Path planning method of cleaning robot for unmanned aerial vehicle to autonomously put photovoltaic modules

Abstract

The invention relates to a path planning method for a cleaning robot for an unmanned aerial vehicle to autonomously launch a photovoltaic module, belongs to the technical field of photovoltaic operation and maintenance, and aims to solve the problem that a transportation unmanned aerial vehicle cannot autonomously plan a launching path due to the fact that the cleaning position of the photovoltaic module is uncertain. The method specifically comprises the following steps: calibrating pan-tilt camera parameters of the inspection unmanned aerial vehicle, installing a camera below the gravity center of the unmanned aerial vehicle, and ensuring that a lens is vertically downward; the shot image is stored in an SD card and is transmitted back to the ground; detecting dust deposition and bird droppings defects on the photovoltaic module by using an improved YOLOv8 model, marking a defect area and calculating a central point of the defect area; through RTK positioning error compensation, mapping is carried out by combining an image coordinate system and an ENU coordinate system, and latitude and longitude coordinates of the defect area in the photovoltaic module are solved; and finally, generating a delivery path of the transportation unmanned aerial vehicle based on the coordinates, and completing the autonomous delivery function of the cleaning robot. The experimental result verifies the reliability of the method in the path planning of the cleaning robot for autonomously putting the photovoltaic module in the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to a path planning method for a drone-based autonomous deployment robot for cleaning photovoltaic modules, belonging to the field of photovoltaic operation and maintenance technology. Background Technology

[0002] Against the backdrop of global energy transition, the photovoltaic (PV) industry, as a crucial component of clean energy, has experienced rapid growth in installed capacity in recent years and has secured a core position in the energy structure. However, during the operation of PV power plants, issues such as dust accumulation, bird droppings, or shading often occur on the surface of PV modules. These problems not only affect the photoelectric conversion efficiency of PV modules but may also accelerate equipment aging and reduce power generation efficiency. Therefore, efficient and clean operation and maintenance of PV modules are particularly important.

[0003] Currently, photovoltaic module cleaning mainly employs methods such as manual cleaning, drone cleaning, and tracked cleaning robots. Manual cleaning is labor-intensive, carries high safety risks, and is inefficient; while drone cleaning boasts strong mobility and terrain-crossing capabilities, it relies on the pilot and ground water supply, has limited range, and insufficient cleaning effect, especially poor at removing stubborn stains; tracked cleaning robots, on the other hand, with their high-efficiency cleaning and strong terrain-crossing capabilities, can adapt to complex environments and flexibly handle large-scale and localized spot cleaning, especially excelling in removing stubborn stains such as bird droppings, overcoming many limitations of manual and drone cleaning.

[0004] However, in common photovoltaic power station application scenarios such as elevated supports, mountain slopes, ponds, swamps, and deserts, tracked cleaning robots face challenges in deployment and retrieval. Using drones for the deployment and retrieval of cleaning robots can effectively overcome this challenge.

[0005] Drones offer significant advantages in transportation due to their ability to traverse complex terrain, overcoming the limitations of traditional roads and landscapes. They have demonstrated immense potential in special scenarios such as disaster relief, mountain transport, and material delivery to remote areas. However, in practical applications, particularly in confined and complex environments like photovoltaic power plants, achieving flexible drone transport and autonomous deployment of tracked cleaning robots remains challenging. The key issue is how to enable transport drones to accurately and autonomously transport and deploy cleaning robots onto defective photovoltaic modules without the need for pilot control. Therefore, precise positioning of defective photovoltaic modules and drone path planning are crucial.

[0006] Photovoltaic power plants have complex spatial layouts with limited spacing and dense distribution of photovoltaic modules. A current technological challenge is how to accurately identify photovoltaic modules covered in dust and bird droppings within these densely distributed arrays, and ensure that drones can avoid obstacles during flight and accurately fly to a specific height above the target modules to prepare for subsequent deployment of cleaning robots.

[0007] Therefore, accurately identifying the latitude and longitude coordinates of defective photovoltaic modules and using them as input for drone delivery path planning, in order to successfully plan a reasonable delivery path, is a prerequisite for drones to autonomously deliver photovoltaic module cleaning robots. Summary of the Invention

[0008] Technical Problem: This invention relates to a path planning method for autonomously deploying a photovoltaic module cleaning robot by a drone, aiming to solve the problem in existing technologies where the uncertain location of the photovoltaic modules to be cleaned prevents the transport drone from autonomously planning its deployment path. Existing methods rely on manual inspection to determine the location of the photovoltaic modules requiring cleaning, followed by the drone operator manually deploying the cleaning robot to the target photovoltaic modules, thus failing to achieve autonomous deployment by the drone.

[0009] Technical Solution: To achieve the above objectives, the present invention provides the following technical solution:

[0010] The present invention provides a path planning method for a drone-based autonomous photovoltaic module cleaning robot, comprising the following steps:

[0011] S1: Calibrate the internal and external parameters of the inspection drone's gimbal camera; adjust the drone's flight altitude and path according to the photovoltaic module's installation angle to ensure the gimbal camera lens is pointing vertically downwards to capture a large number of images of the photovoltaic module surface. The captured photovoltaic module images are stored on an SD card and then transmitted back to the ground server.

[0012] S2: Integrate a large number of photovoltaic module surface images obtained by inspection drones and open source data obtained from the Roboflow website into a new photovoltaic module defect image dataset, and preprocess the dataset, including data augmentation and image size adjustment, and divide the preprocessed dataset into training set, validation set and test set;

[0013] S3: An improved YOLOv8 model is used to detect dust and bird droppings defects on photovoltaic (PV) modules. Defect regions on the PV modules are bounded and their center points are calculated. First, channel attention and spatial attention mechanisms (CBAM) are added to the backbone feature extraction network to enhance attention to key features. Next, an RFB module is introduced to enhance the extraction capability of multi-scale features. Then, four multi-scale feature maps are fused, and the ASFF adaptive spatial feature fusion method is used to form a four-detection-head structure. Next, the original YOLOv8 loss function is modified to incorporate a Focal Loss mechanism in the bounding box IoU loss calculation to more effectively address the positive and negative sample imbalance problem in PV module bird droppings detection, thereby improving the detection accuracy of small-target bird droppings and further optimizing model performance. Finally, the pixel coordinates of the center point of the defect region detection box in the PV module image are calculated.

[0014] S4: By compensating for RTK positioning errors and based on the mapping relationship between the image coordinate system and the ENU coordinate system, the latitude and longitude coordinates of the center point of the defect area are accurately calculated. First, the RTK positioning error of the inspection drone is compensated by analyzing the RTK error ellipse. Then, the compensated RTK positioning data is matched with the center point of the photovoltaic module image in the image coordinate system. Next, the coordinates of the center point of the defect area detection box are mapped from the image coordinate system to the ENU coordinate system, and finally the latitude and longitude coordinates of the defective photovoltaic module are accurately calculated.

[0015] S5: Based on the obtained latitude and longitude coordinates of the center point of the defect area, a delivery path for the transport drone is generated to ensure that the drone can accurately deliver the cleaning robot. First, the take-off point and delivery point of the transport drone are determined according to the obtained latitude and longitude coordinates of the center point of the defect area. Second, the flight altitude and path of the drone are set to ensure that there are no obstacles during the flight. Finally, the optimal flight path is generated using a path planning algorithm, and the delivery path is integrated with the flight control system to ensure that the drone flies accurately to a specific height above the target photovoltaic module according to the predetermined path to complete the delivery task of the cleaning robot.

[0016] S6: Conduct detailed analysis and testing of the delivery path planning results to verify the feasibility and reliability of this method in the autonomous delivery of photovoltaic module cleaning robots by drones. Ensure that the delivery path generated based on these coordinates can ensure that the transport drone accurately completes the delivery path planning task, effectively supporting the autonomous delivery function of the cleaning robot.

[0017] Furthermore, in step S1, the intrinsic and extrinsic parameters of the inspection drone's gimbal camera are calibrated and photovoltaic module images are acquired. This prepares for mapping the defective photovoltaic module images to the ENU coordinate system and solving for the true latitude and longitude coordinates of the defective photovoltaic module. This includes the following steps:

[0018] S1.1: Initialize the calibration board, i.e., the checkerboard image, and complete the calibration of camera intrinsic parameters and distortion coefficients. The specific intrinsic parameter calibration method is as follows: First, prepare a checkerboard of known size, using each corner point on the calibration board as a reference in the 3D world coordinate system; capture multiple images of the calibration board; then detect and optimize the corner points of the calibration board; next, perform 3D-2D point pairing; finally, calculate the camera's intrinsic parameter matrix and distortion coefficients using the calibration algorithm.

[0019] S1.2: Using the takeoff point as the origin of the ENU coordinate system, calculate the camera's extrinsic parameters relative to the ENU coordinate system, including the rotation matrix (R) and translation vector (T), to describe the camera's position and orientation in space. Specifically, the extrinsic parameter calibration method is as follows: Using the takeoff point as the origin of the ENU coordinate system, calculate the rotation matrix of the body coordinate system relative to the ENU coordinate system based on the UAV's attitude angles; further, calculate the rotation matrix of the gimbal coordinate system relative to the UAV coordinate system based on the gimbal's attitude angles relative to the UAV; obtain the final rotation matrix of the camera coordinate system relative to the ENU coordinate system through these two rotation matrices; then, calculate the translation T using the camera's translation relative to the fuselage and the fuselage's position in the ENU. Finally, determine the camera's extrinsic parameters relative to the ENU coordinate system, including the rotation matrix (R) and translation vector (T), thus describing the camera's position and orientation in the ENU coordinate system.

[0020] S1.3: Adjust the flight altitude and path of the inspection drone according to the installation angle of the photovoltaic modules to ensure that the gimbal camera lens is pointing vertically downwards to capture a large number of images of the photovoltaic module surface. The captured photovoltaic module images are stored on an SD card and manually copied back to the ground server.

[0021] Furthermore, in step S2, the construction, data augmentation, and image preprocessing of the photovoltaic module defect image dataset include the following steps:

[0022] S2.1: Store the images of the photovoltaic modules captured by the inspection drone to an SD card and then transmit them back to the ground server. Specifically, using a DJI M300 as the inspection drone, the gimbal camera is positioned at a pre-set angle and shooting angle to ensure it is shooting vertically downwards at the surface of the photovoltaic modules. The captured images serve as a test set.

[0023] S2.2: A photovoltaic module defect image dataset with a unified label format will be generated by integrating public datasets from Roboflow. Specifically, the Roboflow open-source dataset will be organized and filtered to extract images and label data related to photovoltaic module defects. The filtered Roboflow dataset will be merged with a dataset of labeled images taken by drones to construct a comprehensive image dataset containing photovoltaic module defects such as bird droppings and accumulated dust, for use in model training and validation. This dataset will be divided into a training set and a validation set in an 8:2 ratio, which, together with the test set, constitute the photovoltaic module defect image dataset used in this invention.

[0024] S2.3: Perform data augmentation operations on the photovoltaic module defect image dataset to improve sample diversity and model robustness. Specifically, data augmentation operations are performed on the image dataset, including image translation, rotation, scaling, and stitching, aiming to enhance data diversity, simulate the actual conditions of photovoltaic modules under different working scenarios, and improve the model's robustness and generalization ability.

[0025] S2.4: The processed dataset is divided, and the images are resized to ensure consistent input image dimensions. Adaptive anchor box calculation is performed at the model's input. Anchor box parameters are continuously adjusted by comparing the differences between the predicted and actual labeled boxes (the size and proportion of the target in the dataset) to improve target detection accuracy.

[0026] Furthermore, in step S3, the improved YOLOv8 model is used to detect dust and bird droppings defects on the photovoltaic module, the defect area on the photovoltaic module is outlined and its center point is calculated, including the following steps:

[0027] S3.1: Build a YOLOv8 model as the basic architecture model for photovoltaic module defect detection, and improve upon it;

[0028] S3.2: A Convolutional Block Attention Module (CBAM) is introduced into the backbone network to alleviate the inherent limitations of traditional convolutional neural networks in representing cross-scale and multi-morphological targets. This mechanism achieves dynamic calibration of the feature map through serial channels and spatial attention sub-modules, enabling it to adaptively enhance discriminative features related to photovoltaic panel defects and suppress interference from irrelevant information, thereby improving the model's feature extraction quality and allowing the network to more accurately focus on key regions.

[0029] S3.3: A receptive field module (RFB) is inserted into the backbone network to improve the model's ability to extract features at multiple scales, addressing the inherent bottleneck of traditional deep convolutional networks in balancing large-scale and small-scale targets during feature extraction. This module employs a multi-branch dilated convolutional structure to construct heterogeneous receptive fields at a single layer in the network, thereby efficiently simulating multi-scale contextual information. This design significantly enhances the model's robust feature extraction capability for the scale heterogeneity of photovoltaic panel surface defects (such as large areas of dust accumulation and localized bird droppings), effectively solving the problem of the difficulty in coordinating macroscopic feature recognition and microscopic feature capture with a single receptive field.

[0030] S3.4: The four multi-scale feature maps output by the backbone feature extraction network are fused, and an adaptive spatial feature fusion method (ASFF) is introduced. Based on this, an ASFF four-detection head is proposed. Specifically, to better fuse small target features in the image, the traditional YOLO three-detection head structure is replaced, and the information from the P2 feature map is fused into the detection network to obtain a feature tensor for four-scale feature fusion for detection. Simultaneously, the traditional detection head module is modified, introducing an adaptive spatial feature fusion method (ASFF) to effectively filter out conflicting information and solve the consistency problem between different feature scales. This significantly improves the network's ability to detect small and multi-scale targets.

[0031] S3.5: The loss function of the YOLOv8 model is optimized by introducing the FocalLoss mechanism in the bounding box IoU loss calculation to better adapt to the problem of extreme imbalance between positive and negative samples in the detection of bird droppings on photovoltaic panels. By reducing the loss value of samples that have been correctly classified, the training focus of the model is placed more on samples that are difficult to classify.

[0032] Set training parameters to ensure that the model can converge stably in the later stages of training.

[0033] Load the pre-trained weights yolov8n.pt as the starting point for training. This accelerates convergence and improves the final performance of the model.

[0034] In each batch, forward propagation calculates a weighted sum of the bounding box regression loss, classification loss, and distribution focus loss. Backpropagation is then performed to calculate the gradient to minimize the loss function.

[0035] Real-time visualization of loss curves on the training and validation sets. Observe the difference between the two to determine if the model is overfitting. Monitor the trends of performance metrics such as mAP50 and mAP50-95 to determine the direction of model training.

[0036] S3.6: Determine the coordinates of the center point of the defect detection box in the photovoltaic module image.

[0037] Furthermore, in step S4, by compensating for RTK positioning errors and based on the mapping relationship between the image coordinate system and the ENU coordinate system, the latitude and longitude coordinates of the center point of the defect area are accurately calculated, including the following steps:

[0038] S4.1: Compensation for RTK errors in inspection drones based on RTK error ellipse analysis. Specifically: By solving the RTK error ellipse and analyzing it based on evaluation indicators such as maximum error, standard deviation, and 2DRMS, the main sources of error and directional deviations are identified. Based on the analysis results, the RTK positioning data is further adjusted to correct directional errors, and corresponding compensation measures are applied to improve the positioning accuracy of the drone and ensure the accuracy and reliability of inspection tasks.

[0039] S4.2: Match the compensated RTK positioning data with the center point of the photovoltaic module image in the image coordinate system to ensure that the latitude and longitude coordinates of the UAV at this time are the center point of the captured photovoltaic module image.

[0040] S4.3: Map the center point coordinates of the defect area detection box from the image coordinate system to the ENU coordinate system, and finally accurately calculate the latitude and longitude coordinates of the defective photovoltaic module in the ENU coordinate system.

[0041] Furthermore, in step S5, based on the obtained latitude and longitude coordinates of the center point of the defect area, the deployment path of the transport drone is generated, including the following steps:

[0042] S5.1: Based on the latitude and longitude coordinates of the center point of the defect area, determine the take-off point and drop point of the transport drone to ensure that the starting and ending positions of the path planning are accurate.

[0043] S5.2: Set the flight altitude and flight path of the transport drone to avoid potential obstacles, ensure flight safety, and meet the stable altitude required for the deployment of the cleaning robot.

[0044] S5.3: Generate the optimal delivery path using a path planning algorithm.

[0045] S5.4: Integrate the delivery path with the flight control system to ensure that the drone flies accurately to the designated height above the target photovoltaic module according to the predetermined path, and successfully completes the delivery task of the cleaning robot.

[0046] Furthermore, in step S6, the delivery path planning results are analyzed and tested in detail to verify the feasibility and reliability of the method in the task of autonomously delivering photovoltaic module cleaning robots by drones, including the following steps:

[0047] S6.1: Conduct detailed simulation tests of the delivery path to evaluate the accuracy and feasibility of the path planning. By simulating flight under different environmental conditions, ensure that the path planning meets the accuracy requirements in actual operation.

[0048] S6.2: Conduct on-site tests to verify the actual performance of the drone when flying along the planned path. By comparing the actual deployment location with the target location, evaluate the accuracy and stability of the path planning to ensure that the drone can accurately reach the designated location.

[0049] S6.3: Analyze the deviations in the test results and further optimize the path planning. By adjusting the parameters of the flight control system and path planning algorithm, the path accuracy is improved, ensuring stable operation of the UAV in complex environments and guaranteeing that it can successfully complete the UAV deployment path planning task under different environments and photovoltaic power station layouts.

[0050] Beneficial effects: Compared with the prior art, the path planning method for a drone-based autonomous photovoltaic module cleaning robot proposed in this invention has the following beneficial effects:

[0051] (1) Enhancing the flexibility and reliability of deployment path planning: Unlike traditional path planning methods, this invention first accurately locates the defective areas of photovoltaic modules and then plans the path based on the latitude and longitude coordinates of the defective modules. This method not only ensures the accuracy of path planning but also optimizes the deployment path, autonomously sets the UAV's flight altitude and optimal flight route, and ensures that there are no obstacles in the flight path, significantly enhancing the flexibility and reliability of path planning.

[0052] (2) Improving the Autonomous Deployment Accuracy of the Photovoltaic Panel Cleaning Robot: This invention further improves the positioning accuracy of defective photovoltaic modules through RTK positioning error compensation technology. Simultaneously, deep learning technology is used for defect detection, effectively solving the problem of false detection caused by fatigue in traditional manual inspection. Compared with manual inspection, the deep learning model has higher defect detection accuracy. By improving the positioning accuracy of the deployment target point, the cleaning robot can accurately deploy the modules, further enhancing the accuracy of autonomous deployment.

[0053] (3) Providing technical support for autonomous transportation and deployment of UAVs: Deployment path planning is the foundation for UAVs to achieve autonomous transportation and deployment of cleaning robots. Through the path planning method of this invention, it is ensured that UAVs can plan deployment tasks based on the precise coordinates of defective photovoltaic modules, providing technical support for autonomous transportation and deployment of UAVs in complex environments.

[0054] (4) Improve the transportation and delivery efficiency of photovoltaic panel cleaning robots: Drones can accurately locate and quickly reach the target modules to perform autonomous delivery operations, eliminating the need for manual inspection and location of defective photovoltaic modules or manual operation by drone pilots. Compared with traditional manual delivery methods, the use of drones for autonomous delivery of cleaning robots significantly improves delivery efficiency.

[0055] (5) Enhancing the intelligence level of photovoltaic power plant cleaning and maintenance: This invention solves the problem that photovoltaic module cleaning robots cannot be widely used for photovoltaic module cleaning due to the difficulty of deployment. The proposed path planning method for autonomous deployment of photovoltaic module cleaning robots by drones realizes the autonomous deployment of cleaning robots. Compared with traditional manual cleaning or drone cleaning methods, the use of photovoltaic module cleaning robots for cleaning significantly improves efficiency and effectiveness, and enhances the intelligence level of photovoltaic power plant cleaning and maintenance.

[0056] (6) Broad Application Value and Market Prospects: my country has a huge installed capacity of photovoltaic power generation, and the demand for clean operation and maintenance is growing daily. By using drones to autonomously transport and deploy photovoltaic panel cleaning robots to defective photovoltaic modules to clean bird droppings, dust, and other foreign objects, the cleaning efficiency of photovoltaic power plants can be significantly improved, operation and maintenance costs reduced, and the long-term stable operation of the power plants ensured. The method proposed in this invention has strong promotion potential and broad market prospects, providing the photovoltaic operation and maintenance industry with an efficient and low-cost cleaning solution, and promoting the rapid development of the photovoltaic clean energy field. Attached Figure Description

[0057] Figure 1 This is a flowchart of a path planning method for an unmanned aerial vehicle (UAV) autonomously deploying a photovoltaic module cleaning robot, as proposed in this invention.

[0058] Figure 2 This is a partial image representation of the dataset used in this invention;

[0059] Figure 3 This is the YOLOv8 benchmark model architecture for detecting defects in photovoltaic modules upon which the method of this invention is based;

[0060] Figure 4 This is a structural diagram of the CBAM module, which introduces the channel attention and spatial attention mechanism into the YOLOv8 backbone feature extraction network in this invention.

[0061] Figure 5 This is a structural diagram of the RFB module introduced into the YOLOv8 backbone feature extraction network in this invention;

[0062] Figure 6 The structure diagram of the ASFF adaptive spatial feature fusion network introduced in this invention is shown below;

[0063] Figure 7 This invention provides an improved YOLOv8 model architecture for detecting defects in photovoltaic modules.

[0064] Figure 8 Table of ablation experiment results for the improved YOLOv8 model;

[0065] Figure 9This is a graph showing the coordinates of the center point of the detection box for the defect area of ​​a photovoltaic module based on the improved YOLOv8 detection method.

[0066] Figure 10 A flowchart illustrating the training approach used to train the photovoltaic module defect detection model proposed in this invention.

[0067] Figure 11 This is an ellipse diagram of the RTK error on which the present invention is based to compensate for the RTK error of the inspection drone.

[0068] Figure 12 The result of matching RTK positioning data with the center point of the photovoltaic module image in the image coordinate system;

[0069] Figure 13 This is a diagram showing the optimal delivery path generated by the present invention for the autonomous delivery of photovoltaic module cleaning robots by drones. Detailed Implementation

[0070] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. However, the described embodiments are only some embodiments of the present invention, and the protection scope of the present invention is not limited to the described embodiments.

[0071] This invention provides a path planning method for an unmanned aerial vehicle (UAV) autonomously deploying a photovoltaic (PV) module cleaning robot. Through efficient image acquisition and processing strategies, it accurately locates defects such as bird droppings and dust accumulation on the PV modules, obtains the latitude and longitude information of the defect area, and uses this information as the endpoint for the UAV's flight. Finally, based on the target deployment point coordinates, a path planning algorithm generates the optimal flight path and integrates it with the flight control system to ensure that the UAV accurately flies to a designated altitude above the target PV module along the predetermined path, completing the autonomous deployment task. The following describes the method in conjunction with... Figures 1-13 This embodiment will be described in detail below.

[0072] S1: UAV gimbal camera parameter calibration. Based on the installation angle of the photovoltaic module, adjust the flight altitude and path of the inspection UAV to ensure that the camera lens is vertically downward to shoot the surface of the photovoltaic module and obtain a large number of photovoltaic module surface images.

[0073] Specifically, it includes the following sub-steps:

[0074] S11: Camera Intrinsic Calibration: Prepare a checkerboard or dot array of known size as a calibration board. The position of each corner point or center point on the calibration board is known and serves as a reference point in the 3D world coordinate system. Take multiple images containing the calibration board from different angles to cover as many viewpoint variations as possible. Use these images to extract feature points for camera intrinsic parameter calibration. Use image processing algorithms (OpenCV's cv2.findChessboardCorners) to detect the corner points or centers of the calibration board in each image and record their pixel coordinates. To improve accuracy, use sub-pixel level optimization methods (such as cv2.cornerSubPix) to further optimize the corner point positions. Based on the known size and number of corner points of the calibration board, prepare the 3D world coordinate points (actual coordinates) and corresponding 2D image points (pixel coordinates) for each image. Use a calibration algorithm (cv2.calibrateCamera) to calculate the camera's intrinsic parameter matrix and distortion coefficients. The algorithm then inputs 3D-2D point pairs from multiple images for global optimization. The accuracy of the calibration results is verified by performing distortion correction and reprojection error analysis on the calibration images.

[0075] S12: Camera extrinsic parameter calibration:

[0076] Based on the UAV's attitude angles, obtain the rotation matrix from the body coordinate system {B} to the ENU coordinate system {E}:

[0077]

[0078] The rotation matrix of the gimbal relative to the body coordinate system {B} is obtained based on the attitude angle {C} of the gimbal camera relative to the drone:

[0079] =

[0080] Therefore, the camera's extrinsic rotation matrix can be obtained. The translation T is used to describe the camera's position and orientation in space:

[0081] The final rotation matrix of the camera relative to the ENU = Rotation matrix of machine body → ENU × Rotation matrix of camera → machine body = × ;

[0082] Translation amount T = Translation of the camera relative to the body + Position of the body in the ENU.

[0083] S13: Adjust the drone's flight altitude and path according to the installation angle of the photovoltaic module, while ensuring that the gimbal camera lens is pointing vertically downwards to photograph the surface of the photovoltaic module; take multiple shots along the preset flight path to collect a sufficient number of images of the photovoltaic module surface, providing a foundation for subsequent dataset construction and analysis.

[0084] S2: Integrate a large number of photovoltaic module surface images obtained by drones and Roboflow open source data into a new photovoltaic module defect image dataset, preprocess the dataset, and then divide it into training set, validation set, and test set.

[0085] Specifically, it includes the following steps:

[0086] S21: Filter image compressed packages related to dust accumulation and bird droppings defects in photovoltaic modules from the Roboflow open-source dataset, and download and export them in YOLOv8 label format.

[0087] S22: By configuring the 'translate', 'scale', 'flipud', 'fliplr', and 'mosaic' parameters in the default hyperparameter file 'hyp.scratch.yaml' of YOLOv8, operations such as translation, scaling, flipping, and mosaicking of image datasets are implemented to enhance the diversity of data, simulate the actual situation of photovoltaic modules under different working scenarios, and improve the robustness and generalization ability of the model.

[0088] S23: The dataset contains 3873 surface images of photovoltaic modules with defects such as dust and bird droppings, divided in an 8:2 ratio as the training and validation sets; 300 images of photovoltaic modules in real-world application scenarios, taken by inspection drones as required, serve as the test set. The training set is used for model training, the validation set for hyperparameter tuning, and the test set for evaluating model performance. A partial image of the dataset is shown below. Figure 2 As shown.

[0089] S24: Standardize the image size to ensure consistent input image dimensions for easier model processing and training; adjust the image size to 640×640 using the command-line parameter `--img-size`; perform adaptive anchor box calculation at the model input. By comparing the difference between the predicted and actual labeled boxes, the anchor box parameters are automatically adjusted to improve object detection accuracy; the `autoanchor.py` script in YOLOv8 is used to generate suitable anchor box parameters based on the labeled boxes in the dataset using the k-means clustering algorithm; the calculation results are then used to replace the `anchors` field in the model configuration file to complete the adaptive anchor box settings.

[0090] S3: Use the improved YOLOv8 model to detect dust and bird droppings defects on photovoltaic modules, outline the defect areas on the photovoltaic modules and calculate their center points.

[0091] Specifically, it includes the following steps:

[0092] S31: Introduce the channel attention and spatial attention mechanism CBAM (Convolutional Block Attention Module) into the backbone feature extraction network.

[0093] CBAM mainly includes channel attention mechanism and spatial attention mechanism, and its structure diagram is as follows: Figure 4 As shown:

[0094] Channel attention mechanism:

[0095] Spatial attention mechanism:

[0096] in Let F be the input feature map, W0 be the sigmoid activation function, and W1 be the weight matrix for increasing dimensionality. For each channel of F, the average of all pixel values ​​is calculated. For each channel of F, take the maximum value of all its pixel values. It is a standard convolutional layer with a kernel size of 7x7. It is the average of the values ​​of the input feature map F across all C channels. It is to find the maximum value of F across C channels.

[0097] A convolutional attention module was introduced after the SPPF module of the backbone network. Figure 7 This mechanism aims to alleviate the inherent limitations of traditional convolutional neural networks in representing cross-scale, multi-morphological targets. Through serial channels and a spatial attention submodule, it dynamically calibrates the feature map, adaptively enhancing discriminative features related to photovoltaic panel defects and suppressing irrelevant information interference. This improves the model's feature extraction quality, enabling the network to more accurately focus on key regions.

[0098] S32: Introduce the RFB module into the backbone feature extraction network to improve the model's ability to extract features at multiple scales.

[0099] The dataset related to this invention exhibits a significant bimodal distribution in scale, with the targets to be detected including both minute defects such as bird droppings and large-area targets such as ash accumulation. When traditional convolutional neural networks process such multi-scale targets, their receptive fields tend to be singular, making it difficult to collaboratively represent subtle textures and macroscopic structures on a unified feature layer. This results in insufficient adaptability of the model to targets of different scales, limiting its detection performance.

[0100] To address the aforementioned issues, this invention optimizes the backbone network structure by integrating a multi-scale receptive field enhancement module after the SPPF module. Figure 7 This module simulates the hierarchical distribution of receptive fields in the human visual system, employing a multi-branch parallel topology. Figure 5 Its core design includes multiple convolutional branches, each employing dilated convolution operations with different dilation rates. This effectively extracts contextual information at different spatial scales while maintaining the feature map resolution. The output features of each branch are aggregated through a concatenation operation, and finally, a 1×1 convolution is used to achieve channel-level fusion and dimensionality reduction.

[0101] Specifically, this module achieves multi-scale context extraction and fusion through the following steps. The RFB module structure is as follows: Figure 5 As shown: (1) Perform multi-path parallel convolution processing on the feature map output by the SPPF module, where each path uses a different convolution kernel configuration (especially by adjusting the dilation rate of the dilated convolution) to capture the contextual dependencies of different spatial ranges; (2) Concatenate the feature maps output by each parallel branch in the channel dimension to form a fusion feature containing multi-scale information; (3) Perform channel compression and information integration on the concatenated feature through a 1×1 convolution layer to generate an enhanced multi-scale feature representation.

[0102] This design enables the network to simultaneously preserve the accuracy of local details and the integrity of the global context on a single feature layer, providing a more robust feature foundation for subsequent detection networks.

[0103] S33: The four multi-scale feature maps output by the backbone feature extraction network are fused together, and the ASFF adaptive spatial feature fusion method is introduced. Based on this, an ASFF four-detection head is proposed.

[0104] Because the dataset of this invention includes small targets such as bird droppings, the saliency features of these small targets are easily diluted during multiple downsampling processes in traditional deep convolutional neural networks. This results in traditional models exhibiting weak feature responses and high false negative rates when dealing with small targets. This invention extends the baseline feature pyramid network by adding layers. Based on the traditional feature pyramid consisting of P3, P4, and P5 layers, a P2 feature map from a shallow layer of the backbone network is introduced and fused. This layer's feature map has higher spatial resolution and richer fine-grained texture information, effectively preserving the complete pixel-level features of small targets. This compensates for the loss of detail caused by increased network depth, providing a more robust feature representation for small target detection. Furthermore, since the four detection heads contain multi-layered complex semantic information, while traditional object detection networks typically only utilize a single-layer feature map, an adaptive spatial feature fusion method (ASFF) is introduced. Figure 6 By optimizing the feature fusion process, it is specifically designed to integrate feature maps at different levels, thereby improving the accuracy and performance of target detection.

[0105] Its ASFF detection head structure ( Figure 6 The specific operation steps are as follows:

[0106] (1) ASFF-1 node processing (Level-4)

[0107] Input: Level-4 (stride=4) high-resolution features

[0108] Processing: Directly use 1×1 convolution to adjust the channels, generating feature maps X4→1.

[0109] Function: To retain the richest spatial detail information, providing basic features for subsequent fusion.

[0110] (2) ASFF-2 node processing (Level-3)

[0111] Level-4 processing: Perform a 3×3 convolution on the Level-4 features + downsampling with a stride of 2 to obtain X4→2.

[0112] Level-3 processing: Perform a 1×1 convolution on the Level-3 features to generate X3→2.

[0113] Level-2 processing: Perform a 3×3 convolution on the Level-2 features + bilinear upsampling by 2 times to obtain X2→2.

[0114] Function: To achieve cross-scale alignment between high-resolution details and mid-level semantic features.

[0115] (3) ASFF-3 node processing (Level-2)

[0116] Level-4 processing: Perform a 3×3 convolution on the Level-4 features and downsample with a stride of 4 to obtain X4→3.

[0117] Level-3 processing: Perform a 1×1 convolution on the Level-3 features + bilinear upsampling by 2 times to obtain X3→3.

[0118] Level-2 processing: Perform a 1×1 convolution on the Level-2 features to generate X2→3.

[0119] Level-1 processing: Perform a 3×3 convolution on the Level-1 features + bilinear upsampling by 2 times to obtain X1→3

[0120] Function: Integrates contextual information at four different scales, balancing detail and semantics.

[0121] (4) ASFF-4 node processing (Level-1)

[0122] Level-4 processing: Perform a 3×3 convolution on the Level-4 features and downsample with a stride of 8 to obtain X4→4.

[0123] Level-3 processing: Perform a 1×1 convolution on the Level-3 features + bilinear upsampling by 4 times to obtain X3→4.

[0124] Level-2 processing: Perform a 1×1 convolution on the Level-2 features + bilinear upsampling by 2 times to obtain X2→4.

[0125] Level-1 processing: Perform a 1×1 convolution on the Level-1 features to generate X1→4.

[0126] Function: To unify multi-scale features to the lowest possible resolution, thereby enhancing high-level semantic representation capabilities.

[0127] Adaptive weighted fusion

[0128] After completing the feature transformation and alignment described above, learnable weights are used to perform weighted fusion of features at each level. Taking ASFF-4 as an example, its output features are determined by the following formula:

[0129]

[0130] Weight This is achieved by performing a 1×1 convolution on the feature maps from level 1 to level 4, followed by concat and then softmax to ensure that the range is [0,1] and the sum is 1. The formula is shown below.

[0131]

[0132] S34: Improve the target bounding box regression loss function and further propose the Focal-Loss loss function.

[0133] Because bird droppings datasets are characterized by a large amount of background information and a limited amount of foreground information, traditional cross-entropy loss functions often fail to adequately identify foreground targets due to the dominance of background samples in such extremely imbalanced datasets. The Focal-Loss loss function is introduced, which involves a probability modulation factor with an adjustable focusing parameter to dynamically scale the loss weights of samples. This design significantly suppresses gradients generated by a large number of easily classified negative samples, redistributing the model's learning focus to challenging, difficult samples. The implementation of this method effectively enhances the model's sensitivity and robustness in detecting small-scale targets and low-salience targets in complex scenes.

[0134] The formula for the Focal-Loss loss function is as follows:

[0135]

[0136] in: It is the probability that the model predicts a sample as type t; Let be the weight coefficient of category t. These are the model's hyperparameters, used to control the changes in weights; To adjust the factor, the loss value of samples that have been correctly classified is reduced.

[0137] S35: Determine the coordinates of the center point of the defect detection box in the photovoltaic module image.

[0138] Finding the center point:

[0139] (1) Symbols and Input - Input Image The color space is BGR; For height, Width. - The detection model outputs a bounding box for each defect. ,satisfy , (Ideal case). - Let the pixel coordinates on the integer grid be denoted as . All drawing and annotation are performed on an integer grid.

[0140] (2) Bounding box coordinate legalization

[0141] Map the bounding box coordinates to an integer grid:

[0142]

[0143] Perform image domain constraints (squeeze mapping) to eliminate out-of-bounds errors:

[0144]

[0145]

[0146]

[0147]

[0148] The squeeze operator is defined as follows: .

[0149] If a degenerate box appears The box then undergoes a minimum magnitude correction to restore its geometric meaning. .

[0150] (3) Determining the coordinates of the center point (geometric definition and integer meshing)

[0151] Theoretical geometric center:

[0152]

[0153] Stable integerization is achieved by rounding down, resulting in the center point on the integer grid:

[0154]

[0155] Image domain constraints ensure that the annotation base points are located within the domain:

[0156]

[0157] S4: By compensating for RTK positioning errors and based on the mapping relationship between the image coordinate system and the ENU coordinate system, the latitude and longitude coordinates of the center point of the defect area are accurately solved.

[0158] Specifically, it includes the following steps:

[0159] S41: Solve for the RTK error ellipse.

[0160] First, select 5 points and measure the true coordinates of each point using an RTK measuring instrument. Then, obtain 10 RTK positioning data for each point and solve for the RTK error ellipse based on this.

[0161] For each point, calculate the deviation between the RTK measurement result and the true coordinates for each measurement:

[0162]

[0163] in, These are the actual coordinates. These are the coordinates measured by RTK.

[0164] Error data for each point and Perform statistical analysis to calculate the mean and variance of the error.

[0165] Calculate the error covariance matrix for each point:

[0166]

[0167] in and It is the variance of the error. It is covariance.

[0168] Based on covariance matrix Calculate the lengths of the principal axis and semi-axis of the RTK error ellipse, and calculate the direction of the principal axis of the error ellipse. ,Right now:

[0169]

[0170] Calculate the length of the semi-axis of the error ellipse; the length of the principal semi-axis is:

[0171]

[0172] The length of the secondary half-shaft is:

[0173]

[0174] Using the parameters of the error ellipse (major axis length) Secondary shaft length and the direction angle of the ellipse Draw the error ellipse in the plane coordinate system.

[0175] S42: RTK error analysis.

[0176] Maximum error: (Safety tolerance analysis index)

[0177] Definition: For a set coordinate points The maximum error is defined as the maximum Euclidean distance between any two points:

[0178]

[0179] Significance: It represents the most extreme positioning deviation in this set of measurement points, which can be understood as the error performance under the worst case, and is used for safety tolerance analysis.

[0180] Judgment criteria: The smaller the better, indicating that the measurement system has not experienced a serious deviation at any time.

[0181] Standard deviation: (Stability analysis indicator)

[0182] Regarding direction , The standard deviation of the coordinate sample is defined as:

[0183]

[0184] Meaning: Represents the range of fluctuation of a point in a certain direction. It can be used to measure the stability in the east-west (longitude) and north-south (latitude) directions respectively.

[0185] Judgment criterion: The smaller the value, the more stable the system.

[0186] Error ellipse (directional analysis index)

[0187] Mathematical method: Based on the 2D covariance matrix Σ:

[0188]

[0189] Perform eigenvalue decomposition on Σ: the largest eigenvalue λ1 → the major axis radius a = Minimum eigenvalue λ² → Minor axis radius b = The corresponding eigenvector direction → ellipse rotation angle θ.

[0190] When drawing, the major axis a = 2 is typically used. minor axis b=2

[0191] Significance: The elliptical coverage area represents the error distribution region with a certain confidence level, indicating the strength and correlation of errors in different directions. The rotation angle indicates whether there is a significant shift in the principal direction of the error.

[0192] Judgment criteria: The smaller the ellipse, the more concentrated and stable the positioning; the shorter the major axis of the ellipse / the lower the flattening, the more uniform the error and the no directional deviation; the ellipse is elongated or deviates in a certain direction, indicating the existence of directional systematic error.

[0193] 2DRMS: (Positioning accuracy index)

[0194] Assume the standard deviation of the coordinate error (relative to the center) of a set of points is ,but:

[0195]

[0196] Meaning: This indicates that there is a 95% probability that the measured point falls within the two-dimensional error circle of this radius.

[0197] Judgment criterion: The smaller the value, the higher the positioning accuracy.

[0198] S43: RTK error compensation.

[0199] The true coordinate values ​​and corresponding RTK positioning data are represented in the ENU coordinate system with the RTK base station as the origin. The eastward and northward errors of each set of measurements relative to the true values ​​are calculated, and the average value of the measurement errors in each direction is taken as the system average deviation. The system average deviation in the eastward and northward directions is subtracted from the original RTK measurement values ​​to obtain the compensated ENU coordinates. Then, the compensated RTK measurement values ​​and ENU coordinates are converted into latitude and longitude coordinates to obtain high-precision positioning data.

[0200] S44: Match the compensated RTK positioning data with the center point of the photovoltaic module image in the image coordinate system to ensure that the RTK information of the UAV can accurately describe the latitude and longitude position information of the center point of the photovoltaic module image.

[0201] The compensated RTK positioning data is displayed in conjunction with the photovoltaic module image in the image coordinate system. Figure 12 The blue dots represent the localization points in the image before RTK compensation, while the red dots represent the localization points after RTK compensation. The latitude and longitude information corresponding to the image center point is also calculated. It can be seen that the blue and red dots largely overlap, proving that the RTK positioning error is relatively small.

[0202] S45: Map the center point coordinates of the defect detection box in the photovoltaic module image to the ENU coordinate system, and solve for the latitude and longitude coordinates of the photovoltaic module with defects.

[0203] (1) Pixel coordinate system -> Imaging plane coordinate system:

[0204]

[0205] Represents coordinates in the pixel coordinate system. Represents the coordinates in the imaging plane coordinate system. This represents the coordinates of the principal point in the imaging plane coordinate system in the pixel coordinate system. , Indicates the camera is in Effective focal length in the direction.

[0206] (2) Imaging plane coordinate system -> Camera coordinate system:

[0207]

[0208] Represents coordinates in the pixel coordinate system. Represents the coordinates in the camera coordinate system. This indicates the camera's focal length.

[0209] (3) Camera coordinate system -> World coordinate system (ENU)

[0210]

[0211] Represents the coordinates in the camera coordinate system. Represents coordinates in the northeast-northeast coordinate system. Denotes the extrinsic parameter matrix of the camera, where For rotation matrix, It is a translation vector.

[0212] (4) Combine the coordinates in the ENU coordinate system with the coordinates and latitude and longitude of the origin in the ECEF coordinate system to calculate the latitude and longitude of the target point.

[0213] ENU->ECEF:

[0214]

[0215]

[0216] This represents the coordinates of the origin of the ENU coordinate system in the ECEF coordinate system. This indicates the coordinates of the target point in the ECEF coordinate system. This represents the coordinates of the target point in the ENU coordinate system. ϕ is the rotation matrix from the ENU coordinate system to the ECEF coordinate system, and ϕ and λ are the latitude and longitude of the ENU origin.

[0217] ECEF coordinates converted to latitude and longitude

[0218] Based on the (X,Y,Z) information in the ECEF coordinate system, the latitude and longitude information of the target point is obtained:

[0219]

[0220]

[0221] Decoupling yields:

[0222]

[0223]

[0224] In the formula, For the latitude and longitude information of the target point, , Let be the radius of curvature of the zonal loop at that point. , This is the semi-major axis of the ellipsoid corresponding to this geodetic coordinate system. This is the minor semi-axis of the ellipsoid corresponding to this geodetic coordinate system. This represents the first eccentricity of the ellipsoid corresponding to this geodetic coordinate system. Note: The accuracy L is calculated using the arctan2 function to distinguish between quadrants, preventing a 180° discrepancy from the actual value when x is less than 0.

[0225] S5: Based on the obtained latitude and longitude coordinates of the center point of the defect area, generate the delivery path of the transport drone to ensure that the drone can accurately deliver the cleaning robot.

[0226] Specifically, it includes the following steps:

[0227] S51: Define the starting and target points of the UAV flight and set key flight altitude parameters;

[0228] (1) Starting point The preset safe takeoff area is taken as the starting point, and its coordinates and ground altitude are defined as follows: ( )and o

[0229] (2) Target point ( Take the latitude and longitude coordinates of the defective photovoltaic module calculated in step S4. And set it as the horizontal target point for the task deployment.

[0230] (3) Safe cruising altitude Set a safe cruising altitude. This altitude must be above all potential obstacles within the photovoltaic power station (such as photovoltaic array supports and monitoring poles) to ensure that the drone has no risk of collision during level flight.

[0231] (4) Drop preparation height Set the deployment preparation height directly above the target point. This height is designed to balance deployment accuracy and safety distance; it needs to be low enough to ensure accurate deployment while maintaining a safe distance from the surface of the photovoltaic modules.

[0232] S52: Generates a sequence of flight commands based on waypoint control to guide precise flight from the starting point to the target point. This process avoids complex traditional path planning algorithms and adopts a two-point straight-line flight control mechanism. The specific command sequence is as follows:

[0233] Command 1: Vertical Climb. The drone performs a vertical climb from the starting point (Lon_start, Lat_start) until it reaches the safe cruising altitude H_cruise.

[0234] Command 2: Horizontal Flight. The UAV maintains altitude H_cruise and flies along the shortest straight path to the airspace above the target point (Lon_target, Lat_target).

[0235] Command 3: Hover over the target point. After arriving in the airspace directly above the target point, the drone enters a hovering state to prepare for descent.

[0236] Command 4: Vertical Descent. The drone descends vertically from its safe cruise altitude H_cruise, and after reaching the preset deployment preparation altitude H_deploy, it enters a stable hovering state again.

[0237] Command 5: Execute Deployment. At the deployment preparation height H_deploy, the drone triggers the deployment mechanism to complete the precise deployment of the photovoltaic module cleaning robot, causing it to land on the target photovoltaic module below.

[0238] S53: Integrates waypoint command sequences with the flight control system; it packages waypoint command sequences containing origin, destination, and key altitude information into a mission file that the flight control system can recognize. This mission file is then sent to the transport UAV's flight control module. The flight control module automatically parses the file and executes the flight commands sequentially. The entire process encompasses autonomous takeoff, straight-line navigation, hovering over the target point, vertical descent to the delivery altitude, and finally, delivery, all without human intervention.

[0239] S6: Analyze and test the delivery path planning results to verify the reliability of the method in the task of autonomously delivering photovoltaic module cleaning robots by drones.

[0240] Specifically, it includes the following steps:

[0241] S61: Evaluation of a photovoltaic module defect detection model based on improved YOLOv8

[0242] (1) Accuracy:

[0243] It measures the proportion of actual positive samples among those predicted by the model. It focuses on the accuracy of the prediction results. Its calculation formula is:

[0244]

[0245] in:

[0246] TP (True Positive): The number of targets correctly detected by the model.

[0247] FP (False Positive): False positives, which is the number of false positives (misclassifying background or non-target objects as targets) by the model.

[0248] High accuracy means fewer false positives from the model, resulting in more reliable predictions.

[0249] (2) Recall: Measures the proportion of all true positive samples that are successfully predicted by the model. It focuses on the model's ability to detect the target. Its calculation formula is:

[0250]

[0251] FN (False Negative) refers to the number of targets that the model misses (actually exist but are not detected).

[0252] (3) Average accuracy:

[0253] Precision and recall alone often fail to reflect the overall performance of a model. A comprehensive evaluation is typically achieved by plotting a Precision-Recall (PR) curve. Average Precision (AP) is the area under the PR curve, providing a comprehensive picture of the model's performance on a single class. A higher AP value indicates that the model is both accurate and comprehensive on that class. Its calculation formula is as follows:

[0254]

[0255] Mean, average precision:

[0256] Mean Average Precision (mAP) is the average AP value across multiple classes and is a core metric for measuring the overall performance of a model on the entire dataset.

[0257] mAP@50: refers to the mAP at an intersection-union (IU) threshold of 0.5. mAP@50:95: refers to the average mAP calculated at 10 different IU thresholds ranging from 0.5 to 0.95 (with a step size of 0.05). This metric is more stringent, requiring higher localization accuracy in the predicted bounding boxes, thus providing a more comprehensive evaluation of the model's overall performance. It is currently recognized as an authoritative metric in both academic research and industry.

[0258] In evaluating the series of models for detecting dust accumulation and bird droppings on photovoltaic panels proposed in this invention, a comprehensive performance comparison analysis was conducted. The baseline model (YOLOv8) performed best in terms of accuracy (P=91.3%). Figure 8 This indicates that its false alarm rate is the lowest.

[0259] However, considering the practical application scenarios of detecting dust accumulation and bird droppings on photovoltaic panels, this invention has specifically prioritized various indicators. In this task, the cost of missed detections far outweighs that of false positives. If a dusty area or bird droppings is not detected, it may lead to a continuous decline in the efficiency of the photovoltaic panel. False positives, on the other hand, can be eliminated through subsequent review processes. Therefore, recall (R) becomes a key indicator for measuring the model's practicality, directly reflecting its ability to detect defects.

[0260] Based on this consideration, the present invention analyzes the performance of each model: Model M-5 performs excellently in the core metrics. Although its precision (P=88.3%) is slightly lower than the baseline model, its recall (R=89.7%) is the highest among all models, proving that the model can effectively reduce missed detections and identify potential defects on photovoltaic panels to the greatest extent, perfectly meeting the task requirements.

[0261] Model M-5 also exhibits superior overall localization accuracy and robustness. Mean Average Precision (mAP@50:95) is a more stringent evaluation metric than mAP@50, requiring not only accurate classification but also higher localization precision in the predicted bounding boxes. M-5 achieves an mAP@50:95 of 60.8%, the highest among all compared models. This result demonstrates that M-5 exhibits more stable and accurate detection performance across different overlap thresholds, possessing stronger generalization ability and practicality.

[0262] In summary, simply pursuing a single metric (such as the precision of the baseline model) cannot meet the actual needs of industrial testing. This invention, by introducing modules such as CBAM, RFB, and Four-ASFF, and optimizing the loss function, achieves a final model M-5 that not only maintains high precision but also significantly improves the key metric recall (R) and the overall performance metric mAP@50:95. This successfully optimizes performance in specific application scenarios, fully validating the effectiveness and advancement of the technical solution presented in this invention.

[0263] S62: Using the default image center point coordinates as the target delivery point, a photovoltaic power station actually used on the roof of a central building is selected as the test scenario to conduct a detailed analysis and test of the delivery path planning results.

[0264] Specifically, it includes the following steps:

[0265] Multiple tests and analyses of the delivery path planning results were conducted to verify the feasibility and reliability of the proposed method in the autonomous delivery of photovoltaic module cleaning robots by drones. The delivery path generated based on these coordinates ensures that the transport drone accurately completes the delivery path planning task, effectively supporting the autonomous delivery function of the cleaning robot. Experimental test results show that the proposed method can be well used for path planning of autonomous delivery of photovoltaic module cleaning robots by drones. The test results are as follows: Figure 13 As shown.

[0266] Path element explanation ( Figure 13 ):

[0267] (1) Takeoff point: The green dot is located in the lower left safe zone and is marked "takeoff point".

[0268] (2) Vertical climb: The blue vertical arrow indicates that the vehicle climbs to the safe cruising altitude H_cruise.

[0269] (3) Horizontal flight: Blue horizontal arrow, fly straight from the end of the climb to the airspace directly above the target (hovering point).

[0270] (4) Vertical descent: Orange vertical arrow, descends vertically from H_cruise to the deployment preparation height H_deploy.

[0271] (5) Deployment location: Red target point (near the center of the image), with the target latitude and longitude text marked on the side; "Deployment / Deployment" is marked at the endpoint.

[0272] The above path planning satisfies the following constraints:

[0273] (1) Safety height constraint: Set H_cruise higher than potential obstacles such as photovoltaic brackets, monitoring poles, and cables; direct diagonal lines may cross obstacles at low heights, and the risk is uncontrollable.

[0274] (2) Drop accuracy: After hovering directly above the target, “vertical descent → drop” can minimize lateral speed and lateral drift caused by wind; when arriving at an angle, there is often still horizontal speed, which increases the drop deviation.

[0275] (3) Mission compliance and verifiability: Segmented waypoints (climb / level / descent) facilitate strict execution and recording of flight control, and make it easier to conduct safety audits and fault tolerance; straight free flight is not conducive to safety boundary control.

[0276] (4) Rotor underwash control: Set H_deploy to maintain a safe distance from the component surface to reduce the impact of underwash on the cleaning robot and components.

[0277] The above path planning result diagram is a schematic diagram of the "shortest feasible" straight route plus vertical segment, which meets the safety and accuracy requirements under the obstacle-free constraint; if there are obstacles or no-fly zones, it can be switched to the constrained optimal path and obstacle avoidance waypoints can be marked on the diagram.

Claims

1. A path planning method for an unmanned aerial vehicle autonomously launching a photovoltaic module cleaning robot, characterized in that, The method comprises the following steps: S1: UAV gimbal camera parameter calibration, for subsequent defect photovoltaic module image mapping to ENU coordinate system, solving the real latitude and longitude coordinates of the defect photovoltaic module makes preparation; according to the installation angle of the photovoltaic module, the flight height and path of the inspection UAV are adjusted to ensure that the camera lens is vertically downward to shoot the surface of the photovoltaic module, and a large number of photovoltaic module surface images are obtained; S2: a large number of photovoltaic module surface images obtained by the UAV and open source data integrated new photovoltaic module defect image data set downloaded from Roboflow are combined, the data set is preprocessed, and then divided into training set, verification set; S3: the improved YOLOv8 model is used to detect the dust and bird droppings defects on the photovoltaic module, and the center point of the defect area on the photovoltaic module is calculated; S4: the RTK positioning error is compensated, and based on the mapping relationship between the image coordinate system and the ENU coordinate system, the latitude and longitude coordinates of the center point of the defect area are accurately solved; S5: based on the obtained latitude and longitude coordinates of the center point of the defect area, the delivery path of the transport UAV is generated to ensure that the UAV can accurately deliver the cleaning robot; S6: the delivery path planning result is analyzed and tested to verify the reliability of the method in the task of the UAV autonomously delivering the photovoltaic module cleaning robot.

2. The path planning method for autonomously launching a photovoltaic module cleaning robot from a UAV according to claim 1, wherein, The step S1 is for subsequent defect photovoltaic module image mapping to ENU coordinate system, solving the real latitude and longitude coordinates of the defect photovoltaic module makes preparation, comprising: S1.1: initialize the calibration board, i.e. the checkerboard image, complete the camera intrinsic parameter and distortion coefficient calibration; S1.2: taking the take-off point as the origin of the ENU coordinate system, calculating the camera external parameters relative to the ENU coordinate system, including rotation matrix (R) and translation vector (T), describing the position and direction of the camera in space; S1.3: according to the installation angle of the photovoltaic module, the flight height and path of the UAV are adjusted, and at the same time, the gimbal camera lens is ensured to be vertically downward to shoot the surface of the photovoltaic module. 3.The path planning method for autonomously launching a photovoltaic module cleaning robot by a UAV according to claim 1, wherein, The step S2 comprises: S2.1: store the photovoltaic module images shot by the inspection UAV to the SD card and return to the ground server; S2.2: generate the photovoltaic module defect image data set containing uniform label format by integrating the public data set on Roboflow; S2.3: perform data enhancement operation on the photovoltaic module defect image data set to improve the diversity of samples and the robustness of the model; S2.4: divide the processed data set.

4. The path planning method for autonomously launching a photovoltaic module cleaning robot from a UAV according to claim 1, wherein, The step S3 comprises: S3.1: build YOLOv8 model as the basic architecture model of photovoltaic module defect detection, and improve it on this basis; S3.2: introduce channel attention and spatial attention mechanism CBAM in the backbone feature extraction network; S3.3: introduce RFB module in the backbone feature extraction network to improve the ability of the model to extract multi-scale features; S3.4: fuse four multi-scale feature maps output by the backbone feature extraction network, introduce ASFF adaptive spatial feature fusion method, and propose ASFF four detection heads on this basis; S3.5: optimize YOLOv8 loss function; S3.6: Obtain the center point coordinates of the defect area detection frame in the photovoltaic module image.

5. The path planning method for autonomously launching a photovoltaic module cleaning robot from a UAV according to claim 1, wherein, The step S4 comprises: S4.1: Based on the analysis of RTK error ellipse, compensate the RTK error of the inspection unmanned aerial vehicle; S4.2: Match the compensated RTK positioning data with the center point of the photovoltaic module image in the image coordinate system; S4.3: Map the center point coordinates of the defect area detection frame in the photovoltaic module image to the ENU coordinate system, and solve the longitude and latitude coordinates of the photovoltaic module with defects.

6. The path planning method for autonomously launching a photovoltaic module cleaning robot from a UAV according to claim 1, wherein, The step S5 comprises: S5.1: According to the longitude and latitude coordinates of the photovoltaic module with defects, determine the take-off point and drop point of the transport unmanned aerial vehicle; S5.2: Set the flight height and flight path of the transport unmanned aerial vehicle, and ensure that there is no obstacle in the flight path; S5.3: Generate the optimal drop path of the transport unmanned aerial vehicle; S5.4: Integrate the drop path with the flight control system to ensure that the unmanned aerial vehicle accurately flies to the specified height position above the photovoltaic module that needs to be cleaned according to the predetermined path.

7. The path planning method for autonomously launching a photovoltaic module cleaning robot from a UAV according to claim 1, wherein, The step S6 comprises: S6.1: Perform detailed simulation test of the drop path to evaluate the accuracy and feasibility of the path planning; by simulating the flight under different environmental conditions, ensure that the path planning meets the accuracy requirements in actual operation; S6.2: Perform on-site measurement to verify the actual performance of the unmanned aerial vehicle when flying according to the planned path; by comparing the actual drop position with the target position, evaluate the accuracy and stability of the path planning, and ensure that the unmanned aerial vehicle can accurately reach the specified position; S6.3: Analyze the deviation in the test results and further optimize the path planning; by adjusting the parameters of the flight control system and path planning algorithm, improve the path accuracy, ensure the stable operation of the unmanned aerial vehicle in complex environment, and ensure that it can well complete the unmanned aerial vehicle drop path planning task under different environments and photovoltaic power station layouts.

Images