Automobile underbody fault detection robot system based on double models
The dual-model collaborative automotive undercarriage fault detection robot system, combining the ITSCMN model and the FR-YOLOv12 model, solves the problems of low efficiency and insufficient accuracy in undercarriage fault detection, achieving high-precision and intelligent detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN JIAOTONG UNIVERSITY
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies for detecting faults under vehicles suffer from low efficiency, insufficient accuracy, and susceptibility to environmental influences, making it particularly difficult to achieve intelligent and high-precision detection in complex environments.
A dual-model-based automotive undercarriage fault detection robot system is adopted, which combines an improved ITSCMN model for mapping and navigation with an FR-YOLOv12 model for target detection. By integrating a power supply unit, an environmental perception unit, a motion control unit, a visual sensing unit, and an AI voice interaction unit, automated and high-precision detection is achieved.
It improves the accuracy and real-time performance of detection, simplifies the operation process, reduces labor costs, significantly improves detection efficiency and accuracy, and realizes intelligent detection of undercarriage faults.
Smart Images
Figure CN121898376A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent vehicle fault detection technology, and more specifically, to a dual-model-based robot system for detecting vehicle undercarriage faults. Background Technology
[0002] With the rapid development of the automotive industry and the continuous increase in vehicle ownership, the undercarriage structure is constantly exposed to a complex environment characterized by humidity, dust, and high corrosion, making it highly susceptible to various malfunctions such as loosening, rust, cracks, and foreign object adhesion. Failure to detect and address these malfunctions in a timely manner will directly affect vehicle stability and seriously threaten driving safety. Therefore, accurate detection of undercarriage malfunctions is of great significance for ensuring travel safety.
[0003] Traditional vehicle undercarriage inspection methods rely on manual lifting of the vehicle for visual inspection. This is not only inefficient and labor-intensive, but also susceptible to subjective factors such as the inspector's experience and sense of responsibility, leading to a high rate of missed or false faults. Although semi-automatic inspection robots have emerged, alleviating the pressure of manual inspection to some extent, the results are still unstable due to the confined space under the vehicle, uneven lighting, complex component textures, and mutual obstruction, making it difficult to achieve intelligent and high-precision inspection.
[0004] Early vehicle undercarriage fault detection algorithms primarily relied on manually designed features, using sliding windows combined with classifiers such as SVM and Adaboost to achieve fault identification. However, these algorithms had limited feature representation capabilities in complex lighting and noisy environments under vehicles, making it difficult to accurately capture fault-related features and resulting in poor detection performance. In recent years, deep learning algorithms have been widely used in the field of object detection, with the YOLO series algorithms becoming the mainstream detection solution due to their end-to-end architecture and high real-time performance. Among them, two-stage methods such as Faster R-CNN have high detection accuracy but slow operation speed, making it difficult to meet real-time detection requirements; single-stage methods such as YOLOv5, YOLOv7, and YOLOv8 balance accuracy and efficiency, making them suitable for embedded deployment scenarios, but they still have significant shortcomings in complex environments under vehicles: weak feature extraction capabilities under low lighting conditions, poor recognition of periodic structures and small defects, insufficient multi-scale feature fusion, and the detection system and navigation system are independent with insufficient coordination, making them unable to adapt to the special needs of vehicle undercarriage detection.
[0005] To address the challenges of detection in complex environments, various improvement schemes have been proposed, but many limitations remain. In low-light scenarios, some schemes improve detection performance by introducing low-light enhancement networks, but they rely excessively on image preprocessing and lack optimization for feature fusion. Improved algorithms for hazy environments can achieve defogging enhancement or improve detection accuracy, but they suffer from complex model structures and excessive computational load, affecting real-time performance. Furthermore, some schemes are only applicable to small sample scenarios and lack generalization ability; some focus on traffic video detection or long-distance small target recognition, and have poor adaptability to the detection of complex structures under vehicles at close range; others do not consider the problem of occlusion in narrow spaces, or only use single-sensor vision schemes without involving multi-sensor fusion, making it difficult to solve core issues such as sensor collaboration and real-time performance.
[0006] Therefore, there is an urgent need for a dual-model-based automotive undercarriage fault detection robot system that can overcome the pain points of detection in complex environments through the collaborative innovation of mapping and navigation with target detection models, and achieve intelligent and efficient detection of undercarriage faults, providing reliable technical support for the field of intelligent automotive detection. Summary of the Invention
[0007] This invention provides a dual-model-based robot system for detecting faults under vehicles, which aims to solve the problems of low accuracy and efficiency of manual inspection in the geometric measurement of modular concrete room units due to spatial complexity and task diversity.
[0008] Furthermore, existing point cloud-based building component quality inspection technologies rely on manual intervention to set the measurement start and end points, making it difficult to achieve automated geometric positioning of measurement tasks. Therefore, this invention aims to achieve high-precision automated measurement of complex room structures by combining planar segmentation of point cloud data, graph structure reasoning, and measurement task chains, thereby improving the flexibility, interpretability, and accuracy of measurements and meeting the needs of high-precision quality inspection of modular buildings.
[0009] To achieve the above objectives, this invention provides a dual-model-based automotive undercarriage fault detection robot system, specifically comprising: The robot entity includes: an integrated power supply unit, a model deployment and display unit, an environmental perception unit, a motion control unit, a vision sensing unit, and an AI voice interaction unit; The remote backend connects to the robot entity via a data communication unit. The data communication unit is integrated into the robot body to enable data interaction between the robot body and the remote backend. The model deployment and display unit in the robot entity includes the improved spatiotemporal collaborative mapping and navigation model ITSCMN and the frequency domain-rectangular self-calibrated YOLOv12 model FR-YOLOv12. The ITSCMN model is used to improve the accuracy of mapping and navigation, providing the robot with more accurate environmental perception capabilities, thereby constructing a more accurate navigation map and providing the robot with better mapping and localization in complex environments under the vehicle. The automotive undercarriage fault detection robot system achieves automated and high-precision detection of automotive undercarriage faults through the collaboration of dual models and various units.
[0010] Preferably, the integrated power supply unit integrates the power supply drive interfaces required by all vehicle-mounted units, realizing integrated power supply for multiple units, greatly improving power supply efficiency and saving vehicle interior space resources; The model deployment and display unit is used for model execution and result visualization; The environmental perception unit acquires information about the surrounding environment and constructs a high-precision navigation map; The motion control unit performs robot movement and posture control; The visual sensing unit is used to collect image data of the vehicle's underside; the AI voice interaction unit is used to realize human-machine voice command interaction.
[0011] Preferably, the functions implemented by the remote backend include: issuing vehicle undercarriage inspection task instructions, viewing the real-time condition of the vehicle undercarriage and inspection progress, performing fault analysis and repair quotation, and providing data storage space for the robot entity; The data communication unit integrates multiple transmission methods such as wireless communication module and serial communication module, and supports multiple communication protocols; at the same time, it takes into account communication redundancy and communication link switching mechanism, so that even in low-quality communication environment, it can still ensure good communication between the robot and the remote back-end and host computer.
[0012] Preferably, the detection process for vehicle undercarriage faults includes: S1. After the system powers on and performs a self-test, the model deployment and display unit starts the ITSCMN model. The environmental perception unit sends the real-time data from the lidar and IMU to the ITSCMN model via the data communication unit to complete the high-precision mapping of the under-vehicle environment and the initial positioning of the robot. S2. The remote backend issues the detection task through the data communication unit. The motion control unit plans the path based on the high-precision mapping and robot initial positioning provided by the ITSCMN model, and drives the robot to autonomously drive to the designated starting position at the bottom of the vehicle to be inspected. S3. After reaching the starting position, the AI voice interaction unit announces "Start detection", the motion control unit adjusts the robot's posture, and the vision sensing unit continuously collects multi-view images of the underside of the vehicle under the power supply unit and sends the image data back to the model deployment and display unit. S4. The model deployment and display unit calls the FR-YOLOv12 model to perform real-time inference on the image, outputting the fault category, location coordinates and confidence level, forming a structured detection result; S5. The detection results are visualized locally by the model deployment and display unit, and the AI voice interaction unit simultaneously announces "Detection complete, a certain type of fault has been found"; S6. The detection results are uploaded to the remote backend through the data communication unit. The remote backend completes data storage, fault analysis and maintenance decision-making, realizing automated and high-precision detection of vehicle undercarriage faults.
[0013] Preferably, the ITSCMN model includes: an improved Cartographer algorithm and an improved anisotropic kriging interpolation algorithm; S11. The improved Cartographer algorithm is used to achieve more accurate perception of complex environments under vehicles and to construct two-dimensional grid maps. It uses the timestamp alignment-online time offset calibration mechanism TAOTOC to construct message preprocessing nodes and improves mapping accuracy through the temporal and spatial collaboration of multi-source data. S12. The improved anisotropic kriging interpolation algorithm is proposed based on the improved Cartographer algorithm to obtain input data. According to the dynamic weighting principle of the correlation structure of adjacent points, it can achieve accurate interpolation of pose and efficiently remove motion distortion caused by the difficulty in matching the sampling frequency of the robot body pose with the sampling frequency of the radar laser beam.
[0014] Preferably, the improved Cartographer algorithm specifically includes: S111. Initialize the data queues of each active sensor and input the data collected by each active sensor into the corresponding data queue; The sensors include lidar and inertial measurement units (IMUs). S112. If only a single lidar has data input, only the corresponding data queue is updated, and the synchronization process is not executed; if both lidars collect data, not only are the corresponding data sequences updated, but the synchronization process is also started. S113. During the synchronization process, compare the timestamps of the data to be synchronized with the latest historical input data of each sensor, select only the data with updated timestamps to participate in the synchronization, and filter out outdated data. S114. Construct a message preprocessing node based on the TAOTOC mechanism; S115. After time synchronization is completed, the data is sent to the Cartographer backend module to participate in real-time positioning and map building calculations, generating a high-precision two-dimensional raster map of the vehicle's underside environment.
[0015] Preferably, the specific content of S114 includes: S1141. Set two LiDARs with consistent timing as reference sensors, and establish a unified reference timestamp reference_timestamp with the first data timestamp. S1142. Remap the timestamps of the data from the non-reference sensor according to the reference_timestamp and its own sampling rate, and then send them to the mapping function node cartographer_ros to participate in the calculation of the real-time positioning and map building system. S1143. Using the lidar timestamp as a unified reference time axis, the timing consistency of multi-sensor data during robot motion is analyzed online. The time offset of non-reference sensors relative to the reference lidar is dynamically estimated and compensated to ensure that the data of the dual lidars and IMU are aligned in the time dimension.
[0016] Preferably, the improved anisotropic kriging interpolation algorithm includes: S121, Solving for the pose function: S1211, Regarding the sampling time Obtain the true robot pose value output by the pose measurement sensor. ; S1212. To achieve high-precision interpolation of robot pose samples, the robot pose change during the scanning period is considered as a random process, and Kriging interpolation is used to obtain... pose estimate at time 1 ; S1213. Define pose estimation error and derive the formula for pose estimation error variance. S1214, Let the variance of the pose estimation error be correlated with the Kriging weight vector. The partial derivatives are 0, and the target weight vector is obtained by solving. S1215. Based on the target weight vector, calculate respectively Time Robot Location, Location and orientation angle pose estimation value , , ; S122, Laser point motion compensation: S1221. Let the reference time be... For any laser beam at time Collected laser points ( Based on the estimated pose values, a true pose matrix is constructed. ; S1222. Project the laser point onto the reference time coordinate system to obtain the true coordinates after distortion correction. Intra-frame motion compensation is completed.
[0017] Preferably, in S1211, the expression for the true value of the robot pose is: ; in, Represents the sampling time The actual robot pose value is obtained from the pose measurement sensor output. , , Representing the sampling time respectively Lowering the robot's pose Projection in the main spatial direction, projection in the secondary spatial direction, and robot orientation angle; In S1212, the formula for calculating the pose estimate is: ; in, represent The pose estimate corresponding to the radar laser scanning point at any given time, i.e. the pose point to be interpolated; Kriging weights represent known pose sample points; In S1213, the pose estimation error is defined as follows: ; in, for The true pose value corresponding to the radar laser scanning point at any time; The formula for the variance of the pose estimation error is: ; in, Let Variance be the pose estimation error. This represents the Kriging weight vector to be solved. 3D matrix For known pose sample points , , The covariance vector between them , representing the position of the point to be interpolated and the known position sample point. , , The covariance vector between them represent exist Variance over time; In S1214, let Then there is ,in, Representative to Find the partial derivative; The calculation formula in S1215 is: ; Preferably, the and The proposed improved spatiotemporal anisotropic covariance function satisfies: ; in, Indicates the baseline variance. This represents the relevant scale parameters; simultaneously, it considers the sampling delay of the pose measurement sensor within the sampling period, i.e., the time between two known pose sample points. and Sampling delay between and satisfy: in, Represents a known pose sample point and The spatiotemporal correlation length function between them Represents a known pose sample point With the pose point to be interpolated The spatiotemporal correlation length function between them express and The distance between the main and secondary directions in space, represent Projection in the principal and secondary directions of space; Represents a known pose sample point and the pose point to be interpolated The distance between the main and secondary directions in space, represent Time and The difference in time; The length related to the main direction. For the secondary direction related length, The time-scale normalization coefficient is... Indicates the robot's movement speed. This indicates the sampling delay of the pose measurement sensor within the sampling period.
[0018] Preferably, the true pose matrix in S1221 The expression is: ; The actual coordinates in S1222 The expression is: ; in, Indicates laser point The true coordinates at the unified reference time after distortion correction. Representing the inverse pose at the reference time, this process transforms each sampling point from the actual scanning time to the reference coordinate system, realizing intra-frame motion compensation, and directly applying the interpolated pose to temporal point cloud correction, forming a motion distortion removal model driven by spatiotemporal covariance, thus avoiding the geometric errors of traditional linear interpolation in the turning stage.
[0019] Preferably, the construction process of the FR-YOLOv12 model includes: S401. In the YOLOv12s model, frequency domain attention is inserted into the original rectangular self-calibration attention structure RCA of the self-calibration module RCM to achieve deep fusion of spatial and frequency domain features, thus forming a frequency domain feature-enhanced rectangular self-calibration attention mechanism RCA_Freq to enhance the multidimensional perception capability of feature representation. The original RCA in RCM is replaced with the proposed RCA_Freq to obtain the improved rectangular self-calibration module IRCM; While maintaining the ability to extract local spatial features, the IRCM module can perceive global spectral information and demonstrate stronger modeling capabilities for periodic patterns, texture details and long-range dependencies in images, thereby significantly improving feature extraction and self-calibration performance.
[0020] S402. Further, the proposed IRCM is integrated with the Concat and Up-sample operations to propose a dual-branch feature concatenation and fusion module (DBFCFM), which is placed in the Neck part of the YOLOv12 network to achieve the concatenation and fusion of the output features of the A2C2f and C3K2 branches. Based on DBFCFM, an FR-YOLOv12 model is constructed. The IRCM significantly enhances the global perception capability of the FR-YOLOv12 model, thereby achieving higher accuracy in vehicle undercarriage fault detection. The A2C2f branch is a feature enhancement branch built based on the A2C2f module, and the C3K2 branch is a feature extraction branch built based on the C3K2 module.
[0021] Therefore, the present invention employs the above-mentioned dual-model-based automotive undercarriage fault detection robot system, which has the following advantages compared with the prior art: (1) The ITSCMN model proposed in this invention can perform fine modeling and compensation of the temporal and spatial errors of the two-dimensional lidar point cloud during the robot's motion process, effectively reduce motion distortion caused by high-speed or non-uniform motion conditions, construct a high-precision two-dimensional grid map, and provide reliable environmental data support for subsequent navigation path planning and detection tasks. (2) The FR-YOLOv12 model proposed in this invention can accurately detect faults under the car, significantly improve the accuracy and real-time performance of the detection, and greatly improve the detection efficiency; (3) This invention supports high-quality human-computer interaction, simplifies the operation process, and allows users to control the robot more easily and understand the undercarriage fault situation in real time through voice feedback. The robot has both online detection and voice interaction functions, which enriches the robot's operation methods, improves the user experience, and makes the robot more intelligent and convenient. This partially replaces manual inspection of undercarriage faults, effectively liberates productivity, significantly reduces labor costs, and greatly improves detection efficiency and accuracy.
[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0023] Figure 1 This is an architecture diagram of a dual-model-based automotive undercarriage fault detection robot system according to the present invention. Figure 2 The flowchart shows the improved anisotropic Kriging interpolation algorithm in a dual-model-based automotive undercarriage fault detection robot system of the present invention. Figure 3 The diagram shows the FR-YOLOv12 network structure proposed in this invention for a dual-model-based automotive undercarriage fault detection robot system. Figure 4 This is a structural diagram of an IRCM in a dual-model-based automotive undercarriage fault detection robot system according to the present invention. Figure 5 The image shows the training loss results of FR-YOLOv12 in a dual-model automotive undercarriage fault detection robot system of the present invention; where (a) is the classification loss curve, (b) is the bounding box loss curve, and (c) is the distribution focus loss curve. Figure 6 The image shows the detection results of rust and foreign object-related faults by a dual-model-based automotive undercarriage fault detection robot system of the present invention. Figure 7 The image shows the detection results of different types of key components by a dual-model-based automotive undercarriage fault detection robot system according to the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0025] Example 1 This embodiment provides the ITSCMN model mapping and navigation process of the present invention. The vehicle to be detected is parked in a designated detection area. The space under the vehicle is narrow, the lighting is uneven, and there are environmental interferences such as oil stains and pipelines on the ground. The robot needs to achieve high-precision autonomous navigation and modeling of the vehicle's underside environment.
[0026] After the ITSCMN model is started, the data queues of the LiDAR and IMU are first initialized using the improved Cartographer algorithm, and the vehicle undercarriage environment data and attitude data collected by the two types of sensors are received in real time. When both LiDARs have data input, the system starts the synchronization process, filters outdated data and builds a message preprocessing node based on the TAOTOC mechanism, establishes a unified reference timestamp with the two LiDARs as reference sensors, and remaps and compensates for time offset of the IMU data timestamp to ensure that the multi-source data is accurately aligned in the time dimension. The synchronized data is sent to the Cartographer backend module to generate a high-precision two-dimensional grid map.
[0027] Subsequently, an improved anisotropic kriging interpolation algorithm was introduced to obtain the true robot pose value at the sampling time. The pose change within the scanning period was treated as a random process, and the pose estimate at any time was solved by kriging interpolation. Combined with the improved spatiotemporal anisotropic covariance function to optimize the weight calculation, motion compensation was performed on the laser points to remove point cloud distortion. Finally, accurate under-vehicle environment modeling results and real-time positioning information were provided for the robot, ensuring that the robot can still autonomously plan its path and smoothly drive to the designated starting position in the inspection area under complex interference. Figure 2 A detailed algorithm flowchart for improved anisotropic kriging interpolation is provided.
[0028] The improved Cartographer algorithm verifies the sensor data preprocessing process by synchronizing sensor data timestamps. To evaluate preprocessing performance, this embodiment measures the execution time of IMU data timestamp synchronization. The measurement results show an average execution time of approximately 0.55 milliseconds, indicating that the preprocessing scheme has high operating efficiency. It can be seen that the entire system can maintain smooth detection even under complex lighting and confined space conditions, with an average single detection time of 105 seconds, which is about 5 times more efficient than traditional manual detection. The system demonstrates good stability and environmental adaptability during long-term operation testing.
[0029] Example 2 This embodiment provides experimental verification and performance comparison of the FR-YOLOv12 model of the present invention.
[0030] First, an FR-YOLOv12 model is built based on the YOLOv12s framework. The RCM module is replaced with IRCM, and the DBFCFM module is embedded into the Neck section. The FR-YOLOv12 network structure is as follows: Figure 3 As shown, the IRCM module structure is as follows: Figure 4 As shown.
[0031] Data augmentation techniques such as random flipping and brightness adjustment were employed for iterative training. The training results are as follows: Figure 5 As shown, all three losses decrease steadily with the number of iterations, indicating that the model converges well and provides a basis for subsequent detection accuracy data.
[0032] To verify the effectiveness of the FR-YOLOv12 model of this invention in the field of vehicle chassis fault detection, the robotic system collected undercarriage samples from 50 different car models, including 10 types of key undercarriage components and 2 types of faults: rust and foreign objects, totaling more than 8,000 images. A comparative experiment was conducted between the FR-YOLOv12 model proposed in this invention and the YOLOv12 improved by the rectangular self-calibration module RCM (RCM-YOLOv12).
[0033] Mean Average Precision (mAP@0.5), Precision (P), and Recall (R) were used as specific evaluation metrics. AP@0.5 refers to a prediction being considered correct when the Intersection over Union (IoU) between the predicted and ground truth bounding boxes is ≥ 0.5. mAP@0.5 is a performance metric for evaluating models in object detection, calculated by first averaging the average precision (AP) for each category, and only when the IoU between the predicted and ground truth bounding boxes is ≥ 0.5 is the prediction considered correct. The experimental results are shown in Tables 1, 2, and 3.
[0034] Table 1 Comparison of experimental results for different models
[0035] As shown in Table 1, under the same dataset and training conditions, the P, R, and mAP@0.5 of the original YOLOv12s model are all at a low level, at 76.9%, 74.8%, and 77.6%, respectively. This indicates that the unimproved original model still has insufficient feature representation ability and detection accuracy in the vehicle undercarriage fault detection task.
[0036] After introducing the RCM module, the detection performance of the RCM-YOLOv12 model was improved to a certain extent. Its detection precision increased from 76.9% to 79.4%, mAP@0.5 increased from 77.6% to 79.7%, and recall increased from 74.8% to 77.3%. This shows that the RCM module can enhance the feature extraction capability of the model and reduce false positives and false negatives to a certain extent. However, the improvement in precision, recall and mAP@0.5 of YOLOv12 by using only RCM is still relatively limited.
[0037] In contrast, this invention proposes an IRCM module, and based on this, designs an improved model FR-YOLOv12 with a DBFCFM structure. It achieves optimal performance in all indicators, with a detection accuracy of 80.4%, mAP@0.5 improved to 81.6%, and recall rate improved to 79.6%. The overall performance is significantly improved compared to the original YOLOv12s, and is also superior to RCM-YOLOv12.
[0038] Table 2 Comparison of Fault Experiment Results for Different Models
[0039] Table 3 Comparison of experimental results for ten key components of different models
[0040] Analysis of the experimental results in Tables 2 and 3 shows that the FR-YOLOv12 model achieves the best AP@0.5 performance across various undercarriage faults and key components. Specifically, in the two typical undercarriage faults (foreign objects and rust) shown in Table 2, the FR-YOLOv12 model exhibits higher detection accuracy compared to RCM-YOLOv12 and YOLOv12s, indicating its stronger fault identification capability. Furthermore, it demonstrates a more significant detection advantage in structurally complex components such as half-shafts, fuel tanks, and suspension arms, as shown in Table 3, demonstrating that the proposed DBFCFM effectively enhances multi-scale feature fusion capabilities. These results clearly demonstrate that the IRCM module further improves the global perception capability of the FR-YOLOv12 model, while the DBFCFM built based on IRCM achieves effective fusion of global and local feature information, thereby significantly improving the model's detection accuracy and robustness, verifying the effectiveness of the improved method of this invention.
[0041] Therefore, the improved YOLOv12 demonstrates significant advantages in automotive undercarriage fault detection tasks: the model can more accurately identify difficult-to-detect targets such as rust and abnormal half-shaft conditions, while maintaining high stability under complex backgrounds and varying lighting conditions. Compared to the previous version, overall detection accuracy and robustness have been improved, significantly reducing false positives and false negatives, providing more reliable technical support for the automated detection of undercarriage faults.
[0042] Example 3 This embodiment provides the complete fault detection process for the robot system according to the present invention, following... Figure 1 The operation and detection process are carried out using the architecture diagram: S1. After the robot is powered on, it sequentially completes the self-test of the power supply and communication link, the verification of the sensor working status, and the loading of initial parameters. After initialization, the ITSCMN model is run in the model deployment and display unit. It runs based on the Cartographer framework. By synchronizing the time and coordinates of pose measurement sensors such as LiDAR, IMU, and odometer, and by connecting the improved anisotropic Kriging interpolation LiDAR distortion removal algorithm as a front-end node, it performs online SLAM mapping and pose estimation on the corrected point cloud data to build a high-precision feature map of the vehicle underside environment, and realizes real-time localization and navigation support for the vehicle underside fault points.
[0043] S2. After receiving the vehicle undercarriage inspection task from the remote backend, the robot performs path planning and autonomous motion control based on the navigation system and real-time mapping and positioning results, driving the robot to smoothly travel under the vehicle to be inspected. After reaching the target area, the robot controls the six-degree-of-freedom drive device to perform multi-degree-of-freedom posture adjustment according to the vehicle undercarriage structure characteristics, so that the high-definition vision sensor is aligned with different chassis component areas, thereby collecting a large number of vehicle undercarriage fault images under multi-view and multi-scale conditions, providing high-quality visual data support for subsequent fault detection and analysis.
[0044] The S3 and FR-YOLOv12 models perform real-time inference calculations on the acquired undercarriage images within the robot's model deployment and display unit, generating corresponding fault detection results. These results include fault category, location coordinates, and confidence level information, and are displayed in real-time on the robot's onboard screen for on-site personnel to view intuitively. Simultaneously, the detection results are transmitted synchronously to a host computer via a data communication unit and stored to support subsequent fault analysis, historical data tracing, and maintenance decisions.
[0045] Real-time detection results of vehicle undercarriage faults Figure 6 and Figure 7As shown in the figure, the proposed automotive undercarriage fault detection robot system demonstrates excellent system stability and overall coordination under actual inspection conditions. Leveraging the high-precision mapping and localization capabilities of the ITSCMN model, the system maintains stable operation in narrow and complex undercarriage environments, providing reliable spatial positioning and attitude assurance for fault detection. Furthermore, the FR-YOLOv12 model achieves stable and highly consistent online detection and accurate localization of fault targets in undercarriage images. Detection results are visually labeled with bounding boxes, simultaneously providing corresponding category and confidence information. The detection results highly match the actual fault areas, significantly reducing false positives and false negatives. Even under typical undercarriage conditions such as complex backgrounds, uneven lighting, and component occlusion, the system maintains good detection accuracy and robustness for faults such as rust and foreign object adhesion, as well as for complex components like mufflers, suspension arms, exhaust pipes, half-shafts, and shock absorbers. Simultaneously, the detection results can be displayed in real-time and stably transmitted online to the backend system via the designed data communication unit, demonstrating the real-time collaborative operation of detection, display, and communication. The above results fully verify the comprehensive advantages of this invention in terms of detection accuracy, environmental adaptability, system stability and online application capabilities, and provide reliable technical support for the automated and intelligent real-time detection of vehicle undercarriage faults.
[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A robot system for detecting vehicle undercarriage faults based on a dual-model approach, characterized in that, include: The robot entity includes: an integrated power supply unit, a model deployment and display unit, an environmental perception unit, a motion control unit, a vision sensing unit, and an AI voice interaction unit; The remote backend connects to the robot entity via a data communication unit. The data communication unit is integrated into the robot body to enable data interaction between the robot body and the remote backend. The model deployment and display unit in the robot entity includes the improved spatiotemporal collaborative mapping and navigation model ITSCMN and the frequency domain-rectangular self-calibration YOLOv12 model FR-YOLOv12; The automotive undercarriage fault detection robot system achieves automated and high-precision detection of automotive undercarriage faults through the collaboration of dual models and various units.
2. The automotive undercarriage fault detection robot system based on a dual-model approach according to claim 1, characterized in that, The process of detecting faults under a car includes: S1. After the system powers on and performs a self-test, the model deployment and display unit starts the ITSCMN model. The environmental perception unit sends the real-time data from the lidar and IMU to the ITSCMN model via the data communication unit to complete the high-precision mapping of the under-vehicle environment and the initial positioning of the robot. S2. The remote backend issues the detection task through the data communication unit. The motion control unit plans the path based on the high-precision mapping and robot initial positioning provided by the ITSCMN model, and drives the robot to autonomously drive to the designated starting position at the bottom of the vehicle to be inspected. S3. After reaching the starting position, the AI voice interaction unit announces "Start detection". The motion control unit adjusts the robot's posture. The vision sensing unit continuously collects multi-view images of the underside of the vehicle under the power supply unit and sends the image data back to the model deployment and display unit. S4. The model deployment and display unit calls the FR-YOLOv12 model to perform real-time inference on the image, outputting the fault category, location coordinates and confidence level, forming a structured detection result; S5. The detection results are visualized locally by the model deployment and display unit, and the AI voice interaction unit simultaneously announces "Detection complete, a certain type of fault has been found"; S6. The detection results are uploaded to the remote backend through the data communication unit. The remote backend completes data storage, fault analysis and maintenance decision-making, realizing automated and high-precision detection of vehicle undercarriage faults.
3. A dual-model-based automotive undercarriage fault detection robot system according to claim 1 or 2, characterized in that, The ITSCMN model includes: an improved Cartographer algorithm and an improved anisotropic Kriging interpolation algorithm; S11. The improved Cartographer algorithm utilizes the timestamp alignment-online time offset calibration mechanism TAOTOC to construct message preprocessing nodes, thereby improving mapping accuracy through the temporal and spatial collaboration of multi-source data. S12. The improved anisotropic kriging interpolation algorithm is proposed based on the improved Cartographer algorithm to obtain input data. It efficiently removes motion distortion caused by the difficulty in matching the sampling frequency of the robot body pose with the sampling frequency of the radar laser beam.
4. The automotive undercarriage fault detection robot system based on a dual-model according to claim 3, characterized in that, The improved Cartographer algorithm specifically includes: S111. Initialize the data queues of each active sensor and input the data collected by each active sensor into the corresponding data queue; The sensors include lidar and inertial measurement units (IMUs). S112. If only a single lidar has data input, only the corresponding data queue is updated, and the synchronization process is not executed; if both lidars collect data, not only are the corresponding data sequences updated, but the synchronization process is also started. S113. During the synchronization process, compare the timestamps of the data to be synchronized with the latest historical input data of each sensor, select only the data with updated timestamps to participate in the synchronization, and filter out outdated data. S114. Construct a message preprocessing node based on the TAOTOC mechanism; S115. After time synchronization is completed, the data is sent to the Cartographer backend module to participate in real-time positioning and map building calculations, generating a high-precision two-dimensional raster map of the vehicle's underside environment.
5. The automotive undercarriage fault detection robot system based on a dual-model approach according to claim 4, characterized in that, The specific content of S114 includes: S1141. Set two LiDARs with consistent timing as reference sensors, and establish a unified reference timestamp reference_timestamp with the first data timestamp. S1142. Remap the timestamps of the data from the non-reference sensor according to the reference_timestamp and its own sampling rate, and then send them to the mapping function node cartographer_ros to participate in the calculation of the real-time positioning and map building system. S1143. Using the lidar timestamp as a unified reference time axis, the timing consistency of multi-sensor data during robot motion is analyzed online. The time offset of non-reference sensors relative to the reference lidar is dynamically estimated and compensated to ensure that the data of the dual lidars and IMU are aligned in the time dimension.
6. The automotive undercarriage fault detection robot system based on a dual-model according to claim 3, characterized in that, The improved anisotropic kriging interpolation algorithm includes: S121, Solving for the pose function: S1211, Regarding the sampling time Obtain the true robot pose value output by the pose measurement sensor. ; S1212. Treating the robot pose change during the scanning cycle as a random process, and obtaining the result through Kriging interpolation. pose estimate at time 1 ; S1213. Define pose estimation error and derive the formula for pose estimation error variance. S1214, Let the variance of the pose estimation error be correlated with the Kriging weight vector. The partial derivatives are 0, and the target weight vector is obtained by solving. S1215. Based on the target weight vector, calculate respectively Time Robot Location, Location and orientation angle pose estimation value , , ; S122, Laser point motion compensation: S1221. Let the reference time be... For any laser beam at time Collected laser points ( Based on the estimated pose values, a true pose matrix is constructed. ; S1222. Project the laser point onto the reference time coordinate system to obtain the true coordinates after distortion correction. Intra-frame motion compensation is completed.
7. The automotive undercarriage fault detection robot system based on a dual-model according to claim 6, characterized in that, In S1211, the expression for the true value of the robot pose is: ； in, Represents the sampling time The actual robot pose value is obtained from the pose measurement sensor output. , , Representing the sampling time respectively Lowering the robot's pose Projection in the main spatial direction, projection in the secondary spatial direction, and robot orientation angle; In S1212, the formula for calculating the pose estimate is: ； in, represent The pose estimate corresponding to the radar laser scanning point at any given time, i.e. the pose point to be interpolated; Kriging weights represent known pose sample points; In S1213, the pose estimation error is defined as follows: ； in, for The true pose value corresponding to the radar laser scanning point at any time; The formula for the variance of the pose estimation error is: ； in, Let Variance be the pose estimation error. This represents the Kriging weight vector to be solved. 3D matrix For known pose sample points , , The covariance vector between them , representing the position of the point to be interpolated and the known position sample point. , , The covariance vector between them represent exist Variance over time; In S1214, let Then there is ,in, Representative to Find the partial derivative; The calculation formula in S1215 is: ； The true pose matrix in S1221 The expression is: ； The actual coordinates in S1222 The expression is: ； in, Indicates laser point The true coordinates at the unified reference time after distortion correction. The reverse pose at the reference time.
8. The automotive undercarriage fault detection robot system based on a dual-model according to claim 7, characterized in that, The and The proposed improved spatiotemporal anisotropic covariance function satisfies: ； in, Indicates the baseline variance. This represents the relevant scale parameters; simultaneously, it considers the sampling delay of the pose measurement sensor within the sampling period, i.e., the time between two known pose sample points. and Sampling delay between and satisfy: in, Represents a known pose sample point and The spatiotemporal correlation length function between them Represents a known pose sample point With the pose point to be interpolated The spatiotemporal correlation length function between them express and The distance between the main and secondary directions in space, represent Projection in the principal and secondary directions of space; Represents a known pose sample point and the pose point to be interpolated The distance between the main and secondary directions in space, represent Time and The difference in time; The length related to the main direction. For the secondary direction related length, The time-scale normalization coefficient is... Indicates the robot's movement speed. This indicates the sampling delay of the pose measurement sensor within the sampling period.
9. A dual-model-based automotive undercarriage fault detection robot system according to claim 1 or 2, characterized in that, The construction process of the FR-YOLOv12 model includes: S401. In the YOLOv12s model, frequency domain attention is inserted into the original rectangular self-calibration attention structure RCA of the self-calibration module RCM to form a rectangular self-calibration attention mechanism RCA_Freq with enhanced frequency domain features, and the improved rectangular self-calibration module IRCM is obtained. S402. Integrate IRCM with Concat and Up-sample operations to form the dual-branch feature splicing and fusion module DBFCFM; S403. Embed DBFCFM into the Neck part of the YOLOv12s model to achieve the splicing and fusion of the output features of the A2C2f branch and the C3K2 branch, forming the FR-YOLOv12 model as a whole. Among them, the A2C2f branch is a feature enhancement branch built based on the A2C2f module, and the C3K2 branch is a feature extraction branch built based on the C3K2 module.