Unmanned aerial vehicle intelligent inspection method based on lightweight model and storage medium

Abstract

The invention discloses an unmanned aerial vehicle intelligent inspection method based on a lightweight model, and the method comprises the steps: enabling an unmanned aerial vehicle to carry a camera and a sensor, obtaining the image data of a power transmission line in real time, processing the image data through a student model, and obtaining the image data from a teacher model through a decoupling distillation technology, and the student model outputs a bounding box, a category label and a confidence score for the image data, and judges whether a hidden danger exists or not. And if the hidden danger is found, processing is performed according to the confidence score, and if the confidence is low, the unmanned aerial vehicle is guided to fly to the hidden danger position through an automatic path-finding technology, and secondary data acquisition is performed. The collected data is then transmitted to a ground control center, and secondary judgment is carried out by a teacher model to confirm the type and position of the hidden danger. The method can effectively improve the hidden danger detection precision and real-time performance in the unmanned aerial vehicle inspection task, reduces the potential safety hazard and work intensity of manual inspection, and is suitable for efficient inspection in a complex environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of transmission line detection, and particularly relates to an intelligent inspection method for unmanned aerial vehicles based on a lightweight model and a storage medium. Background Art

[0002] With the continuous expansion and complexity of the power grid, the traditional manual inspection method can no longer meet the requirements of efficient, accurate, and real-time monitoring of transmission line equipment. Especially during the inspection of high-voltage transmission lines, due to the wide distribution, complex terrain, and variable environmental conditions of the transmission lines, manual inspection not only poses great safety hazards but also faces problems such as high work intensity and low efficiency. To improve the efficiency and safety of transmission line inspection, more and more power companies have begun to explore the use of unmanned aerial vehicles for inspection.

[0003] The application of unmanned aerial vehicle inspection technology in transmission lines has made certain progress, especially in terms of efficient data collection. Unmanned aerial vehicles can quickly reach complex and dangerous geographical environments and use the installed cameras and sensors to collect images and videos of transmission lines to monitor the operating status of equipment in real time. However, the existing unmanned aerial vehicle inspection technology still has the following problems in practical applications, which affect the accuracy and efficiency of inspection.

[0004] Although unmanned aerial vehicles can collect a large amount of image data, how to accurately and quickly identify the fault points and potential hazard locations of power equipment from these images is still a technical problem. Existing deep learning models perform well on large-scale data sets, but these models have a high computational complexity and require a large amount of computing resources. On the computing platform of unmanned aerial vehicles, due to limited computing power, traditional high-precision deep learning models are difficult to apply to real-time inspection tasks, thus affecting the accuracy and real-time performance of inspection results. Especially under the large model technology gradually established and applied by power grid enterprises, although these large models have significant advantages in high precision and high efficiency, due to their huge computing requirements, it is difficult to directly apply them to the edge devices of unmanned aerial vehicles.

[0005] The computing platform of unmanned aerial vehicles is usually relatively lightweight, and its computing power is much lower than that of server-level hardware. Traditional deep learning models, such as convolutional neural networks (CNNs), although excellent in image processing and object detection, have a large amount of computation and cannot run in real time on the edge devices of unmanned aerial vehicles. This results in the inability of unmanned aerial vehicles to timely perform real-time processing and analysis of the collected images, thereby affecting the efficiency and accuracy of inspection.

[0006] Transmission line inspection often needs to be carried out under different environmental conditions, including complex scenarios such as rainy and foggy weather, night inspection, strong light irradiation, etc. The target recognition ability of existing UAV inspection systems usually depends on clear visual images, which makes them perform poorly in adverse environments. Especially in low light, haze or other adverse weather conditions, the image quality deteriorates, resulting in a significant reduction in the accuracy of target detection and fault diagnosis. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the existing technologies described above and provide an intelligent UAV inspection method based on a lightweight model.

[0008] The purpose of the present invention can be achieved through the following technical solutions:

[0009] On the one hand, the present invention provides an intelligent UAV inspection method based on a lightweight model, including the following steps:

[0010] The UAV obtains real-time transmission line image data through a camera and inputs the image data into a student model for processing. The student model is carried on the UAV and obtained from a teacher model through decoupled distillation technology;

[0011] The student model processes the image data and outputs the bounding box, class label and confidence score of the image;

[0012] Judge whether there are potential hazards according to the output class label;

[0013] If there are no potential hazards, continue the inspection;

[0014] If there are potential hazards and the confidence score is within the first preset interval, record the GPS position of the image and label the image, and save it as potential hazard data;

[0015] If there are potential hazards and the confidence score is within the second preset interval, fly the UAV to the potential hazard position through the bounding box and automatic pathfinding technology, and hover in front of the potential hazard position for secondary data collection, and label the secondary data as suspected potential hazard data;

[0016] After the UAV inspection is completed, transmit the collected suspected potential hazard data to the ground control center, and the teacher model in the ground control center makes a secondary judgment on the suspected potential hazard data to judge the position and type of the potential hazard.

[0017] Furthermore, the teacher model is a teacher model based on Faster R-CNN, and the student model is a student model based on Tiny-YOLO.

[0018] Furthermore, the obtaining from the teacher model through decoupled distillation technology specifically includes:

[0019] Use a pre-trained teacher model to detect potential hazards in the input image, generating bounding boxes, class labels, and corresponding confidence scores for the potential hazards;

[0020] Use the class labels generated by the teacher model as the training target for the student model, and optimize the student model by calculating the class loss function;

[0021] Obtain the feature maps output by the teacher model in the intermediate layer during potential hazard detection, use the feature maps output by the teacher model in the intermediate layer as the training target for the student model, and optimize the student model by calculating the feature loss function;

[0022] Optimize the student model according to the calculated class loss and feature loss until the student model converges, and obtain the trained student model.

[0023] Furthermore, the input image includes transmission line potential hazard images of various potential hazard types.

[0024] Furthermore, the class loss function is:

[0025]

[0026] where L cls is the class loss function of the student model, N is the number of input images, y i is the class label generated by the teacher model, is the class label generated by the student model.

[0027] Furthermore, the feature loss function is:

[0028]

[0029] where L feat is the feature loss function of the student model, C, H, and W are respectively the number of channels, height, and width of the feature map, is the feature value at the h, w position of the C-th channel in the student model, is the feature value at the h, w position of the C-th channel in the teacher model.

[0030] Furthermore, the secondary data acquisition specifically includes: obtaining secondary data of the potential hazard location through a high-definition camera and a 3D camera mounted on the drone, and the secondary data includes two-dimensional image data and depth data.

[0031] Furthermore, flying the drone to the potential hazard location through the bounding box and automatic pathfinding technology specifically includes:

[0032] Determine the relative coordinates of the potential hazard with respect to the drone according to the bounding box output by the student model;

[0033] Calculate the optimal flight path from the current UAV position to the hidden danger position using a path planning algorithm based on the relative UAV coordinates of the hidden danger;

[0034] Control the UAV to fly autonomously according to the calculated optimal flight path, obtain environmental information in real time, dynamically detect obstacles in the flight path using a vision sensor or lidar sensor, and adjust the flight route according to the detection results;

[0035] During the flight, the bounding box of the hidden danger is output in real time through the student model on the UAV, and the relative UAV coordinates of the hidden danger are updated in real time through the bounding box, and the optimal flight path is updated;

[0036] When the UAV flies to the target hidden danger position, the UAV is kept hovering in front of the target hidden danger position through the control system.

[0037] Further, determining the relative UAV coordinates of the hidden danger according to the bounding box output by the student model specifically includes:

[0038] According to the bounding box information output by the student model and the preset camera internal parameters of the UAV camera, convert the image coordinates of the bounding box from the pixel space to the relative camera coordinates through an image geometric transformation algorithm;

[0039] Calculate the shooting angle of the UAV by combining the attitude angles provided by the inertial measurement unit and the flight control system on the UAV, and combine the flight altitude of the UAV to convert the relative camera coordinates to the relative UAV coordinates.

[0040] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements a UAV intelligent inspection method based on a lightweight model as described in any one of the above.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] (1) The present invention adopts a decoupled distillation technique to extract knowledge from a pre-trained teacher model and transfer it to a student model. The student model reduces the computational complexity through lightweight design. This enables the UAV to process image data in real time on an edge computing platform with limited resources, thus solving the problems of complex calculations and high requirements for computing resources of traditional deep learning models. Through this method, while ensuring high-precision target detection, the student model greatly improves the real-time performance and efficiency during the inspection process.

[0043] (2) The present invention combines the bounding box information output by the student model with the automatic pathfinding technology to achieve real-time adjustment and optimization of the UAV flight path. Through the path planning algorithm, the optimal flight path of the UAV from the current position to the potential hazard position is calculated to ensure rapid and accurate positioning of the potential hazard position during the inspection process. This method effectively improves the autonomy and accuracy of the UAV flight. Especially in the case of complex environments and dynamic obstacles, it can ensure that the UAV safely and stably reaches the potential hazard position.

[0044] (3) When a potential hazard is detected, the present invention conducts secondary data collection through the UAV, including the combination of two-dimensional images and depth data, and transmits the data to the ground control center for secondary judgment. The teacher model conducts secondary judgment on the potential hazard data, further improving the accuracy of the potential hazard type and position judgment. This mechanism reduces the probability of misjudgment, ensures the precise positioning and type identification of potential hazards, and improves the reliability and effectiveness of the inspection.

[0045] (4) The UAV intelligent inspection method of the present invention enhances the adaptability of the UAV under different environmental conditions by integrating high-precision image processing and depth data analysis. Even under complex and adverse weather conditions, the system can still effectively identify and process the potential hazard information of the transmission line, thus ensuring the continuity and efficiency of the transmission line inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment 1:

[0049] A UAV intelligent inspection method based on a lightweight model aims to improve the efficiency, accuracy, and ability to adapt to different environmental conditions of the transmission line inspection by applying the decoupling and distillation technology to the UAV image processing task. And by making full use of the advantages of the large model of the power grid enterprise, a lightweight student model is constructed. As Figure 1 shown, the method of this embodiment includes the following steps:

[0050] Step S1: The UAV obtains real-time transmission line image data through a camera and inputs the image data into the student model for processing. The student model is carried on the UAV and obtained from the teacher model through the decoupling and distillation technology.

[0051] Step S2: The student model processes the image data, outputs the bounding box, class label, and confidence score of the image, and determines whether there is a potential hazard based on the output class label.

[0052] Step S3: If there is no potential hazard, continue the inspection; if there is a potential hazard and the confidence score is within the first preset interval, record the GPS position of the image and label the image, and save it as potential hazard data; if there is a potential hazard and the confidence score is within the second preset interval, use the bounding box and automatic pathfinding technology to fly the UAV to the potential hazard position and hover in front of the potential hazard position for secondary data collection, and label the secondary data as suspected potential hazard data.

[0053] Step S4: After the UAV inspection is completed, transmit the collected suspected potential hazard data to the ground control center. The teacher model in the ground control center makes a secondary judgment on the suspected potential hazard data to determine the location and type of the potential hazard.

[0054] The specific content includes:

[0055] The UAV first obtains the image data of the transmission line in real time through the equipped high-resolution camera. The image data is processed on the UAV by a lightweight student model. The student model is the Tiny-YOLO model, and the student model is obtained from a more powerful teacher model (such as Faster R-CNN) through decoupled distillation technology. Among them, the teacher model (such as Faster R-CNN) is obtained from the large model of the power grid enterprise. The Tiny-YOLO model is relatively concise in design and is suitable for being deployed on a UAV platform with limited computing resources. It can process image data in real-time tasks. This design can greatly reduce the consumption of computing resources, while maintaining high computing efficiency, adapting to the computing power of the UAV platform, ensuring that the inspection task can be carried out smoothly under limited computing resources, enabling the UAV to work efficiently and quickly in complex environments, and improving the real-time performance of the inspection task.

[0056] In this embodiment, the process of obtaining the student model includes:

[0057] Select a teacher model with excellent performance and high computational complexity, such as a teacher model based on Faster R-CNN, as a reference model for deep learning tasks. The teacher model is usually well-trained on a large-scale dataset and can identify potential hazards in images with high precision, generating bounding boxes, class labels, and confidence scores. Select a network structure with low computational complexity as the student model, such as a student model based on Tiny-YOLO, to achieve fast and efficient inference capabilities, suitable for drone devices with limited computational resources. To enable the student model to achieve performance close to that of the teacher model, knowledge is obtained from the teacher model through decoupled distillation technology. The specific process is as follows: Use the pre-trained teacher model to detect potential hazards in the input image, generating bounding boxes, class labels, and corresponding confidence scores for the potential hazards. Take the class labels output by the teacher model as the training target of the student model, calculate the class loss between the student model and the teacher model, and optimize the classification ability of the student model by minimizing this loss function. In addition, by extracting the feature maps of the intermediate layers of the teacher model and comparing them with the corresponding intermediate layer features of the student model, the feature loss is calculated, and the spatial feature extraction ability of the student model is further optimized by minimizing the feature loss. By repeatedly calculating the class loss and the feature loss, the parameters of the student model are gradually adjusted to make it as close as possible to the teacher model in performance, while having low computational complexity and high inference efficiency. After the above process, the student model can be trained and optimized to run on a drone platform with limited computational resources, achieving fast and accurate potential hazard detection and providing corresponding bounding boxes, class labels, and confidence scores. This enables the student model to provide performance close to that of the teacher model while maintaining low computational overhead, providing effective support for the intelligent inspection tasks of drones.

[0058] Among them, the class loss function is:

[0059]

[0060] Among them, L cls is the class loss function of the student model, N is the number of input images, y i is the class label generated by the teacher model, is the class label generated by the student model.

[0061] The feature loss function is:

[0062]

[0063] Among them, L feat is the feature loss function of the student model, C, H, and W are the number of channels, height, and width of the feature map respectively, is the feature value of the C-th channel of the student model at the h, w position, It is the eigenvalue of the C-th channel in the teacher model at the h, w position.

[0064] During the data processing, the student model processes the input image data and outputs the bounding box, class label, and confidence score of the image. The bounding box determines the location where potential hazards may exist in the image, and the class label and confidence score provide an assessment of what kind of hazard it is at that location and its credibility. The class label includes a hazard label and a non-hazard label. Through this step, the drone can automatically identify potential hazards on the transmission line and make a preliminary judgment. The application of this technical feature greatly reduces manual intervention, improves the automation level of inspection, reduces the work intensity of manual inspection, and at the same time improves the accuracy and efficiency of hazard detection.

[0065] If the student model determines that there are no hazards in the image, that is, the class label is a non-hazard label, the drone will continue to perform the inspection task along the predetermined trajectory without being disturbed and can continuously and efficiently conduct large-scale inspection work. This step effectively avoids false alarms and improves the inspection efficiency. If it is determined that there is a hazard, that is, the class label is a hazard label, the model will further judge the severity of the hazard based on the output confidence score. If the confidence score is within the first preset interval, the first preset interval is 80%-100%, the image of the hazard location will be marked, and the GPS location of the image will be recorded and saved as hazard data. This technical feature ensures the accuracy and integrity of the hazard data, facilitates subsequent analysis and processing, and improves the reliability of hazard identification and the intelligent level of the system.

[0066] If the confidence score of the hazard is within the second preset interval, the second preset interval is 0%-80%, the drone will fly to the hazard location through the bounding box and automatic pathfinding technology. The bounding box can help the drone determine the position of the hazard relative to the drone, and the automatic pathfinding technology calculates the optimal flight path through a path planning algorithm to ensure that the drone can accurately fly to the hazard location. The automatic pathfinding system can obtain environmental information in real time, use visual sensors or lidar to detect obstacles in the flight path, and dynamically adjust the flight path according to the detection results. This technical feature can significantly improve the autonomous flight ability of the drone in a dynamic environment, ensure that the task execution will not be affected by the presence of obstacles during the inspection process, optimize the inspection efficiency, and ensure that the drone can accurately fly to the hazard point for processing.

[0067] In this embodiment, the specific process of flying the UAV to the potential hazard location based on the bounding box and the automatic pathfinding technology is achieved through multiple steps, combining the computer vision, path planning, environmental perception, and dynamic adjustment technologies of the UAV. When the UAV collects image data through the camera and transmits it to the student model (such as Tiny-YOLO) for processing, the model will output a bounding box containing potential hazard information. The bounding box defines the area in the image where potential hazards may exist and calibrates the relative position of the potential hazards. Each bounding box is represented by four coordinates (the pixel coordinates of the upper left corner and the lower right corner), indicating the position of the potential hazard area in the image.

[0068] After obtaining the bounding box, it is necessary to convert the pixel coordinates in the image into the relative coordinates of the UAV. First, using the internal parameters of the UAV camera, perform image geometric transformation to convert the pixel coordinates in the image into three-dimensional coordinates in the camera coordinate system. The camera internal parameters include information such as focal length and optical center, which are used to map two-dimensional image coordinates to three-dimensional space. Next, the attitude information of the UAV also needs to be considered. The attitude of the UAV is determined by the data provided by the inertial measurement unit (IMU) and the flight control system. Through these data, the current flight angle of the UAV can be calculated. Combining the flight altitude information, the camera coordinates are converted into the relative coordinates of the UAV. This process ensures that the UAV can understand the relative position of the potential hazard and provides accurate positioning data for subsequent flight path planning.

[0069] Using the path planning algorithm, based on the current position of the UAV and the relative coordinates of the potential hazard location, calculate the optimal flight path for the UAV to fly towards the potential hazard point. Common path planning algorithms such as the A algorithm, Dijkstra algorithm, or RRT algorithm can calculate the shortest or safest path according to the real-time requirements of the flight. These algorithms consider factors such as the flight speed of the UAV, the complexity of the flight environment, and possible obstacles, and generate a feasible flight route to ensure that the UAV can reach the potential hazard location smoothly.

[0070] During the execution of the flight path by the UAV, environmental perception, especially obstacle detection, will be continuously carried out during the flight. The UAV uses the onboard vision sensor (such as the front camera) or lidar (LiDAR) sensor to scan the flight path in real time to detect possible obstacles. The obstacle information will be fed back to the flight control system, and according to the real-time detection results, the flight path will be dynamically adjusted. For example, if obstacles such as trees, wires, or buildings appear in the flight path, the system will immediately adjust the flight route to avoid the obstacles while ensuring that the UAV continues to move towards the potential hazard location. Through these real-time obstacle detections and path optimizations, it can ensure that the UAV flies unobstructed in a complex environment.

[0071] When the drone approaches the location of the potential hazard, the flight control system will automatically control the drone to decelerate and make it precisely reach the front of the potential hazard point for hovering. During the hovering process, the flight control system will utilize information such as GPS positioning, visual sensors, and inertial measurement units to ensure that the drone can stably hover in front of the target potential hazard point. This process requires the drone to perform fine attitude adjustments and position control to avoid position drift caused by wind speed changes or other external factors.

[0072] Once the drone hovers in front of the potential hazard, the high-definition camera and 3D camera will be activated for secondary data collection. These secondary data include high-resolution two-dimensional image data and depth data. Through these data, the ground control center can obtain detailed information about the potential hazard and further conduct potential hazard assessment and processing.

[0073] Throughout the process, the combination of the bounding box and the automatic pathfinding technology enables the drone to fly precisely according to the target position and accurately hover for data collection after reaching the potential hazard point. The collaborative work of the automatic path planning and obstacle detection system enables the drone to fly stably in a complex environment, avoid obstacles, and ensure efficient task execution. Through the application of these technical features, the present invention can significantly improve the navigation accuracy and inspection efficiency of the drone in a dynamic environment and ensure the smooth completion of the inspection task.

[0074] After the drone flies to the location of the potential hazard, the drone is kept hovering in front of the potential hazard location through the control system. In the hovering state, the high-definition camera and 3D camera are combined for secondary data collection. Through the secondary collection, more refined two-dimensional image data and depth data can be obtained, and the acquired secondary data is marked as suspected potential hazard data to ensure the accurate judgment and positioning of the potential hazard. This technical feature effectively improves the accuracy of the inspection by increasing the accuracy and depth of the data, making the inspection results more reliable, and can provide more detailed data support for the ground control center, facilitating the formulation of subsequent potential hazard assessment and repair plans.

[0075] In one solution of this embodiment, after the drone inspection is completed, the collected suspected potential hazard data will be transmitted to the ground control center. The teacher model in the ground control center will make a secondary judgment on these suspected potential hazard data to determine the specific location and type of the potential hazard. This process makes full use of the powerful computing power and deep learning technology of the teacher model, making the potential hazard judgment more accurate and enabling optimization and improvement in subsequent inspection tasks.

[0076] In another solution of this embodiment, after the drone inspection is completed, the collected suspected hidden danger data will be transmitted to the ground control center. The ground control center extracts features from the two-dimensional image data through a convolutional neural network (CNN), captures information such as the shape and texture in the image, and identifies possible hidden danger features, such as broken wires, corroded metals, etc. At the same time, the depth data provides the distance information between each pixel and the camera, helping the system understand the position of the object in the three-dimensional space, reducing misjudgment and improving the judgment accuracy. The image features and depth information can be fused through multimodal learning methods, and a deep learning model is used to jointly analyze these two data sources, thereby improving the accuracy of hidden danger recognition. In addition, traditional image processing methods such as morphological operations and edge detection can also be combined with the deep learning model to further enhance the ability to identify abnormal shapes in the image. Finally, the system comprehensively analyzes the image and depth data, outputs the category and location of the hidden danger, helps to judge whether there is a hidden danger, and provides a basis for subsequent processing.

[0077] Through the decoupled distillation technology, the student model can be trained by means of the feature maps of the intermediate layers of the teacher model. While maintaining a low computational complexity, the performance of the model in the hidden danger recognition task is improved. This technical feature can improve the accuracy of hidden danger detection while ensuring the real-time nature of the inspection task, and avoids the excessive dependence of traditional deep learning models on computing resources. The application of the decoupled distillation technology enables the student model to effectively execute complex image processing tasks under the limited computing resources of the drone, greatly improving the performance of the drone intelligent inspection system.

[0078] Generally speaking, this embodiment combines the decoupled distillation technology with a lightweight model, successfully overcomes the problem of insufficient computing resources in traditional drone inspection systems, and at the same time improves the accuracy and efficiency of hidden danger recognition. The drone can accurately execute the inspection task in various complex environments, reduce manual intervention, lower the inspection cost, and improve the safety and reliability of the inspection.

[0079] Embodiment 2:

[0080] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0081] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An intelligent inspection method for drones based on a lightweight model, characterized in that, It includes the following steps: The drone obtains real-time image data of the transmission line through a camera and inputs the image data into a student model for processing. The student model is carried on the drone and obtained from a teacher model through decoupled distillation technology; The student model processes the image data and outputs the bounding box, class label, and confidence score of the image; Judge whether there are potential hazards according to the output class label; If there are no potential hazards, continue the inspection; If there are potential hazards and the confidence score is within the first preset interval, record the GPS position of the image, label the image, and save it as potential hazard data; If there are potential hazards and the confidence score is within the second preset interval, use the bounding box and automatic pathfinding technology to fly the drone to the potential hazard position and hover in front of the potential hazard position for secondary data collection. Label the secondary data as suspected potential hazard data; After the drone inspection is completed, transmit the collected suspected potential hazard data to the ground control center. The teacher model in the ground control center makes a secondary judgment on the suspected potential hazard data to judge the position and type of the potential hazard.

2. The method for intelligent inspection of unmanned aerial vehicles based on a lightweight model according to claim 1, wherein The teacher model is a teacher model based on Faster R-CNN, and the student model is a student model based on Tiny-YOLO.

3. The method for intelligent inspection of unmanned aerial vehicles based on a lightweight model according to claim 1, characterized in that, The obtaining from the teacher model through decoupled distillation technology specifically includes: Use the pre-trained teacher model to detect potential hazards in the input image and generate the bounding box, class label, and corresponding confidence score of the potential hazard; Use the class label generated by the teacher model as the training target of the student model, and optimize the student model by calculating the class loss function; Obtain the feature map output by the teacher model in the middle layer during potential hazard detection, use the feature map output by the teacher model in the middle layer as the training target of the student model, and optimize the student model by calculating the feature loss function; Optimize the student model according to the calculated class loss and feature loss until the student model converges to obtain the trained student model.

4. The method for intelligent inspection of an unmanned aerial vehicle based on a lightweight model according to claim 3, wherein, The input image includes transmission line potential hazard images of various potential hazard types.

5. The method for intelligent inspection of an unmanned aerial vehicle based on a lightweight model according to claim 3, wherein The class loss function is: Among them, L cls is the class loss function of the student model, N is the number of input images, and y i is the class label generated by the teacher model, and is the class label generated by the student model.

6. The method for intelligent inspection of an unmanned aerial vehicle based on a lightweight model according to claim 3, wherein The feature loss function is: Among them, L feat is the feature loss function of the student model, where C, H, and W are the number of channels, height, and width of the feature map respectively, is the feature value of the C-th channel of the student model at the h, w position, is the feature value of the C-th channel of the teacher model at the h, w position.

7. A method for intelligent inspection of drones based on a lightweight model according to claim 1, characterized in that The secondary data collection specifically includes: obtaining secondary data of the potential hazard position through a high-definition camera and a 3D camera carried on the drone. The secondary data includes two-dimensional image data and depth data.

8. A method for intelligent inspection of drones based on a lightweight model according to claim 1, characterized in that, The flying the drone to the potential hazard position through the bounding box and automatic pathfinding technology specifically includes: Determine the relative drone coordinates of the potential hazard according to the bounding box output by the student model; Use the path planning algorithm to calculate the optimal flight path from the current drone position to the potential hazard position through the relative drone coordinates of the potential hazard; Control the drone to fly autonomously according to the calculated optimal flight path, obtain environmental information in real time, dynamically detect obstacles in the flight path using a vision sensor or a lidar sensor, and adjust the flight route according to the detection result; During the flight, the student model on the drone outputs the bounding box of the potential hazard in real time, updates the relative drone coordinates of the potential hazard in real time through the bounding box, and updates the optimal flight path; When the drone flies to the target potential hazard position, keep the drone hovering in front of the target potential hazard position through the control system.

9. The method for intelligent inspection of an unmanned aerial vehicle based on a lightweight model according to claim 8, characterized in that, Determining the relative UAV coordinates of potential hazards based on the bounding boxes output by the student model specifically includes: According to the bounding box information output by the student model and the preset camera internal parameters of the UAV camera, the image coordinates of the bounding box are converted from the pixel space to the relative camera coordinates through the image geometric transformation algorithm; By combining the attitude angles provided by the inertial measurement unit and the flight control system on the UAV, the shooting angle of the UAV is calculated, and in combination with the flight altitude of the UAV, the relative camera coordinates are converted to relative UAV coordinates.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements a UAV intelligent inspection method based on a lightweight model as described in any one of claims 1 to 9.

Images