A method and system for intelligent body inspection for railway communication signal room

Abstract

The application discloses a body intelligent inspection method and system for railway communication signal machine rooms, relates to the field of robust artificial intelligence technology, and comprises the following steps: collecting and preprocessing machine room inspection data to generate a standardized machine room inspection data set; extracting a machine room space topology to generate a machine room body inspection task graph; performing inspection pose closure control to generate equipment inspection samples; performing view angle fitting and state unit registration to generate equipment state identification samples; improving StarNet reconstruction state features to generate equipment state candidate results; ByteTrack cross-frame associated state candidates to generate equipment state tracking chains; and abnormity track closure discrimination and task graph back connection to generate body intelligent inspection results. Through the introduction of the improved StarNet network and ByteTrack algorithm, the application realizes accurate identification and abnormity closed loop back connection of the equipment state of the railway communication signal machine room.

Description

Technical Field

[0001] This invention relates to the field of robust artificial intelligence technology, and in particular to an embodied intelligent inspection method and system for railway communication signal equipment rooms. Background Technology

[0002] With the development of railway communication and signaling systems towards centralized and intelligent operation, the status inspection of communication equipment, signaling equipment, power supply equipment, relay equipment, wiring equipment, and environmental monitoring equipment in communication and signaling equipment rooms has a significant impact on railway transportation safety. Currently, railway communication and signaling equipment room inspections mainly rely on manual inspections, fixed camera monitoring, and inspection robots. Manual inspections require maintenance personnel to check equipment indicator lights, terminal connections, cable access, equipment temperature, and the equipment room environment cabinet by cabinet. The quality of inspections is greatly affected by personnel experience, inspection frequency, and work time. Fixed cameras can cover part of the cabinet area, but their fixed installation positions make it difficult to cover the sides of cabinets, parts of terminal blocks, cable access points, and obstructed areas. Inspection robots can carry visible light cameras, infrared thermal imaging equipment, audio pickup devices, and environmental sensors into the equipment room to perform data collection tasks, improving the automation level of equipment room inspections to a certain extent.

[0003] Existing embodied intelligent inspection methods still have shortcomings in railway communication and signaling equipment room scenarios. The lack of a unified mapping between equipment room space, equipment locations, inspection paths, and data acquisition tasks makes inspection robots prone to docking posture deviations, data acquisition perspective deviations, and inaccurate equipment location alignment, resulting in difficulties in stably associating inspection data with specific equipment locations and status units. Existing visual recognition methods often directly detect or classify the current image, making them susceptible to influences from cabinet edges, nameplate borders, ventilation holes, fixed terminal blocks, reflective bright spots, and repetitive textures. This can cause stable structures to be misidentified as abnormal, and can also lead to the loss of status responses from small targets such as indicator lights, terminal connection areas, and cable access areas. The dense arrangement of status units within the equipment room easily leads to candidate box adhesion; low-confidence candidate boxes are easily formed under conditions of low light, reflection, and occlusion. Traditional cross-frame association methods are prone to trajectory breaks, trajectory concatenation, and missed tracking. Existing inspection results are mostly output in the form of anomaly lists or alarm records, lacking a process of connecting the anomaly trajectory closure judgment results with the specific inspection task map of the computer room, which affects the positioning accuracy, stability and closed-loop processing efficiency of the inspection results.

[0004] Therefore, how to provide an embodied intelligent inspection method and system for railway communication signal equipment rooms 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 an embodied intelligent inspection method and system for railway communication signal equipment rooms. This invention utilizes embodied inspection task map construction, robot pose closure control, state unit registration, improved StarNet network state recognition, and ByteTrack cross-frame correlation tracking technology. It details the implementation process of equipment location reachability mapping, autonomous robot inspection and data acquisition, equipment state candidate identification, state trajectory tracking, and anomaly result feedback within the railway communication signal equipment room. This achieves precise correlation between communication signal equipment locations, state units, and inspection tasks. Compared to traditional manual inspection, fixed camera monitoring, and ordinary robot inspection methods, this invention offers advantages such as accurate equipment location mapping, stable acquisition perspective, strong small target state recognition capability, stable tracking of dense state units, low-confidence anomalies that are less likely to be missed, and clear closed-loop feedback of inspection results.

[0006] An embodied intelligent inspection method for railway communication signal equipment rooms according to an embodiment of the present invention includes: Collect inspection data of railway communication signal equipment rooms, preprocess the inspection data of railway communication signal equipment rooms, and generate a standardized equipment room inspection dataset; Based on the standardized data center inspection dataset, the data center spatial topology is extracted, the equipment locations are mapped to reachability during inspection, and a data center-specific inspection task map is generated. According to the specific inspection task diagram of the computer room, the inspection robot performs inspection posture closure control and equipment status acquisition to generate equipment inspection samples. Based on the robot pose, acquisition angle and equipment location in the equipment inspection sample, viewpoint snapping and state unit registration are performed to generate equipment state recognition sample; An improved StarNet network is constructed to reconstruct star operation features on device state recognition samples, snapping together stable structural responses, retaining local star responses, and screening high-response candidates to generate device state candidate results. The ByteTrack algorithm is used to correlate device state candidate results across frames, recover dense state unit trajectories, and perform bifurcation trajectory merging on low-confidence candidate boxes that do not match the main trajectory to generate a device state tracking chain. Using the equipment status tracking chain as the discrimination object, abnormal trajectory closure discrimination is performed, and the closure discrimination result is fed back to the computer room specific inspection task map to generate specific intelligent inspection results.

[0007] Optionally, the railway communication and signal equipment room inspection data specifically includes equipment room spatial layout data, communication and signal equipment location ledger data, inspection robot pose data, inspection task data, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, equipment room environmental monitoring data, and historical inspection and handling data.

[0008] Optionally, the generation of the standardized data center inspection dataset includes: The fields and numbering of the equipment room spatial layout data, communication and signal equipment location ledger data, inspection task data and historical inspection and handling data in the railway communication and signal equipment room inspection data are unified and numbered to form a basic index table for the equipment room. The robot pose data, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, and computer room environment monitoring data are timestamped and calibrated, and an inspection segment index table is formed according to the inspection task number. Distortion correction and image stabilization are performed on visible light inspection images, temperature calibration and field of view registration are performed on infrared thermal imaging data, and frame segmentation and noise clip removal are performed on equipment sound data to form a multimodal inspection record. The multimodal inspection records are linked to the computer room basic index table according to the equipment location, inspection task number and robot pose, and the collection view and inspection segment number are marked to form the equipment location inspection record. The equipment inspection records are supplemented with missing items, abnormal formats are removed, and data formats are standardized to generate a standardized computer room inspection dataset.

[0009] Optionally, the generation of the data center-specific inspection task map includes: Based on the computer room spatial layout data and inspection robot access constraints in the standardized computer room inspection dataset, the computer room passage, equipment cabinets and obstacle areas are divided into topology nodes to generate a computer room spatial topology table. Link the communication signal equipment location ledger data with the computer room space topology table, mark the cabinet, inspection surface and status unit range of the equipment location, and generate an equipment location mapping table; Based on the computer room space topology table and the equipment location mapping table, inspection docking poses and acquisition viewpoint anchor points are generated for the equipment locations, and unreachable locations are eliminated to generate an equipment location reachability mapping table. Based on the inspection task data, the reachable points in the equipment location reachable mapping table are assigned tasks, the inspection sequence, data collection type and review trigger conditions are marked, and a specific inspection task node table is generated. The data center space topology table, equipment location mapping table, equipment location reachability mapping table, and embodied inspection task node table are encapsulated into a graph structure to generate a data center embodied inspection task graph.

[0010] Optionally, the generation of equipment inspection samples includes: According to the inspection sequence in the data center inspection task diagram, the tasks of reachable equipment points, inspection docking positions, data collection view anchor points and data collection types are arranged to generate the data center inspection execution sequence. The inspection robot is driven to move along the topology of the computer room space according to the computer room inspection execution sequence, and the pose deviation between the robot's current pose and the inspection docking pose is calculated to obtain the inspection pose deviation. The chassis position, body orientation, and acquisition gimbal angle of the inspection robot are closed-loop corrected based on the inspection posture deviation to obtain the equipment alignment inspection status. In the equipment alignment and inspection state, visible light, infrared, video, sound and environmental status are collected at the communication signal equipment locations according to the collection angle anchor point to generate equipment inspection segments; The equipment inspection segments are packaged with equipment locations, robot poses, acquisition perspectives, inspection task numbers, and acquisition times to generate equipment inspection samples.

[0011] Optionally, the generation of device status identification samples includes: Based on the equipment locations, robot poses, and acquisition angles in the equipment inspection samples, the equipment inspection segments are matched to the inspection surface coordinates of the corresponding communication signal equipment to generate equipment angle matching records. Based on the equipment's perspective and the recording, the status display area, terminal connection area, and cable access area within the inspection surface are divided into status units to generate equipment status unit frames. Perform perspective correction, scale normalization, and acquisition time alignment on the multimodal inspection segments corresponding to the equipment status unit frames to generate status unit registration segments. The state unit registration segment is matched with the historical normal inspection segment of the same equipment location to generate a state baseline pairing record. The status baseline pairing record is encapsulated with the equipment location, status unit number and inspection task number to generate equipment status identification samples.

[0012] Optionally, the generation of candidate device status results includes: An improved StarNet network was constructed, which included a baseline star accumulation residual stripping block, a small target star response fidelity bypass, and an abnormal channel return verification block. The current state segment and baseline state segment in the equipment status identification sample are sent into the StarNet star operation encoding process. The corresponding position and corresponding channel features in the same state unit are multiplied and interacted to generate the current inspection star features and baseline star features. The baseline star-product residual stripping block receives the current inspection star characteristics and the baseline star characteristics, performs star-product coupling on the current inspection response and the normal baseline response at the same equipment location, strips the stable structural response, and generates equipment state residual characteristics. The small target star response fidelity bypass receiver status residual characteristics are used to perform shallow buffering, position preservation and channel backfeeding on the local star response of the status display area, terminal connection area and cable access area to generate small target fidelity status characteristics. The abnormal channel back-back verification block receives the small target fidelity state features, filters high-response channels and back-backs the high-response region to the shallow structural features of the corresponding state unit, eliminates false candidates based on the co-occurrence relationship between the equipment structure boundary response and the state change response, and generates valid state response features. Based on the valid status response characteristics, generate device status candidate boxes, status categories, status response values, and status unit numbers, and encapsulate them into device status candidate results; The improved StarNet network was trained using manually labeled communication signal device state boxes, state categories, and state unit numbers as supervision labels. The network parameters were updated by using a joint loss consisting of device state classification error, candidate box localization error, baseline residual constraint error, small target star response preservation error, and abnormal backtracking verification error. When the average localization deviation of device state candidate boxes in the validation set was less than 2 pixels and the state category recognition accuracy reached 98.0%, the improved StarNet network was considered to have completed training.

[0013] Optionally, the generation of the device state tracking chain includes: The candidate results of device status are sorted into frames according to the status unit number, device location and acquisition time, and the candidate boxes are divided into high confidence candidate boxes and low confidence candidate boxes. The ByteTrack algorithm is used to associate high-confidence candidate boxes with the main trajectory of device status for the first time, generating the main trajectory matching results and unmatched candidate boxes; When the same unmatched candidate box covers more than two device status main trajectory prediction areas at the same time, the unmatched candidate box is marked as a stuck candidate box, and the stuck candidate box splitting and reconnection process is performed according to the trajectory center arrangement direction of the previous frame, the position of the status unit and the response peak inside the candidate box to generate the dense status unit trajectory recovery result. A bifurcation trajectory pool is established for low-confidence candidate boxes that do not match the main trajectory. Low-confidence bifurcation trajectory merging is performed according to the position offset of consecutive frames, consistency of state category, and trend of response value change to generate low-confidence bifurcation trajectory processing results. The main trajectory matching results, dense state unit trajectory recovery results, and low-confidence bifurcation trajectory processing results are encapsulated according to the device location and state unit number to generate a device status tracking chain.

[0014] Optionally, generating embodied intelligent inspection results includes: Using the equipment status tracking chain as the discrimination object, the continuity verification, status consistency verification and trajectory closure discrimination are performed on the status unit trajectory corresponding to the equipment point, and abnormal trajectory discrimination results are generated. The abnormal trajectory identification results are fed back to the computer room's specific inspection task map. The abnormal equipment locations, status unit numbers, inspection task numbers, and review tasks are associated and mapped to generate specific inspection task feedback results. The results of the embodied inspection task are packaged according to the equipment location number to generate embodied intelligent inspection results.

[0015] An embodied intelligent inspection system for railway communication signal equipment rooms according to an embodiment of the present invention includes the following modules: The inspection data standardization module is used to collect and preprocess railway communication signal equipment room inspection data to generate a standardized equipment room inspection dataset. The embodied task map construction module is used to perform inspection reachability mapping of equipment locations and generate an embodied inspection task map of the computer room. The robot inspection execution module is used to control the inspection robot to perform specific inspection tasks and generate equipment inspection samples; The status recognition sample generation module is used to perform perspective snapping and status unit registration on equipment inspection samples to generate equipment status recognition samples. The device state candidate generation module is used to identify the state of device state recognition samples by improving the StarNet network and generate device state candidate results. The device status tracking module is used to perform cross-frame correlation of device status candidate results using the ByteTrack algorithm to generate a device status tracking chain. The inspection result feedback module is used to identify abnormal trajectory closure and feedback the task map of the equipment status tracking chain, and generate personalized intelligent inspection results.

[0016] The beneficial effects of this invention are: This invention proposes an embodied intelligent inspection method and system for railway communication and signaling equipment rooms. By constructing a standardized equipment room inspection dataset, an embodied inspection task map, equipment inspection samples, and equipment status identification samples, a continuous processing flow is formed, from equipment room space, equipment locations, inspection tasks to robot data collection actions. Compared to manual inspection, fixed camera monitoring, and ordinary robot inspection methods, this invention can uniformly arrange the equipment room space topology, equipment locations, inspection docking postures, and data collection viewpoint anchor points, reducing data mismatch problems caused by docking deviations, viewpoint deviations, and inaccurate point matching, and improving the accuracy of the correlation between inspection data and specific equipment locations and status units.

[0017] This invention constructs an improved StarNet network to reconstruct star operation features from equipment status identification samples. It also introduces a baseline star-product residual stripping block, a small target star response fidelity bypass, and an abnormal channel return verification block to generate candidate equipment status results. The baseline star-product residual stripping block weakens stable background response interference from cabinet edges, nameplate borders, heat dissipation holes, and fixed terminal structures. The small target star response fidelity bypass preserves local star responses in the status display area, terminal connection area, and cable access area. The abnormal channel return verification block filters out false candidates formed by reflective bright spots, repetitive textures, and local noise, improving the robustness of communication signal equipment status identification and the reliability of candidate results.

[0018] This invention employs the ByteTrack algorithm to perform cross-frame association of device status candidate results. It generates a device status tracking chain through processes such as splitting and reconnecting adhered candidate boxes and merging low-confidence bifurcated trajectories. This mitigates issues caused by densely packed status units, candidate box adhesion, low-light reflection occlusion, and low-confidence candidate boxes, leading to trajectory breaks, trajectory concatenation, and missed tracking. By using abnormal trajectory closure discrimination and data center-specific inspection task map reconnection, this invention clearly associates abnormal device locations, status unit numbers, inspection task numbers, and verification tasks, improving the positioning accuracy, trajectory stability, and closed-loop processing efficiency of inspection results. Attached Figure Description

[0019] 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 flowchart of an embodied intelligent inspection method for railway communication signal equipment rooms proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved StarNet network for an embodied intelligent inspection method for railway communication signal equipment rooms proposed in this invention. Figure 3 This is a schematic diagram of the structure of an embodied intelligent inspection system for railway communication signal equipment rooms proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figure 1 and Figure 2 An embodied intelligent inspection method for railway communication signal equipment rooms includes: Collect inspection data of railway communication signal equipment rooms, preprocess the inspection data of railway communication signal equipment rooms, and generate a standardized equipment room inspection dataset; Based on the standardized data center inspection dataset, the data center spatial topology is extracted, the equipment locations are mapped to reachability during inspection, and a data center-specific inspection task map is generated. According to the specific inspection task diagram of the computer room, the inspection robot performs inspection posture closure control and equipment status acquisition to generate equipment inspection samples. Based on the robot pose, acquisition angle and equipment location in the equipment inspection sample, viewpoint snapping and state unit registration are performed to generate equipment state recognition sample; An improved StarNet network is constructed to reconstruct star operation features on device state recognition samples, snapping together stable structural responses, retaining local star responses, and screening high-response candidates to generate device state candidate results. The ByteTrack algorithm is used to correlate device state candidate results across frames, recover dense state unit trajectories, and perform bifurcation trajectory merging on low-confidence candidate boxes that do not match the main trajectory to generate a device state tracking chain. Using the equipment status tracking chain as the discrimination object, abnormal trajectory closure discrimination is performed, and the closure discrimination result is fed back to the computer room specific inspection task map to generate specific intelligent inspection results.

[0022] In this embodiment, the railway communication signal equipment room inspection data specifically includes equipment room spatial layout data, communication signal equipment location ledger data, inspection robot pose data, inspection task data, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, equipment room environmental monitoring data, and historical inspection and handling data.

[0023] In this embodiment, generating a standardized data center inspection dataset includes: The fields and numbering of the equipment room spatial layout data, communication and signal equipment location ledger data, inspection task data and historical inspection and handling data in the railway communication and signal equipment room inspection data are unified and numbered to form a basic index table for the equipment room. The pose data of the inspection robot, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, and computer room environmental monitoring data are timestamped and calibrated. An inspection segment index table is then created according to the inspection task number, including: Perform timestamp calibration, specifically as follows: Using the clock of the inspection robot controller as the reference clock, the synchronization trigger time between the robot controller and each data acquisition device is recorded when the inspection task is issued, the equipment location is reached, data acquisition is started, and data acquisition ends. Divide the base time interval of the start and end phases of the inspection task by the local time interval of the equipment to obtain the clock drift correction factor. Divide the local timestamp of the equipment by the local start timestamp, multiply the time difference by the clock drift correction factor, and then add the base start timestamp to obtain the unified base timestamp. According to the unified reference timestamp, the visible light inspection images, infrared thermal images, equipment status videos, equipment sounds and computer room environment monitoring data are converted to the unified reference time axis, and the robot pose records before and after the unified reference timestamp are located. When the unified reference timestamp is located between two adjacent robot pose records, the robot position, body orientation and gimbal attitude are linearly interpolated according to the ratio of the time difference before and after. When it is impossible to locate an adjacent robot pose record with a time difference of no more than 50 milliseconds, the corresponding inspection data is removed. The inspection data that has been calibrated with timestamps are merged according to the inspection task number. The start time and end time of the data collection are used as the start and end times of the inspection segment, respectively. The visible light image range, infrared thermal image range, video frame range, sound frame range, environmental monitoring record range and robot pose record range corresponding to each inspection segment are recorded to form an inspection segment index table. Distortion correction and image stabilization are performed on visible light inspection images; temperature calibration and field-of-view registration are performed on infrared thermal imaging data; and frame segmentation and noise clip removal are performed on equipment audio data to form a multimodal inspection record, including: Multimodal inspection records are generated, specifically as follows: Based on the focal length, principal point coordinates, and radial distortion parameters of the visible light camera, the distorted pixel coordinates in the visible light inspection image are converted into corrected pixel coordinates. Stable feature points are extracted from the cabinet edge, panel corners, and equipment panel outline. The lateral displacement, longitudinal displacement, and rotational deviation between adjacent images are calculated. The current image is then mapped in reverse to the coordinates of the first frame inspection image to complete the image stabilization process. Based on the temperature calibration coefficient of the infrared thermal imaging equipment, the grayscale values ​​of infrared pixels are converted into temperature values. The corresponding cabinet edges and panel corners are selected as registration points in the visible light inspection image and the infrared thermal image. The perspective mapping parameters are calculated based on the coordinate correspondence of the registration points, and the infrared temperature pixels are mapped to the corresponding visible light inspection image coordinates. The device sound data is overlapped and framed according to a frame length of 0.5 seconds and a frame shift of 0.25 seconds. The short-time energy of each sound frame is calculated. Sound frames that coincide with the start and stop time of the inspection robot chassis and whose short-time energy exceeds twice the average short-time energy of adjacent sound frames are removed. The processed visible light inspection images, infrared temperature data, valid sound frames, equipment status video frames, and computer room environment monitoring data are time-series aligned and recorded according to the calibration timestamp, and associated with the inspection task number and robot pose to form a multimodal inspection record. The multimodal inspection records are linked to the computer room basic index table according to equipment location, inspection task number, and robot pose, and the acquisition viewpoint and inspection segment sequence number are marked to form equipment location inspection records, including: Mark the acquisition viewpoint and inspection segment number, specifically as follows: The target equipment location is determined based on the inspection task number. The center of the equipment inspection surface and the normal of the inspection surface are extracted from the basic index table of the computer room. The planar distance between the inspection robot and the target equipment location, the angle between the camera optical axis and the normal of the inspection surface, and the pixel offset of the center of the image relative to the center of the inspection surface are calculated. The acquisition angle is determined based on the planar distance, the optical axis angle, and the pixel offset. When the planar distance does not exceed 0.15 meters, the optical axis angle does not exceed 5 degrees, and the pixel offset does not exceed 5% of the image width, it is marked as a normal viewing angle. When the conditions for determining the frontal viewing angle are not met, but the plane distance does not exceed 0.20 meters, the optical axis angle does not exceed 15 degrees, and the pixel offset does not exceed 15% of the image width, the optical axis deviation direction is marked as left-biased viewing angle, right-biased viewing angle, top-down viewing angle, or bottom-up viewing angle. Records that exceed any maximum limit range are marked as invalid viewing angle records and are discarded. Valid multimodal inspection records are merged according to the inspection task number and equipment location number. They are sorted according to the calibration timestamp, and records with adjacent time intervals of no more than 2 seconds are divided into the same inspection segment. The inspection segment sequence number is marked according to the segment start time to form the equipment location inspection record. The equipment inspection records are supplemented with missing items, abnormal formats are removed, and data formats are standardized to generate a standardized computer room inspection dataset.

[0024] In this embodiment, the generation of the data center-specific inspection task map includes: Based on the computer room spatial layout data and inspection robot access constraints in the standardized computer room inspection dataset, topology nodes are divided for computer room aisles, equipment racks, and obstacle areas, generating a computer room spatial topology table, where: The topology nodes are partitioned as follows: Based on the vehicle width and safe passage distance of the inspection robot, the boundaries of walls, equipment cabinets and obstacle areas in the computer room spatial layout data are expanded outward, and the area within the effective surface of the computer room that is not occupied by the expanded boundaries is determined as the area that the robot can pass through. Set up topology nodes at intersections, turns, ends of equipment racks, entrances and exits of the computer room, and locations where the width of the passage changes, and record the node number, spatial coordinates, and node type of each topology node; For adjacent topology nodes, a passage verification is performed. When the node connection path does not cross the extended obstacle area, and the effective width of the channel and the turning space meet the requirements of the inspection robot's body width, minimum turning radius and safe passage distance, the node connection relationship is established, and the connection distance, channel width and passage direction are recorded. Summarize the topology nodes, node connection relationships, and corresponding data center channels, equipment cabinets, and obstacle area markers to generate a data center space topology table; Link the communication signal equipment location ledger data with the computer room space topology table, mark the cabinet, inspection surface and status unit range of the equipment location, and generate an equipment location mapping table; Based on the data center space topology table and equipment location mapping table, inspection docking poses and acquisition viewpoint anchor points are generated for each equipment location. Unreachable locations are removed, and an equipment location reachability mapping table is generated, including: Generate a device location reachability mapping table, specifically as follows: The center of the equipment inspection surface is calculated based on the boundary coordinates of the equipment inspection surface, and the normal of the inspection surface is determined based on the spatial orientation of the equipment inspection surface. Combined with the effective imaging distance of the camera pre-configured by the inspection robot, the acquisition distance is determined under the condition that the equipment inspection surface is completely within the camera's field of view. Candidate docking positions are determined along the outward normal of the inspection surface facing the side of the computer room passage, and the inspection robot is made to face the equipment inspection surface to generate candidate inspection docking poses. The acquisition center is determined based on the state unit range. The line connecting the candidate docking position and the acquisition center is taken as the direction of the camera optical axis. Combined with the camera installation height and gimbal rotation range pre-configured by the inspection robot, the gimbal pitch angle and yaw angle are determined to generate candidate acquisition viewpoint anchor points. From the topology table of the computer room space, select accessible topology nodes that are not blocked by obstacles between them and the candidate docking positions. Connect the accessible topology node with the shortest connection distance to the candidate docking position and verify the connection path from the topology node to the candidate docking position, the docking area, the turning space, and the candidate acquisition view anchor point. When the connection path is connected to the robot's accessible area, the docking area and the turning space meet the inspection robot's passage constraints, and the gimbal pitch angle and yaw angle are within the gimbal rotation range, mark the corresponding equipment point as a reachable point. Remove the equipment points that fail the verification, determine the candidate inspection docking pose and candidate acquisition view anchor point corresponding to the reachable point as the inspection docking pose and acquisition view anchor point respectively, and summarize the equipment point number, inspection docking pose, acquisition view anchor point and attached topology node to generate the equipment point reachability mapping table. Based on the inspection task data, the reachable points in the equipment location reachable mapping table are assigned tasks, the inspection sequence, data collection type and review trigger conditions are marked, and a specific inspection task node table is generated. The data center space topology table, equipment location mapping table, equipment location reachability mapping table, and embodied inspection task node table are encapsulated into a graph structure to generate a data center embodied inspection task graph, in which: The graph structure is encapsulated, specifically as follows: The topology nodes in the computer room space topology table, the device points in the device point mapping table, the inspection docking poses in the device point reachability mapping table, and the inspection task nodes in the embodied inspection task node table are respectively encapsulated as space nodes, device nodes, docking nodes, and task nodes. Establish spatial access edges based on the connection relationship of topological nodes, establish equipment mapping edges based on the location attachment relationship of equipment points, establish docking association edges based on the correspondence between equipment points and inspection docking poses, establish task attachment edges based on the task attachment relationship between inspection task nodes and equipment points, and establish task timing edges according to the inspection sequence. Write the spatial coordinates, equipment location number, inspection docking posture, data collection view anchor point, inspection sequence, data collection type and review triggering conditions into the corresponding node, and write the connection distance, channel width, travel direction and task sequence into the corresponding connection edge. Combine various nodes and connecting edges into a graph structure, and establish an index relationship between equipment location number, inspection task number and graph node number to generate a data center-specific inspection task graph.

[0025] In this embodiment, the generation of equipment inspection samples includes: According to the inspection sequence in the data center inspection task diagram, the tasks of reachable equipment points, inspection docking positions, data collection view anchor points and data collection types are arranged to generate the data center inspection execution sequence. The inspection robot is driven to move along the topology of the computer room space according to the inspection execution sequence, and the pose deviation between the robot's current pose and the inspection docking pose is calculated to obtain the inspection pose deviation amount, where: The inspection pose deviation is obtained as follows: Extract the inspection docking position, docking direction and data collection view anchor point corresponding to the current equipment location from the data center inspection execution sequence, and obtain the current position, body orientation and gimbal angle of the inspection robot; The coordinate difference between the inspection docking position and the robot's current position is converted to the robot's body coordinate system to obtain the longitudinal position deviation and the lateral position deviation. The docking distance deviation is then calculated based on the longitudinal position deviation and the lateral position deviation. The angle difference between the parking heading and the current fuselage orientation is calculated to obtain the fuselage heading deviation. The gimbal pitch angle and yaw angle in the collected viewpoint anchor point are respectively subtracted from the current gimbal angle to obtain the gimbal pitch deviation and gimbal yaw deviation. The longitudinal position deviation, lateral position deviation, docking distance deviation, fuselage heading deviation, gimbal pitch deviation, and gimbal yaw deviation are summarized to obtain the inspection attitude deviation. Based on the inspection posture deviation, the chassis position, body orientation, and acquisition gimbal angle of the inspection robot are corrected for closure to obtain the equipment alignment inspection status, where: The equipment alignment and inspection status is obtained as follows: The chassis adjustment direction and movement amount are determined based on the longitudinal and lateral position deviations, driving the inspection robot to make steering and forward / backward adjustments. The chassis rotation is controlled based on the body heading deviation to correct the inspection robot's docking position and body orientation. Based on the pitch and yaw deviations of the gimbal, adjust the pitch and yaw angles of the acquisition gimbal so that the camera optical axis is oriented toward the acquisition center corresponding to the acquisition viewpoint anchor point. After completing a single calibration, the current position, body orientation, and gimbal angle of the inspection robot are reacquired, and the inspection posture deviation is recalculated. The chassis, body, and acquisition gimbal calibration are then performed according to the recalculated inspection posture deviation. When the docking distance deviation, fuselage heading deviation, gimbal pitch deviation and gimbal yaw deviation are all within the corresponding allowable deviation range, the correction is stopped and the inspection robot chassis and data acquisition gimbal are locked to obtain the equipment alignment inspection status. In the equipment alignment and inspection state, visible light, infrared, video, sound, and environmental status data are collected at the communication signal equipment locations according to the acquisition angle anchor points to generate equipment inspection segments, including: Generate equipment inspection segments, specifically: Based on the anchor point of the acquisition view, locate the communication signal equipment and the corresponding acquisition area, and according to the acquisition type in the computer room inspection execution sequence, start the corresponding visible light camera, infrared thermal imaging equipment, video acquisition equipment, audio pickup equipment or environmental sensor; Using the data acquisition start time as the segment start time, the inspection data output by the started acquisition device is collected synchronously, and the robot pose, gimbal angle and device location number corresponding to various types of inspection data are recorded. When the set collection time is reached or a collection completion marker is received at the current device location, data collection is stopped. All types of data from the collection start time to the collection end time are merged according to a unified timeline to generate a device inspection segment. The equipment inspection segments are packaged with equipment locations, robot poses, acquisition perspectives, inspection task numbers, and acquisition times to generate equipment inspection samples.

[0026] In this embodiment, the generation of device status identification samples includes: Based on the equipment locations, robot poses, and acquisition perspectives in the equipment inspection samples, the equipment inspection segments are overlaid onto the inspection surface coordinates of the corresponding communication signal equipment to generate an equipment perspective overlay record, in which: Generate device view snap-in records, specifically: Based on the equipment location, determine the inspection surface of the corresponding communication signal equipment, and establish the inspection surface coordinate axis along the horizontal and vertical directions of the inspection surface, using the boundary corner of the inspection surface as the coordinate reference. The shooting direction of the equipment inspection segment is determined based on the robot's pose and the acquisition angle. The cabinet edge and panel corner points are extracted from the visible light image, infrared thermal image and equipment status video frame, and the position correspondence is established with the edge corner points of the inspection surface. Based on the position correspondence, the mapping parameters from each image frame to the inspection surface coordinate system are calculated, and the visible light image, infrared thermal image and equipment status video frame are matched to the unified inspection surface coordinates. The corresponding equipment sound and computer room environment status are associated according to the acquisition time. Record the equipment location, robot pose, acquisition perspective, image snapping results in the inspection surface coordinate system, as well as the temporal correlation between equipment sound and computer room environment status, and generate equipment perspective snapping records. Based on the equipment's perspective and tracking records, the status display area, terminal connection area, and cable access area within the inspection surface are divided into status units, generating equipment status unit frames, where: The device status cell is generated as follows: Based on the status unit range in the equipment location mapping table, locate the status display area, terminal connection area, and cable access area in the inspection surface coordinate system of the equipment view snap-in record; Extract the boundary coordinates of each region, determine the top left corner coordinates, bottom right corner coordinates, and frame size of the state unit frame according to the outer boundary of the region, and mark the corresponding state unit type; Assign status unit numbers to each status unit frame according to the equipment location number and the order of the area arrangement, and generate equipment status unit frames; Perspective correction, scale normalization, and acquisition time alignment are performed on the multimodal inspection segments corresponding to the equipment status unit frames to generate status unit registration segments, wherein: Generate state unit registration segments, specifically: According to the device status unit frame, the corresponding status unit area is cropped from the visible light image, infrared thermogram and device status video frame, and local perspective correction is performed based on the deviation between the status unit frame boundary and the standard rectangle boundary. The state unit area that has completed perspective correction is scaled to a uniform width and height, and the visible light image, infrared thermal image and device status video frame are resampled according to the same state unit frame coordinate range. Based on the calibration timestamp, the visible light image, infrared thermal image and equipment status video frame are time-matched, and the equipment sound and the room environment status within the same acquisition period are associated to generate a status unit registration segment. The state unit registration segment is matched with the historical normal inspection segment of the same equipment location to generate a state baseline pairing record, wherein: Generate state baseline pairing records, specifically as follows: According to the equipment location and status unit number, the corresponding status unit area is extracted from the historical normal inspection segments, and perspective correction, scale normalization and acquisition time alignment are performed. Normal status unit segments with the same acquisition type are selected. Historical normal baseline segments are determined according to the acquisition angle and image clarity. Among them, the historical normal inspection segments are historical equipment inspection segments selected from the standardized computer room inspection dataset, whose inspection results are marked as normal and have no associated abnormal handling records. Using the boundary of the device status unit frame as the snapping reference, the status unit registration segment and the historical normal baseline segment are translated, scaled and aligned at the boundary. The registered segments of the snapped-together state units are paired with historical normal baseline segments, and the device location, state unit number, acquisition time and baseline segment number are recorded to generate a state baseline pairing record. The status baseline pairing record is encapsulated with the equipment location, status unit number and inspection task number to generate equipment status identification samples.

[0027] In this embodiment, the generation of candidate device status results includes: An improved StarNet network is constructed, which includes a baseline star-integral residual stripping block, a small target star response fidelity bypass, and an anomaly channel return verification block, wherein: The StarNet network is constructed and improved as follows: In the traditional StarNet network's star-product block, the processing path that directly inputs the bi-branch features into the projection convolution after element-wise star-product is changed to accessing the baseline star-product residual stripping block. The star-product features of the state unit registration segment and the historical normal baseline segment are calculated separately, and the residuals of the two are subtracted before being fed into the original projection convolution. A small target star response fidelity bypass is led out from the output end of the baseline star accumulation residual stripping block to perform position preservation and channel mapping on the local features of the status display area, terminal connection area and cable access area, and the bypass output is superimposed with the equipment status residual features; An abnormal channel back-off verification block is added between the end of each stage and the next stage downsampling layer. The candidate abnormal channels in the stage output are fed back to the corresponding star-integral branch for response comparison. The response of the candidate abnormal channels is adjusted according to the comparison results, and the adjusted channel features are fused with the original stage output to obtain the improved StarNet network. The current state segment and baseline state segment from the device status identification samples are fed into the StarNet star operation encoding process. Multiplication interactions are performed on corresponding positions and corresponding channel features within the same state unit to generate current inspection star features and baseline star features, where: Generate the current inspected star features and baseline star features, specifically as follows: According to the equipment location and status unit number, the current status segment and baseline status segment are extracted from the equipment status identification sample and sent to the StarNet star operation coding branch with shared parameters respectively. Perform depthwise convolution on the current state segment and the baseline state segment respectively to extract local location features within the same state unit. Then, input the two sets of local location features into two linear mapping branches to obtain the first branch features and the second branch features with the same channel dimension. The two branch features corresponding to the current state segment are multiplied element-wise according to the same spatial location and the same channel, and the two branch features corresponding to the baseline state segment are multiplied element-wise according to the same spatial location and the same channel. Perform projection convolution on the product features corresponding to the current state segment and the baseline state segment respectively, and then perform residual superposition on the projection convolution result and the input features of the corresponding star block to obtain the current inspection star features and the baseline star features; The baseline star-product residual stripping block receives the current inspection star characteristics and the baseline star characteristics, performs star-product subtraction on the current inspection response and the normal baseline response at the same equipment location, strips the stable structural response, and generates equipment state residual characteristics, where: Baseline star-shaped residual stripping blocks include: Dual-path satellite feature cache queue: caches the features of the currently inspected satellites and the baseline satellite features; Point-to-Channel Alignment Table: Records the correspondence between equipment points, status units, spatial locations, and characteristic channels; Star-shaped response latch: Pairs and latches responses from the same position and channel; Steady Response Gating: Generates the retention coefficients corresponding to the baseline response; Baseline Residual Stripper: Subtracts the weighted baseline response from the current inspection star characteristics; Residual Feature Output Register: Collects and outputs device status residual features; The processing procedure for baseline star-cumulative residual stripping blocks is as follows: The dual-path satellite feature cache queue receives the current inspection satellite feature and the baseline satellite feature respectively. Let the current inspection satellite feature be X and the baseline satellite feature be B. The size of both sets of input features is H×W×C, where H is the feature height, W is the feature width, and C is the number of feature channels. Paired cache indexes are established according to the device point number and status unit number. The point channel alignment table, based on the paired cache indexes, makes a one-to-one correspondence between the spatial positions and feature channels in X and B, and obtains the current inspection star features and the alignment baseline star features, both of which are of size H×W×C. The star-shaped response latch extracts the current inspection response and the alignment baseline response according to the same spatial position and the same feature channel, calculates the absolute amplitude difference between the current inspection response and the alignment baseline response to obtain the amplitude difference tensor A, and multiplies the current inspection response and the alignment baseline response element by element. When the product is not less than 0, the direction consistency mark is set to 1, and when the product is less than 0, the direction consistency mark is set to 0. The size of the amplitude difference tensor and the direction consistency mark tensor are both H×W×C. The stable response gate normalizes the amplitude difference tensor according to the maximum and minimum amplitudes of each feature channel to obtain the normalized amplitude difference tensor. Then, it multiplies the difference between the normalized amplitude difference and the gate slope by the difference between the amplitude difference threshold and the normalized amplitude difference, and then adds the product of the direction weight and the direction consistency label. The result is input into the Sigmoid gate function to obtain the baseline retention coefficient. The size of the baseline retention coefficient is H×W×C, and the value range is 0 to 1. The baseline residual stripper multiplies the baseline retention coefficient element-wise with the aligned baseline star feature to obtain the weighted baseline response, and generates the equipment state residual tensor according to the calculation rule of subtracting the weighted baseline response from the current inspection star feature. The size of the equipment state residual tensor is kept to be H×W×C. The residual feature output register performs mean and variance normalization on the C channels corresponding to each spatial location in the device state residual tensor, and restricts the normalized device state residual tensor to the interval from negative λ to λ. When the residual response is higher than λ, it is set to λ, and when the residual response is lower than negative λ, it is set to negative λ. The output device state residual feature has a size of H×W×C. The parameters of the baseline star product residual stripping block include feature height H, feature width W, number of feature channels C, amplitude difference threshold, orientation weight, sigmoid gate slope, and residual truncation threshold λ. Among them, the sigmoid gate slope and orientation weight are set as trainable parameters, the amplitude difference threshold is initialized according to the amplitude difference distribution between historical normal star features corresponding to the same device point and state unit, and the residual truncation threshold λ is configured according to the 99th percentile of the absolute value of the normalized residual after baseline subtraction of historical normal star features. The small target star response fidelity bypass receiver status residual characteristics are used to perform shallow buffering, position preservation, and channel backfeeding on the local star responses in the status display area, terminal connection area, and cable access area to generate small target fidelity status characteristics, including: Small target satellite response fidelity bypass includes: Shallow response cache queue: Local star response in the residual characteristics of cache device state; Small target location index table: records the spatial location of the status display area, terminal connection area, and cable access area; Local response pruner: Extracts local features of the corresponding region; Position anchor register: stores the original position coordinates of local features; Channel alignment mapper: performs channel mapping and response weighting on local features; Channel reflow accumulator: Writes back local features and superimposes them onto the corresponding spatial location; Fidelity feature output buffer: outputs fidelity state features of small targets; The process of handling the small target satellite response fidelity bypass is as follows: The shallow response buffer queue receives device state residual features of size H×W×C, where H is the feature height, W is the feature width, and C is the number of feature channels, and caches the corresponding features according to the device point number and state unit number. The small target location index table converts the horizontal and vertical coordinates of the status unit frame according to the ratio between the input image size and the device status residual feature size, respectively, to obtain the start and end row coordinates, start and end column coordinates and region number of the status display area, terminal connection area and cable access area in the H×W feature plane. Let the number of local regions be K. The local response cutter extracts K local features from the device state residual features according to the small target location index table, and resamples local features of different sizes to a uniform h×w×C, where h and w are the configuration height and configuration width of the bypass local features, respectively. The location anchor register synchronously stores the original spatial coordinates, resampling ratio, and region number of each local feature, and generates a location anchor sequence according to the local region cropping order; The channel alignment mapper performs channel mapping on each local feature through 1×1 convolution, converts the trainable channel weights into channel fidelity weights in the range of 0 to 1 through the Sigmoid function, and multiplies each channel response by the corresponding channel fidelity weight to obtain a weighted local feature of size h×w×C. The channel backfeed accumulator backsamples the weighted local features to the corresponding original region size according to the location anchoring sequence, and writes them back to the original spatial position of the equipment status residual features. When multiple local regions overlap, the backfeed features at the overlapping positions are summed and divided by the number of overlaps, and then superimposed with the equipment status residual features at the corresponding positions. The fidelity feature output buffer collects the device state residual features that have completed the local feature backfeeding and outputs small target fidelity state features with a size of H×W×C. The parameters of the small target satellite response fidelity bypass include the number of local regions K, the local feature height h, the local feature width w, the number of feature channels C, and the channel fidelity weight. Among them, K is determined according to the number of device state unit boxes, and h and w are configured according to the minimum effective feature size of various state unit regions. The size of the bypass input feature and the output feature are both H×W×C. The abnormal channel backtracking verification block receives the high-fidelity state features of small targets, filters high-response channels, and backtracks the high-response regions to the shallow structural features of the corresponding state units. Based on the co-occurrence relationship between the equipment structural boundary response and the state change response, false candidates are eliminated, generating valid state response features, where: The abnormal channel return verification block includes: High-response channel sorting queue: Candidate abnormal channels are sorted according to their response intensity; High-response region index table: records the state unit number and spatial location corresponding to the candidate abnormal channel; Shallow structure return buffer: Synchronously buffers the shallow structure characteristics of the corresponding state unit; Structural boundary response comparator: calculates the structural boundary response of equipment in shallow local structural features; State co-occurrence checker: calculates the co-occurrence relationship between structural boundary response and state change response; False candidate suppression gating: Removes false candidate responses based on co-occurrence verification results; Valid Response Output Register: Collects and outputs valid status response characteristics; The processing procedure for the abnormal channel return verification block is as follows: The high-response channel sorting queue receives the fidelity state features of small targets with size H×W×C. The mean and maximum absolute values ​​of the responses in each channel are calculated. The mean response is multiplied by the mean weight and the maximum response is multiplied by the maximum weight, and then added together to obtain the channel response score. The mean weight and the maximum weight are both set as trainable parameters, and their sum is 1. The top M channels are selected as candidate abnormal channels according to the channel response scores from high to low. The high response region index table normalizes the absolute value of the response within each channel according to the maximum and minimum response values ​​of each candidate abnormal channel. The positions whose normalization results are not lower than the region screening threshold are marked as high response positions. Adjacent high response positions are merged into connected regions, and the channel number, state unit number, and start and end row and column coordinates of each high response region are recorded. The shallow structure return buffer synchronously caches the shallow structure features formed by the corresponding state units at the output of the StarNet star-block deep convolution. Let the size of the shallow structure feature be Hs×Ws×Cs, where Cs is configured according to the number of channels at the output of the corresponding deep convolution. According to the size ratio between H and Hs, and W and Ws, the coordinates of the high response region are mapped to the shallow structure feature plane, and the corresponding shallow local structure features are clipped. The structural boundary response comparator calculates the response difference between adjacent positions in the horizontal and vertical directions of each channel of the shallow local structural feature. After taking the absolute value of the response difference in the two directions, the average value is calculated along Cs channels to obtain a two-dimensional structural boundary response map. Positions not lower than the boundary threshold are marked as structural boundary positions. The state co-occurrence verifier subtracts the starting row and starting column coordinates of the shallow local structural features from the mapped high response locations, converting them into local locations within the shallow local structural features. It then compares these locations with the structural boundary locations and divides the number of overlaps between the two types of locations by the number of high response locations to obtain the co-occurrence coefficient. The structural boundary location is extended into a neighborhood according to the boundary extension width to form a structural boundary neighborhood. The boundary extension ratio is obtained by dividing the number of local high response locations outside the structural boundary neighborhood by the total number of local high response locations. When the co-occurrence coefficient is not lower than the co-occurrence threshold and the boundary expansion ratio is not higher than the expansion threshold, the corresponding high response region is marked as a pseudo-candidate region; otherwise, the corresponding high response region is marked as an effective state change region. The pseudo-candidate suppression gate sets the gate value corresponding to the pseudo-candidate region to 0, sets the gate value corresponding to the effective state change region to 1, multiplies the candidate abnormal response in each high response region with the corresponding gate value element by element, and replaces the corresponding high response region in the original feature copy of the candidate abnormal channel. The response in the candidate abnormal channel that is not marked as a high response position remains unchanged. The effective response output register combines the M candidate abnormal channels that have completed region replacement with the original features that have not entered the candidate abnormal channels, according to the original channel order, and outputs an effective state response feature with a size of H×W×C. The parameters of the abnormal channel return verification block include the number of candidate abnormal channels M, mean weight, maximum weight, region screening threshold, boundary threshold, co-occurrence threshold, boundary expansion width, outward expansion threshold, and the number of shallow structure feature channels Cs. Among them, M is determined according to the configuration ratio of the number of feature channels C. The mean weight and maximum weight are set as trainable parameters and are normalized to keep their sum at 1. The region screening threshold is configured according to the normalized response distribution of historical normal small target fidelity state features. The boundary threshold is configured according to the boundary response distribution of historical normal shallow structure features. The boundary expansion width is configured as 1 to 3 feature positions according to the spatial resolution of the shallow structure feature plane. The co-occurrence threshold and outward expansion threshold are configured according to the overlap ratio of high response positions and structural boundary positions and the outward expansion ratio distribution in historical normal samples. Based on the valid state response characteristics, a device state candidate box, state category, state response value, and state unit number are generated and encapsulated into a device state candidate result, wherein: The candidate results for the packaging device status are as follows: The effective state response features are mapped to state categories to obtain the maximum category response value at each spatial location. The locations where the maximum category response value is not lower than the state candidate determination threshold are marked as candidate state locations. The state candidate determination threshold is configured according to the lower quartile of the category response value corresponding to the labeled state region in the training and validation samples. The positions of adjacent candidate states are merged according to the state category to obtain the device state candidate region. The coordinates of the candidate region are converted into device state candidate boxes according to the size ratio between the effective state response features and the current state segment. Calculate the mean response of each state category within the candidate region, determine the category with the largest mean response as the state category, and determine the corresponding mean response as the state response value; Compare the positions of the candidate device status box and the unit device status box, divide the candidate device status box by the area of ​​the candidate device status box to obtain the unit ownership ratio, and determine the device status with the largest unit ownership ratio as the status unit number by the number corresponding to the overlapping area unit box. The device status candidate box, status category, status response value and status unit number are associated and encapsulated to obtain the device status candidate result; This invention structurally optimizes the traditional StarNet network and proposes a feature modeling method for identifying the status of railway communication signal equipment rooms. To address the problem that fixed contours and stable texture responses account for a high proportion of current inspection features, a baseline star product residual stripping block is introduced. By aligning and matching the current inspected star features with historical normal baseline star features, the differential responses corresponding to changes in equipment status are highlighted. To address the issues of small scale in the status display area, terminal connection area, and cable access area, and the tendency for local responses to attenuate during deep transmission, a small target star response fidelity bypass is introduced. Through local feature caching, position preservation, and channel backfeeding, the continuous expression of small-scale status information is enhanced. To address the issue that abnormally high responses can easily form at the inherent structural boundaries of equipment, an abnormal channel back-off verification block is introduced. This block back-offs the high-response region to shallow structural features for boundary co-occurrence verification, eliminating pseudo-candidate responses that have not formed independent state changes. The improved StarNet network not only retains the original lightweight star feature modeling capability, but also realizes the collaborative fusion between the current inspection response, historical normal baseline, local response of small targets and shallow structure information of equipment, thereby improving the accuracy and robustness of equipment status identification and candidate positioning in railway communication signal equipment rooms. The improved StarNet network was trained using manually labeled communication signal device state boxes, state categories, and state cell numbers as supervision labels. A joint loss was constructed, comprising device state classification error, candidate box localization error, baseline residual constraint error, small target star response preservation error, and anomaly backtracking verification error, to update the network parameters. The improved StarNet network was considered complete after training when the average localization deviation of the device state candidate boxes in the validation set was less than 2 pixels and the state category recognition accuracy reached 98.0%. The improved StarNet network is trained as follows: The equipment status recognition samples were divided into training and validation sets in an 8:2 ratio. Manually labeled communication signal equipment status boxes, status categories, and status unit numbers were used as supervision labels. The current status segment, baseline status segment, and equipment status unit box were input into the improved StarNet network for each batch of 16 samples. The network outputs equipment status candidate boxes, status category probabilities, equipment status residual features, small target fidelity status features, and abnormal return gating results. The cross-entropy between the predicted state category probability and the labeled state category is calculated to obtain the equipment state classification error. The smoothing absolute value error between the equipment state candidate box and the labeled state box is calculated to obtain the candidate box localization error. The mean value of the absolute value of the equipment state residual response outside the labeled state box is taken to obtain the baseline residual constraint error. The attenuation difference of the local response before and after bypass processing within the labeled state box is calculated to obtain the small target star response preservation error. The back-gating marker is generated based on the overlap relationship between the high response region and the labeled state box, and the binary cross-entropy between the back-gating marker and the abnormal back-gating result is calculated to obtain the abnormal back-gating verification error. The device state classification error, candidate box localization error, baseline residual constraint error, small target star response preservation error, and abnormal return verification error are multiplied by 0.30, 0.25, 0.15, 0.15, and 0.15 respectively, and then summed to obtain the joint loss value. The improved StarNet network parameters are updated in reverse at a learning rate of 0.001. After each training round, the average absolute difference between the boundary coordinates of the device state candidate box and the corresponding coordinates of the labeled state box is calculated on the validation set, and the state category recognition accuracy is calculated. When the average localization deviation is less than 2 pixels and the state category recognition accuracy reaches 98.0%, the improved StarNet network is considered to have completed training.

[0028] In this embodiment, the generation of the device state tracking chain includes: The candidate device status results are sorted by status unit number, device location, and acquisition time, and the candidate boxes are divided into high-confidence candidate boxes and low-confidence candidate boxes, where: The distinction between high-confidence and low-confidence candidate boxes is as follows: The candidate results of equipment status are merged according to the equipment location and status unit number, and arranged from first to last according to the collection time. The candidate boxes corresponding to the same equipment location and status unit are labeled with consecutive frame numbers. Extract the coordinates, state category, state response value, and state unit number of each candidate box; Candidate boxes with state response values ​​not lower than the high confidence threshold are classified as high confidence candidate boxes; candidate boxes with state response values ​​lower than the high confidence threshold but not lower than the low confidence threshold are classified as low confidence candidate boxes; and candidate boxes with state response values ​​lower than the low confidence threshold are removed from the current frame tracking input. The high confidence threshold is set to 0.6, and the low confidence threshold is set to 0.1. The high confidence candidate box sequence and the low confidence candidate box sequence are encapsulated according to the frame number. The ByteTrack algorithm is used to perform the first association between high-confidence candidate boxes and the main trajectory of the device status, generating the main trajectory matching results and unmatched candidate boxes, where: Generate the main trajectory matching results and unmatched candidate boxes, specifically as follows: High-confidence candidate boxes are extracted frame by frame according to the device location and acquisition time, and the main trajectory of the device in the tracking state at the corresponding device location in the previous frame is obtained. When there is no device status main trajectory in the current frame, the device status main trajectory is established with the high confidence candidate box of the current frame, the established device status main trajectory is written into the main trajectory matching result, and the first association of the current frame ends. When a device state main trajectory exists in the current frame, the Kalman filter in the ByteTrack algorithm is used to predict the trajectory region of the current frame based on the historical candidate box positions of each device state main trajectory, and the intersection-union ratio between the trajectory prediction region and the high-confidence candidate box is calculated. When the state unit numbers are inconsistent, the corresponding device state main trajectory and high confidence candidate box are marked as not associated. When the state unit numbers are consistent, the association cost is obtained by subtracting the intersection-union ratio from 1. The Hungarian algorithm is used for the first matching based on the association cost. The trajectory-candidate box combination with an intersection-union ratio of not less than 0.3 is determined as a valid match. The coordinates, state category, state response value and acquisition time of the matching candidate box are used to update the corresponding device state main trajectory. The obtained matching device status main trajectory and corresponding high confidence candidate box are encapsulated as the main trajectory matching result. The device status main trajectory that does not obtain a match is marked, and the high confidence candidate box that does not form a valid match with the device status main trajectory is retained as an unmatched candidate box. The corresponding device location, status unit number, candidate box coordinates, status category and status response value are recorded. When the same unmatched candidate box simultaneously covers two or more device state main trajectory prediction regions, the unmatched candidate box is marked as a stuck candidate box. Based on the trajectory center arrangement direction of the previous frame, the state unit position, and the response peak inside the candidate box, a stuck candidate box splitting and reconnection process is performed to generate a dense state unit trajectory recovery result, where: Perform the process of splitting and reconnecting the stuck candidate boxes, specifically as follows: Determine whether the center point of the main trajectory prediction area of ​​each device status is located inside the unmatched candidate box. When the same unmatched candidate box contains the center points of two or more main trajectory prediction areas of the device status, mark the unmatched candidate box as an adhered candidate box and record the corresponding device status main trajectory number. Extract the trajectory center of the main trajectory of the corresponding device status in the previous frame, select the two trajectory centers with the largest Euclidean distance from each trajectory center, determine the direction of the line connecting the two trajectory centers as the trajectory center arrangement direction, and project the remaining trajectory centers onto the line connecting them, and sort the trajectory centers according to the projection position. Map the corresponding device status unit box to the inside of the adhesion candidate box, and extract the local response peak from the effective status response features corresponding to the adhesion candidate box. According to the arrangement order of the trajectory center and the position of the device status unit, assign each local response peak to the corresponding device status main trajectory. The response changes between adjacent local response peaks are statistically analyzed along the trajectory center arrangement direction. The lowest response position between adjacent local response peaks is determined as the response valley position. A split line perpendicular to the trajectory center arrangement direction is generated through the response valley position. The adhesion candidate box is divided into sub-candidate boxes with the same number of main trajectories of the coverage device status according to each split line. The status category, status response value and status unit number of each sub-candidate box are determined respectively. Each sub-candidate box is matched with the corresponding device status main trajectory prediction area. The candidate box coordinates, status category, status response value and acquisition time of the corresponding device status main trajectory are updated using the matched sub-candidate boxes. The results are then encapsulated to generate dense state unit trajectory recovery results. For unmatched high-confidence candidate boxes that are not marked as adhered candidate boxes, new equipment status trajectories are established according to the equipment location and status unit number, and the new equipment status trajectories are written into the main trajectory matching results; A bifurcation trajectory pool is established for low-confidence candidate boxes that do not match the main trajectory. Low-confidence bifurcation trajectory merging is then performed based on continuous frame position offset, state category consistency, and response value change trends to generate low-confidence bifurcation trajectory processing results, where: Perform low-confidence fork trajectory merging processing, specifically as follows: Based on the equipment location and status unit number, low-confidence candidate boxes that overlap with the predicted area of ​​the main trajectory of the unmatched equipment status are selected, and a bifurcation trajectory pool is established with the corresponding main trajectory number of the unmatched equipment status. Based on the low-confidence candidate boxes in the bifurcation trajectory pool associated with the acquisition time, a position offset vector is generated according to the horizontal and vertical offsets of the center of the candidate boxes in adjacent frames. When the inner product of two position offset vectors formed by three consecutive frames is not less than 0, and the absolute values ​​of the horizontal offset and the vertical offset do not exceed the width and height of the corresponding trajectory prediction area, the corresponding low-confidence candidate boxes are connected to form a bifurcation trajectory position sequence. Extract the state category of the last 3 consecutive frames of the bifurcation trajectory position sequence. When the state category of at least 2 frames is consistent with the main trajectory of the unmatched device, generate a state category consistency marker. The response change trend of increasing, decreasing or stabilizing is determined based on the difference between the adjacent state response values ​​of the three consecutive frames at the end of the bifurcation trajectory and the main trajectory of the unmatched device state. When the bifurcation trajectory meets the conditions of continuous position offset, has a consistent status category marker, and the response change trend is consistent with the main trajectory of the unmatched device status, the bifurcation trajectory is merged back to the main trajectory of the corresponding device status, and the candidate box coordinates, status category, status response value and acquisition time in the bifurcation trajectory are added. The main trajectory of the device status that has been merged, the low-confidence fork trajectories retained in the fork trajectory pool, and the low-confidence candidate boxes that have not formed fork trajectories are classified and encapsulated to generate the low-confidence fork trajectory processing result. The main trajectory matching results, dense state unit trajectory recovery results, and low-confidence bifurcation trajectory processing results are encapsulated according to the device location and state unit number to generate a device status tracking chain.

[0029] In this embodiment, generating the embodied intelligent inspection result includes: Using the equipment status tracking chain as the discrimination object, the continuity verification, status consistency verification, and trajectory closure discrimination are performed on the trajectory of the status unit corresponding to the equipment point, generating abnormal trajectory discrimination results, among which: The abnormal trajectory discrimination results are generated as follows: The equipment status tracking chain is merged according to the equipment location and status unit number, and the candidate boxes, status categories and status response values ​​of each frame are arranged according to the acquisition time to form a status unit trajectory sequence. Compare the frame sequence number and candidate box center position of adjacent trajectory records, connect trajectory records with consecutive frame sequence and candidate box center located in the corresponding state unit box into continuous trajectory segments, otherwise segment the trajectory segments. Within each continuous trajectory segment, continuous trajectory records with the same state category are merged according to the acquisition time to form a state-consistent trajectory segment, and the state category and number of consecutive frames of each state-consistent trajectory segment are recorded. According to the acquisition time, the trajectory segments with the same state in adjacent states are compared. When the state category of the trajectory segment with the same state in the next state is different from that of the trajectory segment with the same state in the previous state, and the number of consecutive frames is not less than 3 frames, the trajectory segment with the same state in the next state is marked as an abnormal trajectory, and the acquisition time of the first record of the next state category is determined as the abnormal start time. When the abnormal trajectory recovers to the state category before the change and remains there for 2 consecutive frames, the abnormal trajectory is marked as a closed abnormal trajectory, and the acquisition time of the last record before recovery is determined as the abnormal termination time. When an abnormal trajectory is interrupted or the device status tracking chain ends and is not restored, the abnormal trajectory is marked as an unclosed abnormal trajectory, and the acquisition time of the last record is determined as the abnormal termination time. The device location, status unit number, status category before and after the change, abnormal start and end time, number of consecutive frames, average status response value and trajectory closure mark are associated and encapsulated to generate abnormal trajectory discrimination results. The abnormal trajectory identification results are fed back to the data center's specific inspection task map. An association mapping is performed on the abnormal equipment locations, status unit numbers, inspection task numbers, and review tasks to generate specific inspection task feedback results, including: The generated embodied inspection task feedback results are as follows: Based on the equipment location and status unit number, locate the corresponding inspection task node in the computer room inspection task map, and extract the inspection task number, inspection pose and data collection type. Write the state category before and after the change, the start and end time of the abnormality, the average state response value and the trajectory closure mark in the abnormal trajectory discrimination results into the corresponding inspection task node to establish the association between the abnormal trajectory and the original inspection task. Based on the inspection pose and acquisition type of the original inspection task, a review task is generated. The original inspection task number, abnormal equipment location, status unit number, review pose, review acquisition type and execution order are recorded. The review tasks corresponding to the unclosed abnormal trajectory are arranged before the review tasks corresponding to the closed abnormal trajectory. The inspection task number, abnormal equipment location, status unit number, abnormal trajectory information and review task are associated and encapsulated to generate a specific inspection task feedback result. The results of the embodied inspection task are packaged according to the equipment location number to generate embodied intelligent inspection results.

[0030] refer to Figure 3 An embodied intelligent inspection system for railway communication signal rooms includes the following modules: The inspection data standardization module is used to collect and preprocess railway communication signal equipment room inspection data to generate a standardized equipment room inspection dataset. The embodied task map construction module is used to perform inspection reachability mapping of equipment locations and generate an embodied inspection task map of the computer room. The robot inspection execution module is used to control the inspection robot to perform specific inspection tasks and generate equipment inspection samples; The status recognition sample generation module is used to perform perspective snapping and status unit registration on equipment inspection samples to generate equipment status recognition samples. The device state candidate generation module is used to identify the state of device state recognition samples by improving the StarNet network and generate device state candidate results. The device status tracking module is used to perform cross-frame correlation of device status candidate results using the ByteTrack algorithm to generate a device status tracking chain. The inspection result feedback module is used to identify abnormal trajectory closure and feedback the task map of the equipment status tracking chain, and generate personalized intelligent inspection results.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to an inspection cycle of a railway communication and signal equipment room. The system received inspection data collected by an inspection robot. The equipment room contained 16 rows of equipment cabinets, including communication equipment cabinets, signal equipment cabinets, power supply equipment cabinets, and wiring equipment cabinets, totaling 124 equipment points and 686 status units. The status units were mainly distributed in the status display area, terminal connection area, patch panel port area, and cable access area. The inspection robot was equipped with a mobile chassis, a rotatable pan-tilt head, a visible light camera, an infrared thermal imaging device, a sound pickup device, and environmental sensors. During one inspection cycle, a total of 1860 visible light images, 620 infrared thermal images, 38 equipment status videos, 38 equipment audio recordings, 4120 robot pose records, 124 inspection task records, and 1860 environmental monitoring records were collected. The visible light image resolution is 1920×1080 pixels, the infrared thermal image resolution is 640×512 pixels, and the device status video sampling rate is 25 frames per second. 7.8% of the raw images exhibit image jitter, 5.2% of the infrared thermal images show field of view shift, and 3.6% of the audio clips contain chassis start-stop noise.

[0032] The system standardizes fields in data such as computer room layout, communication signal equipment location logs, inspection task data, and historical inspection handling data, converting equipment cabinet numbers, equipment location numbers, status unit numbers, and inspection task numbers into unified codes. Robot pose, visible light images, infrared thermal images, video, sound, and environmental monitoring data are calibrated to the same timeline. After distortion correction and image stabilization, the average edge offset of visible light images is reduced from 11.6 pixels to 2.8 pixels; after temperature calibration and field-of-view registration, the average overlap error between infrared thermal images and visible light images is reduced from 8.4 pixels to 2.1 pixels; sound data is framed in 0.5-second intervals, and 1468 valid sound frames are retained after removing 52 chassis start / stop noise frames. Multimodal inspection records are linked with equipment locations, inspection task numbers, and robot poses to form 124 equipment location inspection records, generating a standardized computer room inspection dataset.

[0033] The system divides the room into 96 spatial nodes and 138 walkable edges based on the server room layout and robot access constraints, and attaches 124 device points to the corresponding rack inspection surfaces. A total of 137 candidate docking poses are generated around the device points. Eight poses are eliminated due to insufficient aisle width, and five poses are eliminated due to rack door obstruction or field-of-view deviation, ultimately retaining 124 reachable device points. The system arranges these reachable device points into four inspection sub-paths according to the inspection task, each containing 29 to 33 device points, with a total path length of 286.4 meters. Spatial nodes, device points, inspection docking poses, acquisition viewpoint anchor points, and inspection tasks are encapsulated into a embodied server room inspection task map.

[0034] The inspection robot proceeded to a signal equipment cabinet location according to the task map. The target's stopping coordinates were 8.42 meters horizontally and 3.16 meters vertically, with a target body orientation angle of 89.0 degrees and a gimbal pitch angle of 6.5 degrees. Upon initial arrival, the robot's actual coordinates were 8.31 meters horizontally and 3.06 meters vertically, with a positional deviation of 0.148 meters, a body orientation deviation of 3.6 degrees, and a 34-pixel offset between the center of the status display area and the acquisition viewpoint anchor point. The system performed three rounds of closed-loop correction based on the inspection posture deviation. After the first round of correction, the positional deviation was reduced to 0.083 meters, the second to 0.041 meters, and the third to 0.026 meters. Ultimately, the orientation deviation was reduced to 0.8 degrees, and the acquisition viewpoint offset was reduced to 7 pixels. The robot then acquired 15 visible light images, 5 infrared thermal images, an 8-second video clip, an 8-second audio clip, and 80 sets of environmental monitoring data, forming an inspection segment for that equipment location.

[0035] The system overlays equipment inspection segments onto the coordinates of the equipment inspection surface and divides the status display area, terminal connection area, and cable access area into 28 status unit frames. After perspective correction, the tilt angle of the equipment cabinet edge is reduced from 4.7 degrees to 0.6 degrees; after scale normalization, the status unit frames are uniformly 96×96 pixels; after time alignment, the maximum time deviation between visible light frames, infrared frames, and video frames is reduced from 124 milliseconds to 18 milliseconds. After point overlaying the current status segment with historical normal inspection segments, the edge overlap rate increases from 82.6% to 96.8%. The normal average brightness of one indicator light status unit is 184, while the current average brightness of consecutive frames is 78; the normal edge spacing of one terminal connection status unit is 12 pixels, while the current edge spacing has increased to 19 pixels; the contour offset of one cable access status unit is 3 pixels, which does not reach the abnormal judgment range. Based on this, the system generates status baseline pairing records and equipment status identification samples.

[0036] This embodiment constructs 12,600 training samples, including 6,840 normal state samples, 1,940 indicator light anomaly samples, 1,460 terminal connection anomaly samples, 980 cable connection anomaly samples, 720 local temperature rise anomaly samples, and 660 interference samples involving reflection, shadow, occlusion, and repetitive textures. Manual annotations include device status bounding boxes, status categories, and status unit numbers. The width of the status bounding boxes is mainly distributed between 12 and 56 pixels, with small target samples (width not exceeding 24 pixels) accounting for 41.3%. After training, the average positioning deviation of the device status candidate bounding boxes in the validation set is 1.7 pixels, and the status category recognition accuracy is 98.3%.

[0037] After the equipment status identification samples enter the improved StarNet network, the current status segment and the baseline status segment are respectively encoded with star operation. The baseline star product residual stripping block reduces the average response of the cabinet edge, nameplate frame, heat dissipation hole, and fixed terminal structure from 0.73 to 0.24, while keeping the response of the abnormal indicator area at 0.81 and the response of the terminal offset area at 0.77. The small target star response fidelity bypass buffers and re-feeds shallow local high responses, improving the average response of small targets with a width of less than 24 pixels from 0.46 to 0.79, and improving the boundary integrity of the abnormal indicator from 74.2% to 94.6%. The abnormal channel back-check block performs back-check on 7 high response candidates, eliminating 3 reflective pseudo-candidates and 2 duplicate texture pseudo-candidates, and retaining 1 candidate for abnormal indicator brightness and 1 candidate for abnormal terminal connection. The center positioning deviations of the two candidate boxes are 1.5 pixels and 1.9 pixels, respectively, and the status response values ​​are 0.93 and 0.88, respectively.

[0038] After the device status candidate results enter the ByteTrack algorithm, in frame 47, two adjacent indicator lights are merged into a single candidate box with a width of 58 pixels due to overexposure. Based on the trajectory center alignment direction of the previous frame, the position of the status unit, and the two response peaks within the candidate box, the system splits the candidate box into two sub-candidate boxes and reconnects them to the original main trajectory. From frames 63 to 65, the candidate boxes for abnormal indicator lights are affected by local occlusion, with confidence levels decreasing to 0.34, 0.29, and 0.36, respectively. The system writes the three low-confidence candidate boxes into the bifurcation trajectory pool. The maximum center position offset between the three frames is 3.8 pixels, and the status category remains consistent. In frame 66, the status response value recovers to 0.82, and the bifurcation trajectory is merged back into the main trajectory for abnormal indicator lights. Ultimately, the abnormal indicator light trajectory continuously covers 186 frames, and the abnormal terminal connection trajectory continuously covers 172 frames, with no trajectory number switching occurring between the two trajectories.

[0039] The system performs continuity checks, status consistency checks, and trajectory closure determination on the equipment status tracking chain. The continuity rate for abnormal indicator light trajectories is 98.9%, and the status category consistency rate is 100%; the continuity rate for abnormal terminal connection trajectories is 97.4%, and the status category consistency rate is 99.4%. Both trajectories are determined to be closed abnormal trajectories and are reconnected to the specific inspection task map in the machine room. The robot collects data again according to the generated review task. The average brightness of the indicator lights in the review segment is 79, and the terminal edge spacing is 18 pixels, with differences of 4.8% and 5.3% respectively compared to the initial inspection results. The system maintains the abnormality determination.

[0040] In 2800 independent test samples: The accuracy rate of device location association using traditional methods is 91.8%, while that of this invention is 98.7%. The accuracy of status recognition improved from 87.1% to 97.6%; The recall rate for small targets increased from 79.4% to 95.1%; The false detection rate for reflective and repetitive textures decreased from 7.6% to 1.9%; The accuracy rate of recovering stuck candidate boxes improved from 71.8% to 94.2%; The low-confidence trajectory recovery rate increased from 62.7% to 91.3%; The trajectory breakage rate decreased from 10.8% to 2.5%.

[0041] When completing the inspection of 124 equipment points: The traditional method takes an average of 23 minutes and 6 seconds, while the present invention takes an average of 19 minutes and 38 seconds. The average time from anomaly detection to task generation has been reduced from 10 minutes and 54 seconds to 4 minutes and 21 seconds.

[0042] Example 1 shows that the present invention can reduce the docking deviation of inspection robots, the offset of the acquisition viewpoint, the interference of stable structures, the loss of small target response, the adhesion of candidate boxes and the breakage of low confidence trajectories, and improve the accuracy and stability of equipment status identification, anomaly tracking and task reconnection in railway communication signal equipment rooms.

[0043] 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 method for physical intelligent inspection of a railway communication signal room, characterized in that, include: Collect inspection data of railway communication signal equipment rooms, preprocess the inspection data of railway communication signal equipment rooms, and generate a standardized equipment room inspection dataset; Based on the standardized data center inspection dataset, the data center spatial topology is extracted, the equipment locations are mapped to reachability during inspection, and a data center-specific inspection task map is generated. According to the specific inspection task diagram of the computer room, the inspection robot performs inspection posture closure control and equipment status acquisition to generate equipment inspection samples. Based on the robot pose, acquisition angle and equipment location in the equipment inspection sample, viewpoint snapping and state unit registration are performed to generate equipment state recognition sample; An improved StarNet network is constructed to reconstruct star operation features on device state recognition samples, snapping together stable structural responses, retaining local star responses, and screening high-response candidates to generate device state candidate results. The ByteTrack algorithm is used to correlate device state candidate results across frames, recover dense state unit trajectories, and perform bifurcation trajectory merging on low-confidence candidate boxes that do not match the main trajectory to generate a device state tracking chain. Using the equipment status tracking chain as the discrimination object, abnormal trajectory closure discrimination is performed, and the closure discrimination result is fed back to the computer room specific inspection task map to generate specific intelligent inspection results.

2. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The railway communication and signal equipment room inspection data specifically includes equipment room spatial layout data, communication and signal equipment location ledger data, inspection robot pose data, inspection task data, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, equipment room environmental monitoring data, and historical inspection and handling data.

3. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated standardized data center inspection dataset includes: The fields and numbering of the equipment room spatial layout data, communication and signal equipment location ledger data, inspection task data and historical inspection and handling data in the railway communication and signal equipment room inspection data are unified and numbered to form a basic index table for the equipment room. The robot pose data, visible light inspection images, infrared thermal imaging data, equipment status video data, equipment sound data, and computer room environment monitoring data are timestamped and calibrated, and an inspection segment index table is formed according to the inspection task number. Distortion correction and image stabilization are performed on visible light inspection images, temperature calibration and field of view registration are performed on infrared thermal imaging data, and frame segmentation and noise clip removal are performed on equipment sound data to form a multimodal inspection record. The multimodal inspection records are linked to the computer room basic index table according to the equipment location, inspection task number and robot pose, and the collection view and inspection segment number are marked to form the equipment location inspection record. The equipment inspection records are supplemented with missing items, abnormal formats are removed, and data formats are standardized to generate a standardized computer room inspection dataset.

4. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated data center inspection task map includes: Based on the computer room spatial layout data and inspection robot access constraints in the standardized computer room inspection dataset, the computer room passage, equipment cabinets and obstacle areas are divided into topology nodes to generate a computer room spatial topology table. Link the communication signal equipment location ledger data with the computer room space topology table, mark the cabinet, inspection surface and status unit range of the equipment location, and generate an equipment location mapping table; Based on the computer room space topology table and the equipment location mapping table, inspection docking poses and acquisition viewpoint anchor points are generated for the equipment locations, and unreachable locations are eliminated to generate an equipment location reachability mapping table. Based on the inspection task data, the reachable points in the equipment location reachable mapping table are assigned tasks, the inspection sequence, data collection type and review trigger conditions are marked, and a specific inspection task node table is generated. The data center space topology table, equipment location mapping table, equipment location reachability mapping table, and embodied inspection task node table are encapsulated into a graph structure to generate a data center embodied inspection task graph.

5. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated equipment inspection sample includes: According to the inspection sequence in the data center inspection task diagram, the tasks of reachable equipment points, inspection docking positions, data collection view anchor points and data collection types are arranged to generate the data center inspection execution sequence. The inspection robot is driven to move along the topology of the computer room space according to the computer room inspection execution sequence, and the pose deviation between the robot's current pose and the inspection docking pose is calculated to obtain the inspection pose deviation. The chassis position, body orientation, and acquisition gimbal angle of the inspection robot are closed-loop corrected based on the inspection posture deviation to obtain the equipment alignment inspection status. In the equipment alignment and inspection state, visible light, infrared, video, sound and environmental status are collected at the communication signal equipment locations according to the collection angle anchor point to generate equipment inspection segments; The equipment inspection segments are packaged with equipment locations, robot poses, acquisition perspectives, inspection task numbers, and acquisition times to generate equipment inspection samples.

6. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated device status identification sample includes: Based on the equipment locations, robot poses, and acquisition angles in the equipment inspection samples, the equipment inspection segments are matched to the inspection surface coordinates of the corresponding communication signal equipment to generate equipment angle matching records. Based on the equipment's perspective and the recording, the status display area, terminal connection area, and cable access area within the inspection surface are divided into status units to generate equipment status unit frames. Perform perspective correction, scale normalization, and acquisition time alignment on the multimodal inspection segments corresponding to the equipment status unit frames to generate status unit registration segments. The state unit registration segment is matched with the historical normal inspection segment of the same equipment location to generate a state baseline pairing record. The status baseline pairing record is encapsulated with the equipment location, status unit number and inspection task number to generate equipment status identification samples.

7. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated device status candidate results include: An improved StarNet network was constructed, which included a baseline star accumulation residual stripping block, a small target star response fidelity bypass, and an abnormal channel return verification block. The current state segment and baseline state segment in the equipment status identification sample are sent into the StarNet star operation encoding process. The corresponding position and corresponding channel features in the same state unit are multiplied and interacted to generate the current inspection star features and baseline star features. The baseline star-product residual stripping block receives the current inspection star characteristics and the baseline star characteristics, performs star-product coupling on the current inspection response and the normal baseline response at the same equipment location, strips the stable structural response, and generates equipment state residual characteristics. The small target star response fidelity bypass receiver status residual characteristics are used to perform shallow buffering, position preservation and channel backfeeding on the local star response of the status display area, terminal connection area and cable access area to generate small target fidelity status characteristics. The abnormal channel back-back verification block receives the small target fidelity state features, filters high-response channels and back-backs the high-response region to the shallow structural features of the corresponding state unit, eliminates false candidates based on the co-occurrence relationship between the equipment structure boundary response and the state change response, and generates valid state response features. Based on the valid status response characteristics, generate device status candidate boxes, status categories, status response values, and status unit numbers, and encapsulate them into device status candidate results; The improved StarNet network was trained using manually labeled communication signal device state boxes, state categories, and state unit numbers as supervision labels. The network parameters were updated by using a joint loss consisting of device state classification error, candidate box localization error, baseline residual constraint error, small target star response preservation error, and abnormal backtracking verification error. When the average localization deviation of device state candidate boxes in the validation set was less than 2 pixels and the state category recognition accuracy reached 98.0%, the improved StarNet network was considered to have completed training.

8. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated device state tracking chain includes: The candidate results of device status are sorted into frames according to the status unit number, device location and acquisition time, and the candidate boxes are divided into high confidence candidate boxes and low confidence candidate boxes. The ByteTrack algorithm is used to associate high-confidence candidate boxes with the main trajectory of device status for the first time, generating main trajectory matching results and unmatched candidate boxes; When the same unmatched candidate box covers more than two device status main trajectory prediction areas at the same time, the unmatched candidate box is marked as a stuck candidate box, and the stuck candidate box splitting and reconnection process is performed according to the trajectory center arrangement direction of the previous frame, the position of the status unit and the response peak inside the candidate box to generate the dense status unit trajectory recovery result. A bifurcation trajectory pool is established for low-confidence candidate boxes that do not match the main trajectory. Low-confidence bifurcation trajectory merging is performed according to the position offset of consecutive frames, consistency of state category, and trend of response value change to generate low-confidence bifurcation trajectory processing results. The main trajectory matching results, dense state unit trajectory recovery results, and low-confidence bifurcation trajectory processing results are encapsulated according to the device location and state unit number to generate a device status tracking chain.

9. The embodied intelligent inspection method for railway communication signal equipment rooms according to claim 1, characterized in that, The generated embodied intelligent inspection results include: Using the equipment status tracking chain as the discrimination object, the continuity verification, status consistency verification and trajectory closure discrimination are performed on the status unit trajectory corresponding to the equipment point, and abnormal trajectory discrimination results are generated. The abnormal trajectory identification results are fed back to the computer room's specific inspection task map. The abnormal equipment locations, status unit numbers, inspection task numbers, and review tasks are associated and mapped to generate specific inspection task feedback results. The results of the embodied inspection task are packaged according to the equipment location number to generate embodied intelligent inspection results.

10. A personalized intelligent inspection system for railway communication signal equipment rooms, comprising the personalized intelligent inspection method for railway communication signal equipment rooms as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The inspection data standardization module is used to collect and preprocess railway communication signal equipment room inspection data to generate a standardized equipment room inspection dataset. The embodied task map construction module is used to perform inspection reachability mapping of equipment locations and generate an embodied inspection task map of the computer room. The robot inspection execution module is used to control the inspection robot to perform specific inspection tasks and generate equipment inspection samples; The status recognition sample generation module is used to perform perspective snapping and status unit registration on equipment inspection samples to generate equipment status recognition samples. The device state candidate generation module is used to identify the state of device state recognition samples by improving the StarNet network and generate device state candidate results. The device status tracking module is used to perform cross-frame correlation of device status candidate results using the ByteTrack algorithm to generate a device status tracking chain. The inspection result feedback module is used to identify abnormal trajectory closure and feedback the task map of the equipment status tracking chain, and generate personalized intelligent inspection results.

Images