An intelligent inspection and maintenance method for door adjustment of a metro vehicle based on an intelligent inspection trolley

Abstract

The present application relates to the field of urban rail vehicle door maintenance, and particularly relates to an intelligent maintenance method for urban rail vehicle door installation and adjustment based on an intelligent inspection trolley. The scheme comprises the following steps: collecting door picture data under different illuminations, different openings and similar color differences by a camera on the intelligent inspection trolley, marking the door core window position of the door, training a YOLOv8-nano light model, and identifying the door frame by the trained YOLOv8-nano light model; judging the accuracy of the door data according to the frame, scanning the door by an HL-Scan 3D scanner to obtain three-dimensional point cloud data of the door, obtaining a digital surface model of the complete door, extracting V-type, front swing, rear swing and lower swing data of the door, comparing the extracted V-type, front swing, rear swing and lower swing data of the door with standard data, and generating a maintenance scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban rail vehicle door maintenance, in particular to an intelligent maintenance method for urban rail vehicle door installation and adjustment based on an intelligent inspection trolley. BACKGROUND

[0002] Existing operation companies mainly rely on manual adjustment of vehicle doors. During new construction and maintenance, manual adjustment of vehicle doors takes a long time because there are many measurement data, and there is cross influence between multiple data, which increases the difficulty of adjustment.

[0003] The prior art collects vehicle door data through a sensor, and then compares the data with standard data to determine the adjustment amount. For example, CN118857197A discloses an urban rail door system maintenance system based on PHM technology, which includes a sensor module arranged on the door for collecting various size data of the door; a comparison module for comparing the various size data with the corresponding standard size; and a maintenance module for determining the corresponding maintenance position based on the size data that is inconsistent with the standard size when any one of the size data is inconsistent with the corresponding standard size. The above-mentioned scheme can collect size data of the door system in real time through the sensor module, and compare the size data of the door system with the standard size, so as to timely find abnormal positions. The "planned repair" is changed to "fault repair", which reduces the maintenance workload of the rail transit door system, and can effectively shorten the maintenance time.

[0004] The above-mentioned scheme can collect data in real time through the sensor for feedback, is mainly used for detecting the change amount, and can timely find abnormal positions, but is not suitable for scenes requiring absolute high-precision adjustment. At the same time, the above-mentioned scheme cannot be applied to the closed-loop scene from "data scanning" to "maintenance scheme generation-execution-precise matching of maintenance tools-feedback". For high-precision maintenance, the maintenance precision and efficiency of the above-mentioned scheme are difficult to meet the requirements. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art, and provides an intelligent maintenance method for urban rail vehicle door installation and adjustment based on an intelligent inspection trolley, which greatly improves the maintenance efficiency and precision of urban rail vehicle doors.

[0006] The present application achieves the above-mentioned purpose by adopting the following technical scheme. The present application provides an intelligent maintenance method for urban rail vehicle door installation and adjustment based on an intelligent inspection trolley, which comprises:

[0007] S1, collecting door data through an intelligent inspection trolley camera and determining the accuracy of the door data;

[0008] Collect vehicle door picture data under different illuminations, different openings, and similar color differences through the camera on the intelligent inspection trolley, and mark the door core window position of the vehicle door to determine the vehicle door frame through the door core window position of the vehicle door;

[0009] Train the YOLOv8-nano lightweight model through the labeled vehicle door picture data, and identify the vehicle door frame through the trained YOLOv8-nano lightweight model;

[0010] Compare the position of the vehicle door frame identified by the YOLOv8-nano lightweight model with the position of the vehicle door frame determined through the door core window position of the vehicle door, if the error of the two sets of vehicle door frames is less than the set threshold, it is determined that the data is accurate, and step S2 is entered, otherwise, the vehicle door data is re-collected and determined.

[0011] S2, scan the vehicle door and the vehicle body by using the HL-Scan 3D scanner, extract the V-type, front swing, rear swing, and down swing data of the vehicle door, and generate a maintenance scheme;

[0012] The V-type data refers to the size difference between the distance between the upper protective finger adhesive tapes of the left and right vehicle doors and the distance between the lower protective finger adhesive tapes of the left and right vehicle doors when the vehicle door is opened to a straight position, the vehicle door is opened to a straight position refers to that the distance between the left and right vehicle doors is greater than or equal to 500 mm, the front swing data refers to the distance between the upper front end of the vehicle door and the surface of the vehicle body when the vehicle door is opened to a straight position, the upper front end of the vehicle door refers to the upper edge of the vehicle door, which is 150 mm away from the protective finger adhesive tape, the rear swing data refers to the distance between the upper rear end of the vehicle door and the surface of the vehicle body when the vehicle door is opened to a straight position, the upper rear end of the vehicle door refers to the upper edge of the vehicle door, which is 150 mm away from the outer edge, and the down swing data refers to the distance between the lower front end of the vehicle door and the surface of the vehicle body when the vehicle door is opened to a straight position, the lower front end of the vehicle door refers to the lower outer edge of the vehicle door, which is 150 mm away from the outer edge of the vehicle door;

[0013] Open the vehicle door to a straight position, and scan the vehicle door and the vehicle body by using the HL-Scan 3D scanner, including the upper and lower protective finger adhesive tapes of the left and right vehicle doors, the upper front end of the vehicle door, the upper rear end of the vehicle door, the lower front end of the vehicle door, and the surface of the vehicle body;

[0014] Process the scanned point cloud data, remove outliers, scanning background, and objects irrelevant to the measured gap, the outliers refer to noise floating in the air, and the scanning background refers to the ground and tooling;

[0015] Align the processed point cloud with the 3D CAD model of the vehicle body, and use the best fitting alignment or RPS positioning based on the non-deformed, known geometric features on the vehicle body;

[0016] On the door point cloud, manually select accurate points or fit key position points on the local point cloud, the key position points include points corresponding to the left and right door upper and lower finger protection strips, points corresponding to the front end of the upper part of the door, points corresponding to the rear end of the upper part of the door, and points corresponding to the front end of the lower part of the door, and select the corresponding area on the vehicle body point cloud, and fit the local reference plane in the area;

[0017] Calculate V-shaped data: calculate V-shaped data according to the point cloud corresponding to the left and right door upper and lower finger protection strips;

[0018] Calculate the front swing data, the distance between the point corresponding to the front end of the upper part of the door and the corresponding vehicle body reference plane, which is the front swing data;

[0019] Calculate the rear swing data, the distance between the point corresponding to the rear end of the upper part of the door and the corresponding vehicle body reference plane, which is the rear swing data;

[0020] Calculate the lower swing data, the distance between the point corresponding to the front end of the lower part of the door and the corresponding vehicle body reference plane, which is the lower swing data;

[0021] Compare the calculated V-shaped, front swing, rear swing and lower swing data with the corresponding standard data to obtain the deviation values of the V-shaped, front swing, rear swing and lower swing data, and generate a corresponding maintenance scheme according to the corresponding deviation values, the maintenance scheme including the door parts that need to be adjusted, the corresponding maintenance tools and the corresponding adjustment process.

[0022] S3, path planning and target intelligent cabinet navigation according to the maintenance scheme;

[0023] The intelligent inspection vehicle queries the intelligent cabinet number, the cabinet grid number and the coordinates of the intelligent cabinet in the factory map of the maintenance tool according to the generated maintenance scheme;

[0024] The intelligent inspection vehicle perceives the surrounding environment in real time through its laser radar and / or camera, identifies lane lines and obstacles, and matches with the existing map, plans a safe path according to its own position and the position of the target intelligent cabinet, and the intelligent inspection vehicle travels according to the safe path and reaches the preset parking point in front of the target intelligent cabinet.

[0025] S4, target intelligent cabinet identification and coarse positioning;

[0026] The fixed camera on the intelligent inspection vehicle collects maintenance tool pictures in real time, and inputs the collected pictures into the trained YOLOv8-nano lightweight model, judges whether it is the required maintenance tool according to the class and confidence of the YOLOv8-nano lightweight model output maintenance tool, and outputs the 2D bounding box coordinates of the maintenance tool in the image;

[0027] Get the geometric center pixel coordinates of the 2D bounding box of the maintenance tool, and according to the depth of each tool grid in the intelligent cabinet, the 3D coordinates of the maintenance tool in the fixed camera coordinate system are calculated by inverse projection through the internal parameters of the fixed camera on the intelligent inspection trolley, and the 3D coordinates are converted to the coarse coordinates in the trolley base coordinate system through the hand-eye matrix, and the end of the mechanical arm is moved to a safe preparation position in front of the maintenance tool according to the coarse coordinates.

[0028] S5, laser scanning and final grabbing.

[0029] When the mechanical arm is in place, the laser radar on the intelligent inspection trolley triggers a high-resolution scan of the grid where the maintenance tool is located, and a dense point cloud is obtained;

[0030] The point cloud cluster of the maintenance tool and the point cloud cluster of the maintenance tool itself are segmented from the point cloud, and the tool point cloud cluster is matched with the 3D model in the knowledge base using 3D CNN, and the 6D pose of the maintenance tool in the laser radar coordinate system is output;

[0031] The 6D pose is converted to the mechanical arm base coordinate system by using the radar external parameter matrix, and the final grabbing coordinate is obtained;

[0032] The mechanical arm receives the final grabbing coordinate, queries the weight and gravity center data of the corresponding maintenance tool, calculates the optimal grabbing pose, jaw opening width and required grabbing force, and grabs the maintenance tool;

[0033] After obtaining the maintenance tool, the maintenance point is navigated, and the maintenance tool is handed over to the maintenance personnel, and the maintenance personnel adjusts the vehicle door according to the maintenance scheme.

[0034] The beneficial effects of the present application are:

[0035] The present application adopts laser and visual positioning fusion technology, defines "detection point alignment standard + data transmission protocol", constructs a centimeter-level accurate positioning model, ensures that the scanning data and the actual position of the vehicle door have no deviation, solves the problem of "positioning deviation leading to distorted detection data", and makes the key parameter measurement error of V-shaped ruler size, front and rear ruler size close to zero, and provides accurate data for maintenance scheme generation.

[0036] The present application integrates "intelligent tool cabinet + mechanical arm coordinate mapping" technology, converts the tool 2D coordinates into the physical position of the mechanical arm through "camera calibration position + odometer fusion", automatically identifies the grabbed tool combined with the maintenance steps, successfully reduces the time loss and matching error of manual tool taking and placing, and improves the single vehicle door maintenance efficiency.

[0037] The application constructs a "detection-analysis-solution-tool-feedback" full closed loop system, updates data in real time through a CNN algorithm, generates a step-by-step maintenance solution through a linkage case library, and synchronizes to a mobile terminal and an intelligent cabinet, thereby solving the problem of "data disconnection from execution", enabling a single person to complete maintenance without the need for multiple operators to cooperate, reducing mechanized repetitive operations and reducing labor costs.

[0038] The YOLOv8-nano light model is adopted, so that the vehicle door can be quickly recognized, and the recognition efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is an intelligent maintenance method flow chart of urban rail vehicle door adjustment based on an intelligent inspection trolley provided by the application.

[0040] Figure 2 It is a V-shaped size schematic diagram provided by the application. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0042] The application provides an intelligent maintenance method of urban rail vehicle door adjustment based on an intelligent inspection trolley, as shown in the figure, specifically comprising the following steps. Figure 1

[0043] S1, collecting door data through an intelligent inspection trolley camera and determining the accuracy of the door data;

[0044] Collecting door picture data under different illuminations, different opening degrees and similar color differences through the camera on the intelligent inspection trolley, and marking the door core window position of the door, the door frame of the door is determined through the door core window position of the door, the gap between the door core window and the vehicle body is fixed and unchanged, so the door frame can be determined through the door core window position;

[0045] Similar color difference refers to the color difference value of two color samples calculated through a selected color space and a color difference calculation formula under the same measurement condition being less than or equal to a set tolerance threshold;

[0046] Training the YOLOv8-nano light model through the marked door picture data, and identifying the door frame through the trained YOLOv8-nano light model;

[0047] Comparing the position of the door frame identified by the YOLOv8-nano light model with the position of the door frame determined through the door core window position of the door, if the error of the two sets of door frames is less than a set threshold, it is determined that the data is accurate, and step S2 is entered, otherwise the door data is re-collected and determined.​

[0048] S2, scanning the door and the vehicle body with an HL-Scan 3D scanner, extracting the V-type, front swing, rear swing and lower swing data of the door, and generating a maintenance scheme;

[0049] As shown in Figure 2 , the V-type data refers to the size difference between the distance X between the upper finger guard strips of the left and right doors and the distance Y between the lower finger guard strips of the left and right doors when the door is opened to the straightaway position, the door is opened, and the distance between the left and right doors is greater than or equal to 500 mm, the front swing data refers to the distance between the upper front end of the door and the surface of the vehicle body when the door is opened to the straightaway position, the upper front end of the door refers to the upper edge of the door 150 mm away from the finger guard strip, the rear swing data refers to the distance between the upper rear end of the door and the surface of the vehicle body when the door is opened to the straightaway position, the upper rear end of the door refers to the upper edge of the door 150 mm away from the outer edge, and the lower swing data refers to the distance between the lower front end of the door and the surface of the vehicle body when the door is opened to the straightaway position, the lower front end of the door refers to the outer lower edge of the door 150 mm away from the outer edge of the door;

[0050] When the door is opened to the straightaway position, the door and the vehicle body are scanned with an HL-Scan 3D scanner, including the upper and lower finger guard strips of the left and right doors, the upper front end of the door, the upper rear end of the door, the lower front end of the door and the surface of the vehicle body;

[0051] The scanned point cloud data is processed to remove outliers, scanning background and objects unrelated to the measured gap, outliers refer to noise floating in the air, and scanning background refers to the ground and tooling;

[0052] The processed point cloud is unified in a coordinate system, the door and vehicle body point cloud is aligned with the 3D CAD model of the vehicle body, and the best fitting alignment or RPS positioning based on the non-deformed, known geometric features on the vehicle body is used;

[0053] On the door point cloud, the key position points are manually selected or fitted by local point cloud, the key position points include the points corresponding to the upper and lower finger guard strips of the left and right doors, the points corresponding to the upper front end of the door, the points corresponding to the upper rear end of the door and the points corresponding to the lower front end of the door, and the areas corresponding to the key position points are selected on the vehicle body point cloud, and the local reference planes are fitted in the areas;

[0054] V-type data is calculated: the V-type data is calculated according to the point cloud corresponding to the upper and lower finger guard strips of the left and right doors;

[0055] The front swing data is calculated, which is the distance between the point corresponding to the upper front end of the door and the corresponding vehicle body reference plane;

[0056] The rear swing data is calculated, which is the distance between the point corresponding to the upper rear end of the door and the corresponding vehicle body reference plane;

[0057] The distance between the corresponding point of the lower end of the vehicle door and the corresponding vehicle body reference plane is the under-swing data;

[0058] The V-shaped, front-swing, rear-swing, and under-swing data are compared with the corresponding standard data to obtain the deviation values of the V-shaped, front-swing, rear-swing, and under-swing data, and the corresponding maintenance scheme is generated according to the corresponding deviation values, which includes the vehicle door parts that need to be adjusted, the corresponding maintenance tools, and the corresponding adjustment process.

[0059] S3, path planning and target intelligent cabinet navigation are performed according to the maintenance scheme;

[0060] The intelligent inspection vehicle queries the intelligent cabinet number, the cabinet grid number, and the coordinates of the intelligent cabinet in the factory map according to the generated maintenance scheme;

[0061] The intelligent inspection vehicle senses the surrounding environment in real time through its laser radar and / or camera, identifies lane lines and obstacles, and matches with the existing map to plan a safe path according to its own position and the position of the target intelligent cabinet. The intelligent inspection vehicle travels according to the safe path and reaches the preset parking point in front of the target intelligent cabinet.

[0062] S4, target intelligent cabinet identification and coarse positioning are performed;

[0063] The fixed camera on the intelligent inspection vehicle collects maintenance tool pictures in real time, and inputs the collected pictures into the trained YOLOv8-nano lightweight model. According to the class and confidence of the output maintenance tool of the YOLOv8-nano lightweight model, it is judged whether it is the required maintenance tool, and the 2D bounding box coordinates of the maintenance tool in the image are output;

[0064] The geometric center pixel coordinates of the 2D bounding box of the maintenance tool are obtained, and according to the depth of each tool grid in the intelligent cabinet, the 3D coordinates of the maintenance tool in the fixed camera coordinate system are calculated by using the intrinsic parameters of the fixed camera on the intelligent inspection vehicle through back projection. The 3D coordinates are converted to coarse coordinates in the vehicle base coordinate system through the hand-eye matrix, and the end of the mechanical arm is moved to a safe preparation position in front of the maintenance tool.

[0065] The training process of the YOLOv8-nano lightweight model is as follows:

[0066] The pictures of the maintenance tool under different angles, light, and occlusion conditions are collected through the intelligent cabinet camera;

[0067] Through manual annotation, a bounding box is drawn for each maintenance tool in each picture and the class is labeled;

[0068] The labeled picture is used for training the YOLOv8-nano light model, and after the training is completed, the real-time collected maintenance tool picture is input into the YOLOv8-nano light model, and the category, confidence and boundary box coordinates of the maintenance tool are output.

[0069] The trained YOLOv8-nano light model is integrated on the intelligent inspection trolley.

[0070] Or first integrate the YOLOv8-nano light model on the intelligent inspection trolley, and then collect the pictures of the maintenance tool under different angles, illuminations and shielding conditions through the fixed camera on the intelligent inspection trolley for training.

[0071] S5, laser scanning and final grabbing.

[0072] When the mechanical arm is in place, the laser radar on the intelligent inspection trolley triggers a high-resolution scan on the grid where the maintenance tool is located, and obtains dense point clouds;

[0073] The point cloud cluster of the maintenance tool and the point cloud cluster of the maintenance tool itself are segmented from the point cloud, the tool point cloud cluster is matched with the 3D model in the knowledge base using 3D CNN, and the 6D pose of the maintenance tool in the laser radar coordinate system is output;

[0074] The 6D pose is converted to the base coordinate system of the mechanical arm by using the radar external parameter matrix, and the final grabbing coordinate is obtained;

[0075] The mechanical arm receives the final grabbing coordinate, queries the weight and gravity center data of the corresponding maintenance tool, calculates the optimal grabbing pose, jaw opening width and required grabbing force, and grabs the maintenance tool;

[0076] After obtaining the maintenance tool, the maintenance point is navigated, and the maintenance tool is handed over to the maintenance personnel, and the maintenance personnel adjusts the door according to the maintenance scheme.

[0077] The application will be described below in combination with specific cases.

[0078] For example, frequent opening and closing of the subway vehicle will make the V-shaped exceed or be less than the standard value 2-5mm range. V-shaped adjustment needs to carry the eccentric wheel of the door frame, and through the rotation of the eccentric wheel, the whole door leaf can be slightly rotated clockwise or counterclockwise.

[0079] The V-shaped is the state of the double door leaf, and the adjustment process is to ensure the size of the two single door leaves to ensure the V-shaped of the door.

[0080] First step: single door V-shaped size measurement 1≤X-Y≤2.5;

[0081] Second step: single door V type size unqualified adjustment door frame eccentric wheel, eccentric wheel clockwise, single door whole clockwise rotation. This process needs to measure, stop adjusting eccentric wheel, to ensure that the X-Y standard range;

[0082] Third step: two single door adjustment is completed, the door is opened to the straight position, and the size is finally confirmed.

[0083] The above only is the preferred embodiment of the present application, it should be understood that the present application is not limited to the form disclosed herein, should not be seen as the exclusion of other embodiments, and can be used in various other combinations, modifications and environment, and can be within the scope of the concept described herein, by the above teaching or related art or knowledge of the person skilled in the art to make changes. And the change of the person skilled in the art does not deviate from the spirit and scope of the present application, then all should be within the scope of protection of the appended claims of the present application.

Claims

1. An intelligent inspection method for adjusting the door of a smart inspection trolley urban rail vehicle, characterized in that, The method comprises the following steps: S1, collecting door data through the camera of the intelligent inspection trolley, and determining the accuracy of the door data; Collecting door picture data under different illuminations, different opening degrees and similar color differences through the camera on the intelligent inspection trolley, and marking the door core window position of the door, and determining the door frame through the door core window position of the door, wherein the similar color difference refers to that under the same measurement condition, the color difference value of two color samples calculated through the selected color space and color difference calculation formula is less than or equal to the set tolerance threshold; Training the YOLOv8-nano lightweight model through the marked door picture data, and identifying the door frame through the trained YOLOv8-nano lightweight model; Comparing the position of the door frame identified by the YOLOv8-nano lightweight model with the door frame determined through the door core window position of the door, if the error of the two sets of door frames is less than the set threshold, it is determined that the data is accurate, and step S2 is entered, otherwise, the door data is re-collected and determined; S2, scanning the door and the vehicle body by using the HL-Scan 3D scanner, extracting the V-shaped data, front swing data, rear swing data and lower swing data of the door, and generating a maintenance scheme, specifically comprising: Open the door to the straight position, scan the door and the vehicle body by using the HL-Scan 3D scanner, including the left and right door upper and lower protective finger adhesive tapes, the door upper front end, the door upper rear end, the door lower front end and the surface of the vehicle body; Processing the scanned point cloud data, removing outliers, scanning background and objects irrelevant to the measured gap, wherein the outliers refer to noise points floating in the air, and the scanning background refers to the ground and the tooling; Unifying the coordinate system of the processed point cloud, aligning the door and vehicle body point cloud with the 3D CAD model of the vehicle body, and using the best fitting alignment or RPS positioning based on the non-deformed and known geometric features on the vehicle body; Manually selecting or fitting the key position points on the door point cloud, the key position points including the points corresponding to the left and right door upper and lower protective finger adhesive tapes, the door upper front end, the door upper rear end and the door lower front end, and selecting the corresponding areas on the vehicle body point cloud, fitting the local reference planes in the areas; Calculating the V-shaped data, calculating the V-shaped data according to the point cloud corresponding to the left and right door upper and lower protective finger adhesive tapes; Calculating the front swing data, the distance between the door upper front end corresponding point and the corresponding vehicle body reference plane is the front swing data; Calculating the rear swing data, the distance between the door upper rear end corresponding point and the corresponding vehicle body reference plane is the rear swing data; Calculating the lower swing data, the distance between the door lower front end corresponding point and the corresponding vehicle body reference plane is the lower swing data; Comparing the calculated V-shaped, front swing, rear swing and lower swing data with the corresponding standard data to obtain the deviation values of the V-shaped, front swing, rear swing and lower swing data, and generating the corresponding maintenance scheme according to the corresponding deviation values, wherein the maintenance scheme includes the door parts that need to be adjusted, the corresponding maintenance tools and the corresponding adjustment process; S3, path planning and target intelligent cabinet navigation according to the maintenance scheme; S4, target intelligent cabinet recognition and coarse positioning are performed; S5, laser scanning and maintenance tool grabbing.

2. The intelligent inspection method for door adjustment of smart inspection trolley-based urban rail vehicle according to claim 1, characterized in that, In step S2, the V-shaped data refers to the size difference between the distance between the upper left and right door finger protection strips and the distance between the lower left and right door finger protection strips when the door is opened to the straightaway position. The straightaway position refers to the state where the door is opened and the distance between the left and right doors is greater than or equal to 500 mm. The front swing data refers to the distance between the upper front end of the door and the surface of the vehicle body when the door is opened to the straightaway position. The upper front end of the door refers to the upper edge of the door 150 mm away from the finger protection strip. The rear swing data refers to the distance between the upper rear end of the door and the surface of the vehicle body when the door is opened to the straightaway position. The upper rear end of the door refers to the upper edge of the door 150 mm away from the outer edge. The lower swing data refers to the distance between the lower front end of the door and the surface of the vehicle body when the door is opened to the straightaway position. The lower front end of the door refers to the lower outer edge of the door 150 mm away from the outer edge of the door.

3. The intelligent inspection method for door adjustment of smart inspection trolley-based urban rail vehicle according to claim 1, characterized in that, Step S3 specifically includes: The intelligent inspection trolley queries the intelligent cabinet number, the cabinet grid number, and the coordinates of the intelligent cabinet in the factory map according to the generated maintenance plan. The intelligent inspection trolley senses the surrounding environment in real time through its laser radar and / or camera, identifies lane lines and obstacles, and matches with the existing map. A safe path is planned according to the position of the intelligent inspection trolley and the position of the target intelligent cabinet. The intelligent inspection trolley travels along the safe path and reaches the preset parking point in front of the target intelligent cabinet.

4. The intelligent inspection method for door adjustment of smart inspection trolley-based urban rail vehicle according to claim 3, characterized in that, Step S4 specifically includes: The fixed camera on the intelligent inspection trolley collects maintenance tool pictures in real time, and inputs the collected pictures into the trained YOLOv8-nano lightweight model. The YOLOv8-nano lightweight model outputs the category and confidence of the maintenance tool, and determines whether it is the required maintenance tool. The 2D bounding box coordinates of the maintenance tool in the image are also outputted. The geometric center pixel coordinates of the 2D bounding box of the maintenance tool are obtained. According to the depth of each tool grid in the intelligent cabinet, the 3D coordinates of the maintenance tool in the fixed camera coordinate system are calculated by using the intrinsic parameters of the fixed camera on the intelligent inspection trolley through back projection. The 3D coordinates are converted to coarse coordinates in the trolley base coordinate system through the hand-eye matrix. The end of the mechanical arm is moved to a safe preparatory position in front of the maintenance tool according to the coarse coordinates.

5. The intelligent inspection method for door adjustment of smart inspection trolley-based urban rail vehicle according to claim 4, characterized in that, Step S5 specifically includes: When the mechanical arm is in place, the laser radar on the intelligent inspection trolley is triggered to perform a high-resolution scan on the grid where the maintenance tool is located, and a dense point cloud is obtained. The tool point cloud cluster and the 3D model in the knowledge base are matched using 3D CNN to output the 6D pose of the maintenance tool in the laser radar coordinate system. The 6D pose is converted to the mechanical arm base coordinate system using the radar extrinsic matrix to obtain the final grabbing coordinates. The mechanical arm receives the final grabbing coordinates, queries the weight and center of gravity data of the corresponding maintenance tool, calculates the optimal grabbing pose, jaw opening width, and required grabbing force, and grabs the maintenance tool. After obtaining the maintenance tool, the intelligent inspection trolley navigates to the maintenance point, and the maintenance tool is handed over to the maintenance personnel, and the maintenance personnel adjusts the vehicle door according to the maintenance scheme.

Images