Rail transit station intelligent inspection system

Abstract

The application discloses a kind of rail transit station intelligent inspection systems, it is related to rail transit intelligent inspection technical field, comprising: data acquisition layer, integration laser radar, visible light / infrared camera and environmental sensor, and realize multi-source data space-time alignment by pre-processing;Map construction and positioning module, three-dimensional point cloud and two-dimensional grid map are constructed by laser inertial navigation fusion SLAM, combined iterative closest point registration realizes centimeter-level positioning, and adopts layered path planning autonomous navigation;Target detection and state identification module, YOLOv8 and Vision Transformer cascade model are used to detect and state fine discrimination for door equipment, while identifying leftovers;Abnormal analysis processing module, based on the correlation analysis of multi-modal abnormal vector on the architecture of Transformer, the optimal disposal suggestion is generated after compliance verification and the evolution of model is supported.The application realizes station environment high-precision perception, equipment state accurate identification and abnormal event closed-loop disposal, significantly improves the inspection efficiency and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for rail transit, and more specifically, to an intelligent inspection system for rail transit stations. Background Technology

[0002] As the backbone of urban public transportation, urban rail transit is characterized by short headways, numerous lines and stations, and high passenger volume, making operational safety a top priority. Over long-term operation, factors such as geological deformation, vehicle impact and vibration, and human construction inevitably cause wear and tear on lines, stations, and train facilities. Therefore, conducting intelligent inspections and anomaly analysis of urban rail transit is of great significance for improving transportation efficiency and ensuring the quality of passenger travel.

[0003] Currently, the inspection of rail transit facilities is still mainly manual. Since the daily maintenance window for station equipment and facilities is typically only a few hours from midnight to 4 AM, manual inspection is time-consuming, inefficient, slow, has many blind spots, and suffers from significant differences in visual standards, easily leading to a backlog of station-related tasks. Furthermore, prolonged nighttime operations are prone to inspection errors, making it difficult to meet the rapid inspection needs of the entire line's infrastructure. At the management level, although some inspection anomaly databases exist, inconsistent adherence to standards by inspection personnel and reliance on individual experience make it difficult to establish standardized inspection operation procedures. At the data level, while anomaly data can be collected, the lack of a stable fusion and analysis process further increases the learning costs for inspection personnel.

[0004] In recent years, there have been some explorations into power scene inspection methods based on robotic dogs. For example, patent CN121043971A proposes using a quadruped robotic dog integrating radar and cameras for substation inspection; CN120996081A uses a large visual language model to realize substation environment updates and intelligent inspection; and CN120975757A completes power equipment inspection through multimodal data acquisition and intelligent analysis. However, these achievements are mainly geared towards the power industry and have not yet provided an inspection equipment that can take into account the diverse facility and service needs of rail transit stations, and there is a lack of real-time analysis and closed-loop handling systems for various abnormal states.

[0005] The closest prior art, CN121764108A, discloses a rail transit inspection device and method, which mainly acquires images of the track and surrounding environment through an image acquisition module and uses a target detection algorithm to identify anomalies. This solution achieves automated inspection to a certain extent, but it has the following shortcomings: First, relying solely on visual perception makes it difficult to achieve stable and reliable spatial positioning and equipment status identification in station environments with column occlusion and variable lighting. Second, it does not involve deep fusion and data preprocessing of multi-source heterogeneous sensors (such as lidar and gas sensors), making it unable to achieve three-dimensional perception of the spatial structure and environmental safety status. Third, it lacks a refined discrimination model for the status of key equipment such as vehicle doors and platform doors, easily leading to false alarms and missed alarms. Fourth, anomaly information only remains at the alarm level, lacking the ability to automatically generate anomaly correlation analysis and compliance handling suggestions based on large models, thus failing to form a closed-loop management from "discovery" to "handling."

[0006] Therefore, there is an urgent need to build an intelligent inspection system for rail transit stations, which integrates multi-source heterogeneous data to achieve high-precision map construction and positioning, accurate identification of equipment status, and intelligent analysis and standardized handling of abnormal events, providing scientific and quantitative decision support tools for the operation and management of urban rail transit. Summary of the Invention

[0007] The purpose of this invention is to solve the problems mentioned in the background art, and to propose an intelligent inspection system for rail transit stations.

[0008] The technical solution adopted by this invention to solve its technical problem is: A smart inspection system for rail transit stations includes: The data acquisition layer, located on the inspection robot dog, includes lidar, visible light camera, infrared camera, and environmental sensors; The map building and positioning module is configured to build a three-dimensional point cloud map and a two-dimensional grid map of the station based on the fusion SLAM of the lidar and the inertial measurement unit, and to match the real-time collected lidar point cloud with the three-dimensional point cloud map for positioning correction by using the point cloud registration method. The target detection and state recognition module is configured to detect and locate door-type devices from the video image data, and to perform state judgment on the detected door-type device areas to obtain the state information of the door-type devices; the door-type devices include vehicle doors, platform doors, and equipment doors; The anomaly analysis and handling module is configured as follows: Acquire and structure the anomaly description data, which includes at least visual anomaly features, environmental anomaly features, and spatial location information; The correlation analysis unit learns the correlation relationships between multiple anomaly description data; Based on the results of relational learning, a decision generation unit generates suggestions for handling abnormal events. The proposed actions are then verified against operational standards by the compliance verification unit before being output. The model evolution unit continuously optimizes the correlation analysis unit and the decision generation unit based on historical anomaly description data and corresponding handling results. The data acquisition layer, the map construction and positioning module, the target detection and status recognition module, and the anomaly analysis and processing module are connected in communication.

[0009] Furthermore, the anomaly description data includes anomaly vector sequences. ; The k-th anomaly vector is defined as ;in For visual anomaly feature vectors, This is an environmental anomaly feature vector. For spatial feature location, As a time feature, This refers to the abnormal label and the handling result.

[0010] Furthermore, the association analysis unit uses a self-attention mechanism to process the anomaly vector sequence, as shown below:

[0011] in, , , These represent the query, key, and value matrices, respectively, where d is the dimension of the key vector. It is a normalized exponential function.

[0012] Furthermore, the decision generation unit generates decisions based on predefined exception types through a conditional generation method. A i Corresponding set of handling actions The process determines a disposal recommendation, which satisfies a constrained generative expression. ;in, This is the context representation obtained by encoding the current anomaly vector sequence using a self-attention mechanism. In a given anomaly vector and context The probability of action r under certain conditions.

[0013] The compliance verification unit uses a preset rule constraint function.

[0014] The proposed actions are filtered, and the suggestion with the highest probability that satisfies the constraints is taken as the final output. .

[0015] Furthermore, the model evolution unit optimizes the association analysis unit and the decision generation unit according to the following objective function: ; in For abnormal prediction loss, Losses are generated as a result of the proposed solutions. These are the weighting coefficients.

[0016] Furthermore, the target detection and state recognition module is set to time... The acquired images are: , The target set of categories within the station is defined as follows: ,in This represents the nth category of equipment.

[0017] The gate state set is defined as follows:

[0018] Mapping is constructed using the YOLOv8 object detection model. , in, For the first Bounding boxes of each category of target; The type of door (vehicle door, platform door, or equipment door). The confidence level for target detection; Sub-images are obtained by cropping the detected categories of equipment regions. The VisionTransformer model is used for state recognition.

[0019] in, , This represents the probability that the door is in the open state. The probability of an abnormal state. This represents the probability that the door is in a closed state. The door status determination rule is set as follows:

[0020] When the system detects that a vehicle door, platform door, or equipment door is not in a properly closed state, it automatically stores the abnormal status information in a structured manner and uploads it to the backend analysis and processing module.

[0021] The visual anomaly feature vector is defined as follows: in This represents the change in the percentage of pixels in the door seam area. Target detection confidence; The YOLOv8 model is used to identify the categories of the remains.

[0022] Furthermore, the map building and positioning module constructs a 3D point cloud map in the following manner. , in For a moment The original laser point cloud; The spatial transformation matrix is ​​determined by the pose state; and a two-dimensional raster map is generated in the following manner;

[0023] in Indicates the first The occupancy status of each grid cell For points in a 3D point cloud, Let the coordinates of the point be 3D coordinates. This is a height threshold used to distinguish between passable areas and obstacles; The point cloud registration employs an iterative nearest-point algorithm, which minimizes the objective function.

[0024] in, Let be a rotation matrix. It is a translation vector; , For nearest neighbor pairs; The optimal pose correction amount is obtained through iterative optimization: ,in This represents pose composition operations. To predict the state and achieve centimeter-level positioning accuracy.

[0025] The map building and positioning module adopts a pose estimation model that integrates laser and inertial navigation. Assume the robot dog is at time... The state vector is: , in This indicates the robot's position in the subway station coordinate system; Represents attitude quaternions; Indicates speed; , These represent the biases of the accelerometer and gyroscope, respectively. State prediction is achieved through continuous integration using an IMU: ,in The IMU observations are used to correct the predicted state through point cloud matching errors, thereby obtaining pose estimation.

[0026] Furthermore, the map building and positioning module employs hierarchical path planning: The map building and positioning module abstracts the interior space of the subway station into a weighted directed graph. , in: A collection of inspection points; The set of inspection path edges Bianquan (wij) indicates that the robot dog starts from the inspection point v. i Move to point v j The weight.

[0027] Inspection points are identified by fusing radar point clouds with visual information and are defined as follows:

[0028] in The corresponding inspection task type for this location (door status monitoring, foreign object detection, environmental monitoring); The global path planning layer generates inspection paths covering all inspection points based on the two-dimensional grid map, and the path cost function is: , in: Indicates the path length cost; Indicates the distance cost to the obstacle; These are the weighting coefficients; The local path planning layer combines real-time laser point cloud perception information to perform dynamic obstacle avoidance, with the obstacle avoidance constraints being: , in: Location of dynamic obstacles; This indicates the robot dog's current position. The safe distance threshold; The underlying motion control layer translates the planned path into motion control commands. ,in Linear velocity, ω is the angular velocity.

[0029] Furthermore, the environmental sensor includes a gas sensor and a temperature and humidity sensor; the environmental anomaly vector is defined as:

[0030] To determine the real-time concentration of the i-th type of gas, This is the safe threshold for the gas within the station.

[0031] When the real-time concentration of any gas Exceeding its safety threshold At that time, that is

[0032] At that time, an environmental anomaly was identified.

[0033] Furthermore, the data acquisition layer also includes a data preprocessing unit, configured to: perform noise reduction, downsampling and coordinate alignment processing on lidar point cloud data; perform distortion correction, brightness equalization and resolution unification processing on video and image data; and perform filtering, outlier removal and time synchronization processing on environmental sensor data to align the time and spatial scales of multi-source sensor data.

[0034] Compared with the prior art, the beneficial effects of the present invention are: 1. Multi-sensor deep fusion perception: By integrating LiDAR, visible light / infrared cameras and environmental sensors into the inspection robot dog, and performing unified denoising, time synchronization and format standardization preprocessing on multi-source data, a highly consistent multimodal perception dataset is constructed. This effectively solves the mismatch problem of different sensors in time, space and scale, and realizes a highly reliable perception of the station environment in three dimensions.

[0035] 2. High-precision equipment status recognition: The YOLOv8 model is used to quickly detect and locate vehicle doors, platform doors, equipment doors, and leftover objects. Furthermore, the Vision Transformer model is introduced to perform global feature modeling and refined status discrimination (closed, open, abnormal) for door-type equipment areas. This significantly improves the accuracy and robustness of equipment status anomaly and leftover object recognition under complex lighting and occlusion environments, and reduces the risk of false alarms and missed alarms.

[0036] 3. Centimeter-level Autonomous Navigation: A large-scale 3D point cloud map of the station is constructed using laser inertial navigation fusion SLAM technology, and a 2D raster map is generated in parallel, balancing 3D accuracy with 2D computational efficiency. A point cloud registration and relocalization method based on iterative nearest point is used to achieve centimeter-level positioning accuracy in environments with dense pillars and repetitive features. A hierarchical path planning strategy organically combines global optimization, local dynamic obstacle avoidance, and low-level motion control, enabling the robot dog to autonomously generate inspection paths and safely complete fixed-point detection tasks.

[0037] 4. Intelligent Anomaly Handling Closed Loop: A large-scale anomaly analysis and processing model based on the Transformer architecture was constructed. Taking multimodal anomaly vector sequences as input, the model learns the correlation between anomalies through a self-attention mechanism. Combined with a predefined set of handling actions and operational rule constraints, it automatically generates compliant handling suggestions, realizing closed-loop management of anomaly information from "discovery to analysis to standardized suggestions to handling". This significantly improves the standardization and intelligence of anomaly handling and reduces reliance on human experience. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the overall processing flow of an intelligent inspection system for rail transit stations according to the present invention. Figure 2 This is a schematic diagram of the hardware structure and multi-sensor layout of the inspection robot dog of the present invention; Figure 3 This is a schematic diagram of the system module architecture and data communication connection of the present invention; Figure 4 This is a schematic diagram of map construction and hierarchical path planning according to the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0040] like Figures 1 to 4 As shown in the figure, this embodiment provides an intelligent inspection system for rail transit stations. Its overall architecture includes a data acquisition layer, a map building and positioning module, a target detection and status recognition module, and an anomaly analysis and processing module, with communication connections between the modules. The specific implementation of the system will be described in detail below, using a real-world subway station inspection scenario as an example.

[0041] I. Multi-sensor-based multi-source heterogeneous data acquisition and preprocessing The inspection robot dog in this embodiment is equipped with multiple heterogeneous sensors to form a data acquisition layer, used to obtain multi-dimensional information within the station: LiDAR: Used to collect 3D point cloud data of spatial structures such as platforms, passages, and equipment areas, supporting environmental modeling and obstacle perception.

[0042] Visible light cameras and infrared cameras: collect visible light video and infrared images, covering inspection areas such as platforms, station halls, and equipment rooms, providing image sources for equipment status identification and the detection of leftover objects.

[0043] Environmental sensors: including gas sensors and temperature and humidity sensors, can detect parameters such as volatile organic compounds, combustible gases, toxic gases, temperature and humidity in the air in real time.

[0044] The data acquisition layer is also equipped with a data preprocessing unit to uniformly process the raw data from each sensor, aligning the temporal and spatial scales of the multi-source data. The specific preprocessing operations are as follows: Denoising, voxel downsampling, and coordinate alignment are performed on lidar point cloud data; Perform distortion correction, brightness equalization, and resolution unification on video and image data; The system performs median filtering and outlier removal on environmental sensor data, and synchronizes it with point cloud and image data based on timestamps.

[0045] The configuration and functions of the above multi-sensor system are shown in Table 1.

[0046] LiDAR 3D spatial perception, mapping and localization 3D point cloud Visible light camera Equipment status image acquisition and object detection Visible light images / videos Infrared camera Abnormal temperature perception at night or under low light conditions Infrared thermal images Gas sensor Detection of flammable / toxic gas concentrations gas concentration value Temperature and humidity sensor Detect ambient temperature and humidity Temperature and humidity values

[0047] After preprocessing, the system further organizes the data in multiple dimensions and constructs sub-databases. Multi-source data is structured and aggregated according to date type (weekdays, non-weekdays, holidays), time granularity (e.g., 1 hour), and station region to form a unified multi-dimensional dataset. Based on this, sub-databases are divided according to line type (e.g., Line 1, Line 2), station number (e.g., Ancheng Street Station, Olympic Sports Center Station), and date type. Each sub-database records equipment status images, environmental parameters, and time-series status data statistically analyzed at hourly, daily, and weekly granularities under specific conditions, providing multi-dimensional input for subsequent model training and analysis.

[0048] II. Map Construction and Positioning To enable the inspection robot dog to navigate autonomously in a large station environment, this embodiment constructs a map and positioning scheme with three-dimensional spatial perception as the core and two-dimensional planning and calculation as the guarantee.

[0049] 2.1 Laser-Inertial Navigation Fusion SLAM Pose Estimation The robot dog acquires environmental and self-motion information through LiDAR and an inertial measurement unit (IMU). Let the robot dog's state vector at time t be:

[0050] in This indicates the robot's position in the subway station coordinate system; Represents attitude quaternions; Indicates speed; , These represent the biases of the accelerometer and gyroscope, respectively.

[0051] State prediction is achieved through continuous integration using an IMU:

[0052] in The observations are from the IMU (Infrared Detection Unit). The lidar point cloud serves as the observation, and the predicted state is corrected through point cloud matching errors, thereby obtaining high-precision pose estimation for large-scale SLAM mapping in subway stations.

[0053] 2.2 Multidimensional Map Construction After SLAM mapping is completed, the system maintains both a 3D point cloud map and a 2D raster map, achieving "dual-map parallelism".

[0054] The point cloud data collected at each time point is mapped to a unified coordinate system based on pose transformation to form a 3D point cloud map:

[0055] in denoted as the original laser point cloud at time t; T(xt) is the spatial transformation matrix determined by the pose state; this 3D map fully preserves the geometric structures of columns, walls, equipment, etc., for subsequent point cloud registration and relocalization.

[0056] To improve path planning efficiency, the 3D point cloud is projected onto the ground to generate a 2D occupancy grid map:

[0057] in Indicates the first The occupancy status of each grid cell For points in a 3D point cloud, Let the coordinates of the point be 3D coordinates. This is a height threshold used to distinguish between passable areas and obstacles.

[0058] During the inspection process, real-time positioning correction is performed through point cloud registration based on Iterative Closest Point (ICP). Let the current point cloud be... The local map point cloud is ICP solves for the pose transformation by minimizing the following objective function:

[0059] in, Let be a rotation matrix. It is a translation vector; , For nearest neighbor pairs; The optimal pose correction amount is obtained through iterative optimization: ,in This represents pose composition operations. For predicting the state This enables centimeter-level positioning accuracy in station environments with dense columns and repetitive features.

[0060] 2.3 Hierarchical Path Planning and Motion Control This embodiment abstracts the interior space of a subway station as a weighted directed graph. This is to optimize the inspection path. For a pre-defined set of inspection points, the edges The weight wij represents the cost of the robot dog moving from vi to vj. Inspection points are calibrated by fusing radar point cloud and visual information, and are defined as follows:

[0061] in The inspection task type corresponding to this location includes door status monitoring, foreign object detection, and environmental monitoring.

[0062] The system adopts a hierarchical path planning strategy: The global path planning layer generates optimized paths covering all inspection points based on a two-dimensional grid map, and its cost function is:

[0063] in For path length cost, The cost of obstacle distance, , These are the weighting coefficients. A typical set of weighting configurations is shown in Table 2.

[0064]

[0065] The local path planning layer combines real-time laser point cloud perception information for dynamic obstacle avoidance, and its safety constraints are as follows:

[0066] in: Location of dynamic obstacles; The safe distance threshold can be set to 0.8 m.

[0067] The underlying motion control layer translates the planned path into motion control commands:

[0068] in: Linear velocity; The angular velocity is used; and closed-loop control enables the robot dog to move stably along the planned path and stop at fixed points to perform inspection tasks.

[0069] III. Target Detection and State Recognition The target detection and status recognition module is responsible for detecting different types of equipment from video image data and determining their status, while also identifying abandoned items in the station.

[0070] 3.1 Target Detection of Various Types of Equipment The image acquired at time t is set as follows: All types of equipment within the station constitute the target set. This includes vehicle doors, platform doors, and equipment doors.

[0071] The gate state space is defined as a finite set:

[0072] A mapping from image to detection result is constructed using the YOLOv8 object detection model: , in, For the first Bounding boxes of each category of target; The type of door (vehicle door, platform door, or equipment door). The confidence level for target detection is given; this step enables accurate positioning and differentiation of multiple types of door equipment in complex station environments.

[0073] 3.2 Refined gate state discrimination For each detected phylum-type target, a sub-image is cropped from the original image:

[0074] Then, the VisionTransformer (ViT) model is input for state recognition:

[0075] Where the posterior probability distribution is: This represents the probability that the door is in the open state. The probability of an abnormal state. The probability that the door is in a closed state; based on a preset threshold. , Perform a status determination:

[0076] In this embodiment, the threshold can be set to =0.7, =0.6, which can be adjusted according to the actual situation on site. When the system detects that a vehicle door, platform door, or equipment door is not in a normally closed state (i.e., open or abnormal), it automatically stores the abnormal state information in a structured manner and uploads it to the backend analysis and processing module.

[0077] Meanwhile, the visual anomaly feature vector is defined as follows:

[0078] in denoted as r, representing the change in the percentage of pixels in the door gap area; r represents the target detection confidence level; used to quantify the degree of visual anomaly.

[0079] 3.3 Identification of Lost Items The YOLOv8 model is also used to identify common abandoned items in subway stations. These categories include, but are not limited to, non-fixed objects such as water bottles, packages, and trash, as well as unusual items that do not fit the station environment. Once an abandoned item is detected, the system uploads the anomaly information in a structured format and generates handling suggestions based on the anomaly type.

[0080] IV. Anomaly Analysis and Processing System The anomaly analysis and processing module is deployed on the backend and is responsible for unified modeling, correlation analysis, generation of handling suggestions and compliance verification of various anomaly description data uploaded by the frontend robot dog, and realizing model evolution through continuous learning.

[0081] 4.1 Exception Database Construction To ensure that outlier data from different sources can be understood by the same model, each outlier sample is represented as a multimodal outlier vector:

[0082] in For visual anomaly feature vectors, This is an environmental anomaly feature vector. For spatial feature location, As a time feature, This includes anomaly labels and handling results (confirmation status, handling method). The system continuously imports historical subway station inspection anomaly data and newly discovered anomaly status data from the robot into the database, forming anomaly vector sequences that can be used for training.

[0083] 4.2 Environmental Anomaly Vector and Judgment The environmental anomaly vector integrates the ratios of various gas concentrations to safety thresholds:

[0084] in To determine the real-time concentration of the i-th type of gas, This is the safe threshold for the gas within the station.

[0085] When any component satisfies: This indicates an environmental anomaly. Table 3 provides examples of safety thresholds for several common gases.

[0086] carbon monoxide 24 ppm methane 10 %LEL hydrogen sulfide 10 ppm Volatile organic compounds 500 ppb

[0087] 4.3 Anomaly Correlation Analysis and Handling Suggestions Generation This embodiment uses a Transformer-based multimodal large model as the analysis foundation, and learns the correlation between anomalies in the anomaly vector sequence through a self-attention mechanism:

[0088] in, , , These represent the query, key, and value matrices, respectively, where d is the dimension of the key vector. It is a normalized exponential function.

[0089] After obtaining the anomaly correlation representation, the model, for each type of anomaly... A i Predefined set of actions:

[0090] The optimal handling suggestion is generated using a condition-based method.

[0091] in, This is a contextual representation obtained by encoding the current anomaly vector sequence using a self-attention mechanism, used to assist in decision-making. In a given anomaly vector and context The probability of action r under certain conditions.

[0092] To avoid generating suggestions that do not comply with operational guidelines, a rule constraint function is introduced for compliance verification:

[0093] The final output should be selected from the candidate actions that satisfy the rules, choosing the one with the highest probability.

[0094] To further illustrate the feasibility of this invention, specific examples of the set of processing actions and rule constraint functions are given here: Define the exception type A i The following is a set of procedures for handling "abnormal opening of platform doors". It may include: Notify the nearest station staff member to confirm; Remotely close the corresponding platform door; : Triggers platform announcements to remind passengers to move away; Report to the control center and record it.

[0095] An Operational RuleSet can contain the following rules: - If the anomaly level is "Severe", then you must not select only... ; - If the current time period is outside of operating hours, selection is not allowed. ; - Any treatment recommendations must include .

[0096] Through the above specific examples, those skilled in the art can clearly understand how rule constraint functions and sets of disposal actions are implemented, and develop systems accordingly.

[0097] 4.4 Model Evolution The model evolution unit optimizes the correlation analysis unit and the decision generation unit using newly collected anomaly samples through a continuous learning mechanism. Its objective function is defined as:

[0098] in To predict losses based on the severity of the anomaly, Losses are generated as a result of the proposed solutions. This is a balancing weighting coefficient. In this embodiment, it can be taken as... =0.5. Through this objective function, the model can progressively improve its ability to analyze new anomalies and the accuracy of its response recommendations.

[0099] V. Beneficial Effects The system described in this embodiment demonstrates significant advantages in actual subway station deployment: 1. Multi-sensor deep fusion perception: By integrating LiDAR, visible light / infrared cameras and environmental sensors, and combining unified preprocessing and multi-dimensional data organization, a three-dimensional and multi-dimensional highly reliable perception of the station environment is achieved, solving the problems of insufficient perception capability of single sensors and inconsistent data.

[0100] 2. High-precision equipment status recognition: By integrating YOLOv8 fast target detection with Vision Transformer fine-grained status discrimination, the accuracy and robustness of identifying abnormal equipment status and leftover objects are significantly improved, effectively reducing false alarms and false negatives, and adapting to complex lighting and occlusion scenarios.

[0101] 3. Centimeter-level autonomous navigation: Based on laser inertial navigation fusion SLAM, parallel maintenance of dual maps and ICP point cloud relocalization, the robot dog achieves centimeter-level positioning accuracy in station environments with dense pillars and repetitive features; hierarchical path planning takes into account both global efficiency and local safety, ensuring the flexibility of inspection paths and the continuity of task execution.

[0102] 4. Intelligent Anomaly Handling Closed Loop: A continuously updated anomaly database and a Transformer-based large-scale analysis model have been built, realizing closed-loop management of anomaly information from "discovery - correlation analysis - standardized handling suggestions - compliance output", which significantly improves the standardization and intelligence level of anomaly handling and reduces the reliance on human experience.

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent inspection system for rail transit stations, characterized in that, include: The data acquisition layer, located on the inspection robot dog, includes lidar, visible light camera, infrared camera, and environmental sensors; The map building and positioning module is configured to build a three-dimensional point cloud map and a two-dimensional grid map of the station based on the fusion SLAM of the lidar and the inertial measurement unit, and to match the real-time collected lidar point cloud with the three-dimensional point cloud map for positioning correction by using the point cloud registration method. The target detection and state recognition module is configured to detect and locate door-type devices from the video image data, and to perform state judgment on the detected door-type device areas to obtain the state information of the door-type devices; the door-type devices include vehicle doors, platform doors, and equipment doors; The anomaly analysis and handling module is configured as follows: Acquire and structure the anomaly description data, which includes at least visual anomaly features, environmental anomaly features, and spatial location information; The correlation analysis unit learns the correlation relationships between multiple anomaly description data; Based on the results of relational learning, a decision generation unit generates suggestions for handling abnormal events. The proposed actions are then verified against operational standards by the compliance verification unit before being output. The model evolution unit continuously optimizes the correlation analysis unit and the decision generation unit based on historical anomaly description data and corresponding handling results. The data acquisition layer, the map construction and positioning module, the target detection and status recognition module, and the anomaly analysis and processing module are connected in communication.

2. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The anomaly description data includes an anomaly vector sequence. ； The k-th anomaly vector is defined as ;in For visual anomaly feature vectors, This is an environmental anomaly feature vector. For spatial feature location, As a time feature, This refers to the abnormal label and the handling result.

3. The intelligent inspection system for rail transit stations according to claim 2, characterized in that, The correlation analysis unit uses a self-attention mechanism to process the abnormal vector sequence, as follows: in, , , These represent the query, key, and value matrices, respectively, where d is the dimension of the key vector. It is a normalized exponential function.

4. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The decision generation unit generates decisions based on predefined exception types using a conditional generation method. A i Corresponding set of handling actions The process determines a disposal recommendation, which satisfies a constrained generative expression. ;in, This is the context representation obtained by encoding the current anomaly vector sequence using a self-attention mechanism. In a given anomaly vector and context The probability of action r under certain conditions. The compliance verification unit uses a preset rule constraint function. The proposed actions are filtered, and the suggestion with the highest probability that satisfies the constraints is taken as the final output. .

5. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The model evolution unit optimizes the correlation analysis unit and the decision generation unit according to the following objective function: ； in For abnormal prediction loss, Losses are generated as a result of the proposed solutions. These are the weighting coefficients.

6. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The target detection and state recognition module is set to time. The acquired images are: ， The target set of categories within the station is defined as follows: ,in This represents the nth category of equipment. The gate state set is defined as follows: Mapping is constructed using the YOLOv8 object detection model. ， in, For the first Bounding boxes of each category of target; The type of door (vehicle door, platform door, or equipment door). The confidence level for target detection; Sub-images are obtained by cropping the detected categories of equipment regions. The VisionTransformer model is used for state recognition. in, , This represents the probability that the door is in the open state. The probability of an abnormal state. This represents the probability that the door is in a closed state. The door status determination rule is set as follows: When the system detects that a vehicle door, platform door, or equipment door is not in a properly closed state, it automatically stores the abnormal status information in a structured manner and uploads it to the backend analysis and processing module. The visual anomaly feature vector is defined as follows: in This represents the change in the percentage of pixels in the door seam area. Target detection confidence; The YOLOv8 model is used to identify the categories of the remains.

7. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The map building and positioning module builds a 3D point cloud map in the following manner. , in For a moment The original laser point cloud; The spatial transformation matrix is ​​determined by the pose state; And generate a two-dimensional raster map in the following manner; in Indicates the first The occupancy status of each grid cell For points in a 3D point cloud, Let the coordinates of the point be 3D coordinates. A height threshold is used to distinguish between passable areas and obstacles; The point cloud registration employs an iterative nearest-point algorithm, which minimizes the objective function. in, For rotation matrix, It is a translation vector; , For nearest neighbor pairs; The optimal pose correction amount is obtained through iterative optimization: ,in This represents pose composition operations. To predict the state and achieve centimeter-level positioning accuracy. The map building and positioning module adopts a pose estimation model that integrates laser and inertial navigation. Assume the robot dog is at time... The state vector is: , in This indicates the robot's position in the subway station coordinate system; Represents attitude quaternions; Indicates speed; , These represent the biases of the accelerometer and gyroscope, respectively. State prediction is achieved through continuous integration using an IMU: ,in The IMU observations are used to correct the predicted state through point cloud matching errors, thereby obtaining pose estimation.

8. The intelligent inspection system for rail transit stations according to claim 7, characterized in that, The map building and positioning module employs hierarchical path planning: The map building and positioning module abstracts the interior space of the subway station into a weighted directed graph. , in: A collection of inspection points; The set of inspection path edges Bianquan (wij) indicates that the robot dog starts from the inspection point v. i Move to point v j The weight. Inspection points are identified by fusing radar point clouds with visual information and are defined as follows: in The corresponding inspection task type for this location (door status monitoring, foreign object detection, environmental monitoring); The global path planning layer generates inspection paths covering all inspection points based on the two-dimensional grid map, and the path cost function is: , in: Indicates the cost of path length; Indicates the distance cost to the obstacle; These are the weighting coefficients; The local path planning layer combines real-time laser point cloud perception information to perform dynamic obstacle avoidance, with the obstacle avoidance constraints being: , in: Location of dynamic obstacles; This indicates the robot dog's current position. The safe distance threshold; The underlying motion control layer translates the planned path into motion control commands. ,in For linear velocity, ω is the angular velocity.

9. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The environmental sensors include gas sensors and temperature and humidity sensors; the environmental anomaly vector is defined as: To determine the real-time concentration of the i-th type of gas, This is the safe threshold for the gas within the station. When the real-time concentration of any gas Exceeding its safety threshold At that time, that is At that time, an environmental anomaly was identified.

10. The intelligent inspection system for rail transit stations according to claim 1, characterized in that, The data acquisition layer also includes a data preprocessing unit, configured to: perform noise reduction, downsampling and coordinate alignment processing on lidar point cloud data; perform distortion correction, brightness equalization and resolution unification processing on video and image data; and perform filtering, outlier removal and time synchronization processing on environmental sensor data to align the time and spatial scales of multi-source sensor data.

Images