Railway outdoor signal equipment monitoring system based on computer vision

Abstract

The application discloses a railway outdoor signal device monitoring system based on computer vision and relates to the technical field of computer vision, comprising: a data synchronization module, which collects device images and synchronously monitors data; a contour evolution module, which constructs an improved Chan-Vese model, generates a device optimized contour set and a device boundary point set; a structure manifold module, which uses the Nystrom method for low-rank approximation and calculates a device structure manifold offset value; a visual coupling module, which generates a visual current coupling offset; a topological drift module, which constructs a device time sequence topological drift graph and generates a device state drift feature; and a state judgment module, which determines a device fault grade result. Through the introduction of the improved Chan-Vese model and the Nystrom method, the application realizes the accurate monitoring and early fault identification of the structural change, electrical change and continuous operation state evolution process of the railway outdoor signal device.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a computer vision-based monitoring system for railway outdoor signaling equipment. Background Technology

[0002] As railway transportation systems develop towards high-speed, intelligent, and networked directions, outdoor railway signaling equipment undertakes the tasks of train operation control, track status transmission, and traffic safety assurance. Signals, track circuits, enclosure equipment, and related auxiliary facilities operate in an outdoor environment for extended periods, exposed to factors such as temperature changes, vibration and shock, rain and snow immersion, equipment aging, and external interference, making them prone to structural deformation, loose connections, status drift, and electrical anomalies. Due to the large number and wide distribution of outdoor railway signaling equipment, traditional manual inspection methods suffer from long inspection cycles, insufficient real-time performance, and significant subjective judgment discrepancies.

[0003] Existing railway outdoor signal equipment monitoring technologies typically employ camera devices to acquire equipment images, which are then combined with track circuit voltage and current, signal status, and environmental monitoring data for joint analysis. Equipment status identification is achieved through target detection, edge extraction, or status threshold judgment. Basic schemes usually first perform contour segmentation on the equipment images, then extract structural features and combine them with monitoring data for status determination. However, these methods primarily rely on single-moment image features or simple numerical changes for analysis, lacking the ability to jointly model the evolution of equipment boundary structures, the correlation between visual and electrical changes, and the continuous temporal state drift process. This makes them prone to misjudgment and missed detection when equipment experiences slow structural shifts, local deformation, or early anomalies.

[0004] Therefore, how to provide a computer vision-based monitoring system for railway outdoor signaling equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a computer vision-based monitoring system for railway outdoor signaling equipment. This invention utilizes images of railway outdoor signaling equipment, 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, enclosure temperature and humidity, water immersion, enclosure opening, and vibration monitoring data. It details the improved Chan-Vese contour evolution process based on dual-level set self-coupling constraints, a full-view cost prediction field, and HJB optimal evolution constraints. Furthermore, it combines edge-preserving asymmetric similarity relationships, optimal transport mapping, and Nystrom low-rank approximation to construct the equipment structural manifold features. Simultaneously, it integrates visual current coupling analysis, equipment temporal topology drift modeling, and multi-source state consistency verification to achieve joint analysis of structural changes, electrical changes, and continuous operational state evolution processes of railway outdoor signaling equipment. Compared with existing technologies, this invention has the advantages of high monitoring accuracy, low false positive rate, strong continuous state tracking capability, and high early anomaly identification capability.

[0006] A computer vision-based railway outdoor signaling equipment monitoring system according to an embodiment of the present invention includes: The data synchronization module collects images and synchronous monitoring data of railway outdoor signaling equipment, performs preprocessing on the synchronous monitoring data and binds it to the images of railway outdoor signaling equipment to generate visual electrical synchronization status data of the equipment. The contour evolution module reads images of railway outdoor signal equipment, establishes an initial contour curve in the target area of ​​the equipment, and constructs an improved Chan-Vese model with HJB optimal evolution constraints to perform contour evolution. It updates the contour evolution direction according to the minimum contour advancement cost and generates an optimized contour set and a set of equipment boundary points. The structural manifold module constructs a similarity matrix of device boundaries based on the set of device boundary points, and uses the Nystrom method to perform low-rank approximation to generate device structural manifold features. It then compares the current device structural manifold features with historical normal structural manifold features to calculate the device structural manifold offset value. The visual coupling module calculates the two-dimensional displacement potential field of the device boundary along the normal direction based on the device optimized contour set, and couples it with the one-dimensional current change sequence to generate visual current coupling offset. The topology drift module reads the device structure manifold offset value and visual current coupling offset according to the sliding time window, constructs the device time-series topology drift diagram, calculates the state trajectory change characteristics, and generates device state drift characteristics. The status determination module performs multi-source status consistency verification on the equipment status drift characteristics based on the equipment's visual electrical synchronization status data, generates a multi-source status consistency score, and determines the equipment fault level result.

[0007] Optionally, the synchronous monitoring data specifically includes 25Hz phase-sensitive track circuit voltage and current data, signal lighting unit voltage and current data, traction return current data, box temperature and humidity data, water immersion status data, box opening status data, and vibration status data.

[0008] Optionally, the generated device visual electrical synchronization status data includes: Read the synchronous monitoring data, perform time alignment according to the collection timestamp, write data with a time difference of less than 0.5 seconds into the same synchronous time window, remove data with empty values, incorrect collection labels and data that exceed the sensor range, and perform resampling at a 1-second sampling interval to generate standardized synchronous monitoring data; The acquisition time, equipment number, and monitoring point number of the railway outdoor signal equipment image are read, and the railway outdoor signal equipment image is bound with standardized synchronous monitoring data within the same equipment number, the same monitoring point number, and the same synchronous time window to generate equipment visual electrical synchronization status data.

[0009] Optionally, the contour evolution module includes: An improved Chan-Vese model is constructed, which includes an initial contour module, a region statistics module, an optimal evolution module, and a boundary output module. The initial contour module reads the image of the railway outdoor signal equipment, establishes the first initial closed contour curve and the second initial closed contour curve in the target area of ​​the equipment, and converts them into the first level set function and the second level set function respectively. It constructs a double level set self-coupling constraint term, exchanges the inner and outer region identifiers of the contour according to the 20-iteration cycle, and generates a double level set contour set. The regional statistics module reads images of railway outdoor signal equipment, divides the images into 32x32 pixel grids, calculates local variance, local entropy, and local gradient variance, and performs weighted summation to generate a full-view cost prediction field. Areas with a value less than the median cost value in the full image are marked as prohibited areas. The optimal evolution module reads the dual-level set of contours, assigns positive-value pixels to the inner region of the contour, and assigns negative-value pixels to the outer region of the contour. It calculates the mean gray value of the inner region and the mean gray value of the outer region respectively, and dynamically updates the weight values ​​of the inner and outer regions based on the total energy change rate. Calculate the gray-scale region difference value, curvature change value, occlusion intensity value, and local noise intensity value. Adjust the proportion of gray-scale region difference value in the contour advancement cost value according to the updated inner and outer region weight values. Perform weighted summation on the gray-scale region difference value, curvature change value, occlusion intensity value, and local noise intensity value to obtain the contour advancement cost value. Combine the full-view cost prediction field to perform HJB optimal evolution constraint, eliminate advancement directions pointing to prohibited entry regions, and determine the optimal contour advancement direction for the current iteration. The boundary output module maps the current contour to the three-dimensional voxel mesh space, calculates the voxel surface tension constraint value and performs contour update. When the difference in the contour advancement cost between two adjacent iterations is less than 0.001, the iteration stops, the zero level set boundary is extracted to generate the device optimized contour set, and the device boundary point set is generated according to a 5-pixel sampling interval. The improved Chan-Vese model was trained using a joint loss value composed of contour advancement cost error, region segmentation error, dual level set consistency error, full view cost prediction error, and boundary reconstruction error as the optimization objective. The parameters in the initial contour module, region statistics module, optimal evolution module, and boundary output module were continuously optimized. Training was stopped when the change in the joint loss value corresponding to 10 consecutive iterations was less than 0.001.

[0010] Optionally, the structural manifold module includes: Based on the set of device boundary points, the pixel coordinates, curvature value, tangent direction angle, boundary gray value and component connection distance of each boundary point are extracted according to the device number. The zero-to-one interval normalization processing is performed on each feature to generate a boundary point feature sequence. Based on the feature sequence of boundary points, calculate the spatial distance difference, curvature difference, orientation angle difference, gray level difference and connection distance difference between any two boundary points, and calculate the boundary affinity value. Determine the high curvature boundary points and low curvature boundary points according to the curvature value, retain the boundary affinity value of the high curvature boundary point pointing to the low curvature boundary point, and set the reverse boundary affinity value to 0 to generate an edge-preserving asymmetric similarity matrix. Based on the edge-preserving asymmetric similarity matrix, initial representative boundary points are selected according to 10% of the number of boundary points. The Nystrom method is used to perform initial low-rank approximation to generate an initial low-dimensional boundary embedding feature matrix. The embedding error hotspot region is determined based on the reconstruction difference between the initial low-dimensional boundary embedding feature matrix and the edge-preserving asymmetric similarity matrix. In the hotspot region of the embedding error, add representative boundary points corresponding to 15% of the number of boundary points, merge them to generate a recursive representative boundary point set, convert the similarity matrix corresponding to the recursive representative boundary point set into an affinity probability distribution, use Sinkhorn-Knopp iterative execution to perform optimal transport mapping, generate transport-weighted representative similarity matrix, and use the Nystrom method to perform low-rank approximation to generate equipment structure manifold features. Read the historical normal structural manifold features corresponding to the same device number, compare the current device structural manifold features with the historical normal structural manifold features item by item according to the corresponding dimensions, and calculate the device structural manifold offset value.

[0011] Optionally, the visual coupling module includes: Read the optimized contour set of the device, extract the current pixel coordinates of the boundary points of each device according to the device number and the acquisition time, and read the coordinates of the historical normal boundary points corresponding to the same device number. Match the current boundary point coordinates with the historical normal boundary point coordinates to generate a boundary point matching sequence. Based on the boundary point matching sequence, the displacement distance and normal direction of each current boundary point relative to the historical normal boundary points are calculated. The displacement distance is projected along the normal direction to obtain the boundary normal displacement value. All boundary normal displacement values ​​are arranged according to the pixel position of the boundary point in the image to generate a two-dimensional displacement potential field. Read the 25Hz phase-sensitive track circuit voltage and current data, signal lighting unit voltage and current data and traction return current data corresponding to the same equipment number from the equipment visual electrical synchronization status data, extract the current component in each data, calculate the absolute difference between the current components of two adjacent acquisition times, and generate a one-dimensional current change sequence. Normalize the boundary normal displacement value and the one-dimensional current change sequence in the two-dimensional displacement potential field to obtain the visual displacement value and the current change value. Multiply the visual displacement value by 0.60 and the current change value by 0.40 and add them together to generate the visual current coupling value. Then, aggregate all visual current coupling values ​​according to the device number to generate the visual current coupling offset.

[0012] Optionally, the topology drift module includes: According to a 60-second window length and a 10-second sliding step, the device structure manifold offset value and visual current coupling offset of the same device number within a continuous time window are read to form a device state vector, and a device state vector sequence is generated according to the acquisition time order. Based on the device state vector sequence, the device state vector is used as a state node. The structural manifold offset difference and the visual current coupling offset difference between two adjacent state nodes are calculated. The structural manifold offset difference is multiplied by 0.55 and the visual current coupling offset difference is multiplied by 0.45 and then added together to obtain the state transition edge weight. Based on the state nodes and the state transition edge weight, the device time-series topology drift graph is constructed. Based on the device time-series topology drift graph, the average value of the device state vector of all state nodes is calculated as the state trajectory center, the distance between the current state node and the historical normal state trajectory center is calculated as the state trajectory center offset value, the growth rate of the adjacent state transition edge weight is calculated as the trajectory divergence value, and the change rate of the state node distribution density within a unit time window is calculated as the state density change value. The state trajectory center offset value, trajectory divergence value, and state density change value are arranged in the order of acquisition time to generate state trajectory change characteristics. These characteristics are then associated with the equipment number, equipment structure manifold offset value, and visual current coupling offset to generate equipment state drift characteristics.

[0013] Optionally, the state determination module includes: Read the equipment's visual electrical synchronization status data and equipment status drift characteristics, extract the status trajectory center offset value, trajectory divergence value and status density change value, and read the 25Hz phase-sensitive track circuit voltage and current, signal light unit voltage and current, traction return current, box temperature and humidity, immersion, unpacking and vibration data to generate a multi-source verification dataset; Normalization is performed on the state trajectory center offset value, trajectory divergence value, and state density change value. The state trajectory center offset value is multiplied by 0.40, the trajectory divergence value is multiplied by 0.35, and the state density change value is multiplied by 0.25, and then added together to generate the structure drift score. Read the multi-source verification dataset, calculate the variation amplitude of 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, box temperature and humidity and vibration data under two adjacent acquisition times, and generate multi-source state change scores by combining water immersion data and box opening data. The structure drift score is multiplied by 0.55 and the multi-source state change score is multiplied by 0.45, and then the results are added together to generate the multi-source state consistency score. When the multi-source status consistency score is less than 0.40, it is determined to be a normal status level; when the multi-source status consistency score is greater than or equal to 0.40 and less than 0.65, it is determined to be a potential fault level; when the multi-source status consistency score is greater than or equal to 0.65 and less than 0.85, it is determined to be an enhanced early warning level; and when the multi-source status consistency score is greater than or equal to 0.85, it is determined to be a fault alarm level.

[0014] The beneficial effects of this invention are: This invention, a computer vision-based railway outdoor signaling equipment monitoring system, compared to existing railway outdoor signaling equipment monitoring technologies, can jointly analyze the equipment status by combining images of railway outdoor signaling equipment, 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, enclosure temperature and humidity, water immersion, unpacking, and vibration monitoring data. By introducing a dual-level set self-coupling constraint term, a full-view cost prediction field, and HJB optimal evolution constraints to construct an improved Chan-Vese model, dynamic contour evolution processing is performed on the boundaries of railway outdoor signaling equipment. Compared to traditional edge extraction and single contour segmentation methods, this reduces the impact of complex backgrounds, local occlusion, and lighting changes on the equipment contour extraction process, improving the equipment boundary positioning accuracy and contour continuity. Simultaneously, by combining boundary normal displacement analysis and visual current coupling offset construction processes, the system enhances the ability to express the correlation between visual and electrical changes in the equipment.

[0015] This invention constructs edge-preserving asymmetric similarity relationships and combines optimal transport mapping with Nystrom low-rank approximation to establish equipment structural manifold features. This preserves key structural change information at equipment boundaries, improving the ability to identify minor structural shifts and local abnormal states. Simultaneously, by utilizing equipment temporal topology drift diagrams, state trajectory change characteristics, and multi-source state consistency scores, it analyzes the continuous operating state of equipment, enabling joint judgment of structural changes, electrical changes, and the evolution of operating states. This improves the ability to identify slow structural shifts, hidden faults, and early abnormal states, reduces misjudgments and omissions, and enhances the accuracy of railway outdoor signal equipment operating status monitoring, continuous state tracking capabilities, and fault early warning reliability. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the computer vision-based monitoring system for railway outdoor signaling equipment proposed in this invention. Figure 2 This is a functional flowchart of the improved Chan-Vese model for the computer vision-based railway outdoor signal equipment monitoring system proposed in this invention. Figure 3 This is a functional flowchart of the Nystrom method for a computer vision-based railway outdoor signal equipment monitoring system proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 , Figure 2 and Figure 3 A computer vision-based monitoring system for railway outdoor signaling equipment includes: The data synchronization module collects images and synchronous monitoring data of railway outdoor signaling equipment, performs preprocessing on the synchronous monitoring data and binds it to the images of railway outdoor signaling equipment to generate visual electrical synchronization status data of the equipment. The contour evolution module reads images of railway outdoor signal equipment, establishes an initial contour curve in the target area of ​​the equipment, and constructs an improved Chan-Vese model with HJB optimal evolution constraints to perform contour evolution. It updates the contour evolution direction according to the minimum contour advancement cost and generates an optimized contour set and a set of equipment boundary points. The structural manifold module constructs a similarity matrix of device boundaries based on the set of device boundary points, and uses the Nystrom method to perform low-rank approximation to generate device structural manifold features. It then compares the current device structural manifold features with historical normal structural manifold features to calculate the device structural manifold offset value. The visual coupling module calculates the two-dimensional displacement potential field of the device boundary along the normal direction based on the device optimized contour set, and couples it with the one-dimensional current change sequence to generate visual current coupling offset. The topology drift module reads the device structure manifold offset value and visual current coupling offset according to the sliding time window, constructs the device time-series topology drift diagram, calculates the state trajectory change characteristics, and generates device state drift characteristics. The status determination module performs multi-source status consistency verification on the equipment status drift characteristics based on the equipment's visual electrical synchronization status data, generates a multi-source status consistency score, and determines the equipment fault level result.

[0019] In this embodiment, the synchronous monitoring data specifically includes 25Hz phase-sensitive track circuit voltage and current data, signal lighting unit voltage and current data, traction return current data, box temperature and humidity data, water immersion status data, box opening status data, and vibration status data.

[0020] In this embodiment, the generation of device visual electrical synchronization status data includes: Read the synchronous monitoring data, perform time alignment according to the collection timestamp, write data with a time difference of less than 0.5 seconds into the same synchronous time window, remove data with empty values, incorrect collection labels and data that exceed the sensor range, and perform resampling at a 1-second sampling interval to generate standardized synchronous monitoring data; The acquisition time, equipment number, and monitoring point number of the railway outdoor signal equipment image are read, and the railway outdoor signal equipment image is bound with standardized synchronous monitoring data within the same equipment number, the same monitoring point number, and the same synchronous time window to generate equipment visual electrical synchronization status data.

[0021] In this embodiment, the contour evolution module includes: An improved Chan-Vese model is constructed, comprising an initial contour module, a region statistics module, an optimal evolution module, and a boundary output module, wherein: The improved Chan-Vese model is constructed as follows: This invention adds a dual-level-set self-coupling process to the initial contour module of the traditional Chan-Vese model. By establishing a first-level-set function and a second-level-set function to replace the traditional single-level-set contour evolution method, the initial contour module of this invention is obtained. Based on the region statistics module of the traditional Chan-Vese model, the gray-scale statistics process of the inner and outer regions of the contour is retained, and a full-view cost prediction field is introduced to obtain the region statistics module. Based on the energy evolution module of the traditional Chan-Vese model, HJB optimal evolution constraints are added, and the contour evolution direction is dynamically adjusted according to the contour advancement cost to obtain the optimal evolution module. Based on the boundary output module of the traditional Chan-Vese model, three-dimensional voxel mesh mapping and voxel surface tension value calculation are added to obtain the boundary output module, thus forming an improved Chan-Vese model. The initial contour module reads images of railway outdoor signaling equipment, establishes a first initial closed contour curve and a second initial closed contour curve within the target area of ​​the equipment, and converts them into a first level set function and a second level set function, respectively. A dual-level set self-coupling constraint term is constructed, and the identifiers of the inner and outer regions of the contour are exchanged according to a 20-iteration cycle to generate a dual-level set contour set, where: The initial contour module includes: Initial contour buffer: stores pixel blocks of the target area of ​​the equipment in the read image of the railway outdoor signal equipment; Dual level set registers: store the first level set function and the second level set function respectively; Self-coupling difference calculator: Stores the difference between the first level set function and the second level set function; Contour Label Exchanger: Exchanges the labels of the inner and outer regions of the contour in 20 iterations per cycle; Contour output buffer: Stores the double-level set of contours after the contour label swap is completed; Establish a first initial closed contour curve and a second initial closed contour curve within the target area of ​​the equipment, specifically as follows: Read the coordinates of the bounding box of the target area of ​​the railway outdoor signal equipment in the image, calculate the center coordinates, width and height of the bounding box, and determine the first contour vertex coordinates by indenting the bounding box by 15% in both width and height, and determine the second contour vertex coordinates by expanding the bounding box by 10% in both width and height, and connect the vertex coordinates in a clockwise order to form the first initial closed contour curve and the second initial closed contour curve. Convert them respectively into first-level set functions and second-level set functions, specifically as follows: Read all pixel coordinates within the target area of ​​the device, calculate the shortest distance from each pixel coordinate to the boundary of the first initial closed contour curve and the boundary of the second initial closed contour curve, assign a positive value to the shortest distance inside the corresponding closed contour, assign a negative value to the shortest distance outside the corresponding closed contour, assign a value of 0 to the distance at the boundary of the corresponding closed contour, and arrange all signed distance values ​​in order of pixel coordinates. Define the signed distance value corresponding to the first initial closed contour curve as the first level set function, and define the signed distance value corresponding to the second initial closed contour curve as the second level set function. The dual-level-set self-coupling constraint is a contour association constraint formed by the difference between the corresponding function values ​​of the first and second level-set functions at the same pixel position, used to represent the positional correspondence between two contours. Constructing a bilevel set self-coupling constraint term, specifically: Read the first level set function and the second level set function, calculate the difference between the first level set function and the second level set function item by item according to the corresponding pixel coordinates, take the absolute value of all differences and perform accumulation, divide the accumulation result by the total number of pixels in the target area of ​​the device to obtain the average difference value of the two level sets, continue to read the difference value corresponding to each pixel coordinate, divide the corresponding difference value by the average difference value of the two level sets respectively to obtain the self-coupling constraint value corresponding to each pixel position, arrange all self-coupling constraint values ​​in the order of pixel coordinates to form the self-coupling constraint term of the two level sets; Generate a set of bilevel set contours, specifically as follows: Read the first level set function, the second level set function, and the dual level set self-coupling constraint term. Exchange the inner and outer region identifiers of the contour in 20 iterations per cycle. After each exchange of inner and outer region identifiers, recalculate the level set function value of the corresponding pixel position. Stop the iteration when the difference between the corresponding level set function values ​​of two adjacent iterations is less than 0.001. Read the pixel coordinates corresponding to the level set function value of 0. Connect the corresponding pixel coordinates in sequence according to the condition that the Euclidean distance between adjacent pixels is less than 2 pixels to form a closed contour curve. Summarize all closed contour curves to generate a dual level set contour set. The regional statistics module reads images of railway outdoor signaling equipment, divides the images into a 32x32 pixel grid, calculates the local variance, local entropy, and local gradient variance, and performs a weighted sum to generate a full-view cost prediction field. Regions with a cost value less than the mean value in the entire image are marked as prohibited areas. The regional statistics module includes: Local mesh generator: stores the 32×32 pixel grid number and grid coordinates corresponding to the target area of ​​the device; Gray-scale variance register: stores the gray-scale variance value corresponding to each pixel grid; Gray-level entropy register: stores the gray-level entropy value corresponding to each pixel grid; Gradient variance register: stores the gradient variance value corresponding to each pixel grid; Cost generator: It performs accumulation by multiplying grayscale variance by 0.35, grayscale entropy by 0.35, and gradient variance by 0.30, and outputs the full-view cost prediction field and prohibited entry region markers; Calculate the local variance, local entropy, and local gradient variance separately, and then perform a weighted summation: Read all pixel grayscale values ​​within each 32×32 pixel grid, calculate the average of all grayscale values, and calculate the average of the squared differences between each grayscale value and the average value to obtain the local variance. Count the occurrence frequency of each grayscale value within the current pixel grid, divide the occurrence frequency of each grayscale value by the total number of pixels in the current pixel grid to obtain the probability of each grayscale value. Multiply the probability of each grayscale value by the corresponding logarithmic probability, take the negative of the result, and sum them to obtain the local entropy. Read the grayscale difference between adjacent pixels as the gradient value, calculate the average of all gradient values, and calculate the average of the squared differences between each gradient value and the average value to obtain the local gradient variance. Multiply the local variance by 0.35, the local entropy by 0.35, and the local gradient variance by 0.30, and then sum them to obtain the cost value of the corresponding pixel grid. Generate the full-view cost prediction field as follows: Read the cost value corresponding to all pixel grids, write the cost value of each pixel grid back to the corresponding area of ​​the railway outdoor signal equipment image according to the original pixel grid spatial position, and after writing all pixel grids, form a continuous cost distribution covering the target area of ​​the equipment, and generate a full-view cost prediction field. Regions with a bit value less than the global value are marked as prohibited areas. Specifically: Read the cost value corresponding to all pixel grids in the full-view cost prediction field, sort them in ascending order according to the value, read the value corresponding to the middle position of all cost values ​​as the median cost value in the whole image, and mark the pixel grids whose cost value is less than the median cost value in the whole image as prohibited areas. The optimal evolution module reads the dual-level contour set, assigns positive-value pixels to the inner contour region, and negative-value pixels to the outer contour region, calculates the mean grayscale value of the inner and outer regions respectively, and dynamically updates the weight values ​​of the inner and outer regions based on the total energy change rate, where: The optimal evolution module includes: Gray-scale mean updater: stores the gray-scale mean of the region inside the contour, the gray-scale mean of the region outside the contour, and the weight values ​​of the inside and outside regions; Contour Advancement Cost Accumulator: Stores grayscale region difference values, curvature change values, occlusion intensity values, and local noise intensity values, and performs accumulation by multiplying the grayscale region difference value by 0.40, the curvature change value by 0.25, the occlusion intensity value by 0.20, and the local noise intensity value by 0.15; Directional blocking device: stores the outline advance direction number corresponding to the prohibited entry area; Evolutionary step size controller: stores the difference in the cost of contour advancement between two adjacent iterations; The pixels with positive values ​​are assigned to the area inside the contour, specifically: Read the level set function values ​​of the corresponding pixel positions in the double level set contour set, and draw the pixel coordinates corresponding to the level set function values ​​greater than 0 into the inner region of the contour. Arrange all the pixels in the inner region in the order of pixel coordinates. Negative-valued pixels are assigned to the region outside the contour, specifically: Read the horizontal set function values ​​of the corresponding pixel positions in the double horizontal set contour set, and draw the pixel coordinates corresponding to the horizontal set function values ​​less than 0 into the outer region of the contour. Arrange all the pixels in the outer region in the order of pixel coordinates. Calculate the internal grayscale mean and the external grayscale mean separately, as follows: Read all pixel grayscale values ​​in the inner region of the contour and accumulate them. Divide the accumulated result by the total number of pixels in the inner region of the contour to obtain the inner grayscale average. Read all pixel grayscale values ​​in the outer region of the contour and accumulate them. Divide the accumulated result by the total number of pixels in the outer region of the contour to obtain the outer grayscale average. The weight values ​​of the inner and outer regions are dynamically updated based on the total energy change rate, specifically as follows: Read the mean grayscale value inside and outside the current iteration, calculate the absolute value of the difference between the mean grayscale value inside and outside the current iteration. When the number of iterations is equal to 1, assign the weight values ​​of the internal region and the external region to 0.50. When the number of iterations is greater than 1, read the absolute value of the difference corresponding to the previous iteration, subtract the absolute value of the difference of the previous iteration from the current absolute value of the difference, and divide by the absolute value of the difference of the previous iteration to obtain the change ratio. When the change ratio is greater than 0.10, increase the weight value of the internal region by 0.05 and decrease the weight value of the external region by 0.05. When the change ratio is less than or equal to 0.10, decrease the weight value of the internal region by 0.05 and increase the weight value of the external region by 0.05. When the weight value of the internal region is greater than 0.80, take 0.80. When the weight value of the external region is less than 0.20, take 0.20. This completes the update of the weight values ​​of the internal and external regions. Calculate the grayscale region difference value, curvature change value, occlusion intensity value, and local noise intensity value. Adjust the proportion of the grayscale region difference value in the contour advancement cost value based on the updated inner and outer region weight values. Perform a weighted summation of the grayscale region difference value, curvature change value, occlusion intensity value, and local noise intensity value to obtain the contour advancement cost value. Combine this with the full-view cost prediction field to apply HJB optimal evolution constraints, eliminating advancement directions pointing to prohibited entry regions, and determining the optimal contour advancement direction for the current iteration. Where: The calculation of grayscale region difference value, curvature change value, occlusion intensity value, and local noise intensity value is as follows: Read the pixels inside the contour, the pixels outside the contour, and the pixels at the contour boundary in the current iteration. Calculate the absolute value of the difference between the mean gray value inside the contour and the mean gray value outside the contour to obtain the gray value difference value. Read the coordinates of three adjacent boundary points on the contour boundary. Calculate the angle between the directions of the two boundary segments before and after. Average all the angles to obtain the curvature change value. Read the number of consecutive missing boundary points at the contour boundary. Divide the number of consecutive missing boundary points by the total number of contour boundary points to obtain the occlusion intensity value. Read the pixel gray values ​​within a range of five pixels on both sides of the contour boundary. Calculate the average value of the gray value difference between adjacent pixels to obtain the local noise intensity value. The proportion of grayscale region difference in the contour advancement cost is adjusted based on the updated inner and outer region weight values, specifically as follows: Read the internal region weight value and the external region weight value, add the internal region weight value and the external region weight value to get the total region weight value, divide the internal region weight value by the total region weight value to get the gray-scale region difference adjustment coefficient, and multiply the gray-scale region difference value by the gray-scale region difference adjustment coefficient to get the adjusted gray-scale region difference value. The weighted summation of grayscale region differences, curvature changes, occlusion intensity, and local noise intensity is as follows: Multiply the adjusted grayscale region difference value by 0.40, the curvature change value by 0.25, the occlusion intensity value by 0.20, and the local noise intensity value by 0.15. Add the four products together to obtain the contour advancement cost. HJB optimal evolution constraint refers to the constraint relationship used to limit the evolution of the contour in the direction of minimum contour advancement cost in the candidate contour advancement direction; The HJB optimal evolution constraint is applied in conjunction with the full-view cost prediction field, specifically as follows: Read the eight candidate propulsion directions (up, down, left, right, upper left, upper right, lower left, and lower right) of the current contour boundary point. Read the cost value of the corresponding position of each candidate propulsion direction in the full-view cost prediction field. Add the corresponding cost value to the current contour propulsion cost value to obtain the instantaneous propulsion cost value of each candidate propulsion direction. Continue to read the cumulative propulsion cost value of the next position of each candidate propulsion direction in the previous iteration. Add the instantaneous propulsion cost value to the cumulative propulsion cost value multiplied by 0.90 to obtain the state value function value of each candidate propulsion direction. Assign the maximum value of 999 to the candidate propulsion direction pointing to the prohibited entry area to complete the HJB optimal evolution constraint. The optimal contour advancement direction for the current iteration is determined as follows: Read the state value function value corresponding to all candidate propulsion directions, compare the size of each state value function value, and select the candidate propulsion direction with the smallest state value function value as the optimal contour propulsion direction for the current iteration. The boundary output module maps the current contour to a 3D voxel mesh space, calculates the voxel surface tension constraint value, and performs contour updates. Iteration stops when the difference in contour advancement cost between two adjacent iterations is less than 0.001. It extracts the zero-level set boundary to generate an optimized device contour set and generates a device boundary point set according to a 5-pixel sampling interval, where: The boundary output module includes: Voxel map converter: stores the correspondence between two-dimensional contour coordinates and three-dimensional voxel mesh coordinates; Voxel surface tension calculator: Stores voxel surface coordinates and corresponding voxel surface tension values; Contour updater: Stores the current contour advancement direction and voxel surface tension value; Zero level set extractor: stores the pixel coordinates where the level set function is equal to 0; Boundary sampler: Stores a set of device boundary points sampled at fixed intervals of 5 pixels; The boundary output module maps the current contour to a 3D voxel mesh space, specifically: Read the horizontal and vertical pixel coordinates of the current contour boundary point, use the horizontal pixel coordinates as the horizontal coordinates of the 3D voxel mesh, use the vertical pixel coordinates as the vertical coordinates of the 3D voxel mesh, use the current iteration number as the depth coordinates of the 3D voxel mesh, and determine the voxel position of the current contour boundary point in the 3D voxel mesh space according to the 3D coordinates. Calculate the voxel surface tension constraint values ​​and perform contour updates, specifically as follows: Read the voxel position corresponding to the current contour boundary point and the positions of its adjacent voxels. Calculate the absolute value of the coordinate difference between the current voxel position and the adjacent voxel position and take the average to obtain the voxel surface tension value. Multiply the lateral displacement and longitudinal displacement corresponding to the current iterative optimal contour advancement direction by 1 pixel step size to obtain the contour update displacement. When the voxel surface tension value is greater than 0.50, multiply the contour update displacement by 0.80 and update the current contour coordinates. When the voxel surface tension value is less than or equal to 0.50, update the current contour coordinates directly according to the contour update displacement. Extracting the zero-level set boundary generates the optimized profile set for the device, specifically: Read the value of the horizontal set function after the contour is updated, extract the pixel coordinates with a value of 0, connect the coordinates of adjacent pixels with an Euclidean distance of less than 2 pixels to form a closed contour curve, and collect all closed contour curves in the same target area of ​​the device according to the device number and the acquisition time to generate the device optimized contour set. The device boundary point set is generated according to a 5-pixel sampling interval, specifically as follows: Read the closed contour curve in the device optimized contour set, start from the starting coordinate point of the closed contour curve and accumulate the Euclidean distance between adjacent coordinate points along the curve direction. When the accumulated distance reaches 5 pixels, retain the current position coordinate point as the device boundary point until the sampling of the closed contour curve is completed. Arrange all the retained device boundary points according to the device number and the acquisition time to generate a device boundary point set. The improved Chan-Vese model was trained using a joint loss value comprised of contour advancement cost error, region segmentation error, bilevel set consistency error, full-view cost prediction error, and boundary reconstruction error as the optimization objective. The parameters in the initial contour module, region statistics module, optimal evolution module, and boundary output module were continuously optimized. Training was stopped when the change in the joint loss value was less than 0.001 for 10 consecutive iterations. The improved Chan-Vese model was trained as follows: Read images of railway outdoor signal equipment, real contour annotations of the target area of ​​the equipment, and real boundary point annotations. Input the initial contour module, regional statistics module, optimal evolution module, and boundary output module to obtain the optimized contour set of the equipment, the set of equipment boundary points, the dual-level set function, the full-view cost prediction field, and the contour advancement cost. The sum of the absolute values ​​of the corresponding pixel differences between the optimized contour set and the real contour is calculated and divided by the total number of pixels in the real contour to obtain the contour advancement cost error. The number of overlapping pixels between the internal region of the optimized contour set and the real labeled region is calculated and divided by the total number of pixels in the entire region to calculate the region overlap ratio. Then, 1 is subtracted from the region overlap ratio to obtain the region segmentation error. The first level set function and the second level set function are read, and the absolute value of their difference is calculated pixel by pixel and accumulated. The accumulated result is divided by the total number of pixels in the target region of the device to obtain the dual level set consistency error. The full-view cost prediction field and the real cost distribution are read, and the square of the cost difference at the corresponding position is calculated pixel by pixel and averaged to obtain the full-view cost prediction error. The device boundary point set and the real boundary points are read, and the Euclidean distance between the corresponding boundary points is calculated and accumulated. The accumulated result is divided by the number of boundary points to obtain the boundary reconstruction error. The contour advancement cost error, region segmentation error, dual-level set consistency error, full-view cost prediction error, and boundary reconstruction error are multiplied by 0.25, and then accumulated to obtain the joint loss value. The parameters in the initial contour module, region statistics module, optimal evolution module, and boundary output module are read, and the joint loss value after increasing the parameter by 0.001 and decreasing the parameter by 0.001 are calculated respectively. When the joint loss value after increasing the parameter is smaller, the parameter is increased by 0.0001, and when the joint loss value after decreasing the parameter is smaller, the parameter is decreased by 0.0001. Training stops when the change in the joint loss value is less than 0.001 for 10 consecutive iterations.

[0022] In this embodiment, the structural manifold module includes: Based on the set of device boundary points, the pixel coordinates, curvature value, tangent direction angle, boundary gray value, and component connection distance of each boundary point are extracted according to the device number. Zero-to-one normalization is then performed on each feature to generate a boundary point feature sequence, where: Extract the pixel coordinates, curvature value, tangent direction angle, boundary grayscale value, and component connection distance of each boundary point according to the device number. Specifically: Read the device number and boundary point coordinates from the device boundary point set, classify the boundary points according to the device number, read the horizontal and vertical pixel coordinates of the current boundary point as pixel coordinates, read the coordinates of one boundary point before and after the current boundary point, calculate the angle between the connecting lines of the two boundary segments and divide it by the corresponding boundary arc length to obtain the curvature value, read the direction angle of the connecting line between the current boundary point and the previous boundary point to obtain the tangent direction angle, read the gray value of the corresponding pixel position of the current boundary point to obtain the boundary gray value, and read the Euclidean distance between the current boundary point and the center coordinates of the device to which it belongs to obtain the component connection distance; Generate the boundary point feature sequence, specifically as follows: Read pixel coordinates, curvature value, tangent direction angle, boundary gray value and component connection distance, calculate the corresponding maximum and minimum value of the feature respectively, subtract the corresponding minimum value from the current feature value and divide by the difference between the maximum and minimum values ​​to complete the normalization process in the zero to one interval, arrange all normalized feature values ​​in the order of pixel coordinates, curvature value, tangent direction angle, boundary gray value and component connection distance, and generate the boundary point feature sequence; Based on the feature sequence of boundary points, the spatial distance difference, curvature difference, orientation angle difference, grayscale difference, and connectivity distance difference between any two boundary points are calculated. Boundary affinity values ​​are also calculated. High-curvature boundary points and low-curvature boundary points are determined according to the magnitude of the curvature values. The boundary affinity values ​​pointing from high-curvature boundary points to low-curvature boundary points are retained, while the reverse boundary affinity values ​​are set to 0. An edge-preserving asymmetric similarity matrix is ​​generated, where: Calculate the spatial distance difference, curvature difference, orientation angle difference, grayscale difference, and connectivity difference between any two boundary points, specifically as follows: Read the pixel coordinates corresponding to any two boundary points, calculate the Euclidean distance between the two pixel coordinates to obtain the spatial distance difference, read the curvature values ​​corresponding to the two boundary points and calculate the absolute value of the difference to obtain the curvature difference, read the tangent direction angles corresponding to the two boundary points and calculate the absolute value of the difference to obtain the direction angle difference, read the boundary grayscale values ​​corresponding to the two boundary points and calculate the absolute value of the difference to obtain the grayscale difference, read the component connection distances corresponding to the two boundary points and calculate the absolute value of the difference to obtain the connection distance difference; The boundary affinity value is calculated as follows: Read the spatial distance difference, curvature difference, orientation angle difference, grayscale difference, and connection distance difference. Multiply the spatial distance difference by 0.30, the curvature difference by 0.25, the orientation angle difference by 0.20, the grayscale difference by 0.15, and the connection distance difference by 0.10, and then perform accumulation. Read the accumulation result and calculate 1 minus the accumulation result to obtain the boundary affinity value. The high curvature boundary points and low curvature boundary points are determined according to the magnitude of the curvature values, specifically as follows: Read all the curvature values ​​corresponding to the boundary points and sort them in descending order according to their numerical values. Calculate the average value of all curvature values. When the curvature value of a boundary point is greater than 1.20 times the average value of all curvature values, the corresponding boundary point is determined as a high curvature boundary point. When the curvature value of a boundary point is less than 0.80 times the average value of all curvature values, the corresponding boundary point is determined as a low curvature boundary. Set the reverse boundary affinity value to 0, specifically: Read the corresponding boundary affinity values ​​between high curvature boundary points and low curvature boundary points, retain the boundary affinity values ​​pointing from high curvature boundary points to low curvature boundary points, and assign 0 to the boundary affinity values ​​pointing from low curvature boundary points to high curvature boundary points. Generate an edge-preserving asymmetric similarity matrix, specifically as follows: Read the corresponding boundary affinity values ​​between all boundary points, arrange the boundary affinity values ​​in the order of the boundary point numbers, use the boundary points as row numbers and column numbers, write the corresponding boundary affinity values ​​at the intersection of the corresponding row numbers and column numbers, and generate an edge-preserving asymmetric similarity matrix after writing all boundary affinity values. Based on the edge-preserving asymmetric similarity matrix, initial representative boundary points are selected according to 10% of the total number of boundary points. The Nystrom method is used to perform an initial low-rank approximation, generating an initial low-dimensional boundary embedding feature matrix. The embedding error hotspot region is determined based on the reconstruction difference between the initial low-dimensional boundary embedding feature matrix and the edge-preserving asymmetric similarity matrix. Initial representative boundary points are selected based on 10% of the total number of boundary points, specifically as follows: Read the total number of boundary points, multiply the total number of boundary points by 0.10 to get the initial representative boundary point number, round up if the calculation result has a decimal, read the curvature value corresponding to all boundary points, sort them in descending order of curvature value, and select the corresponding number of boundary points as the initial representative boundary points in turn; The initial low-rank approximation is performed using the Nystrom method, specifically as follows: Read the corresponding similarity values ​​between the initial representative boundary points and the corresponding similarity values ​​between the initial representative boundary points and all boundary points. Read all the similarity values ​​corresponding to each initial representative boundary point in turn. Square each similarity value, accumulate them, and take the square root to obtain the length value. Divide each similarity value by the length value to obtain the unit direction value. Read the unit direction value and the corresponding similarity value. Multiply the unit direction value and the corresponding similarity value item by item and accumulate them to obtain the corresponding weight value. Continue to read the corresponding similarity values ​​between all boundary points and the initial representative boundary points. Multiply each similarity value by the corresponding weight value and accumulate them to obtain the low-dimensional mapping value corresponding to each boundary point. Arrange all the low-dimensional mapping values ​​in the order of the boundary point numbers to generate the initial low-dimensional boundary embedding feature matrix. The determination of the embedding error hotspot region is as follows: Read the initial low-dimensional boundary embedding feature matrix, read the low-dimensional mapping values ​​corresponding to any two boundary points, calculate the absolute value of the difference between the corresponding low-dimensional mapping values, sum all the absolute values ​​of the difference to obtain the reconstructed similarity value, read the corresponding similar values ​​in the edge-preserving asymmetric similarity matrix, subtract the reconstructed similarity value from the corresponding similarity value and take the absolute value to obtain the reconstructed difference, summarize all the reconstructed differences according to the boundary point number, add the reconstructed differences of each boundary point and divide by the number of boundary points to obtain the average reconstructed difference, and determine the region where the boundary point with a reconstructed difference greater than 1.50 times the average reconstructed difference is located as the embedding error hotspot region; In the embedded error hotspot region, representative boundary points corresponding to 15% of the boundary points are added, and these are merged to generate a recursive representative boundary point set. The similarity matrix corresponding to the recursive representative boundary point set is converted into an affinity probability distribution. The optimal transport mapping is performed using Sinkhorn-Knopp iteration to generate a transport-weighted representative similarity matrix. Finally, the Nystrom method is used to perform low-rank approximation to generate the device structure manifold features, where: In the embedded error hotspot region, add representative boundary points corresponding to 15% of the number of boundary points, specifically: Read the total number of boundary points corresponding to the embedding error hotspot region, multiply the total number of boundary points by 0.15 to obtain the number of additional representative boundary points, read the reconstruction difference corresponding to the boundary points in the hotspot region, sort them in descending order of reconstruction difference, select the corresponding number of boundary points as additional representative boundary points, merge the initial representative boundary points and the additional representative boundary points to generate a recursive representative boundary point set; The recursive representation of the similarity matrix corresponding to the set of boundary points is converted into an affinity probability distribution, specifically: Read the corresponding similarity values ​​between the recursive representative boundary point sets, calculate the sum of all similarity values, divide each similarity value by the sum of all similarity values ​​to obtain the corresponding probability value, arrange all probability values ​​according to the original position of the similarity value, and generate the affinity probability distribution. The optimal transport mapping is performed using the Sinkhorn-Knopp iterative method, specifically as follows: Read the affinity probability distribution, calculate the cumulative result of the probability values ​​of each row and the cumulative result of the probability values ​​of each column. When the difference between the cumulative result of a row and 1 is greater than 0.001, divide all the probability values ​​of the corresponding row by the cumulative result of the corresponding row. When the difference between the cumulative result of a column and 1 is greater than 0.001, divide all the probability values ​​of the corresponding column by the cumulative result of the corresponding column. Alternately perform row normalization and column normalization. Stop the iteration when the difference between the cumulative result of all rows and the cumulative result of all columns and 1 is less than 0.001. Read the probability values ​​and the corresponding similarity values ​​of the recursive representative boundary point set after the iteration is completed. Multiply each probability value by the corresponding similarity value to obtain the transport weighted similarity value. Write each transport weighted similarity value back to the original similarity value position between the corresponding boundary points according to the recursive representative boundary point numbering order. After writing all positions, generate the transport weighted representative similarity matrix. The Nystrom method is used to perform the low-rank approximation, specifically as follows: Read the transport weighted representative similarity matrix and read the similarity value between the recursive representative boundary point and all boundary points. Read the weighted similarity value between each recursive representative boundary point and the other recursive representative boundary points in the transport weighted representative similarity matrix in turn. Square all weighted similarity values, sum them up and take the square root to obtain the direction length value corresponding to the recursive representative boundary point. Divide each weighted similarity value corresponding to the recursive representative boundary point by the direction length value to obtain the corresponding unit direction value. Multiply the unit direction value by the corresponding weighted similarity value item by item and sum them to obtain the direction weight value corresponding to the recursive representative boundary point. Read the similarity value between each ordinary boundary point and the recursive representative boundary point, multiply the similarity value by the corresponding direction weight value and sum them to obtain the low-dimensional mapping value of the ordinary boundary point in the corresponding recursive representative direction. Arrange all low-dimensional mapping values ​​in the order of boundary point number to obtain the low-dimensional boundary embedding feature. Read the first 6 dimensions in the low-dimensional boundary embedding feature and calculate the average value of the low-dimensional mapping value of all boundary points in each dimension. Arrange the 6 average values ​​in the order of dimensions to generate the device structure manifold feature. Read the historical normal structural manifold features corresponding to the same device number, compare the current device structural manifold features with the historical normal structural manifold features item by item according to the corresponding dimensions, and calculate the device structural manifold offset value, where: Historical normal structural manifold features refer to a 6-dimensional feature composed of the average of all feature dimensions in the set of structural manifold features corresponding to the same equipment number under normal operating conditions. The calculation of the equipment structure manifold offset value is as follows: Read the current equipment structure manifold features and the historical normal structure manifold features, subtract them item by item according to the corresponding dimensions and take the absolute value, square all the absolute values ​​of the dimensions and then sum them up, divide the sum by 6 and take the square root to obtain the equipment structure manifold offset value.

[0023] In this embodiment, the visual coupling module includes: Read the optimized contour set of the devices, extract the current pixel coordinates of the boundary points of each device according to the device number and acquisition time, and read the coordinates of the historical normal boundary points corresponding to the same device number. Match the current boundary point coordinates with the historical normal boundary point coordinates to generate a boundary point matching sequence, where: Historical normal boundary point coordinates refer to the pixel coordinates of boundary points saved in the boundary point order in the optimized contour set of the corresponding device under normal operating conditions for the same device number; Generate a boundary point matching sequence, specifically as follows: Read the coordinates of the current device boundary point and the coordinates of the corresponding historical normal boundary points with the same device number. Read the horizontal and vertical coordinate differences between the current boundary point coordinates and all historical normal boundary point coordinates in sequence. Add the square of the horizontal coordinate difference and the square of the vertical coordinate difference and take the square root to obtain the coordinate distance value. Select the historical normal boundary point with the smallest coordinate distance value as the current boundary point matching point. Arrange the coordinates of the current boundary point and the corresponding historical normal boundary point coordinates in the order of boundary points to generate a boundary point matching sequence. Based on the boundary point matching sequence, the displacement distance and normal direction of each current boundary point relative to historical normal boundary points are calculated. The displacement distance is projected along the normal direction to obtain the boundary normal displacement value. All boundary normal displacement values ​​are then arranged according to the pixel position of the boundary point in the image to generate a two-dimensional displacement potential field, where: The displacement distance and normal direction of each current boundary point relative to historical normal boundary points are calculated based on the boundary point matching sequence, specifically as follows: Read the coordinates of the current boundary point and the historical normal boundary point in the boundary point matching sequence. Add the square of the difference between the horizontal coordinates and the square of the difference between the vertical coordinates and take the square root to obtain the displacement distance. Read the coordinates of the previous boundary point and the next boundary point. Calculate the direction of the line connecting the previous and next boundary points to determine the boundary tangent direction. Rotate the boundary tangent direction 90° clockwise to obtain the normal direction. The boundary normal displacement value is obtained as follows: Read the displacement distance and normal direction, read the lateral displacement and longitudinal displacement between the current boundary point coordinates and the historical normal boundary point coordinates, multiply the lateral displacement by the lateral component of the normal direction, multiply the longitudinal displacement by the longitudinal component of the normal direction, add the two product results to obtain the boundary normal displacement value, arrange all boundary normal displacement values ​​in order according to the pixel position of the boundary point in the image, and generate a two-dimensional displacement potential field. The voltage and current data of the 25Hz phase-sensitive track circuit, the voltage and current data of the signal lighting unit, and the traction return current data corresponding to the same equipment number are read from the equipment visual electrical synchronization status data. The current component in each data is extracted, and the absolute difference between the current components at two adjacent acquisition times is calculated to generate a one-dimensional current change sequence, where: Extract the current component from each data point, specifically: Read the voltage and current data of the 25Hz phase-sensitive track circuit, the voltage and current data of the signal lighting unit, and the traction return current data. Extract the corresponding current values ​​of each data and arrange all current values ​​in order of equipment number and acquisition time. Generate a one-dimensional current change sequence, specifically as follows: Read the current components corresponding to two adjacent acquisition times under the same device number, calculate the absolute value of the difference between the current component at the current acquisition time and the current component at the previous acquisition time, arrange all the absolute values ​​of the difference in order of device number and acquisition time, and generate a one-dimensional current change sequence. The boundary normal displacement values ​​and one-dimensional current change sequences in the two-dimensional displacement potential field are normalized to obtain visual displacement values ​​and current change values. The visual displacement values ​​are multiplied by 0.60 and the current change values ​​by 0.40, and then summed to generate visual current coupling values. All visual current coupling values ​​are then aggregated according to the device number to generate the visual current coupling offset, where: The visual displacement value and current change value are obtained as follows: Read all boundary normal displacement values ​​in the two-dimensional displacement potential field. Read the maximum and minimum values ​​of all boundary normal displacement values ​​respectively. Subtract the minimum value from the current boundary normal displacement value and divide by the difference between the maximum and minimum values ​​to obtain the visual displacement value. Read all current change values ​​in the one-dimensional current change sequence. Read the maximum and minimum values ​​of all current change values ​​respectively. Subtract the minimum value from the current current change value and divide by the difference between the maximum and minimum values ​​to obtain the current change value.

[0024] In this embodiment, the topology drift module includes: According to a 60-second window length and a 10-second sliding step, the device structure manifold offset value and visual current coupling offset of the same device number within a continuous time window are read to form a device state vector, and a device state vector sequence is generated according to the acquisition time order. Based on the device state vector sequence, the device state vector is used as a state node. The structural manifold offset difference and the visual current coupling offset difference between two adjacent state nodes are calculated. The structural manifold offset difference is multiplied by 0.55 and the visual current coupling offset difference is multiplied by 0.45 and then added together to obtain the state transition edge weights. Based on the state nodes and the state transition edge weights, a device time-series topology drift graph is constructed, where: The construction of the device timing topology drift graph is as follows: Read the state vectors of each device in the device state vector sequence, assign state node numbers in sequence according to the acquisition time, read the device structural manifold offset value and visual current coupling offset value corresponding to two adjacent state nodes, calculate the absolute value of the difference between the device structural manifold offset values ​​of adjacent state nodes to obtain the structural manifold offset difference value, calculate the absolute value of the difference between the visual current coupling offset values ​​of adjacent state nodes to obtain the visual current coupling offset difference value, multiply the structural manifold offset difference value by 0.55 and the visual current coupling offset difference value by 0.45 and add them together to obtain the state transition edge weight, connect adjacent state nodes in sequence according to the state node number, and write the corresponding state transition edge weight into the connection edge. After completing the connection of all state nodes, generate the device timing topology drift graph. Based on the device time-series topology drift graph, the average device state vector of all state nodes is calculated as the state trajectory center. The distance between the current state node and the historical normal state trajectory center is calculated as the state trajectory center offset. The increase in the weight of adjacent state transition edges is calculated as the trajectory divergence value. The change in state node distribution density within a unit time window is calculated as the state density change value. Where: The historical normal state trajectory center refers to the center coordinates composed of the average value of the equipment structure manifold offset value and the visual current coupling offset value corresponding to the state vectors of all equipment with the same equipment number under normal operating conditions. The growth rate of the transition edge weights between adjacent states is calculated as the trajectory divergence value, specifically: Read the state transition edge weight corresponding to the current state node and the state transition edge weight corresponding to the previous state node. Subtract the previous state transition edge weight from the current state transition edge weight and take the absolute value to obtain the edge weight change value. Read the state transition edge weight corresponding to the previous state node and divide the edge weight change value by the previous state transition edge weight to obtain the state transition edge weight growth rate. Complete the calculation of all state nodes in the order of collection time and use the growth rate of all state transition edge weights as the trajectory divergence value. The change in the distribution density of state nodes within a unit time window is calculated as the state density change value, specifically as follows: Read the total number of all state nodes within a 60-second time window. Read the distance between the current state node and other state nodes. When the distance between nodes is less than 0.80 times the average distance between state nodes, it is considered a neighboring node. Count the total number of neighboring nodes within the current time window and divide it by the total number of state nodes within the current time window to obtain the state node distribution density of the current time window. Read the state node distribution density of the adjacent previous time window. Subtract the state node distribution density of the previous time window from the state node distribution density of the current time window and take the absolute value to obtain the state density change value. The state trajectory center offset value, trajectory divergence value, and state density change value are arranged in the order of acquisition time to generate state trajectory change characteristics. These characteristics are then associated with the equipment number, equipment structure manifold offset value, and visual current coupling offset to generate equipment state drift characteristics.

[0025] In this embodiment, the state determination module includes: Read the equipment's visual electrical synchronization status data and equipment status drift characteristics, extract the status trajectory center offset value, trajectory divergence value and status density change value, and read the 25Hz phase-sensitive track circuit voltage and current, signal light unit voltage and current, traction return current, box temperature and humidity, immersion, unpacking and vibration data to generate a multi-source verification dataset; Normalization is performed on the state trajectory center offset value, trajectory divergence value, and state density change value. The state trajectory center offset value is multiplied by 0.40, the trajectory divergence value is multiplied by 0.35, and the state density change value is multiplied by 0.25, and then added together to generate the structure drift score. Read the multi-source verification dataset, calculate the variation amplitude of 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, box temperature and humidity, and vibration data at two adjacent acquisition times, and combine the immersion data and box opening data to generate a multi-source state change score, including: Generate multi-source state change scores, specifically: Read the 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, box temperature and humidity, and vibration data corresponding to two adjacent acquisition times. Calculate the absolute value of the difference between the values ​​of adjacent acquisition times and perform normalization processing in the range of zero to one. When the immersion data indicates the immersion state, assign a value of 1; otherwise, assign a value of 0. When the box opening data indicates the box opening state, assign a value of 1; otherwise, assign a value of 0. Multiply the normalized value of the 25Hz phase-sensitive track circuit by 0.25, the normalized value of the signal lighting unit by 0.20, the normalized value of the traction return current by 0.20, the normalized value of the box temperature and humidity by 0.10, the normalized value of the vibration by 0.10, the immersion state value by 0.075, and the box opening state value by 0.075, and then add them together to generate a multi-source state change score. The structure drift score is multiplied by 0.55 and the multi-source state change score is multiplied by 0.45, and then the results are added together to generate the multi-source state consistency score. When the multi-source status consistency score is less than 0.40, it is determined to be a normal status level; when the multi-source status consistency score is greater than or equal to 0.40 and less than 0.65, it is determined to be a potential fault level; when the multi-source status consistency score is greater than or equal to 0.65 and less than 0.85, it is determined to be an enhanced early warning level; and when the multi-source status consistency score is greater than or equal to 0.85, it is determined to be a fault alarm level.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a simulated monitoring area for railway outdoor signal equipment. The monitoring objects included one low-profile signal, one 25Hz phase-sensitive track circuit box, one cable entry point, one set of lead wire fixing points, and one outdoor signal box. The image acquisition device covered the signal casing, the box door seam, the cable entry point, and the lead wire fixing area. Inside the box, it collected data on the voltage and current of the 25Hz phase-sensitive track circuit, the voltage and current of the signal lighting unit, traction return current, box temperature and humidity, water immersion, opening, and vibration. The system acquired one set of images and synchronous monitoring data every 5 seconds, and bound them according to the equipment number and acquisition time to form equipment visual electrical synchronization status data.

[0027] A total of 6200 training samples were used, including 3600 normal samples, 1100 samples with slight structural deviations, 650 samples with partial occlusion, 420 samples with water immersion interference, and 430 samples with slow current fluctuations. In the normal samples, the offset of the box door seam boundary ranged from 0.4 pixels to 1.8 pixels, the 25Hz phase-sensitive track circuit voltage ranged from 182.6 volts to 187.9 volts, the current ranged from 1.18 amps to 1.34 amps, and the vibration amplitude ranged from 0.06 mm to 0.18 mm. In the samples with slight structural deviations, the offset of the box door seam boundary ranged from 3.2 pixels to 6.8 pixels, and the offset of the cable inlet edge ranged from 2.7 pixels to 5.9 pixels, but the current was still 1.21 amps to 1.39 amps, which did not reach the traditional electrical alarm threshold. After training, the average contour localization error of the improved Chan-Vese model was reduced from 4.71 pixels in the traditional Chan-Vese model to 1.62 pixels, and the boundary reconstruction error was reduced from 3.88 pixels to 1.31 pixels. Training was stopped when the change in the joint loss value was less than 0.001 after 10 consecutive iterations.

[0028] During a continuous monitoring period, the voltage of the 25Hz phase-sensitive track circuit of a certain box changed from 185.4 volts to 184.7 volts, the current changed from 1.26 amps to 1.31 amps, the voltage of the signal lighting unit changed from 219.6 volts to 220.8 volts, the traction return current changed from 52.4 amps to 56.9 amps, the box temperature changed from 26.8 degrees Celsius to 27.6 degrees Celsius, the humidity changed from 61.5% to 64.2%, the water immersion state was 0, the box opening state was 0, and the vibration amplitude changed from 0.12 mm to 0.19 mm. Traditional electrical threshold methods determine that the box is in a normal state based on the above data. After reading the equipment images during the same monitoring period, the system of this invention found a 4.9 pixel offset along the boundary of the box door seam relative to the historical normal boundary point, a 3.7 pixel offset on the left boundary of the cable inlet, and a 2.8 pixel offset near the lead wire fixing point. Traditional edge detection generates 9 pseudo-boundary points in the ballast reflective area, and the contour break length in the box door seam area is 31 pixels. This invention optimizes the contour set by using a dual-level set self-coupling constraint term, a full-view cost prediction field, and an HJB optimal evolution constraint extraction device, generating only 2 pseudo-boundary points in the same area, and reducing the contour break length to 7 pixels.

[0029] The system calculates the structural manifold offset based on the set of device boundary points. The current total number of boundary points is 480. The system selects 48 initial representative boundary points at a 10% ratio, and adds 72 representative boundary points in the embedding error hotspot region, forming 120 recursive representative boundary points. The curvature value of the sharp corner boundary point is 0.74, and the average curvature of all boundary points is 0.41; the system identifies this point as a high-curvature boundary point. The curvature value of the straight edge of the box is 0.22; the system identifies this point as a low-curvature boundary point. The boundary affinity value pointing from the high-curvature boundary point to the low-curvature boundary point is retained, and the reverse boundary affinity value is set to 0. After 12 Sinkhorn-Knopp iterations, the maximum difference between the sum of the row direction probability values ​​and 1 is 0.0008, and the maximum difference between the sum of the column direction probability values ​​and 1 is 0.0009, satisfying the stopping condition of less than 0.001. The system obtained the following top 6 dimensions of the current equipment structure manifold features: 0.382, 0.417, 0.246, 0.193, 0.155, and 0.128. The top 6 dimensions of the historical normal structure manifold features were 0.301, 0.346, 0.218, 0.151, 0.122, and 0.104. The calculated equipment structure manifold offset value is 0.053. Traditional image methods only yield an edge change score of 0.031, which does not reach the alarm threshold of 0.080.

[0030] In visual-current coupling analysis, the system calculated the average displacement distance along the top edge of the box door seam to be 4.6 pixels, the boundary normal displacement value projected along the normal direction to be 3.9 pixels, the average displacement distance at the cable inlet edge to be 3.3 pixels, and the boundary normal displacement value to be 2.8 pixels. The system extracted current components from the 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, and traction return current data, calculating the absolute difference between current components at two adjacent acquisition times. The current change value for the 25Hz phase-sensitive track circuit was 0.05 A, for the signal lighting unit it was 0.012 A, and for the traction return current it was 4.5 A. After normalizing the boundary normal displacement value and the one-dimensional current change sequence, the system calculated the visual-current coupling value by multiplying the visual displacement value by 0.60 and the current change value by 0.40. The visual-current coupling offset for this box was 0.612, while the average value for normal samples was 0.284, indicating that the equipment boundary displacement had synchronized with the slow current change.

[0031] Within eight consecutive sliding time windows, the device structure manifold offset values ​​were 0.021, 0.026, 0.031, 0.038, 0.046, 0.053, 0.061, and 0.068, respectively, and the visual current coupling offset values ​​were 0.301, 0.336, 0.372, 0.428, 0.509, 0.612, 0.664, and 0.701, respectively. The system constructed a device temporal topology drift graph with a 60-second window length and a 10-second sliding step. The calculated state transition edge weights were 0.018, 0.019, 0.029, 0.040, 0.049, 0.027, and 0.021, respectively. The state trajectory center offset value increased from 0.086 to 0.244, the trajectory divergence value reached a maximum of 0.225, and the state density change value reached 0.137. The status determination module calculates a structural drift score of 0.642 and a multi-source status change score of 0.438. Multiplying the structural drift score by 0.55 and the multi-source status change score by 0.45 yields a multi-source status consistency score of 0.550, which is then used to determine the equipment fault level as a potential fault level. The traditional electrical threshold method, calculating a comprehensive score of 0.318 for the same time period, still classifies it as a normal state level.

[0032] The validation was performed using 1800 test samples, including 900 normal samples, 360 early structural offset samples, 210 partial occlusion samples, 150 reflective interference samples, and 180 samples with slow current changes. Traditional methods identified 196 early structural offset samples (54.4% accuracy), while this invention identified 323 (89.7% accuracy). Traditional methods misclassified 46 reflective interference samples (30.7% misclassification), while this invention misclassified 17 (11.3% misclassification). Overall, the traditional method achieved an accuracy of 70.6%, a false negative rate of 23.4%, and a false positive rate of 18.1% across the 1800 samples. This invention achieved an accuracy of 90.3%, a false negative rate of 7.2%, and a false positive rate of 8.6%. The simulation results demonstrate that this invention can identify potential fault states in advance based on changes in the equipment's external structure, visual current coupling offset, and equipment timing topology drift diagram, even before electrical parameters show obvious abnormalities.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computer vision-based monitoring system for railway outdoor signaling equipment, characterized in that, include: The data synchronization module collects images and synchronous monitoring data of railway outdoor signaling equipment, performs preprocessing on the synchronous monitoring data and binds it to the images of railway outdoor signaling equipment to generate visual electrical synchronization status data of the equipment. The contour evolution module reads images of railway outdoor signal equipment, establishes an initial contour curve in the target area of ​​the equipment, and constructs an improved Chan-Vese model with HJB optimal evolution constraints to perform contour evolution. It updates the contour evolution direction according to the minimum contour advancement cost and generates an optimized contour set and a set of equipment boundary points. The structural manifold module constructs a similarity matrix of device boundaries based on the set of device boundary points, and uses the Nystrom method to perform low-rank approximation to generate device structural manifold features. It then compares the current device structural manifold features with historical normal structural manifold features to calculate the device structural manifold offset value. The visual coupling module calculates the two-dimensional displacement potential field of the device boundary along the normal direction based on the device optimized contour set, and couples it with the one-dimensional current change sequence to generate visual current coupling offset. The topology drift module reads the device structure manifold offset value and visual current coupling offset according to the sliding time window, constructs the device time-series topology drift diagram, calculates the state trajectory change characteristics, and generates device state drift characteristics. The status determination module performs multi-source status consistency verification on the equipment status drift characteristics based on the equipment's visual electrical synchronization status data, generates a multi-source status consistency score, and determines the equipment fault level result.

2. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The synchronous monitoring data specifically includes 25Hz phase-sensitive track circuit voltage and current data, signal lighting unit voltage and current data, traction return current data, box temperature and humidity data, water immersion status data, box opening status data, and vibration status data.

3. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The generated device visual electrical synchronization status data includes: Read the synchronous monitoring data, perform time alignment according to the collection timestamp, write data with a time difference of less than 0.5 seconds into the same synchronous time window, remove data with empty values, incorrect collection labels and data that exceed the sensor range, and perform resampling at a 1-second sampling interval to generate standardized synchronous monitoring data; The acquisition time, equipment number, and monitoring point number of the railway outdoor signal equipment image are read, and the railway outdoor signal equipment image is bound with standardized synchronous monitoring data within the same equipment number, the same monitoring point number, and the same synchronous time window to generate equipment visual electrical synchronization status data.

4. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The contour evolution module includes: An improved Chan-Vese model is constructed, which includes an initial contour module, a region statistics module, an optimal evolution module, and a boundary output module. The initial contour module reads the image of the railway outdoor signal equipment, establishes the first initial closed contour curve and the second initial closed contour curve in the target area of ​​the equipment, and converts them into the first level set function and the second level set function respectively. It constructs a double level set self-coupling constraint term, exchanges the inner and outer region identifiers of the contour according to the 20-iteration cycle, and generates a double level set contour set. The regional statistics module reads images of railway outdoor signal equipment, divides the images into 32x32 pixel grids, calculates local variance, local entropy, and local gradient variance, and performs weighted summation to generate a full-view cost prediction field. Areas with a value less than the median cost value in the full image are marked as prohibited areas. The optimal evolution module reads the dual-level set of contours, assigns positive-value pixels to the inner region of the contour, and assigns negative-value pixels to the outer region of the contour. It calculates the mean gray value of the inner region and the mean gray value of the outer region respectively, and dynamically updates the weight values ​​of the inner and outer regions based on the total energy change rate. Calculate the gray-scale region difference value, curvature change value, occlusion intensity value, and local noise intensity value. Adjust the proportion of gray-scale region difference value in the contour advancement cost value according to the updated inner and outer region weight values. Perform weighted summation on the gray-scale region difference value, curvature change value, occlusion intensity value, and local noise intensity value to obtain the contour advancement cost value. Combine the full-view cost prediction field to perform HJB optimal evolution constraint, eliminate advancement directions pointing to prohibited entry regions, and determine the optimal contour advancement direction for the current iteration. The boundary output module maps the current contour to the three-dimensional voxel mesh space, calculates the voxel surface tension constraint value and performs contour update. When the difference in the contour advancement cost between two adjacent iterations is less than 0.001, the iteration stops, the zero level set boundary is extracted to generate the device optimized contour set, and the device boundary point set is generated according to a 5-pixel sampling interval. The improved Chan-Vese model was trained using a joint loss value composed of contour advancement cost error, region segmentation error, dual level set consistency error, full view cost prediction error, and boundary reconstruction error as the optimization objective. The parameters in the initial contour module, region statistics module, optimal evolution module, and boundary output module were continuously optimized. Training was stopped when the change in the joint loss value corresponding to 10 consecutive iterations was less than 0.

001.

5. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The structural manifold module includes: Based on the set of device boundary points, the pixel coordinates, curvature value, tangent direction angle, boundary gray value and component connection distance of each boundary point are extracted according to the device number. The zero-to-one interval normalization processing is performed on each feature to generate a boundary point feature sequence. Based on the feature sequence of boundary points, calculate the spatial distance difference, curvature difference, orientation angle difference, gray level difference and connection distance difference between any two boundary points, and calculate the boundary affinity value. Determine the high curvature boundary points and low curvature boundary points according to the curvature value, retain the boundary affinity value of the high curvature boundary point pointing to the low curvature boundary point, and set the reverse boundary affinity value to 0 to generate an edge-preserving asymmetric similarity matrix. Based on the edge-preserving asymmetric similarity matrix, initial representative boundary points are selected according to 10% of the number of boundary points. The Nystrom method is used to perform initial low-rank approximation to generate an initial low-dimensional boundary embedding feature matrix. The embedding error hotspot region is determined based on the reconstruction difference between the initial low-dimensional boundary embedding feature matrix and the edge-preserving asymmetric similarity matrix. In the hotspot region of the embedding error, add representative boundary points corresponding to 15% of the number of boundary points, merge them to generate a recursive representative boundary point set, convert the similarity matrix corresponding to the recursive representative boundary point set into an affinity probability distribution, use Sinkhorn-Knopp iterative execution to perform optimal transport mapping, generate transport-weighted representative similarity matrix, and use the Nystrom method to perform low-rank approximation to generate equipment structure manifold features. Read the historical normal structural manifold features corresponding to the same device number, compare the current device structural manifold features with the historical normal structural manifold features item by item according to the corresponding dimensions, and calculate the device structural manifold offset value.

6. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The visual coupling module includes: Read the optimized contour set of the device, extract the current pixel coordinates of the boundary points of each device according to the device number and the acquisition time, and read the coordinates of the historical normal boundary points corresponding to the same device number. Match the current boundary point coordinates with the historical normal boundary point coordinates to generate a boundary point matching sequence. Based on the boundary point matching sequence, the displacement distance and normal direction of each current boundary point relative to the historical normal boundary points are calculated. The displacement distance is projected along the normal direction to obtain the boundary normal displacement value. All boundary normal displacement values ​​are arranged according to the pixel position of the boundary point in the image to generate a two-dimensional displacement potential field. Read the 25Hz phase-sensitive track circuit voltage and current data, signal lighting unit voltage and current data and traction return current data corresponding to the same equipment number from the equipment visual electrical synchronization status data, extract the current component in each data, calculate the absolute difference between the current components of two adjacent acquisition times, and generate a one-dimensional current change sequence. Normalize the boundary normal displacement value and the one-dimensional current change sequence in the two-dimensional displacement potential field to obtain the visual displacement value and the current change value. Multiply the visual displacement value by 0.60 and the current change value by 0.40 and add them together to generate the visual current coupling value. Then, aggregate all visual current coupling values ​​according to the device number to generate the visual current coupling offset.

7. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The topology drift module includes: According to a 60-second window length and a 10-second sliding step, the device structure manifold offset value and visual current coupling offset of the same device number within a continuous time window are read to form a device state vector, and a device state vector sequence is generated according to the acquisition time order. Based on the device state vector sequence, the device state vector is used as a state node. The structural manifold offset difference and the visual current coupling offset difference between two adjacent state nodes are calculated. The structural manifold offset difference is multiplied by 0.55 and the visual current coupling offset difference is multiplied by 0.45 and then added together to obtain the state transition edge weight. Based on the state nodes and the state transition edge weight, the device time-series topology drift graph is constructed. Based on the device time-series topology drift graph, the average value of the device state vector of all state nodes is calculated as the state trajectory center, the distance between the current state node and the historical normal state trajectory center is calculated as the state trajectory center offset value, the growth rate of the adjacent state transition edge weight is calculated as the trajectory divergence value, and the change rate of the state node distribution density within a unit time window is calculated as the state density change value. The state trajectory center offset value, trajectory divergence value, and state density change value are arranged in the order of acquisition time to generate state trajectory change characteristics. These characteristics are then associated with the equipment number, equipment structure manifold offset value, and visual current coupling offset to generate equipment state drift characteristics.

8. The computer vision-based railway outdoor signal equipment monitoring system according to claim 1, characterized in that, The state determination module includes: Read the equipment's visual electrical synchronization status data and equipment status drift characteristics, extract the status trajectory center offset value, trajectory divergence value and status density change value, and read the 25Hz phase-sensitive track circuit voltage and current, signal light unit voltage and current, traction return current, box temperature and humidity, immersion, unpacking and vibration data to generate a multi-source verification dataset; Normalization is performed on the state trajectory center offset value, trajectory divergence value, and state density change value. The state trajectory center offset value is multiplied by 0.40, the trajectory divergence value is multiplied by 0.35, and the state density change value is multiplied by 0.25, and then added together to generate the structure drift score. Read the multi-source verification dataset, calculate the variation amplitude of 25Hz phase-sensitive track circuit voltage and current, signal lighting unit voltage and current, traction return current, box temperature and humidity and vibration data under two adjacent acquisition times, and generate multi-source state change scores by combining water immersion data and box opening data. The structure drift score is multiplied by 0.55 and the multi-source state change score is multiplied by 0.45, and then the results are added together to generate the multi-source state consistency score. When the multi-source status consistency score is less than 0.40, it is determined to be a normal status level; when the multi-source status consistency score is greater than or equal to 0.40 and less than 0.65, it is determined to be a potential fault level; when the multi-source status consistency score is greater than or equal to 0.65 and less than 0.85, it is determined to be an enhanced early warning level; and when the multi-source status consistency score is greater than or equal to 0.85, it is determined to be a fault alarm level.

Images