Unmanned aerial vehicle inspection method and system suitable for daily operation and maintenance work of rail transit vehicle

Abstract

The invention belongs to the technical field of unmanned aerial vehicle inspection, and particularly relates to an unmanned aerial vehicle inspection method and system suitable for daily operation and maintenance work of rail transit vehicles. According to the method, firstly, a composite unmanned aerial vehicle system is researched and developed according to the inspection requirements of rail transit vehicles; then, a deep learning technology based on an attention mechanism is introduced, and the technology enhances the recognition capability of the model by focusing key features and assists in fault positioning. And then, the flight stability of the unmanned aerial vehicle in a complex environment is enhanced by optimizing the aerodynamic configuration design and the control algorithm of the unmanned aerial vehicle. Then, a multi-source SLAM technology and an advanced motion planning algorithm are developed. And finally, a comprehensive management system is constructed, and collaborative operation among the equipment is realized. The problem of unstable fuselage control caused by airflow disturbance in a narrow space and the problem of degradation of simultaneous positioning and mapping in a low-texture long straight channel environment in the inspection process of an existing unmanned aerial vehicle are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of unmanned aerial vehicle (UAV) inspection, and more specifically, relates to a UAV inspection method and system suitable for the daily operation and maintenance of rail transit vehicles. Background Technology

[0002] Routine vehicle inspection is one of the main businesses of rail transit vehicle operation and maintenance services. For jobs with safety hazards and low efficiency, "machine replacement" has become a powerful means for enterprises to improve efficiency, reduce costs, improve product technology content, and expand their competitiveness. At present, similar research includes video acquisition systems and through-type detection systems.

[0003] Currently, urban rail transit depots still primarily rely on manual inspections, supplemented by video capture systems. The basic principle is image acquisition; the main method involves personnel using handheld cameras to capture video / images of relevant inspection points (including information on roof, body, and undercarriage components) during inspections. These videos / images are then transmitted to a backend system where specialists zoom in, analyze, and record relevant faults. The advantage of this method is that the inspection personnel no longer need to focus on whether each item is normal; they only need to focus on capturing good video. The diagnostic and identification process is delegated to backend staff, creating a complementary workflow. On the one hand, personnel in relatively harsh environments are less affected by environmental factors, while on the other hand, under better conditions, they can focus more intently on analysis and promptly annotate and archive faults. However, this method essentially still involves manual fault identification and does not fundamentally change aspects such as establishing unified quality standards, reducing human influence, or reducing labor intensity.

[0004] Currently, in industrial fields such as power transmission and substation, drones can replace manual labor for line inspection and data analysis. However, current issues include unstable drone control due to airflow disturbances in confined spaces and degradation in localization and mapping in low-texture, long, straight passage environments. Therefore, it is necessary to develop an intelligent drone system for the daily inspection of urban rail transit vehicles. This system would meet the complex inspection needs of various items on the undercarriage and sides of rail transit vehicles, reduce labor costs, achieve intelligent detection and early warning, and ensure the operational safety of urban rail transit. Summary of the Invention

[0005] This invention aims to overcome at least one of the defects of the prior art and provide a drone inspection method suitable for the daily operation and maintenance of rail transit vehicles. This method addresses the problems of unstable fuselage control caused by airflow disturbance in confined spaces and degradation of localization and mapping in low-texture long straight passage environments during drone inspections.

[0006] The present invention also discloses a drone inspection system suitable for the daily operation and maintenance of rail transit vehicles.

[0007] The detailed technical solution of this invention is as follows: A drone inspection method applicable to the daily operation and maintenance of rail transit vehicles, the method comprising: S1. The drone management platform receives inspection instructions from operators through a human-machine interface and then transmits the instructions to the intelligent scheduling system for task analysis and allocation. S2. The intelligent scheduling system selects the appropriate drone to perform the task based on the task requirements and the drone's status. S3. After receiving the mission instructions, the selected UAV will fly autonomously along the preset route and carry out inspection operations. S4. During the inspection process, the UAV uses multi-source SLAM technology and motion planning algorithm for degraded scenarios to integrate sensor data to achieve high-precision positioning and intelligent motion path planning, so as to ensure effective detection of key parts of rail transit vehicles. S5. During the inspection process, airflow disturbance is reduced by suppressing airflow disturbance, and adaptive control algorithms are used to adjust flight parameters to ensure stable flight in complex environments. S6. Using image recognition and fault analysis technology based on attention mechanism, the attention mechanism in deep learning is used to improve the accuracy of image recognition and assist in fault analysis. By constructing a model based on convolutional neural network, key feature regions are automatically focused, and the UAV completes vehicle item identification and fault detection, and collects relevant data in real time. S7. After the inspection task is completed, the drone automatically generates an inspection report and uploads the report to the control platform for subsequent viewing and analysis.

[0008] According to a preferred embodiment of the present invention, the UAV management platform in step S1 combines a UAV, a three-axis self-stabilized camera and sensors to replace manual inspection tasks, enabling the UAV to complete the inspection according to a preset route and transmit data to the background for analysis and processing in real time. The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

[0009] According to a preferred embodiment of the present invention, the specific steps of step S4 are as follows: S41. Design a SLAM system architecture that integrates LiDAR point cloud data and visual camera image data collected by UAV sensors. Extract the environmental 3D point cloud constructed by LiDAR through high-precision distance measurement, and the dynamic texture information that is complementary to the environmental 3D point cloud captured by the visual camera through optical flow estimation and ORB feature matching. S42. Develop a multimodal data alignment algorithm based on deep feature learning. Design a cross-modal feature matching network for low-texture scenes. The cross-modal feature matching network includes a LiDAR point cloud encoder and a visual feature extractor. The LiDAR point cloud is encoded using a PointNet++ structure to extract local geometric features and obtain LiDAR feature vectors. The visual image is processed by a ResNet-50 backbone network to extract high-level semantic features and obtain visual feature vectors. Then, an attention mechanism is used to perform cross-modal alignment between LiDAR features and visual features, incorporating a contrastive learning loss function. (1) in , These represent the lidar and visual feature vectors, respectively. Temperature parameter controls the sharpness of the distribution; This represents the k-th feature vector in the visual feature set; S43. Perform joint optimization and pose estimation. Adopt a joint optimization framework based on factor graphs to unify the multi-source observation data, such as the 3D point cloud acquired by LiDAR and the dynamic texture information acquired by the visual camera, and optimize the objective function as shown in the formula: (2) In equation (2), For pose variables, These are sensor observations. For the observation function, For robust kernel functions, For prior weights, This is based on prior pose information. Through this optimization process, the system can achieve high-precision pose estimation and map construction, maintaining robustness even in degraded scenarios such as low texture and low lighting. S44. Construct a robust motion planning and path optimization module, employing a hierarchical planning framework to achieve collaborative decision-making between global and local planners. The global planner generates a reference trajectory based on a hybrid A* algorithm, incorporating terrain slope and energy consumption constraints. The local planner uses the Dynamic Window Method (DWA) to adjust the velocity vector in real time by introducing obstacle potential field functions. (3) in The distance to the obstacle. This is the attenuation coefficient.

[0010] According to a preferred embodiment of the present invention, the specific steps of step S5 are as follows: S51. In order to suppress airflow disturbance and enhance flight stability, we carried out aerodynamic shape optimization design of UAV: ​​First, we adopted lightweight and high-strength composite materials to replace traditional metal structures, reducing the weight of the fuselage while ensuring structural strength. S52. Then, a biomimetic propeller structure is introduced into the blade design, and the blade twist angle distribution is optimized through parametric modeling, as shown in the following formula: (4) In equation (4), r is the radius, R is the total length of the blade, and n is the exponential parameter. This represents the twist angle at a position r away from the center on the blade. It is a fundamental parameter called the blade root twist angle; S53. Construct an adaptive flight control system based on MPC model predictive control. This system consists of three core modules: a six-degree-of-freedom dynamic model, a robust state observer, and a dynamic compensation controller.

[0011] According to a preferred embodiment of the present invention, the six-degree-of-freedom dynamic model is used to describe the complete motion of the UAV in space and explicitly considers environmental disturbances, providing a predictive model for MPC, which includes state equations containing aerodynamic disturbance terms: (5) In equation (5), Let u be the state vector, d be the control input, and d be the environmental disturbance. denoted by the first derivative of the state vector, f represents the system dynamics function, and t represents time; The robust state observer uses extended Kalman filtering to fuse multi-source data from IMU, barometer, and GPS. The dynamic compensation controller generates the optimal control quantity by solving a quadratic programming problem online. (6) In equation (6), and These represent the control input increment and the control input increment at step k, respectively. They are the variables that MPC needs to optimize, representing the change in control quantity between adjacent time steps, and are used to achieve smooth control. Indicates the prediction time domain, Indicates control over the time domain; This represents the predicted output of the system at step k, which is the object that is expected to be controlled. The reference output is represented by Q, which is the expected target value. Q and R are the weight matrices for the output error and control increment, respectively.

[0012] According to a preferred embodiment of the present invention, the specific steps of step S6 are as follows: S61. First, construct a multi-class image dataset containing normal and faulty samples, and label each sample with its class label and fault location information. Use image preprocessing techniques to improve data quality, then use the Dataset class in the PyTorch framework to standardize the data. Define the input image as X, the corresponding class label as Y, and the fault region coordinates as R. Then, standardize each group... To constitute a complete training sample; S62. Input the training samples into a convolutional neural network architecture that integrates a multi-layer attention mechanism for training. The architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses ResNet-50 as the backbone network and obtains general image representation capabilities through transfer learning of the pre-trained model; The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial locations. The formula is expressed as: (7) Where Q, K, and V are the query, key, and value matrices, respectively. Where T is the dimension scaling factor, and T is the transpose operation; The fault localization branch generates pixel-level attention heatmaps through deconvolution operations, enabling the visual localization of key feature regions. S63. Construct a composite loss function to optimize the performance of a convolutional neural network architecture, consisting of three parts: cross-entropy loss function. Used for classification tasks and attention consistency loss and eigensignificance loss Total loss function Defined as: (8) in , , The weighting coefficients for each part of the loss are used to determine the optimal combination through grid search.

[0013] S64. A distributed training strategy is adopted to accelerate model convergence, and mixed precision training technology is used to reduce memory usage. The optimizer selected is the AdamW algorithm, with an initial learning rate of [value missing]. A cosine annealing learning rate scheduler is used; during training, an early stopping mechanism (patience=10) is used to prevent overfitting, and TensorBoard is used to record training logs; weight updates follow these rules: (9) in For learning rate, and These are the first-order and second-order momentum terms, respectively. =0.9, =0.999.

[0014] In another aspect of the present invention, an intelligent unmanned aerial vehicle (UAV) inspection system for the daily operation and maintenance of rail transit vehicles is provided, the system comprising: The system includes a drone management platform, an attention-based image recognition and fault analysis module, an airflow disturbance suppression and flight stability control module, a multi-source SLAM localization and path planning module, and an intelligent scheduling system. The drone management platform is used to replace manual inspection tasks, enabling drones to complete inspections along preset routes and transmit data to the backend for analysis and processing in real time; it includes drones, a three-axis self-stabilized camera, and sensors; the sensors include a high-definition camera and an infrared thermal imager; the drone body is made of lightweight materials, the three-axis self-stabilized camera is used to stabilize and capture images, and the sensors are used to collect images and thermal data of rail transit vehicles; The attention-based image recognition and fault analysis module is used to identify and locate faults in images from inspection data, including: A dataset construction unit is used to construct a multi-class image dataset containing normal samples and faulty samples, wherein the dataset is labeled with category labels and faulty region location information; The model training unit adopts a convolutional neural network architecture that integrates a multi-layer attention mechanism. The architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses a ResNet-50 backbone network. The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial locations. The fault localization branch generates pixel-level attention heatmaps through deconvolution operations. The airflow disturbance suppression and flight stability control module is used to enhance the flight stability of the UAV in narrow passages, including: Aerodynamic shape optimization unit, used to optimize the aerodynamic shape of the UAV; The adaptive control unit adopts an adaptive flight control architecture based on model predictive control. The architecture includes a six-degree-of-freedom dynamic model, a robust state observer, and a dynamic compensation controller. The dynamic compensation controller generates the optimal control quantity by solving a quadratic programming problem online. The multi-source SLAM localization and path planning module is used to achieve accurate localization and path planning in complex environments, including: The multi-source sensor fusion unit is used to fuse lidar point cloud data and visual camera image information. It extracts geometric and texture features through feature-level fusion strategies and models multi-source observation data using a factor graph-based joint optimization framework. The cross-modal alignment unit employs a deep learning-based cross-modal data alignment algorithm, which enhances the matching capability between LiDAR and visual features through an attention mechanism. The path planning unit adopts a hierarchical motion planning strategy. The global planner generates a reference trajectory based on the improved A* algorithm, and the local planner uses a dynamic window method to adjust the path in real time to achieve obstacle avoidance. The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention deeply integrates a high-precision image recognition mechanism with an attention-driven fault analysis model, and introduces aerodynamic optimization design and adaptive flight control strategy to improve the UAV's operational capabilities in narrow and complex spaces: By utilizing a deep learning model based on an attention mechanism, the system can automatically focus on key component areas, significantly improving image recognition accuracy and fault location accuracy. Combined with the optimized aerodynamic shape design and the application of lightweight materials, along with the adaptive control algorithm, airflow disturbances in narrow channels are effectively suppressed, ensuring flight stability and control response.

[0016] (2) By constructing a SLAM positioning system and dynamic path planning algorithm based on multi-source sensor fusion, the system enhances its environmental perception and autonomous navigation capabilities in degraded scenarios such as low light and weak texture. The system integrates a lightweight UAV management platform, a three-axis self-stabilized camera, and multimodal sensing devices, enabling efficient acquisition of images and thermal data in typical rail transit vehicle environments. Furthermore, the IoT-based integrated management system supports centralized scheduling and collaborative operation of multiple UAVs, and through task process optimization and emergency response mechanism construction, comprehensively improves the system's intelligence level and execution efficiency.

[0017] (3) This invention establishes an unmanned aerial vehicle inspection system that integrates environmental perception, intelligent decision-making and collaborative execution, which fundamentally breaks through the limitations of traditional inspection methods in terms of efficiency, flexibility and intelligence, and provides an efficient, accurate and reliable solution for the operation and maintenance of urban rail transit vehicles. Attached Figure Description

[0018] Figure 1 This is a flowchart of the UAV inspection method described in this invention.

[0019] Figure 2This is a diagram illustrating the effect of applying the method of the present invention to the inspection of rail transit vehicles in an embodiment of the present invention. Detailed Implementation

[0020] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0021] Example 1 like Figure 1 As shown in the figure, this embodiment provides a drone inspection method suitable for the daily operation and maintenance of rail transit vehicles. The method includes: S1. The drone management platform receives inspection instructions from operators through a human-machine interface and then transmits the instructions to the intelligent scheduling system for task analysis and allocation. The drone management platform combines drones, a three-axis self-stabilized camera, and sensors to replace manual inspection tasks, enabling drones to complete inspections along preset routes and transmit data to the backend for analysis and processing in real time. The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

[0022] S2. The intelligent scheduling system selects the appropriate drone to perform the task based on the task requirements and the drone's status. S3. After receiving the mission instructions, the selected UAV will fly autonomously along the preset route and carry out inspection operations. S4. During the inspection process, the UAV uses multi-source SLAM technology and motion planning algorithms for degraded scenarios, and integrates UAV sensor data to achieve high-precision positioning and intelligent motion path planning, so as to ensure effective detection of key parts of rail transit vehicles. The specific steps of step S4 are as follows: S41. To address the challenges of localization and navigation in degraded scenarios such as poor lighting conditions and low texture, a multi-source sensor fusion SLAM system architecture is designed. This architecture integrates lidar point cloud data collected by UAV sensors and visual camera image data. It extracts the environmental 3D point cloud constructed by lidar through high-precision distance measurement, as well as the dynamic texture information that is complementary to the environmental 3D point cloud captured by the visual camera through optical flow estimation and ORB feature matching. This allows for the acquisition of multimodal raw observations, preparing for feature-level fusion. S42. To address the feature sparsity problem in degraded scenarios such as low texture and improve the reliability of association between different modal data, a multimodal data (laser point cloud data and visual data) alignment algorithm based on deep feature learning is developed. A cross-modal feature matching network is designed for low-texture scenarios. This network includes a laser radar point cloud encoder (PointNet++ structure) and a visual feature extractor (ResNet-50 backbone network). The laser radar point cloud is encoded using the PointNet++ structure to extract local geometric features, obtaining the laser radar feature vector. The visual image is processed by the ResNet-50 backbone network to extract high-level semantic features, obtaining the visual feature vector. Then, an attention mechanism is used to align the laser features and visual features across modalities, enabling them to mutually enhance each other in the feature space. A contrastive learning loss function is introduced. (1) in , These represent the lidar and visual feature vectors, respectively. Temperature parameter controls the sharpness of the distribution; This method represents the k-th feature vector in the visual feature set (i.e., those visual features that are matched and compared with the current laser features). It maps the originally heterogeneous laser and visual data to a unified feature space through a neural network, realizes feature-level fusion, provides high-quality input for subsequent joint optimization, and can effectively improve the reliability of data association in feature-sparse regions.

[0023] S43: Joint optimization and pose estimation are performed using a factor graph-based joint optimization framework. This framework unifies the multi-source observation data, including the 3D point cloud acquired by LiDAR and the dynamic texture information acquired by the visual camera, and optimizes the objective function as shown in the formula: (2) In equation (2), For pose variables, These are sensor observations. For the observation function, For robust kernel functions, For prior weights, This is based on prior pose information. Through this optimization process, the system can achieve high-precision pose estimation and map construction, maintaining robustness even in degraded scenarios such as low texture and low lighting. S44. Construct a robust motion planning and path optimization module, employing a hierarchical planning framework to achieve collaborative decision-making between global and local planners. The global planner generates a reference trajectory based on a hybrid A* algorithm, incorporating terrain slope and energy consumption constraints. The local planner uses the Dynamic Window Method (DWA) to adjust the velocity vector in real time by introducing obstacle potential field functions. (3) in The distance to the obstacle. This is the attenuation coefficient.

[0024] S44. Design a multi-condition verification platform, constructing a simulation environment that includes typical degradation scenarios such as tunnels, warehouses, and underground pipelines. A digital twin system is built using Gazebo, integrating a lidar noise model (Gaussian-Poisson mixed noise) and a camera image degradation module (dynamic blur, low-light simulation). Monte Carlo simulation is used to verify the system's robustness under different illumination intensities and texture densities.

[0025] S45. Establish a multi-dimensional evaluation system covering core indicators such as positioning accuracy (trajectory reprojection error), path feasibility (minimum distance to obstacles), and system real-time performance (frame processing latency). Conduct comparative tests in a laboratory darkroom and in actual industrial scenarios to verify the percentage reduction in positioning error and path planning success rate of the fusion system compared to single-source SLAM in degraded areas. By introducing sensor failure injection tests, demonstrate that the system can still maintain basic navigation functions when some sensors fail.

[0026] S5. During the inspection process, airflow disturbance is reduced by suppressing airflow disturbance, and adaptive control algorithms are used to adjust flight parameters to ensure stable flight in complex environments. The specific steps of step S5 are as follows: S51. In order to suppress airflow disturbance and enhance flight stability, we carried out aerodynamic shape optimization design of UAV: ​​First, we used lightweight high-strength composite materials (such as carbon fiber reinforced polymer) to replace the traditional metal structure, and reduced the weight of the fuselage while ensuring structural strength. S52. Then, a biomimetic propeller structure is introduced into the blade design, and the blade twist angle distribution is optimized through parametric modeling, as shown in the following formula: (4) In equation (4), r is the radius, R is the total length of the blade, and n is the exponential parameter. This represents the twist angle at a position r away from the center on the blade. It is a fundamental parameter called the blade root twist angle; The optimization effect was verified through computational fluid dynamics (CFD) simulation, which confirmed whether the optimized design could reduce the intensity of the wake vortex and reduce the interference of airflow on the UAV. S53. Construct an adaptive flight control system based on MPC model predictive control. This system contains three core modules: 1) A six-degree-of-freedom dynamic model, which describes the complete motion of the UAV in space (including translation and rotation) and explicitly considers environmental disturbances, providing a predictive model for MPC. This model includes state equations for aerodynamic disturbance terms. (5) In equation (5), Let u be the state vector, d be the control input, and d be the environmental disturbance. denoted by the first derivative of the state vector, f represents the system dynamics function, and t represents time; 2) Robust State Observer: It uses extended Kalman filter (EKF) to fuse multi-source data from IMU, barometer and GPS. By comprehensively processing the data from different sensors, it can estimate the real state of the UAV (such as position, speed, attitude, altitude, etc.) in real time, accurately and robustly. Especially in the presence of sensor noise, missing data or external interference, it provides reliable state information support for flight control. 3) Dynamic compensation controller, which generates the optimal control quantity by solving a quadratic programming problem online: (6) In equation (6), and These represent the control input increment and the control input increment at step k, respectively. They are the variables that MPC needs to optimize, representing the change in control quantity between adjacent time steps, and are used to achieve smooth control. Indicates the prediction time domain, Indicates control over the time domain; This represents the predicted output of the system at step k, which is the object that is expected to be controlled. This represents the reference output, i.e., the desired target value; Q and R are the weight matrices for the output error and control increment, respectively. S54. Design a multiphysics coupled simulation verification platform, integrating three-dimensional transient turbulence simulation, structure-aerodynamic-control coupling interface, and virtual environment construction module. Verify the system's stability performance in complex environments by simulating different wind speeds and crosswind angles; the simulation process focuses on flight attitude stability and trajectory tracking capability to ensure the system can adapt to various interference scenarios. S55. Conduct hardware-in-the-loop testing of the flight control system, using the PX4 flight control system and ROS2 framework to build a verification platform. The platform includes real-time data acquisition, control law execution, and fault injection testing modules, verifying system reliability through multi-dimensional scenario simulations. Combined with parameter tuning experiments, optimize the control strategy to improve the system's robustness under abnormal operating conditions; S56. Establish a multi-dimensional performance evaluation system covering core indicators such as aerodynamic performance, flight stability, and environmental adaptability. Through comparative experiments in a laboratory wind tunnel and a field test range, verify the adaptability of the optimized system in typical complex environments. Test results show that the system can effectively improve flight stability and exhibit stronger passability in special scenarios such as narrow passages.

[0027] S6. Utilizing attention-based image recognition and fault analysis technology, the attention mechanism in deep learning is used to improve image recognition accuracy and assist in fault analysis. By constructing a model based on a convolutional neural network, key feature regions are automatically focused, and the UAV completes vehicle item identification and fault detection, while collecting relevant data in real time. The specific steps of step S6 are as follows: S61. To construct an image recognition and fault analysis model based on an attention mechanism, a multi-class image dataset containing normal and faulty samples is first built. The image dataset needs to be labeled with the class label and fault region location information for each sample to support supervised learning of the subsequent model. To enhance the model's sensitivity to key features, image preprocessing techniques such as contrast enhancement and Gaussian noise injection are used to improve data quality. In the engineering implementation, the Dataset class of the PyTorch framework is used to standardize the data. The input image is defined as X, the corresponding class label as Y, and the fault region coordinates as R. Each set (X, Y, R) constitutes a complete training sample.

[0028] S62. Design a convolutional neural network architecture that integrates a multi-layer attention mechanism. This architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses ResNet-50 as the backbone network and obtains general image representation capabilities through transfer learning of a pre-trained model. The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial locations, as expressed by the formula: (7) Where Q, K, and V are the query, key, and value matrices, respectively. is the dimension scaling factor, and T is the transpose operation.

[0029] The fault localization branch generates pixel-level attention heatmaps through deconvolution operations, enabling the visual localization of key feature regions.

[0030] S63. Constructing a composite loss function to optimize the performance of a convolutional neural network architecture, consisting of three parts: 1) Cross-entropy loss function Used for classification tasks to quantify category prediction error; 2) Attention consistency loss The accuracy of the attention distribution is measured by calculating the intersection-over-union (IoU) ratio between the predicted heatmap and the actual fault region; 3) Feature saliency loss Gradient-weighted class activation mapping (Grad-CAM) is used to enhance the response intensity of key feature regions. The total loss function... Defined as: (8) in , , The weighting coefficients for each part of the loss are used to determine the optimal combination through grid search.

[0031] S64. A distributed training strategy is adopted to accelerate model convergence, and mixed precision training technology is used to reduce memory usage. The optimizer selected is the AdamW algorithm, with an initial learning rate of [value missing]. A cosine annealing learning rate scheduler is used; during training, an early stopping mechanism (patience=10) is used to prevent overfitting, and TensorBoard is used to record training logs; weight updates follow these rules: (9) in For learning rate, and These are the first-order and second-order momentum terms, respectively. =0.9, =0.999.

[0032] S65. Establish a three-level evaluation system to verify the effectiveness of the model: 1) Basic classification accuracy (ACC), calculated using the following formula: (10) 2) Fault localization accuracy (PLP), measured by the Dice coefficient, which indicates the degree of overlap between the predicted heatmap and the real area; 3) Attention interpretability score (AIS), where domain experts are invited to score the reasonableness of the model's attention areas. During the testing phase, the TPE (Tree-structured Parzen Estimator) algorithm was used for hyperparameter tuning, ultimately achieving an average classification accuracy of 92.7% and a fault localization Dice coefficient of 89.3% on the NIH ChestX-ray14 dataset.

[0033] S7. After the inspection task is completed, the drone automatically generates an inspection report and uploads the report to the control platform for subsequent viewing and analysis.

[0034] For the application scenarios of drones in the inspection of rail transit vehicles, Figure 2Detailed statistics were compiled for each key indicator. The item coverage rate refers to the proportion of all predetermined key parts or areas that the UAV can cover and inspect in a single inspection mission, and its value is 88.21%. High coverage means more comprehensive detection and reduced omissions. Fault identification accuracy characterizes the UAV's ability to correctly identify faults using its onboard sensors and algorithms; the value for this invention's method is 97.25%. High accuracy ensures timely detection of potential problems and avoids safety accidents. Fault item recall rate refers to the proportion of actual faults that the UAV can successfully detect during inspection; the fault item recall rate for this invention's method is 98.80%. High recall rate indicates that the system has good sensitivity and is less prone to missing faults. Positioning accuracy characterizes the UAV's ability to accurately locate specific fault locations on rail vehicles, which is crucial for subsequent maintenance work; the positioning accuracy rate for this invention's method is 99.47%. Obstacle avoidance success rate refers to the UAV's ability to successfully avoid obstacles encountered during mission execution and continue the mission. This includes both natural and man-made obstacles; this indicator was tested and found to be 100%, with no collisions observed. The improvement in inspection efficiency refers to the increased work efficiency achieved by using drones for inspection compared to traditional manual inspection. The method of this invention is measured by the time required to complete the inspection. This method can reduce the time of conventional manual inspection from 45 minutes to 20 minutes, improving efficiency by 55.56%.

[0035] Example 2 This embodiment provides an intelligent unmanned aerial vehicle (UAV) inspection system for the daily operation and maintenance of rail transit vehicles. The system includes: The system includes a drone management platform, an attention-based image recognition and fault analysis module, an airflow disturbance suppression and flight stability control module, a multi-source SLAM localization and path planning module, and an intelligent scheduling system. The drone management platform is used to replace manual inspection tasks, enabling drones to complete inspections along preset routes and transmit data to the backend for analysis and processing in real time. It includes drones, a three-axis self-stabilized camera, and sensors. The sensors include a high-definition camera and an infrared thermal imager. The drone body is made of lightweight materials. The three-axis self-stabilized camera is used to stabilize and capture images. The sensors are used to collect images and thermal data of rail transit vehicles.

[0036] The attention-based image recognition and fault analysis module is used to identify and locate faults in images from inspection data, including: A dataset construction unit is used to construct a multi-class image dataset containing normal samples and faulty samples, wherein the dataset is labeled with category labels and faulty region location information; The model training unit adopts a convolutional neural network architecture that integrates a multi-layer attention mechanism. The architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses a ResNet-50 backbone network. The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial locations. The fault localization branch generates pixel-level attention heatmaps through deconvolution operations.

[0037] The airflow disturbance suppression and flight stability control module is used to enhance the flight stability of the UAV in narrow passages, including: Aerodynamic shape optimization unit, used to optimize the aerodynamic shape of the UAV; The adaptive control unit adopts an adaptive flight control architecture based on model predictive control. The architecture includes a six-degree-of-freedom dynamic model, a robust state observer, and a dynamic compensation controller. The dynamic compensation controller generates the optimal control quantity by solving a quadratic programming problem online.

[0038] The multi-source SLAM localization and path planning module is used to achieve accurate localization and path planning in complex environments, including: The multi-source sensor fusion unit is used to fuse lidar point cloud data and visual camera image information. It extracts geometric and texture features through feature-level fusion strategies and models multi-source observation data using a factor graph-based joint optimization framework. The cross-modal alignment unit employs a deep learning-based cross-modal data alignment algorithm, which enhances the matching capability between LiDAR and visual features through an attention mechanism. The path planning unit adopts a hierarchical motion planning strategy. The global planner generates a reference trajectory based on the hybrid A* algorithm, while the local planner uses a dynamic window method to adjust the path in real time to achieve obstacle avoidance.

[0039] The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

[0040] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A drone inspection method suitable for daily operation and maintenance of rail transit vehicles, characterized in that, The method includes: S1. The drone management platform receives inspection instructions from operators through a human-machine interface and then transmits the instructions to the intelligent scheduling system for task analysis and allocation. S2. The intelligent scheduling system selects the appropriate drone to perform the task based on the task requirements and the drone's status. S3. After receiving the mission instructions, the selected UAV will fly autonomously along the preset route and carry out inspection operations. S4. During the inspection process, the UAV uses multi-source SLAM technology and motion planning algorithm for degraded scenarios to integrate sensor data to achieve high-precision positioning and intelligent motion path planning, so as to ensure effective detection of key parts of rail transit vehicles. S5. During the inspection process, airflow disturbance is reduced by suppressing airflow disturbance, and adaptive control algorithms are used to adjust flight parameters to ensure stable flight in complex environments. S6. Using image recognition and fault analysis technology based on attention mechanism, the attention mechanism in deep learning is used to improve the accuracy of image recognition and assist in fault analysis. By constructing a model based on convolutional neural network, key feature regions are automatically focused, and the UAV completes vehicle item identification and fault detection, and collects relevant data in real time. S7. After the inspection task is completed, the drone automatically generates an inspection report and uploads the report to the control platform for subsequent viewing and analysis.

2. The unmanned aerial vehicle (UAV) inspection method for daily operation and maintenance of rail transit vehicles according to claim 1, characterized in that, The drone management platform in step S1 combines drones, a three-axis self-stabilized camera, and sensors to replace manual inspection tasks, enabling drones to complete inspections along preset routes and transmit data to the backend for analysis and processing in real time. The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

3. The unmanned aerial vehicle (UAV) inspection method for daily operation and maintenance of rail transit vehicles according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Design a SLAM system architecture that integrates LiDAR point cloud data and visual camera image data collected by UAV sensors. Extract the environmental 3D point cloud constructed by LiDAR through high-precision distance measurement, and the dynamic texture information that is complementary to the environmental 3D point cloud captured by the visual camera through optical flow estimation and ORB feature matching. S42. The LiDAR point cloud is encoded using a PointNet++ architecture to extract local geometric features and obtain the LiDAR feature vector. The visual image is processed through a ResNet-50 backbone network to extract high-level semantic features and obtain the visual feature vector. Then, an attention mechanism is used to align the LiDAR features and visual features across modalities, incorporating a contrastive learning loss function. （1） In equation (1), , These represent the lidar and visual feature vectors, respectively. Temperature parameter controls the sharpness of the distribution; This represents the k-th feature vector in the visual feature set; S43. Perform joint optimization and pose estimation. Adopt a joint optimization framework based on factor graphs to unify the multi-source observation data, such as the 3D point cloud acquired by LiDAR and the dynamic texture information acquired by the visual camera, and optimize the objective function as shown in the formula: （2） In equation (2), For pose variables, These are sensor observations. For the observation function, For robust kernel functions, For prior weights, Prior pose information; S44. The global planner generates a reference trajectory based on a hybrid A* algorithm, combined with terrain slope and energy consumption constraints; the local planner uses a dynamic window method to adjust the velocity vector in real time, by introducing the obstacle potential field function. （3） In equation (3), The distance to the obstacle. This is the attenuation coefficient.

4. The unmanned aerial vehicle (UAV) inspection method for daily operation and maintenance of rail transit vehicles according to claim 1, characterized in that, The specific steps of step S5 are as follows: S51: Lightweight and high-strength composite materials are used to replace the traditional metal structure, reducing the weight of the fuselage while ensuring structural strength. S52. Then, a biomimetic propeller structure is introduced into the blade design, and the blade twist angle distribution is optimized through parametric modeling. S53. Construct an adaptive flight control system based on MPC model predictive control. This system consists of three core modules: a six-degree-of-freedom dynamic model, a robust state observer, and a dynamic compensation controller.

5. A drone inspection method for daily maintenance of rail transit vehicles according to claim 4, characterized in that, The six-degree-of-freedom dynamic model is used to describe the complete motion of the UAV in space and explicitly considers environmental disturbances, providing a predictive model for MPC. It includes state equations for aerodynamic disturbances. （4） In equation (4), Let u be the state vector, d be the control input, and d be the environmental disturbance. denoted by the first derivative of the state vector, f represents the system dynamics function, and t represents time; The robust state observer uses extended Kalman filtering to fuse multi-source data from IMU, barometer, and GPS. The dynamic compensation controller generates the optimal control quantity by solving a quadratic programming problem online. （5） In equation (5), and These represent the control input increment and the control input increment at step k, respectively. They are the variables that MPC needs to optimize, representing the change in control quantity between adjacent time steps, and are used to achieve smooth control. Indicates the prediction time domain, Indicates control over the time domain; This represents the predicted output of the system at step k, which is the object that is expected to be controlled. The reference output is represented by Q, which is the expected target value. Q and R are the weight matrices for the output error and control increment, respectively.

6. A drone inspection method for daily maintenance of rail transit vehicles according to claim 5, characterized in that, The specific steps of step S6 are as follows: S61. First, construct a multi-class image dataset containing normal and faulty samples, and label each sample with its class label and fault location information. Use image preprocessing techniques to improve data quality, then use the Dataset class in the PyTorch framework to standardize the data. Define the input image as X, the corresponding class label as Y, and the fault region coordinates as R. Then, standardize each group... To constitute a complete training sample; S62. Input the training samples into a convolutional neural network architecture that integrates a multi-layer attention mechanism for training. The architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses ResNet-50 as the backbone network and obtains general image representation capabilities through transfer learning of the pre-trained model; The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial location; The fault localization branch generates pixel-level attention heatmaps through deconvolution operations, enabling the visual localization of key feature regions. S63. Construct a composite loss function to optimize the performance of a convolutional neural network architecture, consisting of three parts: cross-entropy loss function. Used for classification tasks and attention consistency loss and eigensignificance loss ; S64. A distributed training strategy is adopted to accelerate model convergence, mixed precision training technology is used to reduce memory usage, the AdamW algorithm is selected as the optimizer, and the initial learning rate is set to... A cosine annealing learning rate scheduler is used; during training, an early stopping mechanism is employed to prevent overfitting, and TensorBoard is used to record training logs; weight updates follow these rules: （6） in For learning rate, and These are the first-order and second-order momentum terms, respectively. =0.9, =0.

999.

7. An intelligent unmanned aerial vehicle (UAV) inspection system for the daily operation and maintenance of rail transit vehicles, the system comprising: The system includes a drone management platform, an attention-based image recognition and fault analysis module, an airflow disturbance suppression and flight stability control module, a multi-source SLAM localization and path planning module, and an intelligent scheduling system. The drone management platform is used to replace manual inspection tasks, enabling drones to complete inspections along preset routes and transmit data to the backend for analysis and processing in real time; it includes drones, a three-axis self-stabilized camera, and sensors; the sensors include a high-definition camera and an infrared thermal imager; the drone body is made of lightweight materials, the three-axis self-stabilized camera is used to stabilize and capture images, and the sensors are used to collect images and thermal data of rail transit vehicles; The attention-based image recognition and fault analysis module is used to identify and locate faults in images from inspection data, including: A dataset construction unit is used to construct a multi-class image dataset containing normal samples and faulty samples, wherein the dataset is labeled with category labels and faulty region location information; The model training unit adopts a convolutional neural network architecture that integrates a multi-layer attention mechanism. The architecture includes a feature extraction layer, a multi-head self-attention module, and a fault localization branch. The feature extraction layer uses a ResNet-50 backbone network. The multi-head self-attention module dynamically allocates feature weights by calculating the correlation between channels and spatial locations. The fault localization branch generates pixel-level attention heatmaps through deconvolution operations. The airflow disturbance suppression and flight stability control module is used to enhance the flight stability of the UAV in narrow passages, including: Aerodynamic shape optimization unit, used to optimize the aerodynamic shape of the UAV; The adaptive control unit adopts an adaptive flight control architecture based on model predictive control. The architecture includes a six-degree-of-freedom dynamic model, a robust state observer, and a dynamic compensation controller. The dynamic compensation controller generates the optimal control quantity by solving a quadratic programming problem online. The multi-source SLAM localization and path planning module is used to achieve accurate localization and path planning in complex environments, including: The multi-source sensor fusion unit is used to fuse lidar point cloud data and visual camera image information. It extracts geometric and texture features through feature-level fusion strategies and models multi-source observation data using a factor graph-based joint optimization framework. The cross-modal alignment unit employs a deep learning-based cross-modal data alignment algorithm, which enhances the matching capability between LiDAR and visual features through an attention mechanism. The path planning unit adopts a hierarchical motion planning strategy. The global planner generates a reference trajectory based on the improved A* algorithm, and the local planner uses a dynamic window method to adjust the path in real time to achieve obstacle avoidance. The intelligent scheduling system is used to intelligently schedule multiple drones to achieve collaborative operation between devices. It also includes operation information management and emergency response mechanisms to ensure the safe and efficient execution of inspection tasks.

Images