Unmanned aerial vehicle nest site selection planning method, system and related device

Abstract

The invention provides an unmanned aerial vehicle nest site selection planning method and system and a related device, and relates to the technical field of unmanned aerial vehicles, and the method comprises the steps: converting the latitude and longitude of a tower and the altitude into a 3D voxel grid; extracting spatial features of a tower point group from the 3D voxel grid by using a 3D convolutional neural network; a density peak detection algorithm is applied to the spatial features to determine a clustering center, a decision value is calculated according to the local density and the minimum distance between the local density and the high-density point, and the optimal clustering point is dynamically identified based on the decision value to serve as initial site selection point distribution; and on the basis of the initial site selection point distribution, optimizing the site selection point distribution by using a reinforcement learning model. According to the technical scheme of the invention, the problems of time and labor consumption and difficulty in processing large-scale tower point group data in manual site selection in the prior art can be effectively solved; a traditional method is difficult to comprehensively consider multi-dimensional factors such as signal coverage, economic cost and geographical constraints; and a site selection strategy cannot be dynamically adjusted, and complex geographical environments and task requirements are difficult to deal with.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically to a method, system, and related apparatus for UAV nest site selection and planning. Background Technology

[0002] In current power system operation and maintenance, the inspection of overhead distribution line equipment faces challenges due to its wide distribution, complex geographical environment, and ever-changing task requirements. Traditional tower location methods often rely on manual experience when dealing with large-scale tower point data, which is time-consuming, labor-intensive, and difficult to handle complex datasets, resulting in low efficiency in the location selection process. The limitations of existing methods become even more apparent when considering multiple dimensions such as signal coverage, economic costs, and geographical constraints.

[0003] Traditional UAV nesting site selection and planning methods typically only consider the latitude and longitude of the pole / tower location, neglecting important factors such as the pole / tower's altitude, economic costs, and signal attenuation. In practical applications, such a site selection strategy may lead to low inspection efficiency or even mission failure due to poor signal quality, excessive costs, or geographical limitations. Furthermore, existing methods lack dynamic adjustment mechanisms, making it difficult to adapt quickly to complex and changing geographical environments and mission requirements, thus limiting the flexibility and efficiency of UAV inspections.

[0004] In practice, due to the unique characteristics of overhead power distribution lines, such as irregular paths, differences in distance between towers, and fluctuations in altitude, traditional drone nesting methods often struggle to accurately capture these complex spatial features, thus affecting the rationality of the site selection. Especially in scenarios involving multi-drone collaborative inspections, an effective coverage network between drone nests is required to ensure continuous communication and rapid response capabilities for drones performing inspection tasks; traditional methods often fall short of this requirement.

[0005] Currently, existing methods for selecting and planning cell sites in power distribution networks still suffer from the following problems: 1. Low efficiency. Manual site selection is time-consuming and labor-intensive, and it is difficult to handle large-scale data on tower clusters; 2. Poor accuracy. Traditional methods struggle to comprehensively consider multiple factors such as signal coverage, economic cost, and geographical constraints; 3. Poor adaptability. They cannot dynamically adjust site selection strategies, making it difficult to cope with complex geographical environments and task requirements. To address these issues, this application proposes a cell site selection method based on 3D convolutional neural networks and reinforcement learning, which can automatically and efficiently generate optimal site selection methods. This method can automatically extract the spatial features of tower clusters, adaptively determine cluster centers, and optimize the distribution of site selection points through reinforcement learning, thereby maximizing coverage, minimizing economic cost, and satisfying signal coverage and geographical constraints. Summary of the Invention

[0006] In view of the above problems, this application proposes a method, system and related device for unmanned aerial vehicle (UAV) nest location planning, which aims to solve the problems of low efficiency, insufficient accuracy and poor adaptability in the prior art.

[0007] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a method for unmanned aerial vehicle (UAV) nesting planning based on 3D convolution and reinforcement learning, including: Convert the latitude, longitude, and altitude of the tower into a 3D voxel mesh; Spatial features of tower point groups are extracted from 3D voxel meshes using 3D convolutional neural networks; In spatial features, a density peak detection algorithm is applied to determine cluster centers, and decision values ​​are calculated based on local density and minimum distance to high-density points. Based on the decision values, the optimal cluster points are dynamically identified as the initial site selection point distribution. Based on the initial site distribution, a reinforcement learning model is used to optimize the site distribution. The training process of the reinforcement learning model includes: designing a reinforcement learning environment, training the reinforcement learning model using the PPO algorithm, adding, deleting or moving site points, and a reward function based on coverage, signal attenuation, economic cost and distance constraints.

[0008] As a further improvement to this application, the conversion of the pole's latitude, longitude, and altitude into a 3D voxel mesh is as follows: during the coordinate transformation process, the PyProj library is used to convert the pole's latitude, longitude, and altitude to a unified 3D voxel mesh, and the coordinate data is normalized to the range of [0, 1] to standardize the input data.

[0009] As a further improvement to this application, the conversion of the tower's latitude, longitude, and altitude into a 3D voxel mesh includes: The spatial distribution of tower points is represented by a voxel grid. The voxel grid transforms the spatial distribution of tower points into a regular grid structure, which is used by 3D convolutional neural networks to extract spatial features and serves as the input to the 3D convolutional neural network. A voxel mesh divides three-dimensional space into regular cubic units, where each voxel is a cubic unit representing a region in three-dimensional space and containing spatial information; the size of the voxel mesh is... , where G is the resolution of the grid, normalizes the latitude, longitude and altitude coordinates of the tower points to the range of [0, 1]; and then maps the normalized tower points to the voxel grid.

[0010] As a further improvement to this application, the 3D convolutional neural network includes two convolutional layers, two pooling layers, and two fully connected layers; The first convolutional layer of the 3D convolutional neural network includes 32 convolutional kernels of size 3×3×3, and the activation function is ReLU, which is used for spatial feature extraction in the initial stage. The second convolutional layer of the 3D convolutional neural network includes 64 3×3×3 convolutional kernels, which are used to extract spatial features and enhance the recognition of the model. The pooling layer of the 3D convolutional neural network adopts the max pooling strategy, with the pooling kernel size set to 2×2×2 and the stride to 2, in order to effectively compress the feature map and improve processing efficiency. The first convolutional layer uses 32 convolutional kernels to extract features from the input data, and then the resulting feature map is fed into the first pooling layer. The first pooling layer uses max pooling to process the features, and then the processed features are fed into the second convolutional layer. The second convolutional layer uses 64 convolutional kernels to further extract features, and then feeds them into the second pooling layer. The second pooling layer processes the features in the same way as the first pooling layer, and then feeds the processed features into the fully connected layer. The first fully connected layer uses 256 neurons to process the features, and then feeds the result into the second fully connected layer. The output of the second fully connected layer is a 128-dimensional feature result.

[0011] As a further improvement to this application, the extraction of spatial features of tower point groups from a 3D voxel mesh using a 3D convolutional neural network includes: The convolutional kernel of a 3D convolutional neural network slides along the width, height, and depth dimensions to extract features in three-dimensional space; the formula is:

[0012] Where F(x,y,z) is the value of position (x,y,z) in the output feature map, V(x+i,y+j,z+l) is the value of position (x+i,y+j,z+l) in the input voxel grid, and K(i,j,l) is the value.

[0013] As a further improvement to this application, the step of applying a density peak detection algorithm to determine cluster centers in spatial features, calculating decision values ​​based on local density and minimum distance to high-density points, and dynamically identifying optimal cluster points based on these decision values ​​as the initial site selection point distribution includes: Density peak detection is used as an adaptive clustering method; the specific steps of density peak detection are as follows: 1) For each point i, calculate its local density. The formula is:

[0014] in, It is the distance between point i and point j. It is a broadband parameter that controls the smoothness of density calculation; 2) For each point i, calculate its minimum distance to points with higher density. For the point with the highest density, Defined as the maximum distance between this point and other points; 3) Calculate the decision value Points with larger decision values ​​are used as cluster centers; 4) For points that are not cluster centers, assign them to the cluster containing the nearest higher density points, and then use the best cluster point as the initial location point distribution.

[0015] As a further improvement to this application, the method of optimizing the distribution of site selection points using a reinforcement learning model includes: The framework of reinforcement learning models includes: 1) State: Represents the characteristics of the site selection points and tower point groups selected by adaptive clustering. The state space D is defined as S={L, P}, where L is the set of adaptive clustering points and P is the set of tower point groups. 2) Actions: Represent operations performed on selected points, including adding, deleting, and moving selected points; the action space is defined as follows: Where a=0 indicates adding a location point, a=1 indicates deleting a location point, and a=2 indicates moving a location point; 3) Reward: Used to evaluate the effectiveness of the action, taking into account coverage, signal attenuation, economic cost, and distance constraints. The expression for the reward function is:

[0016] in, These are weighting coefficients used to balance various constraints; Coverage(s,a) is the coverage reward; SignalLoss(s,a) is the signal attenuation loss; DistancePenalty(s,a) is the distance penalty; after obtaining the reward function, a grid search method is used to determine the weighting coefficients of each sub-loss.

[0017] As a further improvement to this application, the coverage bonus represents the number of pole points covered by the selected site, and its mathematical expression is:

[0018] in, These are the coordinates of point j on the tower. is the coordinate of the selected point j, D is the maximum nest coverage distance, and Indicator is the instruction function; Signal attenuation loss represents the impact of the altitude of the site selection point on signal quality. The elevation of site i:

[0019] The distance penalty is the sum of the minimum distance constraint between site selection points and the maximum distance constraint from a site selection point to its tower. Its purpose is to impose a penalty when the distance between site selection points is less than the minimum interval distance; and to impose a penalty when the distance from a site selection point to its corresponding tower exceeds the maximum coverage distance. The minimum distance constraint is:

[0020] in, It is the minimum distance interval. It is the distance between site selection points i and j; the maximum distance constraint from the site selection point to the tower point is:

[0021] In this context, Indicator is the instruction function, which is 1 when the condition is true and 0 otherwise, and R is the maximum coverage distance.

[0022] As a further improvement to this application, a result visualization step is also included, which includes: The matplotlib library was used to create 3D images of the distribution of tower sites and the final site selection points to visually demonstrate the site selection planning results.

[0023] Secondly, this application provides a UAV nesting planning system based on 3D convolution and reinforcement learning, comprising: A data preprocessor is used to convert the latitude, longitude, and altitude of the towers into 3D voxel meshes; The feature extractor, which integrates a 3D convolutional neural network, is used to extract spatial features of tower point groups from a 3D voxel grid. An adaptive clusterer is used to determine cluster centers by applying a density peak detection algorithm to spatial features, and to calculate decision values ​​based on local density and minimum distance to high-density points. Based on the decision values, the optimal cluster points are dynamically identified as the initial site selection point distribution. A reinforcement learning optimizer is used to optimize the distribution of selected points based on an initial distribution using a reinforcement learning model. The training process of the reinforcement learning model includes: designing a reinforcement learning environment, training the reinforcement learning model using the PPO algorithm, adding, deleting or moving selected points, and a reward function based on coverage, signal attenuation, economic cost and distance constraints.

[0024] Thirdly, this application provides a drone nest for power distribution network inspection, comprising: multiple drone nest units deployed at the optimal location determined by the drone nest site selection and planning method.

[0025] Fourthly, this application provides a drone inspection system based on 3D convolution and reinforcement learning, comprising: Drone swarms are capable of performing grid-based autonomous inspection tasks; And drone nests, used to support the take-off, landing, charging and maintenance of drone swarms, wherein the deployment of the nests is based on the optimal site determined by the drone nest site selection planning method.

[0026] Fifthly, this application provides a drone for intelligent inspection, comprising: The communication module is used to maintain a communication connection with the nest. The navigation module, based on the nest location information provided by the aforementioned UAV nest location planning method, achieves accurate flight route planning; The power management system is used for rapid charging and battery replacement in the drone's nest, ensuring continuous inspection work for the drone.

[0027] Sixthly, this application provides a computer program product for UAV nest location planning based on 3D convolution and reinforcement learning, comprising: a non-volatile computer-readable storage medium storing executable instructions configured to execute the UAV nest location planning method.

[0028] Seventhly, this application provides a UAV gridded inspection system based on 3D convolution and reinforcement learning, comprising: Multiple of the aforementioned drones possess autonomous take-off and landing, flight, and data acquisition capabilities; Multiple drone nests are deployed at the optimal locations determined by the drone nest site selection and planning method. The central control system is used to schedule the inspection operations of the drone swarm, enabling intelligent management of grid-based inspections.

[0029] Compared with the prior art, this application has the following advantages: This method converts the latitude, longitude, and altitude of power poles into a 3D voxel grid; extracts spatial features of the pole point group using a 3D convolutional neural network; determines cluster centers using density peak detection within the spatial features; and optimizes the distribution of site selection points using reinforcement learning to maximize coverage while satisfying signal attenuation, economic cost, and distance constraints. Data preprocessing provides input for feature extraction, feature extraction provides feature representations for adaptive clustering, adaptive clustering provides initial cluster centers for reinforcement learning optimization, and reinforcement learning optimization generates the final site selection points. By combining 3D convolutional neural networks and reinforcement learning, not only can spatial environmental features be accurately extracted, but the site selection strategy can also be dynamically adjusted according to specific needs, ensuring wide coverage, stable signal, low cost, and rapid response for UAV inspections. This achieves intelligent management of the UAV inspection system, enhancing its flexibility and adaptability. It can automatically generate optimal site selection points and comprehensively consider factors such as signal coverage, economic cost, and distance constraints, playing a significant role in improving the efficiency and economic benefits of power distribution network inspection and representing a major innovation in the field of power inspection. This method overcomes the limitations of traditional UAV nesting site selection and planning methods that only consider latitude and longitude and ignore altitude. By combining 3D convolutional neural networks, adaptive clustering, and reinforcement learning, it achieves intelligent nesting site selection for pole and tower clusters. This method effectively improves the rationality, intelligence, and practicality of UAV nesting site selection and planning for overhead power distribution lines, significantly reduces construction and human resource costs, improves the quality and efficiency of UAV-nested collaborative inspection, and enables more flexible UAV-nested collaborative inspection methods for power distribution networks.

[0030] Furthermore, the UAV nesting site selection and planning method involves data preprocessing, feature extraction, adaptive clustering, and reinforcement learning optimization. First, the data preprocessor converts the latitude, longitude, and altitude of the poles into a 3D voxel grid, providing a standardized three-dimensional data representation for subsequent feature extraction and clustering. Next, the feature extractor uses a 3D convolutional neural network to extract spatial features of the pole point groups from the voxel grid, forming feature representations. Then, the adaptive clusterer determines cluster centers based on a density peak detection algorithm, providing an initial distribution of site selection points for reinforcement learning optimization. Finally, the reinforcement learning optimizer designs a reinforcement learning environment, trains the model using the PPO algorithm, and optimizes the site selection points, achieving intelligent and optimized nesting site selection planning. This tightly integrated process forms an efficient and intelligent nesting site selection and planning workflow, significantly improving the efficiency and accuracy of nesting site selection and planning, reducing construction and maintenance costs, and enhancing the quality and efficiency of UAV inspections.

[0031] Furthermore, the UAV nesting site selection and planning method involves a process that includes receiving pole location data, executing a site selection planning algorithm, outputting the site selection planning results, and deploying UAV nests. First, the user inputs the latitude, longitude, and altitude data of the power distribution network poles into the system via a data interface. Next, the system executes a UAV nesting site selection and planning method based on 3D convolution and reinforcement learning, including data preprocessing, feature extraction, adaptive clustering, and reinforcement learning optimization steps. Subsequently, the system outputs the optimal nesting site selection planning results, including the latitude, longitude, altitude, and relevant features of the selected site. Finally, based on the planning results, the user deploys UAV nests in the actual environment, achieving efficient UAV grid-based inspection and multi-UAV collaborative operations. The entire process is simple and clear, significantly improving the efficiency and accuracy of nesting site selection and planning, reducing construction and maintenance costs, and enhancing the quality and effectiveness of UAV inspections.

[0032] Furthermore, the UAV nesting site selection and planning solution involves a workflow that includes receiving customer requirements, executing site selection planning algorithms, outputting planning results, and providing follow-up services. First, the solution provider receives the customer's nesting site selection and planning requirements through a service interface, including pole location data, inspection task descriptions, and budget constraints. Next, the solution provider executes a UAV nesting site selection and planning method based on 3D convolution and reinforcement learning, including data preprocessing, feature extraction, adaptive clustering, and reinforcement learning optimization. Subsequently, the solution provider outputs the optimal nesting site selection and planning results, including the latitude, longitude, altitude, and relevant characteristics of the selected site. Finally, based on the planning results, the solution provider provides customers with follow-up services such as nesting deployment, operation and maintenance training, and fault prediction, achieving full-process support from planning to implementation. This tightly integrated workflow significantly improves the efficiency and accuracy of nesting site selection and planning, reduces construction and operation and maintenance costs, and enhances the quality and effectiveness of UAV inspections.

[0033] Furthermore, the UAV grid-based inspection system, technically speaking, integrates multiple UAVs, multiple UAV nests, and a central control system to form a complete UAV grid-based inspection system. In principle, the UAV swarm can perform autonomous grid-based inspection tasks. The UAV nests support the take-off, landing, charging, and maintenance of the UAV swarm, while the central control system schedules the inspection operations of the UAV swarm, achieving intelligent management of grid-based inspections. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and quality of UAV inspections, reduce construction and maintenance costs, and enhance the intelligence level and economic benefits of power distribution network inspections. In other embodiments, the flexibility and reliability of UAV inspections can be further improved by optimizing the UAV swarm scheduling strategy (such as dynamic task allocation, adaptive route planning, etc.) or introducing more nests (such as mobile nests, backup nests, etc.) to adapt to a wider range of inspection needs and more complex working environments. The technical solution of this application involves a work process including the take-off, inspection, return and charging of the UAV, as well as the deployment, maintenance and upgrading of the UAV nest. The whole process is closely integrated to form an efficient and intelligent UAV grid-based inspection process, which can significantly improve the efficiency and quality of inspection, reduce construction and operation and maintenance costs, and improve the intelligence level and economic benefits of power distribution network inspection. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram illustrating the UAV nest location planning method provided in this application; Figure 2 A schematic diagram of the UAV nesting planning method based on 3D convolution and reinforcement learning provided for this application; Figure 3 This application provides a combined frequency regulation control system for offshore wind, thermal power, and energy storage. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] This application relates to the field of UAV nest selection technology, and in particular to a method and system for nest selection of pole point groups based on 3D convolutional neural networks (3DCNN) and reinforcement learning (RL). This method achieves efficient and intelligent nest selection by combining the feature extraction capability of 3D convolutional neural networks, the adaptive clustering capability of peak density detection, and the optimization capability of reinforcement learning.

[0038] This application will leverage the feature extraction capabilities of 3D convolutional neural networks and the optimization capabilities of reinforcement learning to achieve efficient, intelligent, and adaptive drone nesting planning for distribution network overhead line tower clusters. This will improve the efficiency and quality of drone inspections, reduce maintenance costs, and enhance the intelligence level of power system operation and maintenance. One aspect of this application provides a method for UAV nest location planning based on 3D convolution and reinforcement learning. Technically, this method converts the latitude, longitude, and altitude of poles into a 3D voxel grid through data preprocessing, providing a standardized three-dimensional data representation for subsequent feature extraction and clustering. In principle, a 3D convolutional neural network can effectively extract the spatial features of pole point groups from the voxel grid, while a density peak detection algorithm can adaptively determine cluster centers based on these features. Reinforcement learning further adjusts the distribution of location points by optimizing actions and reward mechanisms to meet multi-dimensional optimization objectives. In terms of effectiveness, the technical solution in this embodiment can significantly improve the rationality, intelligence, and practicality of nest location planning, reduce construction and human resource costs, and improve the quality and efficiency of UAV inspections. In other embodiments, by introducing more dimensional features (such as terrain and wind speed) and employing different deep learning architectures (such as ResNet and Inception), the nest location problem in specific environments can be solved, further improving the accuracy and flexibility of location planning.

[0039] Specifically, this application presents a nesting method based on 3D convolutional neural networks and reinforcement learning, which can automatically generate optimal nesting points and comprehensively consider factors such as signal coverage, economic cost, and distance constraints. The technical solution of this method includes: (1) Data preprocessing. The latitude, longitude, and altitude of the towers are converted into 3D voxel meshes; The data preprocessor uses the PyProj library to convert the latitude, longitude, and altitude of the towers to a unified 3D voxel grid during coordinate transformation, and normalizes the coordinate data to the range [0, 1]. Technically, the use of the PyProj library ensures the accuracy and consistency of coordinate transformation, while normalization makes the data suitable for the input requirements of deep learning models. In principle, the process of converting latitude, longitude, and altitude to a voxel grid is essentially the quantization and discretization of spatial data, which is beneficial for the understanding and processing of deep learning models, while normalization avoids the impact of data at different scales on model training. In terms of effectiveness, the technical solution in this embodiment can ensure the quality and format of the input data, providing a reliable foundation for subsequent feature extraction and cluster analysis. In other embodiments, different coordinate transformation libraries (such as Geopy, GDAL, etc.) and normalization methods (such as minimum-maximum scaling, Z-score normalization, etc.) can be used to adapt to the data processing needs in different environments, further improving the efficiency and flexibility of data preprocessing.

[0040] This section primarily converts the latitude, longitude, and altitude of the towers into a 3D voxel grid. In this method, the voxel grid discretizes three-dimensional space into a regular three-dimensional mesh structure to represent the spatial distribution of tower points. The voxel grid transforms the spatial distribution of tower points into a regular mesh structure, facilitating the extraction of spatial features by 3D convolutional neural networks (CNNs), and serves as the input to the 3D CNN. A voxel grid divides three-dimensional space into regular cubic units (called voxels), each voxel representing a region in three-dimensional space and containing certain spatial information (such as the density of tower points). The voxel grid can be viewed as a three-dimensional matrix with a size of... Where G is the grid resolution, which determines the size of the voxels. Higher resolution results in smaller voxels and more refined spatial information. Specifically, the latitude, longitude, and elevation coordinates of the tower points are normalized to the range [0, 1] for consistent processing. The normalization formula is as follows:

[0041]

[0042]

[0043] in , , These are the original coordinates of the tower points, and the corresponding min and max values ​​represent the maximum and minimum values ​​of each tower point in the x, y, and z directions, respectively. The normalized tower points are then mapped onto a voxel mesh.

[0044]

[0045] Where V(x,y,z) is the value of (x,y,z) in the voxel mesh, and N is the total number of towers. Indicator represents the indicator function, which is 1 when the condition is true and 0 otherwise. Floor represents the floor function.

[0046] (2) Data feature extraction. Spatial features of the tower point group are extracted using a 3D convolutional neural network; The first convolutional layer of the 3D convolutional neural network comprises 32 3×3×3 convolutional kernels, using ReLU as the activation function and a stride of 1 for feature extraction. Technically, this configuration captures preliminary spatial features from a 3D voxel grid, laying a solid foundation for subsequent deep feature extraction. In principle, the ReLU activation function introduces non-linearity, helping the network better fit complex spatial distribution patterns, while the 3×3×3 kernel size and 1 stride design extract local features while maintaining high resolution. In terms of effectiveness, the technical solution in this embodiment ensures the accuracy and efficiency of feature extraction, providing high-quality feature representations for subsequent adaptive clustering and reinforcement learning optimization. In other embodiments, the performance of feature extraction can be further improved by changing the size, number, and stride of the convolutional kernels, or by using different activation functions (such as Leaky ReLU, ELU, etc.) to adapt to pole point group data of different scales and complexities.

[0047] The second convolutional layer of the 3D convolutional neural network includes 64 3×3×3 convolutional kernels, further refining spatial features and enhancing the model's recognition and generalization capabilities. Technically, this convolutional layer can perform deeper spatial information mining based on the features extracted by the first convolutional layer, improving the level of feature abstraction. In principle, increasing the number of convolutional kernels while maintaining the same kernel size helps capture more complex feature combinations, improving the model's recognition accuracy and generalization ability, enabling the model to better understand the spatial distribution patterns of tower point groups. In terms of effectiveness, the technical solution in this embodiment significantly improves the depth and accuracy of feature extraction, providing richer feature representations for subsequent adaptive clustering and reinforcement learning optimization. In other embodiments, the number, size, and type of convolutional kernels can be adjusted, or mechanisms such as residual connections and skip connections can be introduced to address a wider range of scenarios and datasets, further enhancing the model's recognition and generalization capabilities.

[0048] In this method, 3D convolution is one of the core components, used to extract spatial features from voxel meshes. 3D convolution can capture the local and global features of tower point groups in three-dimensional space, providing effective feature representations for subsequent clustering and optimization. 3D convolution is a method for performing convolution operations on three-dimensional data (such as voxel meshes). Unlike 2D convolution, the convolution kernel of 3D convolution slides along three dimensions (width, height, and depth) to extract features in three-dimensional space. The mathematical formula for 3D convolution is as follows:

[0049] Where F(x,y,z) is the value at position (x,y,z) in the output feature map, V(x+i,y+j,z+l) is the value at position (x+i,y+j,z+l) in the input voxel grid, and K(i,j,l) is the value. The 3D convolution used in this method mainly consists of two convolutional layers, two pooling layers, and two fully connected layers. The first convolutional layer uses 32 convolutional kernels of size 3×3×3, with ReLU activation function and stride 1 to extract features from the input data. The resulting feature map is then fed into the first pooling layer. The first pooling layer uses max pooling for feature processing, with a kernel size of 2×2×2 and stride 2. The processed features are then fed into the second convolutional layer. The second convolutional layer uses 64 convolutional kernels of size 3×3×3 to further extract features, which are then fed into the second pooling layer. The second pooling layer processes the features in the same way as the first pooling layer, and the processed features are then fed into the fully connected layer. The first fully connected layer uses 256 neurons of size 8×8×8 to process the features, and the result is fed into the second fully connected layer. The second fully connected layer outputs a 128-dimensional feature result.

[0050] The pooling layer of the 3D convolutional neural network employs max pooling, with a 2×2×2 kernel size and a stride of 2. Technically, max pooling effectively reduces the size of the feature map, lowers computational complexity, and preserves key feature information. In principle, pooling eliminates redundant information and improves the abstraction level of features by aggregating features within local regions, which is beneficial for subsequent classification and regression tasks. In terms of effectiveness, the technical solution in this embodiment ensures the efficiency and quality of feature extraction, providing refined feature representations for subsequent adaptive clustering and reinforcement learning optimization. In other embodiments, other pooling strategies such as average pooling and global pooling can be used, or the kernel size and stride can be adjusted to adapt to the feature extraction needs of different scenarios, further improving the flexibility and efficiency of feature extraction.

[0051] (3) Adaptive clustering. Density peak detection is used to determine cluster centers based on spatial features; The adaptive clustering step employs a density peak detection algorithm to determine cluster centers, simultaneously calculating the local density of each point and the minimum distance to high-density points. Technically, the density peak detection algorithm automatically identifies cluster centers without requiring manual setting of the number of clusters, making it suitable for adaptive requirements in tower site selection. In principle, by calculating local density and the minimum distance to high-density points, the density peak detection algorithm can identify relatively independent regions with high density as potential cluster centers. In terms of effectiveness, the technical solution in this embodiment ensures the accuracy and efficiency of adaptive clustering, providing a reasonable initial distribution of site selection points for subsequent reinforcement learning optimization. In other embodiments, other clustering algorithms such as DBSCAN and HDBSCAN can be used to handle different types of tower point distributions, further improving the flexibility and robustness of clustering.

[0052] The goal of adaptive clustering is to automatically generate cluster centers from a group of pole points without pre-specifying the number of clusters. This algorithm uses Density Peak Clustering (DPC) as the adaptive clustering method. DPC is based on two assumptions: first, the density of cluster centers is higher than that of their surrounding points, and they are relatively far from points with even higher density; second, the density of non-cluster centers is lower, or they are relatively close to points with even higher density. The specific steps are: 1) For each point i, calculate its local density. The formula is shown below:

[0053] in, It is the distance between point i and point j. 1) It is a broadband parameter that controls the smoothness of the density calculation; 2) For each point i, calculate its minimum distance to higher density points. For the point with the highest density, Defined as the maximum distance between this point and other points; 3) Calculate the decision value. 4) For points that are not cluster centers, assign them to the cluster containing the nearest higher density point.

[0054] Technically, this state representation accurately reflects the distribution and related characteristics of current site selection points, providing comprehensive information support for the reinforcement learning model's decision-making. In principle, the state space construction is based on adaptive clustering results, including the location information of site selection points and the feature representation of tower point groups, helping the reinforcement learning model understand the current environment and task requirements. In terms of effectiveness, the technical solution in this embodiment ensures that the reinforcement learning model's decisions are based on comprehensive information of the current environment, achieving refined control over nest site selection planning. In other embodiments, more state variables (such as weather and time) can be introduced to adapt to a wider range of application scenarios and more complex problems, further enhancing the model's decision-making and generalization capabilities.

[0055] (4) Reinforcement learning optimization. Reinforcement learning methods are used to optimize the distribution of selected locations, maximizing coverage while satisfying signal attenuation, economic cost, and distance constraints. A reinforcement learning environment is designed, including a state space, action space, and reward function. The PPO algorithm is used to train the model and optimize the location distribution. Reinforcement learning optimization is used to further optimize the distribution of site selection points based on the initial solution generated by adaptive clustering, so that it meets the constraints of coverage, signal attenuation, economic cost, and distance. Reinforcement learning is a machine learning method that learns the optimal policy through interaction with the environment. In this algorithm, the framework of reinforcement learning is as follows: 1) State. The state represents the characteristics of the site selection points and the pole point group selected by adaptive clustering. The state space D is defined as S={L, P}, where L is the set of adaptive cluster points and P is the set of pole point groups; 2) Action. Action represents the operation on the site selection points, including adding a site selection point (adding a new site selection point at a random location), deleting a site selection point (randomly deleting an existing site selection point), and moving a site selection point (randomly moving an existing site selection point). The action space is defined as... Where a=0 indicates adding a location point, a=1 indicates deleting a location point, and a=2 indicates moving a location point; 3) Reward. The reward function is used to evaluate the effect of the action, taking into account coverage, signal attenuation, economic cost, and distance constraints. The expression of the reward function is:

[0056] in, These are weighting coefficients used to balance various constraints; Coverage(s,a) is the coverage reward; SignalLoss(s,a) is the signal attenuation loss; DistancePenalty(s,a) is the distance penalty. After obtaining the reward function, a grid search method is used to determine the weighting coefficients of each sub-loss, and the value range of each weighting coefficient is [0.1, 1.0].

[0057] Coverage bonus represents the number of pole points covered by the selected site. The mathematical expression is:

[0058] in, These are the coordinates of point j on the tower. is the coordinates of the selected point j, D is the maximum nest coverage distance (5km), and Indicator is the instruction function, which is 1 when the condition is true and 0 otherwise.

[0059] Signal attenuation loss represents the impact of site elevation on signal quality, meaning that the higher the elevation of the site, the worse the signal. Signal attenuation loss encourages the selection of sites at lower elevations to improve signal quality. It is the altitude of site i.

[0060]

[0061] The distance penalty is the sum of the minimum distance constraint between site selection points and the maximum distance constraint from a site selection point to its corresponding tower. This penalty is applied when the distance between site selection points is less than the minimum spacing distance, and when the distance from a site selection point to its corresponding tower exceeds the maximum coverage distance. The minimum distance constraint is:

[0062] in, It is the minimum distance interval (5km). This is the distance between site selection points i and j. The maximum distance constraint from the site selection point to the tower point is:

[0063] In this context, Indicator is the instruction function, which is 1 when the condition is true and 0 otherwise, and R is the maximum coverage distance (5km).

[0064] The specific implementation of this method is as follows: 1. Data preprocessing. Input the latitude and longitude coordinates of the tower points, use the PyProj library to convert the latitude and longitude to UTM coordinates, and normalize the coordinates and altitude to the range [0, 1]. Then, map the normalized data into a 3D voxel grid, where each voxel represents a spatial region; 2. Feature extraction. Use a 3D convolutional neural network to extract the spatial features of the tower points; 3. Adaptive clustering. Use density peak detection to determine the cluster centers based on the spatial features; 4. Reinforcement learning optimization. Design a reinforcement learning environment, including a state space, action space, and reward function, train the model using the PPO algorithm, and optimize the point selection distribution; The reward function weight coefficients of the reinforcement learning optimizer provided in this embodiment are optimized using a grid search method, with their values ​​limited to the range [0.1, 1.0]. This method balances the optimization of the reward function by considering coverage, signal attenuation, economic cost, and distance constraints. Technically, the grid search method can systematically explore the optimal combination of weight coefficients, ensuring the fairness and effectiveness of the reward function. In principle, by limiting the range of weight coefficient values, any single factor can be prevented from excessively influencing the optimization result, ensuring the balance of multi-objective optimization. In terms of effect, the technical solution in this embodiment can ensure that the reinforcement learning model fully considers all constraints during training, achieving comprehensive optimization of nest location. In other embodiments, more advanced optimization strategies such as Bayesian optimization and genetic algorithms can be used to further improve the optimization effect of the reward function parameters and solve more complex problems.

[0065] Technically, the PPO algorithm effectively handles continuous action spaces and complex reward functions, improving the stability and efficiency of model training. In principle, the PPO algorithm improves policy gradient estimation by truncating importance sampling and using a dominance function, avoiding high variance during training and enhancing model training stability. In terms of effectiveness, the technical solution in this embodiment ensures rapid convergence of the reinforcement learning model during training, obtaining a high-quality addressing policy. In other embodiments, other advanced reinforcement learning algorithms such as DDPG and TD3 can be employed to adapt to a wider range of application scenarios and more complex problems, further improving the model's training performance and generalization ability.

[0066] The action space of a reinforcement learning optimizer includes, but is not limited to, optimization methods for adding new site selection points, deleting existing site selection points, and moving site selection point locations. Technically, this flexible action space design can cover all basic operations in nest location planning, improving the diversity and adaptability of optimization strategies. In principle, by allowing the addition, deletion, and movement of site selection points, the reinforcement learning model can explore a broader state space and find better location strategies. In terms of effectiveness, the technical solution in this embodiment ensures the flexibility and intelligence of nest location planning, enabling effective responses to complex geographical environments and task requirements. In other embodiments, more types of actions (such as adjusting nest capacity and type) can be introduced to address more complex problems and needs, further enhancing the richness and flexibility of optimization strategies.

[0067] 5. Visualize the results. Use matplotlib to plot a 3D graph of the tower distribution and site selection.

[0068] The results visualization step uses the matplotlib library to draw 3D images of the distribution of the tower clusters and the final site selection points. Technically, matplotlib provides powerful plotting capabilities, clearly and intuitively displaying the 3D positional relationships between the tower clusters and the site selection points. In principle, by converting the 3D coordinates of the tower clusters and site selection points into visualized images, the results visualization step helps users intuitively understand the site selection planning results and assess their rationality. In terms of effectiveness, the technical solution in this embodiment ensures the intuitive display of the site selection planning results, facilitating user understanding and evaluation, and improving the transparency and credibility of decision-making. In other embodiments, other visualization tools (such as Mayavi, Plotly, etc.) or interactive visualization functions can be used to adapt to different user preferences and needs, further enhancing the flexibility of results display and user experience.

[0069] One aspect of this application provides a UAV nest location planning system based on 3D convolution and reinforcement learning. Technically, this system integrates multiple modules, including data preprocessing, feature extraction, adaptive clustering, and reinforcement learning optimization, forming a complete nest location planning process. In principle, the modules cooperate to complete the nest location planning task. The data preprocessor converts the original pole latitude, longitude, and altitude data into a 3D voxel grid. The feature extractor uses a 3D convolutional neural network to extract spatial features from the voxel grid. The adaptive clusterer determines cluster centers based on a density peak detection algorithm, and the reinforcement learning optimizer uses the PPO algorithm to optimize the distribution of location points, achieving intelligent and optimized nest location planning. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and accuracy of nest location planning, reduce construction and maintenance costs, and improve the quality and efficiency of UAV inspections. In other embodiments, more modules (such as path planning and task allocation) can be introduced, or the algorithms and parameters of each module can be optimized to adapt to a wider range of application scenarios and more complex needs, further improving the system's functionality and performance.

[0070] One aspect of this application provides a drone nest for power distribution network inspection, deployed at the optimal location determined by the aforementioned drone nest location planning method. Technically, the deployment of the drone nest is based on the output of the location planning method, ensuring the scientific and rational nature of the nest location. In principle, the nest location planning method is based on feature extraction from a 3D convolutional neural network and cluster center determination through adaptive clustering steps, further optimized through reinforcement learning steps, ensuring that the nest location meets optimization objectives in terms of coverage, signal attenuation, economic cost, and distance constraints. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and quality of drone inspection, reduce construction and maintenance costs, and enhance the intelligence level of power distribution network inspection. In other embodiments, the hardware configuration of the nest (such as charging facilities, communication equipment, etc.) can be optimized, or more nests (such as mobile nests, backup nests, etc.) can be introduced to adapt to a wider range of inspection needs and more complex working environments, further improving the flexibility and reliability of drone inspection.

[0071] One aspect of this application provides a UAV inspection system based on 3D convolution and reinforcement learning. Technically, this system integrates a UAV swarm, a UAV nest, and a central control system, forming a complete UAV inspection system. In principle, the UAV swarm can perform grid-based autonomous inspection tasks, the UAV nest supports the take-off, landing, charging, and maintenance of the UAV swarm, and the central control system schedules the inspection operations of the UAV swarm, achieving intelligent management of grid-based inspections. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and quality of UAV inspections, reduce construction and maintenance costs, and enhance the intelligence level and economic benefits of power distribution network inspections. In other embodiments, by introducing more UAV types (such as fixed-wing UAVs, unmanned helicopters, etc.) or optimizing the scheduling algorithm of the central control system (such as genetic algorithms, ant colony algorithms, etc.), it is possible to adapt to a wider range of inspection needs and more complex working environments, further improving the flexibility and reliability of UAV inspections.

[0072] One aspect of this application provides a drone for intelligent inspection, including a communication module, a navigation module, and a power management system. Technically, the communication module ensures stable communication between the drone and its nest, the navigation module achieves precise flight route planning based on the nest's location information, and the power management system guarantees continuous inspection work. In principle, the communication module employs advanced wireless communication technology to overcome signal attenuation and interference, ensuring the stability and security of data transmission; the navigation module utilizes multiple positioning technologies such as GPS and inertial navigation, combined with nest location information, to achieve precise flight route planning; and the power management system ensures the drone's continuous operation through fast charging and battery replacement technology. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and quality of drone inspections, reduce construction and maintenance costs, and enhance the intelligence level and economic benefits of power distribution network inspections. In other embodiments, the communication protocol of the communication module can be optimized (e.g., LoRa, NB-IoT), or more sensors can be introduced (e.g., infrared cameras, gas detectors), to adapt to a wider range of inspection needs and more complex working environments, further improving the flexibility and reliability of drone inspections.

[0073] One aspect of this application provides a computer program product for UAV nest location planning based on 3D convolution and reinforcement learning. Technically, this product automates the execution of a UAV nest location planning method based on 3D convolution and reinforcement learning through a non-volatile computer-readable storage medium storing executable instructions. In principle, the executable instructions are configured to perform steps such as data preprocessing, feature extraction, adaptive clustering, and reinforcement learning optimization, forming a complete nest location planning process. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and accuracy of nest location planning, reduce construction and maintenance costs, and improve the quality and efficiency of UAV inspections. In other embodiments, the software architecture can be optimized (e.g., microservices, containerization), or more algorithms and models can be introduced (e.g., deep forests, graph neural networks) to adapt to a wider range of application scenarios and more complex needs, further enhancing the product's functionality and performance.

[0074] One aspect of this application provides a UAV nesting site selection and planning service based on 3D convolution and reinforcement learning. Technically, this service receives power grid tower location data, executes a UAV nesting site selection and planning method based on 3D convolution and reinforcement learning, and outputs the optimal nesting site selection and planning result, providing customers with professional site selection and planning services. In principle, the UAV nesting site selection and planning method based on 3D convolution and reinforcement learning can comprehensively consider multiple dimensions such as signal coverage, economic cost, and geographical constraints to achieve intelligent and optimized nesting site selection. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and accuracy of nesting site selection and planning, reduce construction and operation and maintenance costs, and improve the quality and efficiency of UAV inspections. In other embodiments, more service content (such as on-site surveys, operation and maintenance training, etc.) or optimized service processes (such as online booking, remote collaboration, etc.) can be introduced to adapt to the needs and preferences of different customers, further improving service flexibility and customer satisfaction.

[0075] One aspect of this application provides a UAV nesting site planning solution based on 3D convolution and reinforcement learning. Technically, this solution forms a complete solution system by providing a UAV nesting site planning method, system, computer program products and services based on 3D convolution and reinforcement learning. In principle, the UAV nesting site planning method based on 3D convolution and reinforcement learning can comprehensively consider multiple dimensions such as signal coverage, economic cost, and geographical constraints to achieve intelligent and optimized nesting site selection. The system, computer program products and services provide comprehensive support to customers from hardware, software, and service levels, respectively. In terms of effectiveness, the technical solution in this embodiment can significantly improve the efficiency and accuracy of nesting site planning, reduce construction and maintenance costs, and improve the quality and efficiency of UAV inspections. In other embodiments, more technical solutions (such as edge computing, cloud computing, etc.) can be introduced, or the deployment and maintenance strategies of the solution can be optimized (such as remote monitoring, fault prediction, etc.) to adapt to a wider range of application scenarios and more complex needs, further enhancing the functionality and performance of the solution.

[0076] This embodiment provides a method for UAV nest location planning based on 3D convolutional neural networks and reinforcement learning, aiming to solve the problems of low efficiency, poor accuracy, and poor adaptability in UAV nest location planning during power distribution network inspection. The method includes the following steps: First, the latitude, longitude, and altitude of the power distribution network towers are converted into a 3D voxel grid; second, the spatial features of the tower point groups are extracted from the 3D voxel grid using a 3D convolutional neural network; third, a density peak detection algorithm is used to determine cluster centers in the spatial features, and the local density of each point and the minimum distance to high-density points are calculated; finally, reinforcement learning is used to optimize the distribution of location points to maximize coverage, signal quality, and economic benefits, while satisfying specific distance constraints. This method, through the application of deep learning technology and machine learning optimization algorithms, achieves automation, intelligence, and optimization in location planning, effectively improving inspection efficiency and economy.

[0077] Regarding the implementation details of the first convolutional layer: To extract spatial features more effectively, this embodiment employs a first convolutional layer of a 3D convolutional neural network (CNN). This layer contains 32 convolutional kernels of size 3×3×3, with ReLU activation function and a stride of 1. This setup helps capture basic spatial texture and structural information from the 3D voxel grid in the initial stage, providing a solid foundation for subsequent feature analysis and clustering.

[0078] Regarding the implementation details of the second convolutional layer: In this embodiment, the second convolutional layer is configured with 64 convolutional kernels of size 3×3×3. This step aims to deepen the understanding of features in the 3D voxel mesh, further refine and enhance the features, thereby improving the model's ability to identify and generalize the distribution details of tower point groups.

[0079] Application of the PyProj library in coordinate transformation: This example details how to use the PyProj library to convert the latitude and longitude coordinates of a tower to UTM coordinates and normalize its altitude to the range [0, 1]. This process simplifies data processing, enabling 3D convolutional neural networks to analyze and process standardized coordinate data more directly and efficiently.

[0080] Grid search method optimizes the weight coefficients of the reward function: In order to find the best balance point, this embodiment optimizes the weight coefficients in the reward function of the reinforcement learning optimizer by using the grid search method to ensure that the optimal value is selected in the range of [0.1, 1.0], thereby achieving the best balance between coverage, signal attenuation, economic cost and distance constraints, making the site selection decision more reasonable and practical.

[0081] Model Training Using the PPO Algorithm: This example focuses on the process of training a reinforcement learning model using the PPO (Proximal Policy Optimization) algorithm. The PPO algorithm was chosen for its training stability and fast convergence rate, enabling it to find the optimal strategy in complex nesting environments and ensuring the high quality and effectiveness of the final nesting point distribution.

[0082] Density Peaks Detection Algorithm Applied to Adaptive Clustering: In this embodiment, the implementation details of the Density Peaks Clustering (DPC) algorithm are studied in depth. By accurately calculating the local density of each tower point and the minimum distance to high-density points, the optimal cluster center point can be dynamically identified, thereby optimizing the initial layout of the tower nesting site selection.

[0083] Pooling layer settings for 3D convolutional neural networks: For the pooling layers in 3D convolutional neural networks, this embodiment adopts the max pooling strategy with a pooling kernel size of 2×2×2 and a stride of 2, in order to effectively compress feature maps, reduce computational load, and improve processing efficiency. This is crucial for subsequent clustering analysis and reinforcement learning optimization.

[0084] Flexible adjustment of the action space: This embodiment focuses on the action space design of the reinforcement learning optimizer. By flexibly selecting actions such as adding new location points, deleting existing location points, or moving the location points, it ensures that the nest location strategy can be adjusted in real time according to actual needs and environmental changes, thereby improving the adaptability and intelligence of the overall location scheme.

[0085] State Space Construction and Updates: To enable the model to make more accurate decisions, the state space constructed in this embodiment not only includes the location information of the current site selection point but also integrates the feature representation extracted by the 3D convolutional neural network. As the reinforcement learning process iterates, the state space dynamically adjusts based on action updates and environmental feedback; this is key to the model's ability to continuously optimize the distribution of site selection points.

[0086] Visualizing the Results: This example demonstrates how to use the matplotlib library to plot the final site selection planning results. Through the intuitive display of 3D images, users can clearly see the distribution of the tower clusters and the locations of the selected sites, verifying the rationality and effectiveness of the site selection planning. It also provides a visual basis for future optimization and adjustments.

[0087] Optimization of Cell Tower Distribution under Minimum and Maximum Coverage Distance Constraints: In this embodiment, the focus is on the distance constraints in the cell tower location process. Specifically, the minimum distance interval is set to 5 kilometers. This constraint ensures that the spatial distribution of cell towers is not too dense, avoiding resource waste and signal interference. The maximum coverage distance is also set to 5 kilometers, which ensures that each cell tower can effectively cover the nearby tower cluster, guaranteeing inspection efficiency and signal quality. By strictly adhering to these two distance constraints in the action selection and reward function design during the reinforcement learning optimization process, this embodiment achieves a cell tower distribution scheme that is both widely covered and economically reasonable.

[0088] Through the detailed description of the above embodiments, this application proposes a UAV nesting site selection and planning method based on 3D convolutional neural networks and reinforcement learning. This method achieves automation, intelligence, and optimization of nesting site selection through intelligent data processing, feature extraction, adaptive clustering, and reinforcement learning optimization. It not only solves the problems of low efficiency, poor accuracy, and poor adaptability of traditional site selection methods, but also demonstrates in practical applications that this method can significantly improve the quality and economic benefits of UAV inspections, providing a novel and efficient solution for power distribution network inspection. Through the careful design and ingenious combination of various technical features, this embodiment successfully solves the challenge of nesting site selection in complex geographical environments, promoting technological progress in the field of power inspection.

[0089] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] This application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, readable storage media, optical storage, etc.) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0094] Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort should fall within the scope of protection of this application.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A method for unmanned aerial vehicle (UAV) nesting planning based on 3D convolution and reinforcement learning, characterized in that, include: Convert the latitude, longitude, and altitude of the tower into a 3D voxel mesh; Spatial features of tower point groups are extracted from 3D voxel meshes using 3D convolutional neural networks; In spatial features, a density peak detection algorithm is applied to determine cluster centers, and decision values ​​are calculated based on local density and minimum distance to high-density points. Based on the decision values, the optimal cluster points are dynamically identified as the initial site selection point distribution. Based on the initial site distribution, a reinforcement learning model is used to optimize the site distribution. The training process of the reinforcement learning model includes: designing a reinforcement learning environment, training the reinforcement learning model using the PPO algorithm, adding, deleting or moving site points, and a reward function based on coverage, signal attenuation, economic cost and distance constraints.

2. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The process of converting the latitude, longitude, and altitude of the tower into a 3D voxel mesh involves using the PyProj library during the coordinate transformation process to convert the latitude, longitude, and altitude of the tower into a unified 3D voxel mesh, and normalizing the coordinate data to the range of [0, 1] to standardize the input data.

3. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The process of converting the latitude, longitude, and altitude of the tower into a 3D voxel mesh includes: The spatial distribution of tower points is represented by a voxel grid. The voxel grid transforms the spatial distribution of tower points into a regular grid structure, which is used by 3D convolutional neural networks to extract spatial features and serves as the input to the 3D convolutional neural network. A voxel mesh divides three-dimensional space into regular cubic units, where each voxel is a cubic unit representing a region in three-dimensional space and containing spatial information; the size of the voxel mesh is... , where G is the resolution of the grid, normalizes the latitude, longitude and altitude coordinates of the tower points to the range of [0, 1]; and then maps the normalized tower points to the voxel grid.

4. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The 3D convolutional neural network includes two convolutional layers, two pooling layers, and two fully connected layers. The first convolutional layer of the 3D convolutional neural network includes 32 convolutional kernels of size 3×3×3, and the activation function is ReLU, which is used for spatial feature extraction in the initial stage. The second convolutional layer of the 3D convolutional neural network includes 64 3×3×3 convolutional kernels, which are used to extract spatial features and enhance the recognition of the model. The pooling layer of the 3D convolutional neural network adopts the max pooling strategy, with the pooling kernel size set to 2×2×2 and the stride to 2, in order to effectively compress the feature map and improve processing efficiency. The first convolutional layer uses 32 convolutional kernels to extract features from the input data, and then the resulting feature map is fed into the first pooling layer. The first pooling layer uses max pooling to process the features, and then the processed features are fed into the second convolutional layer. The second convolutional layer uses 64 convolutional kernels to further extract features, and then feeds them into the second pooling layer. The second pooling layer processes the features in the same way as the first pooling layer, and then feeds the processed features into the fully connected layer. The first fully connected layer uses 256 neurons to process the features, and then feeds the result into the second fully connected layer. The output of the second fully connected layer is a 128-dimensional feature result.

5. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The method of extracting spatial features of tower point groups from a 3D voxel grid using a 3D convolutional neural network includes: The convolutional kernel of a 3D convolutional neural network slides along the width, height, and depth dimensions to extract features in three-dimensional space; the formula is: Where F(x,y,z) is the value of position (x,y,z) in the output feature map, V(x+i,y+j,z+l) is the value of position (x+i,y+j,z+l) in the input voxel grid, and K(i,j,l) is the value.

6. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The process of applying a density peak detection algorithm to determine cluster centers in spatial features, calculating decision values ​​based on local density and minimum distance to high-density points, and dynamically identifying optimal cluster points as initial site selection points based on these decision values ​​includes: Density peak detection is used as an adaptive clustering method; the specific steps of density peak detection are as follows: 1) For each point i, calculate its local density. The formula is: in, It is the distance between point i and point j. It is a broadband parameter that controls the smoothness of density calculation; 2) For each point i, calculate its minimum distance to points with higher density. For the point with the highest density, Defined as the maximum distance between this point and other points; 3) Calculate the decision value Points with larger decision values ​​are used as cluster centers; 4) For points that are not cluster centers, assign them to the cluster containing the nearest higher density points, and then use the best cluster point as the initial location point distribution.

7. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 1, characterized in that, The framework of the reinforcement learning model includes: 1) State: Represents the characteristics of the site selection points and tower point groups selected by adaptive clustering. The state space D is defined as S={L, P}, where L is the set of adaptive clustering points and P is the set of tower point groups. 2) Actions: Represent operations performed on selected points, including adding, deleting, and moving selected points; the action space is defined as follows: Where a=0 indicates adding a location point, a=1 indicates deleting a location point, and a=2 indicates moving a location point; 3) Reward: Used to evaluate the effectiveness of the action, taking into account coverage, signal attenuation, economic cost, and distance constraints. The expression for the reward function is: in, These are weighting coefficients used to balance various constraints; Coverage(s,a) is the coverage reward; SignalLoss(s,a) is the signal attenuation loss; DistancePenalty(s,a) is the distance penalty; after obtaining the reward function, a grid search method is used to determine the weighting coefficients of each sub-loss.

8. The method for unmanned aerial vehicle (UAV) nest location planning according to claim 7, characterized in that, The coverage bonus represents the number of pole points covered by the selected site, and its mathematical expression is: in, These are the coordinates of point j on the tower. is the coordinate of the selected point j, D is the maximum nest coverage distance, and Indicator is the instruction function; Signal attenuation loss represents the impact of the altitude of the site selection point on signal quality. The elevation of site i: The distance penalty is the sum of the minimum distance constraint between site selection points and the maximum distance constraint from a site selection point to its tower. Its purpose is to impose a penalty when the distance between site selection points is less than the minimum interval distance; and to impose a penalty when the distance from a site selection point to its corresponding tower exceeds the maximum coverage distance. The minimum distance constraint is: in, It is the minimum distance interval. It is the distance between site selection points i and j; the maximum distance constraint from the site selection point to the tower point is: In this context, Indicator is the instruction function, which is 1 when the condition is true and 0 otherwise, and R is the maximum coverage distance.

9. The method for planning the location of unmanned aerial vehicle (UAV) nests according to any one of claims 1 to 8, characterized in that, It also includes a results visualization step, which includes: The matplotlib library was used to create 3D images of the distribution of tower sites and the final site selection points to visually demonstrate the site selection planning results.

10. A UAV nesting and planning system based on 3D convolution and reinforcement learning, characterized in that, include: A data preprocessor is used to convert the latitude, longitude, and altitude of the towers into 3D voxel meshes; The feature extractor, which integrates a 3D convolutional neural network, is used to extract spatial features of tower point groups from a 3D voxel grid. An adaptive clusterer is used to determine cluster centers by applying a density peak detection algorithm to spatial features, and to calculate decision values ​​based on local density and minimum distance to high-density points. Based on the decision values, the optimal cluster points are dynamically identified as the initial site selection point distribution. A reinforcement learning optimizer is used to optimize the distribution of selected points based on an initial distribution using a reinforcement learning model. The training process of the reinforcement learning model includes: designing a reinforcement learning environment, training the reinforcement learning model using the PPO algorithm, adding, deleting or moving selected points, and a reward function based on coverage, signal attenuation, economic cost and distance constraints.

11. A drone nest for power distribution network inspection, characterized in that, include: Multiple unmanned aerial vehicle (UAV) nesting units are deployed at the optimal location determined by the UAV nesting site selection and planning method according to any one of claims 1 to 9.

12. A drone inspection system based on 3D convolution and reinforcement learning, characterized in that, include: Drone swarms are capable of performing grid-based autonomous inspection tasks; And drone nests, used to support the take-off, landing, charging and maintenance of drone swarms, wherein the deployment of the nests is based on the optimal site determined by the drone nest site selection planning method according to any one of claims 1 to 9.

13. A drone for intelligent inspection, characterized in that, include: The communication module is used to maintain a communication connection with the nest. The navigation module, based on the nest location information provided by the UAV nest location planning method according to any one of claims 1 to 9, achieves accurate flight route planning; The power management system is used for rapid charging and battery replacement in the drone's nest, ensuring continuous inspection work for the drone.

14. A computer program product for UAV nesting and planning based on 3D convolution and reinforcement learning, characterized in that, include: A non-volatile computer-readable storage medium storing executable instructions configured to perform the unmanned aerial vehicle (UAV) nesting planning method as described in any one of claims 1 to 9.

15. A UAV gridded inspection system based on 3D convolution and reinforcement learning, characterized in that, include: Multiple of the aforementioned drones possess autonomous take-off and landing, flight, and data acquisition capabilities; Multiple drone nests are deployed at the optimal locations determined by the drone nest site selection and planning method according to any one of claims 1 to 9; The central control system is used to schedule the inspection operations of the drone swarm, enabling intelligent management of grid-based inspections.

Images