0.1 lux threshold for automatic infrared supplementary lighting by the camera. The test scenario involved two railcars coupling at a speed of 5 km / h. The test results showed that the system could correctly identify the opening and closing status of the coupler tongue, with a locking accuracy rate of 97.3% and an opening accuracy rate of 96.8%; the locking position recognition accuracy rate was 95.6%; the pin insertion depth recognition accuracy rate was 94.2%; the coupler centerline alignment recognition accuracy rate was 93.7%; and the dynamic early warning unit during coupling triggered three levels of voice prompts at 20m, 10m, and 3m, with a 100% prompt timeliness rate.

[0080] In a track defect detection test under conditions of scarce defect samples, a track image dataset provided by a railway bureau was used. This dataset contained 10,000 normal track images and only 15 defect images (rail wear, missing fasteners, and sleeper cracks). The test results showed that, using the meta-learning and self-supervised learning joint framework of this invention, the detection rate for rail wear was 96.7%, for missing fasteners 95.2%, and for sleeper cracks 93.8%. After fine-tuning using only 5 defect samples, the model's rapid adaptability met the needs of practical applications.

[0081] Comparative Example 1 uses a traditional single-task detection method, processing coupler condition detection and track defect detection into two independent models. The overall model parameter size is 125MB, and the single-frame inference latency is 85ms. Under the same test conditions, the rail wear detection rate is 78.3%, the fastener missing detection rate is 72.1%, and the sleeper crack detection rate is 68.5%. Comparative Example 2 uses a purely supervised learning method, training with only 33 defect images and without employing meta-learning and self-supervised learning strategies. Its rail wear detection rate is only 45.2%, the fastener missing detection rate is only 38.7%, and the sleeper crack detection rate is only 32.1%, which cannot meet the needs of practical applications.

[0082] To more clearly demonstrate the technical effects, the table below lists comparative data for key performance indicators: Table 1 Comparison of Key Performance Indicators

[0083]

[0084] In summary, the multi-target real-time perception method and coupler status detection system for rail maintenance vehicles provided by this invention achieve real-time synchronous detection of multiple targets through a lightweight multi-task neural network architecture, solve the detection problem under conditions of scarce defect samples through a joint framework of meta-learning and self-supervised learning, and realize the correlation analysis between local defects and overall structural safety through graph neural networks. The test data from the embodiments and comparative examples fully verify the technical effects and superiority of this invention.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-target real-time perception method for rail-mounted work vehicles, characterized in that: Includes the following steps: Step S1: Use multiple cameras deployed on the rail work vehicle to collect image data of the coupler area and the track structure area in real time; Step S2: The acquired image data is input into a lightweight multi-task neural network architecture for synchronous processing. The lightweight multi-task neural network architecture includes an object detection branch, a key point detection branch, and a semantic segmentation branch. The three branches share a feature extraction backbone network. Specifically, the object detection branch outputs the target bounding box coordinates of the coupler and track structure targets, the key point detection branch outputs the spatial position and state information of the key components of the coupler, and the semantic segmentation branch outputs the pixel-level mask of the coupler gap area and the compression and stretching amount of the buffer device. Step S3: The identification result is compared with the operation process instructions and vehicle speed information in real time. When the detected actual state of the coupler is inconsistent with the expected state corresponding to the current operation instruction, an alarm is triggered. Step S4: Detect track structure defects based on the joint framework of meta-learning and self-supervised learning. In this step, a feature extractor is pre-trained on unlabeled normal track images using self-supervised learning. The parameters of the pre-trained feature extractor are used as the initialization of the meta-learning algorithm and meta-trained, so that the model can detect defects under the condition of a small number of labeled samples in the target domain. Step S5: Construct a topological relationship graph of the track structure and coupler components, and use a graph neural network for global reasoning. The topological relationship graph contains spatial topological edges and temporal correlation edges. The graph neural network combines vehicle speed and odometer information through a message passing mechanism to infer the degree of impact of local defects on the overall structure.

2. The multi-target real-time perception method for track maintenance vehicles according to claim 1, characterized in that: In the lightweight multi-task neural network architecture, the three branches share the multi-scale spatial feature map extracted by the backbone network. Before the shared feature map is input into each branch, it is orthogonalized by the attention mask layer of the specific task. The joint loss function includes a feature orthogonality regularization term. This regularization term is constructed by calculating the inner product of the feature matrices generated by different branches after sharing the feature map, which forces the feature matrices of different tasks to remain orthogonal in order to eliminate gradient interference between branches.

3. The multi-target real-time perception method for track maintenance vehicles according to claim 1, characterized in that: Real-time comparison is achieved by constructing a multi-dimensional Boolean logic state matrix: a mapping relationship between the set of operation instructions and the expected combination of coupler states is established in advance, and the real-time detected coupler state vector is XORed with the expected state vector found according to the current operation instruction. An alarm is triggered when any state bit does not match.

4. The multi-target real-time perception method for track maintenance vehicles according to claim 1, characterized in that: In the joint framework of meta-learning and self-supervised learning, a contrastive learning framework is first used to pre-train a general feature extractor on unlabeled normal track images by constructing positive sample pairs, enabling the feature extractor to learn robust feature representations that are insensitive to changes in illumination and viewpoint. Then, the parameters of this pre-trained feature extractor are used as the initialization parameters of the MAML algorithm, and meta-training is performed on a meta-task consisting of normal samples from multiple source domains, enabling the model to quickly adapt to new tasks. Finally, fine-tuning is performed under the condition that there are only a few labeled defect samples in the target domain to achieve small sample defect detection.

5. The multi-target real-time perception method for a track maintenance vehicle according to claim 1 or 4, characterized in that: In track structure defect detection, a variational autoencoder is also used to train on normal track images to construct the feature manifold of the normal pattern. For the image to be tested, the anomaly score is obtained by calculating its reconstruction error or its Mahalanobis distance in the latent space. When the anomaly score exceeds a preset threshold, it is judged as a defect. In addition, sampling is performed outward along the center of the normal distribution in the latent space to generate reference points that deviate from the normal pattern. After being reconstructed by the decoder, these reference points are used as expanded negative samples to participate in meta-learning training to enhance the model's sensitivity to abnormal patterns.

6. The multi-target real-time perception method for a track maintenance vehicle according to claim 1, characterized in that: In the topological relationship graph, spatial topological edges represent the front-to-back adjacency and left-to-right symmetry relationships between components along the mileage direction, while temporal correlation edges represent the state changes of the same component at different time points, linking the historical state features of the component with the current state features. The graph neural network adopts a graph attention network architecture, where the node features are composed of the local defect features output by visual detection and the historical state features, and the edge features are composed of the physical connection relationship encoding and the temporal correlation relationship encoding.

7. The multi-target real-time perception method for track maintenance vehicles according to claim 6, characterized in that: In the message passing process of the graph neural network, when a node is a coupler component node and its neighboring nodes are track defect nodes, the delay time from the defect location to the coupler location is calculated based on the train's running speed, and dynamic weights are assigned to the temporal correlation edges to achieve spatiotemporal correlation prediction based on train speed and mileage markers, thereby increasing the attention weight to the coupler status at the defect location in advance.

8. The multi-target real-time perception method for a track maintenance vehicle according to claim 7, characterized in that: In global reasoning, risk scores for each node are calculated based on the output of the graph neural network. These risk scores are obtained by weighted summation of the severity of each defect in the node's neighborhood and the detection confidence. When the risk score exceeds a preset threshold, a work order recommendation containing the defect type, location, and suggested maintenance measures is generated.

9. The multi-target real-time perception method for a track maintenance vehicle according to claim 1, characterized in that: In real-time comparison, graded warnings are also implemented based on the relative distance between the two vehicles. The closer the distance, the higher the warning level. The distance measurement adopts a monocular depth estimation method. The network is trained by minimizing the relative gradient between the predicted depth map and the real depth map to ensure the accuracy of close-range measurement.

10. A coupler condition detection system, characterized in that: For implementing the method of any one of claims 1 to 9, comprising: The image acquisition module is used to acquire image data of the coupler area and track structure area during the operation of the rail work vehicle; The edge computing module is used to run a lightweight multi-task neural network inference engine to process image data in real time and simultaneously perform object detection, key point detection and semantic segmentation tasks. The edge computing module also includes a model quantization unit and a model pruning unit to compress the model size and improve inference speed. The coupler status recognition module is used to identify the coupler connection status based on the output of the edge computing module and preset status judgment conditions, and to perform real-time comparison with the operation instructions and hierarchical early warning. The track defect detection module includes a meta-learning unit, a self-supervised learning unit, and an anomaly generation unit, which are used to execute a joint framework of meta-learning and self-supervised learning to realize track structure defect detection under the condition of scarce defect samples. The graph neural network inference module is used to construct a topological relationship graph and perform global graph neural network inference, outputting defect correlation analysis results and risk scores; The communication module is used for data interaction with the ground terminal and supports two modes: event-triggered upload and periodic health reporting. The auxiliary display module is used to provide drivers with a visual aid image that overlays the detection results with annotations.

Images