A rail transit vehicle interior part defect automatic identification and interactive inspection method

Abstract

This invention belongs to the field of automated defect detection technology, specifically a method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles, comprising the following steps: S1, Data preparation: Collecting a large number of images of the interior of subway cars, labeling key targets, and generating a dataset; S2, Model training: Using a YOLO11n pre-trained model, performing transfer learning training on the dataset to obtain an interior component analysis model; S3, Edge deployment: Constructing a complete intelligent detection environment on edge computing hardware, including system environment configuration, model loading, and integration of image processing and defect identification algorithms; S4, Real-time detection: Inputting real-time images, quickly locating interior components through target detection, and performing in-depth analysis of stable targets using a dedicated interior component defect identification algorithm; S5, Interactive inspection: Based on the attributes of the defects and the inspectors, dynamically assigning defect handling tasks to the most suitable inspectors through an intelligent task allocation algorithm.

Description

Technical Field

[0001] This invention belongs to the field of automated defect detection technology, specifically a method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles. Background Technology

[0002] As the main parts that passengers directly contact and observe, the interior components of rail transit vehicles, including electrical control cabinets, safety signs, handrails, displays, fire extinguishers, and seats, directly impact the passenger experience and operational quality due to their appearance and structural safety. During long-term vehicle operation, regular inspections of these interior components are necessary to identify defects and perform targeted repairs or replacements. Achieving accurate identification of defects in various types of interior components, adapting to changes in the complex carriage environment, and efficiently collaborating with on-site inspection personnel have become important directions for technological development in this field.

[0003] Traditional inspection of interior components in rail transit vehicles mainly relies on manual visual inspection and cloud-based video analytics solutions, both of which have significant shortcomings. Manual inspection is inefficient, costly, and highly susceptible to human fatigue and experience, with inconsistent standards leading to missed inspections and misjudgments. Video analytics solutions are mostly deployed in the cloud, requiring the transmission of video streams from the carriages back to the server for processing, resulting in high response latency, strong dependence on vehicle-to-ground communication bandwidth, and susceptibility to failure in areas with unstable networks. Furthermore, the models are typically heavy, making it difficult to achieve real-time, low-power operation on edge devices with limited computing resources, thus failing to meet the inspection requirements for low-cost, highly reliable, and millisecond-level response times within the carriages.

[0004] Therefore, the present invention provides an automatic identification and interactive inspection method for defects in the interior components of rail transit vehicles. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: an automatic identification and interactive inspection method for defects in the interior components of rail transit vehicles, comprising the following steps:

[0007] S1. Data preparation: Collect a large number of images of the interior of subway cars under different lighting and angle conditions, label the key targets in the images, and generate a dataset in YOLO format;

[0008] S2. Model Training: The YOLO11n pre-trained model is trained on the dataset using transfer learning. By adjusting the hyperparameters, an internal component analysis model for identifying defects is obtained, and its format is converted.

[0009] S3. Edge Deployment: Build a complete intelligent inspection environment on edge computing hardware, including system environment configuration, model loading, image processing and defect recognition algorithm integration, forming an embedded inspection unit that can run independently in the carriage and complete the analysis and judgment of the interior component status in real time;

[0010] S4. Real-time detection: Input a real-time image, quickly locate the internal components through target detection, confirm the stability of the target by multi-frame tracking, and then perform in-depth analysis of the stable target through a dedicated internal component defect identification algorithm.

[0011] S5. Interactive Inspection: When an interior component defect is detected, the defect handling task is dynamically assigned to the most suitable inspector based on the defect and the inspector's attributes through an intelligent task allocation algorithm. The inspector arrives at the site, confirms and handles the defect, and simultaneously feeds the results back to the system.

[0012] Preferably, the interior component analysis model in step S2 employs an adaptive spatial attention mechanism:

[0013]

[0014] The interior component analysis model is trained end-to-end using a weighted multi-task loss function:

[0015]

[0016] Preferably, the defect identification algorithm in step S4 includes detecting the closed state of the electrical control cabinet door:

[0017] Within the electrical control cabinet area, door gap detection is performed using the orientation-sensitive Canny edge detection algorithm:

[0018]

[0019] Pair the points on the two parallel edges sequentially or according to the nearest neighbor principle, and calculate the average door gap width based on the distance between each pair of points:

[0020]

[0021] Cabinet door closure status determination:

[0022]

[0023] Preferably, the defect identification algorithm in step S4 further includes safety sign curling edge detection:

[0024] The safety marking area is divided into blocks, and each grid area is fitted to a plane using RANSAC.

[0025] Calculate the overall flatness deviation of the markings:

[0026]

[0027] Analyze the gradient discontinuity between the marker edge and the background:

[0028]

[0029] Calculate the overall probability of the label curling up:

[0030]

[0031] Preferably, the defect identification algorithm in step S4 further includes handrail breakage detection:

[0032] The center line of the armrest is obtained using a morphological skeletonization algorithm:

[0033]

[0034] Calculate the curvature at each point on the center line:

[0035]

[0036] Based on both curvature abrupt change and location information, the overall probability of handrail breakage is calculated:

[0037]

[0038] Preferably, the defect identification algorithm in step S4 further includes display screen crack detection:

[0039] Calculate the radial gradient centered on the suspected rupture point:

[0040]

[0041] Analyzing the confidence level of cracks:

[0042]

[0043] Analysis of orientation field consistency using a spider web crack detection algorithm:

[0044]

[0045] Preferably, the defect identification algorithm in step S4 further includes fire extinguisher safety pin presence detection:

[0046] Morphological hit-and-miss transformation using pin-shaped structural elements:

[0047]

[0048] Verify the spatial relationship between the pin and the fire extinguisher handle:

[0049]

[0050] Preferably, the defect identification algorithm in step S4 further includes seat looseness detection:

[0051] Calculate the alignment deviation between the seat base and the floor mounting holes:

[0052]

[0053] Calculate the tilt angle between the seat plane and the floor plane using the normal vector of the seat plane:

[0054]

[0055] Preferably, the intelligent task allocation algorithm in step S5 is specifically as follows:

[0056] Each inspector is considered an agent with the following attribute vector:

[0057]

[0058] Each defect to be processed is defined as:

[0059]

[0060] A multi-objective optimization allocation strategy is used to dynamically assign defects to inspectors for processing.

[0061] Preferably, the multi-objective optimization allocation strategy is as follows:

[0062] Transform the task allocation problem into a multi-objective decision-making process:

[0063]

[0064] Objective 1: Minimize total response time.

[0065]

[0066] Objective 2: Minimize skill mismatch:

[0067]

[0068]

[0069] The beneficial effects of this invention are as follows:

[0070] 1. The present invention provides an automatic identification and interactive inspection method for defects in interior components of rail transit vehicles. From data acquisition, model training, edge deployment to real-time detection and interactive inspection, a complete automated inspection closed-loop system is constructed. It adopts lightweight models such as YOLO11n combined with edge computing hardware to realize an embedded detection unit that runs independently in the carriage. The system has a high degree of integration, realizes full-process automation and intelligence, supports real-time analysis and judgment, significantly reduces dependence on the central server, and improves the system's response speed and reliability.

[0071] 2. The automatic identification and interactive inspection method for defects in the interior components of rail transit vehicles described in this invention designs a variety of dedicated defect identification algorithms to address the diversity and special characteristics of rail transit interior components. These include cabinet door closure judgment based on direction-sensitive edge detection, safety sign warping detection combining RANSAC plane fitting and gradient analysis, handrail breakage detection based on morphological skeleton and curvature analysis, and display screen breakage recognition incorporating radial gradient and directional field consistency judgment. All these algorithms closely adhere to the physical characteristics and manifestations of actual defects, resulting in high detection accuracy and strong adaptability.

[0072] 3. The automatic identification and interactive inspection method for defects in the interior components of rail transit vehicles described in this invention designs an intelligent task allocation mechanism, modeling each inspector as an intelligent agent with multiple attributes such as location, skills, and load. Combining information such as defect type, location, and urgency, the method dynamically assigns tasks through a multi-objective optimization algorithm. This method not only minimizes the overall response time but also takes into account skill matching, thereby improving the efficiency and quality of defect handling and achieving efficient utilization of inspection resources. Attached Figure Description

[0073] The invention will now be further described with reference to the accompanying drawings.

[0074] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0075] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0076] like Figure 1 As shown, the present invention provides an automatic identification and interactive inspection method for defects in the interior components of rail transit vehicles, comprising the following steps:

[0077] S1. Data preparation: Collect a large number of images of the interior of subway cars under different lighting and angle conditions, label key targets in the images (including electrical control cabinets, safety signs, handrails, displays, fire extinguishers, seats, etc.), and generate a dataset in YOLO format.

[0078] S2. Model Training: Use the YOLO11n pre-trained model to perform transfer learning training on the dataset. By adjusting the hyperparameters, an internal component analysis model for identifying defects is obtained (.pt format).

[0079] Using Rockchip's RKNN-Toolkit2 tool, the trained .pt model is converted into the .rknn format, which can be efficiently inferred on the NPU of the RK3588 chip. This process may include model quantization (such as quantization from FP32 to INT8) to further reduce the model size, improve inference speed, and at the same time maintain accuracy as much as possible.

[0080] S3. Edge Deployment: Build a complete intelligent inspection environment on edge computing hardware, including system environment configuration, model loading, image processing and defect recognition algorithm integration, forming an embedded inspection unit that can run independently in the carriage and complete the analysis and judgment of the interior component status in real time;

[0081] Environment setup: Deploy the corresponding Linux operating system, drivers, and RKNN inference library on the RK3588 development board;

[0082] Program Development: Write a Python application that implements the following functionalities:

[0083] Call the camera interface to obtain a real-time video stream (single frame image);

[0084] Load the converted .rknn model;

[0085] Implement preprocessing (such as image scaling and normalization) and postprocessing (decoding detection boxes and non-maximum suppression NMS);

[0086] It integrates functional detection algorithms for different "inspection items". For example, after detecting a "safety sign", a dedicated algorithm can be run to determine whether the sign has a curled edge.

[0087] S4. Real-time detection: Input a real-time image, which is sent to the RK3588 chip. The NPU efficiently runs the YOLO11n model and outputs all detected targets and their positions (bounding boxes). The internal components are quickly located through target detection. Multi-frame tracking is used to confirm the stability of the targets. Then, for each detected target, the corresponding functional detection algorithm is called for in-depth analysis according to its category.

[0088] S5. Interactive Inspection: When an interior component defect is detected, the defect handling task is dynamically assigned to the most suitable inspector based on the defect and the inspector's attributes through an intelligent task allocation algorithm. The inspector arrives at the site, confirms and handles the defect, and simultaneously feeds the results back to the system.

[0089] Alternatives to object detection models: YOLO11n is preferred in this invention, but other lightweight object detection models (such as YOLOv5s, YOLOv8n, YOLOv10n, SSD-MobileNet, NanoDet, etc.) can also be used to achieve the purpose of this invention. The core idea is "lightweight model + edge deployment".

[0090] Edge computing chip alternatives: The RK3588 is preferred in this invention, but other edge computing chips with NPUs (such as NVIDIA Jetson Nano / Orin NX, Huawei Atlas, and chips adapted to Google Coral TPU) are also feasible alternatives.

[0091] Fine-tuning of deployment mode: It can be "one machine per train" (one main controller for the entire train) or "one machine per carriage" (one RK3588 per carriage), the latter has higher real-time performance and better fault tolerance.

[0092] As one embodiment of the present invention, the internal component analysis model in step S2 adopts an adaptive spatial attention mechanism:

[0093]

[0094] in, , where is the attention weight, representing the intensity of attention assigned to the key at position j by the query at position i; and These are the query vector and the key vector, respectively. The dimension of the key vector; Based on the prior location of the device The learnable bias guides the network to focus on critical device areas;

[0095] The interior component analysis model is trained end-to-end using a weighted multi-task loss function:

[0096]

[0097] in, Total loss; To detect the loss; Loss is categorized by defect type; For segmentation loss; This is the regularization loss; , , , These are the weighting coefficients.

[0098] As one embodiment of the present invention, the defect identification algorithm in step S4 includes detection of the closed state of the electrical control cabinet door:

[0099] Within the electrical control cabinet area, door gap detection is performed using the orientation-sensitive Canny edge detection algorithm:

[0100]

[0101] This formula calculates the projection intensity of the image gradient in a specific direction. Cabinet door seams are typically straight lines; by selecting an appropriate direction… This can enhance the detection effect of door seam edges, vertically ( =90°) is suitable for detecting vertical door gaps, and horizontal ( =0°) is suitable for detecting horizontal door gaps;

[0102] Pair the points on the two parallel edges sequentially or according to the nearest neighbor principle, and calculate the average door gap width based on the distance between each pair of points:

[0103]

[0104] Cabinet door closure status determination:

[0105]

[0106] in, For pixels In direction The edge strength in that direction; a larger value indicates a stronger edge in that direction. Let be the gradient vector of image I; The preset unit direction vector for the door gap; Let be the coordinate vector of the i-th point on the first edge; Let be the coordinate vector of the j-th point on the second edge; , These are the minimum and maximum threshold values ​​for the door gap width.

[0107] As one embodiment of the present invention, the defect identification algorithm in step S4 further includes safety sign warping detection:

[0108] The safety marking area is divided into blocks (e.g., a 4×4 grid), and each grid area is fitted to a plane using RANSAC.

[0109] Calculate the overall flatness deviation of the markings; a larger value indicates a more uneven surface.

[0110]

[0111] Analyze the gradient discontinuity between the marker edge and the background:

[0112]

[0113] Calculate the overall probability of the label curling up:

[0114]

[0115] in, The distance from the i-th pixel within the k-th block to the fitted plane; This is an edge separation index; a positive value indicates that the internal gradient is greater than the external gradient, which may indicate warped edges. Gradient vector field to identify the internal edge region; To identify the gradient vector field of the immediate external region; For very small positive numbers, such as 10 -6 To prevent the denominator from being zero; The Sigmoid function maps linear combinations to interval; , , These are the weighting coefficients; The weight of the number of times the edge has historically curled at this position.

[0116] When the edge of the logo curls up, a height difference is created between the logo edge and the background, resulting in a noticeable brightness change (shadow or highlight) at the inner edge, while the outer adjacent area (background surface) is usually more uniform with a smaller gradient, thus the curled area... , It is positive and relatively large.

[0117] As one embodiment of the present invention, the defect identification algorithm in step S4 further includes handrail breakage detection:

[0118] The center line of the armrest is obtained using a morphological skeletonization algorithm:

[0119]

[0120] The centerline coordinate set is a sequence of N points, each represented by its image coordinates. express;

[0121] Calculate the curvature at each point on the centerline; points where curvature abruptly changes may be breakpoints.

[0122]

[0123] The above derivatives are calculated using the Savitzky-Golay filter;

[0124] Based on both curvature abrupt change and location information, the overall probability of handrail breakage is calculated:

[0125]

[0126] in, Let be the curvature value of the i-th point on the center line. The larger the curvature, the more drastic the change in direction at that point. This is an indicator function. The function value is 1 when the condition in parentheses is true, and 0 otherwise. It is used here to count the number of outliers where the curvature exceeds the threshold. This is the curvature threshold; if the curvature exceeds this threshold, it is considered that the point has undergone a sudden change in direction, which may be a fracture feature. The path length along the centerline from the i-th point on the centerline to the nearest support point (such as the connection between the handrail and the wall or column); This is the distance attenuation coefficient;

[0127] The essence of handrail breakage is the disruption of structural continuity. This algorithm determines the likelihood of breakage by identifying points of sharp bending (high curvature points) in the geometry of the centerline and combining this with the prior physical assumption that "breakage is more likely to occur in the suspended segment between two support points." The higher the value, the more certain the system is that the handrail has a fracture defect.

[0128] As one embodiment of the present invention, the defect identification algorithm in step S4 further includes display screen crack detection:

[0129] Calculate the radial gradient centered on the suspected rupture point:

[0130]

[0131] This formula calculates the intensity of brightness variation in the direction of outward radiation from the suspected fracture point, and is used to capture the characteristics of linear cracks diverging outward from the fracture center.

[0132] Analyzing the confidence level of cracks:

[0133]

[0134] Cracks typically exhibit a significant gradient in one direction and a smaller gradient in the opposite direction. This formula quantifies crack characteristics by comparing the gradient differences in fan-shaped regions in opposite directions.

[0135] Analysis of orientation field consistency using a spider web crack detection algorithm:

[0136]

[0137] Spider web cracks are formed by cracks interwoven in multiple directions, and their directional fields are not uniform, while the directional fields of a single crack are relatively uniform. This formula distinguishes between these two fracture modes by quantifying the degree of directional dispersion.

[0138] in, For pixels The radial gradient intensity at a given point represents the rate of change of image brightness along the radial direction; To point from the suspected rupture center to the current point ; This is a crack confidence index; the higher the value, the higher the probability that a crack exists. , The minimum and maximum radii are the radial ranges to be analyzed; For a fan-shaped area The average radial gradient of all points within the area; This is the sum of the absolute values ​​of the radial gradients at all analysis points; It is a very small positive number; As an index for directional field consistency; The local gradient direction angle of the i-th pixel; The average orientation angle of the analysis area;

[0139] The design logic of the screen crack detection described above is based on the unique geometric shape and texture features of cracks. High-precision identification is achieved through multi-cascade analysis: first, radial gradient analysis is used to capture the linear crack features that diverge outward from the suspected crack point, and the gradient difference of the fan-shaped region in the opposite direction is calculated to quantify the crack confidence; then, orientation field consistency analysis is used to distinguish the different modes of single cracks and spider web cracks, and the degree of dispersion of local gradient directions is used to determine the crack morphology.

[0140] As one embodiment of the present invention, the defect identification algorithm in step S4 further includes fire extinguisher safety pin presence detection:

[0141] The design logic of the fire extinguisher safety pin in-place detection is based on a dual verification mechanism of shape matching and context constraint. Its function is to reliably determine whether the safety pin is installed in place through accurate shape recognition and spatial relationship analysis.

[0142] Morphological hit-and-miss transformation using pin-shaped structural elements:

[0143]

[0144] Verify the spatial relationship between the pin and the fire extinguisher handle:

[0145]

[0146] in, The output of the hit-and-miss transform indicates the region that matches both the foreground and background structures. Structural elements in the shape of a safety pin, such as a long, thin rectangle, simulate the shape of a safety pin; This is the complement of image I (the background region); The structural elements surrounding the safety pin are typically slightly larger than the pin and form a ring. Indicates the use of structural elements Perform an erosion operation on image I; Indicates the use of background structural elements Erosion of the background image; The result is either 0 or 1 to verify the information. This represents the Euclidean distance between the detected center point of the safety pin and the reference point of the fire extinguisher handle. This is the maximum allowed distance threshold; The direction and angle of the detected safety pin; , The minimum and maximum permissible safety pin direction angle range;

[0147] The algorithm first utilizes the hit-and-miss transformation (HMT) morphological method, while simultaneously matching the slender shape features of the safety pin with its surrounding background structure, to achieve initial pin localization. Then, through spatial relationship constraint verification, it ensures that the detected object is within a reasonable distance from the fire extinguisher handle and that its orientation and angle conform to installation specifications, effectively eliminating false detections of similar-shaped objects. This dual verification mechanism integrates local feature recognition and global context understanding, focusing not only on the morphological characteristics of the pin itself but also on its correct positional relationship within the overall structure of the fire extinguisher. Ultimately, it achieves automated and highly reliable detection of the safety pin status in complex vehicle environments, ensuring that the fire extinguishing equipment is in a safe and readily available state.

[0148] As one embodiment of the present invention, the defect identification algorithm in step S4 further includes seat looseness detection:

[0149] The design logic for detecting seat fixation looseness is based on a dual verification mechanism of connection point geometric alignment analysis and planar posture stability assessment. Its function is to accurately determine the degree of looseness of the seat fixation structure by quantifying the positional deviation between the seat mounting point and the floor fixing point and the tilt angle of the seat plane relative to the floor.

[0150] Calculate the alignment deviation between the seat base and the floor mounting holes:

[0151]

[0152] Calculate the tilt angle between the seat plane and the floor plane using the normal vector of the seat plane:

[0153]

[0154] in, Let i be the coordinate vector of the i-th seat mounting point; Let i be the coordinate vector of the i-th floor installation point; Let be the normal vector of the seat plane; This is the normal vector of the floor plane;

[0155] The algorithm described above uses computer vision technology to accurately detect the positions of the four mounting points of the seat and their corresponding fixed points on the floor, and calculates the average alignment deviation. To assess the physical displacement at the connection point; simultaneously, by analyzing the three-dimensional geometric features of the seat surface, its planar normal vector is fitted and compared with the floor reference normal vector to calculate the tilt angle. This dual-indicator system assesses the overall stability of the seat posture. It focuses on both the micro-displacement of local connection points and the macro-posture changes of the overall structure, effectively distinguishing between normal wear and tear and dangerous loosening. This allows for automated and quantitative detection of seat fixation safety without relying on human experience, providing reliable technical support for the structural health monitoring of passenger seats in rail transit vehicles and preventing passenger safety hazards caused by seat loosening.

[0156] As one embodiment of the present invention, the intelligent task allocation algorithm in step S5 is specifically as follows:

[0157] Each inspector is considered an agent with the following attribute vector:

[0158]

[0159] in, Let be the current coordinates of inspector i; The proficiency of inspector i in handling type j defects; The current task load (estimated time required to complete the existing task); Type of agent (primary / intermediate / advanced); Movement speed; Status: Working (Idle / Working / Rest);

[0160] Each defect to be processed is defined as:

[0161]

[0162] in, Let J be the coordinates of the j-th type of defect; Defect type requirement vector; To address the urgency; For detecting timestamps; For historical processing and recording;

[0163] A multi-objective optimization allocation strategy is used to dynamically assign defects to inspectors for processing.

[0164] As one embodiment of the present invention, the multi-objective optimization allocation strategy is specifically as follows:

[0165] Transform the task allocation problem into a multi-objective decision-making process:

[0166]

[0167] in, The decision variable matrix; The first objective function is to minimize the total response time. The second objective function is to minimize the skill mismatch.

[0168] The formula means: to find a set of decision variables. This makes the two objective functions , Simultaneously achieve minimization;

[0169] Objective 1: Minimize total response time.

[0170]

[0171] in, The decision variable is used to indicate whether task j is assigned to inspector i. Allow inspector i the time needed to handle task j (familiarize themselves with the situation, prepare tools, etc.);

[0172] This objective ensures that the system can respond quickly to defect alarms, minimize the total time from task assignment to the arrival of inspectors on site, improve the efficiency of handling emergency defects, and prevent small problems from escalating into major failures.

[0173] Objective 2: Minimize skill mismatch:

[0174]

[0175]

[0176] in, To ensure that inspectors possess basic processing capabilities, the minimum skill requirement is set. The attenuation coefficient controls the severity of the penalty.

[0177] Since the objective function needs to be minimized, the problem of maximizing skill matching degree is transformed into the problem of minimizing mismatch degree by using a negative sign. This objective ensures that tasks are assigned to the inspectors who are best at handling this type of defect, thereby improving the quality and efficiency of processing and avoiding assigning high-difficulty tasks to inspectors with insufficient skills, thus reducing the risk of rework and mishandling.

[0178] This multi-objective optimization allocation strategy abstracts each inspector as an intelligent agent with attributes such as location, skill, current load, and movement speed, and each defect task as a task entity with attributes such as location, skill requirements, processing time, and urgency. It uses mathematical optimization methods to find the optimal allocation scheme that balances multiple competing objectives. While ensuring that all defect tasks are handled by qualified inspectors, it minimizes the overall response time to improve emergency handling efficiency and maximizes skill matching to guarantee processing quality. This strategy forms the intelligent decision-making core of the human-machine collaborative inspection system, enabling dynamic adjustment of the allocation strategy based on real-time status, significantly improving the overall efficiency, quality, and fairness of defect handling for rail transit interior components.

[0179] The terms "front," "back," "left," "right," "top," and "bottom" all refer to the figures in the accompanying drawings. Figure 1 Based on the perspective of the observer, the side of the device facing the observer is defined as the front, the left side of the observer is defined as the left, and so on.

[0180] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0181] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic identification and interactive inspection of defects in interior components of rail transit vehicles, characterized in that: Includes the following steps: S1. Data preparation: Collect a large number of images of the interior of subway cars under different lighting and angle conditions, label the key targets in the images, and generate a dataset in YOLO format; S2. Model Training: Use the YOLO11n pre-trained model to perform transfer learning training on the dataset. By adjusting the hyperparameters, obtain an internal component analysis model for identifying defects, and convert its format. S3. Edge Deployment: Build a complete intelligent inspection environment on edge computing hardware, including system environment configuration, model loading, image processing and defect recognition algorithm integration, forming an embedded inspection unit that can run independently in the carriage and complete the analysis and judgment of the interior component status in real time; S4. Real-time detection: Input a real-time image, quickly locate the internal components through target detection, confirm the stability of the target by multi-frame tracking, and then perform in-depth analysis of the stable target through a dedicated internal component defect identification algorithm. S5. Interactive Inspection: When an interior component defect is detected, the defect handling task is dynamically assigned to the most suitable inspector based on the defect and the inspector's attributes through an intelligent task allocation algorithm. The inspector arrives at the site, confirms and handles the defect, and simultaneously feeds the results back to the system.

2. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The interior component analysis model described in step S2 employs an adaptive spatial attention mechanism: in, and These are the query vector and the key vector, respectively. The dimension of the key vector; Based on the prior location of the device Learnable bias; The interior component analysis model is trained end-to-end using a weighted multi-task loss function: in, To detect the loss; Loss is categorized by defect type; For segmentation loss; This is the regularization loss; , , , These are the weighting coefficients.

3. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 includes detecting the closed state of the electrical control cabinet door: Within the electrical control cabinet area, door gap detection is performed using the orientation-sensitive Canny edge detection algorithm: Pair the points on the two parallel edges sequentially or according to the nearest neighbor principle, and calculate the average door gap width based on the distance between each pair of points: Cabinet door closure status determination: in, For pixels In direction Edge strength; Let be the gradient vector of image I; The preset unit direction vector for the door gap; Let be the coordinate vector of the i-th point on the first edge; Let be the coordinate vector of the j-th point on the second edge; , These are the minimum and maximum threshold values ​​for the door gap width.

4. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 also includes safety sign curling edge detection: The safety marking area is divided into blocks, and each grid area is fitted to a plane using RANSAC. Calculate the overall flatness deviation of the markings: Analyze the gradient discontinuity between the marker edge and the background: Calculate the overall probability of the label curling up: in, The distance from the i-th pixel within the k-th block to the fitted plane; This is an indicator of edge separation. Gradient vector field to identify the internal edge region; To identify the gradient vector field of the immediate external region; It is a very small positive number; , , These are the weighting coefficients; The weight of the number of times the edge has historically curled at this position.

5. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 also includes handrail breakage detection: The center line of the armrest is obtained using a morphological skeletonization algorithm: Calculate the curvature at each point on the center line: Based on both curvature abrupt change and location information, the overall probability of handrail breakage is calculated: in, Let be the curvature value at the i-th point on the center line; Curvature threshold; Let be the path length along the centerline from the i-th point on the centerline to the nearest support point. This is the distance attenuation coefficient.

6. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 also includes display screen crack detection: Calculate the radial gradient centered on the suspected rupture point: Analyzing crack confidence levels: Analysis of orientation field consistency using a spider web crack detection algorithm: in, For pixels Radial gradient intensity at the location; To point from the suspected rupture center to the current point ; This is a confidence index for cracks; , The minimum and maximum radii are the radial ranges to be analyzed; It is a very small positive number; As an index for directional field consistency; The local gradient direction angle of the i-th pixel; This represents the average orientation angle of the analysis area.

7. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 also includes fire extinguisher safety pin presence detection: Morphological hit-and-miss transformation using pin-shaped structural elements: Verify the spatial relationship between the pin and the fire extinguisher handle: in, Structural elements in the shape of a safety pin; Structural elements surrounding the safety pin; This is the maximum allowed distance threshold; The direction and angle of the detected safety pin; , The minimum and maximum permissible safety pin direction angle range.

8. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The defect identification algorithm in step S4 also includes seat looseness detection: Calculate the alignment deviation between the seat base and the floor mounting holes: Calculate the tilt angle between the seat plane and the floor plane using the normal vector of the seat plane: in, Let i be the coordinate vector of the i-th seat mounting point; Let i be the coordinate vector of the i-th floor installation point; This is the normal vector of the seat plane; Let be the normal vector of the floor plane.

9. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 1, characterized in that: The intelligent task allocation algorithm described in step S5 is as follows: Each inspector is considered an agent with the following attribute vector: in, Let be the current coordinates of inspector i; The proficiency of inspector i in handling type j defects; Current task load; The type of intelligent agent; Movement speed; In working status; Each defect to be processed is defined as: in, Let J be the coordinates of the j-th type of defect; Defect type requirement vector; To address the urgency; For detecting timestamps; For historical processing records; A multi-objective optimization allocation strategy is used to dynamically assign defects to inspectors for processing.

10. The method for automatic identification and interactive inspection of defects in the interior components of rail transit vehicles according to claim 9, characterized in that: The multi-objective optimization allocation strategy is specifically as follows: Transform the task allocation problem into a multi-objective decision-making process: in, The decision variable matrix; This is the first objective function; This is the second objective function; Objective 1: Minimize total response time. in, To assign decision variables; Prepare the time required for inspector i to handle task j; Objective 2: Minimize skill mismatch: in, This represents the minimum skill requirement. This is the attenuation coefficient.

Images