A method and system for locating a drone nest

Abstract

The application provides a kind of unmanned plane nest site selection planning method and system, it is related to unmanned plane nest site selection technical field, the method includes: obtaining the UTM coordinates of all towers in tower group;Through feature extraction module, according to UTM coordinates, dimensionality processing is carried out, and the high-dimensional coordinate data of tower group is obtained;Feature extraction is carried out on high-dimensional coordinate data, and the coordinate feature of tower group is obtained;Through clustering module, the probability distribution of tower in each cluster is determined by mapping coordinate feature;According to the probability distribution of tower in each cluster, the cluster number of cluster is adjusted, the cluster center corresponding to each tower in tower group is determined, and the unmanned plane nest planning scheme is obtained.The application improves the limitation of unmanned plane site selection in complex environment through feature extraction module and clustering module, so that it can better adapt to environmental changes, and the layout of unmanned plane nest is optimized by screening clustering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle nest site selection, in particular to an unmanned aerial vehicle nest site selection planning method and system. BACKGROUND

[0002] Unmanned aerial vehicles are increasingly widely used in the fields of power inspection, logistics distribution and environmental monitoring. The rationality of the site selection of unmanned aerial vehicle nests, which are the take-off and landing and storage places of unmanned aerial vehicles, directly affects the efficiency and cost of unmanned aerial vehicle operations. Generally, the mainstream method of current site selection planning is to use clustering algorithms combined with bionic algorithms or artificial experience to select sites.

[0003] In related technologies, due to the particularity of overhead lines in distribution networks, clustering algorithms cannot specify the number of clusters. In the case of complex distribution network tower distribution and other complex environments, it is difficult to use clustering algorithms to cluster all towers, so it cannot effectively handle site selection problems in complex environments, thereby limiting the applicability and flexibility of the site selection method. SUMMARY

[0004] The problem solved by the present application is how to improve the limitations of unmanned aerial vehicle site selection in complex environments.

[0005] To solve the above problems, the present application provides an unmanned aerial vehicle nest site selection planning method and system.

[0006] In a first aspect, the present application provides an unmanned aerial vehicle nest site selection planning method, comprising:

[0007] Obtaining the UTM coordinates of all towers in a tower group;

[0008] Through a feature extraction module, performing dimensionality reduction processing according to the UTM coordinates to obtain high-dimensional coordinate data of the tower group;

[0009] Performing feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group;

[0010] Through a clustering module, mapping the coordinate features to determine the probability distribution of the towers in each cluster;

[0011] According to the probability distribution of the towers in each cluster, adjusting the number of clusters of the clusters to determine the cluster center corresponding to each tower in the tower group, and obtaining an unmanned aerial vehicle nest planning scheme.

[0012] Optionally, the obtaining of the UTM coordinates of all towers in the tower group comprises:

[0013] Obtaining the latitude and longitude coordinates of all towers in the tower group;

[0014] Projecting processing is performed according to the latitude and longitude coordinates, and the UTM coordinates of the tower are determined.

[0015] Optionally, the feature extraction module includes a data embedding layer; the high-dimensional coordinate data of the tower group is obtained by performing dimension increasing processing on the UTM coordinates by the feature extraction module, including:

[0016] The point cloud coordinate matrix of the tower group is obtained according to the UTM coordinates of all towers in the tower group;

[0017] The point cloud coordinate matrix is input into the data embedding layer;

[0018] The initial dimension of the point cloud coordinate matrix is mapped to a preset high-order dimension by the data embedding layer;

[0019] The point cloud coordinate matrix of the preset high-order dimension is taken as the high-dimensional coordinate data.

[0020] Optionally, the feature extraction module includes a sparse attention layer and an intermediate layer, and the coordinate feature of the tower group is obtained by performing feature extraction on the high-dimensional coordinate data, including:

[0021] The high-dimensional coordinate data is input into the sparse attention layer, parallel calculation is performed by sliding window attention and global attention, local attention features and global attention features of the high-dimensional coordinate data are obtained, the local attention features and the global attention features are normalized to obtain a fusion feature representation;

[0022] The fusion feature representation is input into the intermediate layer, feature enhancement is performed on the fusion feature representation by a multi-head attention mechanism to obtain an enhanced feature representation, and residual linking and layer normalization are performed on the enhanced feature representation to obtain the coordinate feature.

[0023] Optionally, the high-dimensional coordinate data is input into the sparse attention layer, parallel calculation is performed by sliding window attention and global attention, local attention features and global attention features of the high-dimensional coordinate data are obtained, including:

[0024] A window with a preset size is slid on the high-dimensional coordinate data to obtain an attention score between each data point in the window and other data points in the window;

[0025] Dot product is performed according to the attention score to obtain an attention feature of each data point in the high-dimensional coordinate data;

[0026] According to the attention features of all the data points in the high-dimensional coordinate data, the local attention features of the high-dimensional coordinate data are obtained.

[0027] Optionally, the inputting the high-dimensional coordinate data into the sparse attention layer is parallel computing by sliding window attention and global attention, and the local attention features and the global attention features of the high-dimensional coordinate data are obtained, and the method further comprises:

[0028] by adding a preset number of global attention points in the sequence of the high-dimensional coordinate data;

[0029] obtaining the attention score between each data point in the high-dimensional coordinate data and the global attention point;

[0030] According to the dot product of the attention score, the global attention features of the high-dimensional coordinate data are obtained.

[0031] Optionally, the intermediate layer comprises an encoder layer, and the encoder layer comprises a multi-head attention network comprising a preset number of attention heads; the inputting the fusion feature representation into the intermediate layer comprises performing feature enhancement on the fusion feature representation by a multi-head attention mechanism to obtain an enhanced feature representation, and the method further comprises:

[0032] inputting the fusion feature representation into the multi-head attention network, performing linear transformation on the fusion feature representation by the attention heads to obtain an attention matrix;

[0033] obtaining an attention score of each attention head according to the attention matrix;

[0034] obtaining a weighted feature representation of each attention head according to the attention score and the attention matrix;

[0035] splicing the weighted feature representations of all the attention heads to obtain the enhanced feature representation.

[0036] Optionally, the clustering module comprises a fully connected layer and a Softmax layer; the mapping the coordinate features by the clustering module to determine the probability distribution of the tower in each cluster comprises:

[0037] mapping the feature dimension of the coordinate features according to the maximum cluster number of the clusters by the fully connected layer to obtain the cluster weight of the tower;

[0038] normalizing the cluster weight of each tower by the Softmax layer to obtain the probability distribution of the tower in each cluster.

[0039] Optionally, the adjusting the cluster number of the cluster according to the probability distribution of the tower in each cluster, determining the cluster center corresponding to each tower in the tower group, and obtaining the UAV nest address comprise:

[0040] According to the probability distribution of the tower in each cluster, the cluster is screened, the cluster number of the cluster is adjusted, and the corresponding tower is allocated to the screened cluster;

[0041] According to the probability distribution corresponding to the tower in each cluster, the cluster center of the cluster is determined by weighted average;

[0042] The cluster center and the cluster allocation result of each tower are taken as the UAV nest planning scheme.

[0043] In a second aspect, a UAV nest site planning system is provided, comprising:

[0044] A coordinate acquisition unit is configured to acquire the UTM coordinates of all towers in a tower group;

[0045] A feature extraction unit is configured to perform dimensionality increasing processing on the UTM coordinates by a feature extraction module to obtain high-dimensional coordinate data of the tower group, and perform feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group;

[0046] A clustering unit is configured to map the coordinate features by a clustering module to determine the probability distribution of the tower in each cluster, adjust the cluster number of the cluster according to the probability distribution of the tower in each cluster, determine the cluster center corresponding to each tower in the tower group, and obtain a UAV nest planning scheme.

[0047] The UAV nesting site selection and planning method and system of the present invention first obtains the UTM coordinates of the towers to ensure the uniformity and accuracy of the coordinate system. Then, a feature extraction module is used to perform dimensionality upscaling on the UTM coordinates to obtain high-dimensional coordinate data, thereby capturing the spatial distribution characteristics of the tower group more comprehensively. Then, by further extracting features from the high-dimensional coordinate data, the coordinate features of the tower group are obtained, providing rich information for cluster analysis. By mapping coordinate features through a clustering module, not only is the probability distribution of towers in each cluster determined, but the number of clusters can also be dynamically adjusted based on these distributions, ensuring that the number of clusters can be updated in real time. This allows for the adaptive determination of the cluster center corresponding to each tower, making the location of UAV nests more scientific and reasonable. This not only reduces the operating costs and energy consumption of UAV inspections but also improves the coverage and efficiency of inspections. Especially in complex environments, such as areas with uneven tower distribution and varied terrain, this invention can utilize tower clustering to adapt the cluster selection to environmental changes, thereby optimizing the layout of UAV nests, improving the limitations of UAV location selection in complex environments, realizing intelligent location selection for power distribution network tower groups, and significantly improving the efficiency and accuracy of UAV nest location selection in complex environments. Attached Figure Description

[0048] Figure 1 This is a flowchart of a drone nest location planning method in one embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of a model structure for unmanned aerial vehicle (UAV) nest location planning in another embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the feature extraction module in another embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the clustering module in another embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the structure of the UAV nesting site planning system in another embodiment of the present invention. Detailed Implementation

[0053] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0054] It should be understood that each of the steps recited in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present application is not limited in this respect.

[0055] The term "comprises" and variations thereof herein, do not necessarily mean "consist of". The term "based on" means "based, at least in part, on". The term "one embodiment" means "at least one embodiment". The term "another embodiment" means "at least one additional embodiment". The term "some embodiments" means "at least some embodiments". The term "optional" means "optional in at least some embodiments". Related definitions are given throughout the detailed description. It should be noted that the concepts mentioned in the present application are merely illustrative and not restrictive, and those skilled in the art should understand that "one" "multiple" modification is illustrative and not restrictive, and unless otherwise explicitly stated in the context, it should be understood as "one or more".

[0056] It should be noted that the "one" "multiple" modification mentioned in the present application is illustrative and not restrictive, and those skilled in the art should understand that unless otherwise explicitly stated in the context, it should be understood as "one or more".

[0057] The names of the messages or information exchanged between the devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0058] To solve the problems of the related art, the present embodiment provides a method and system for planning the location of a UAV nest.

[0059] In combination Figure 1 As shown in the drawings, the method for planning the location of a UAV nest provided by the embodiments of the present application comprises:

[0060] UTM coordinates of all towers in the tower group are obtained.

[0061] Specifically, by obtaining the UTM coordinates (Universal Transverse Mercator Grid System) of all towers in the tower group, the geographic coordinates of the towers can be converted into plane rectangular coordinates. This coordinate grid system and the projection it is based on have been widely used in topographic maps, as a reference grid for satellite imagery and natural resource databases, and other applications requiring precise positioning, thus facilitating subsequent calculation and analysis. The UTM coordinate system can provide more accurate plane position information, which is crucial for the accurate location of the UAV nest.

[0062] The feature extraction module is used to perform dimensionality increasing processing on the UTM coordinates, so as to convert the original two-dimensional coordinate data into data in a high-dimensional feature space, and map the original coordinates to a higher-dimensional space. In this high-dimensional space, the tower point group that can be difficult to distinguish in a two-dimensional plane can be more clearly distinguished, and the high-dimensional space provides more degrees of freedom to express the subtle differences and complex relationships between data points in the high-dimensional coordinate data.

[0063] Specifically, the feature extraction module is used to perform dimensionality increasing processing on the UTM coordinates, so as to convert the original two-dimensional coordinate data into data in a high-dimensional feature space, and map the original coordinates to a higher-dimensional space. In this high-dimensional space, the tower point group that can be difficult to distinguish in a two-dimensional plane can be more clearly distinguished, and the high-dimensional space provides more degrees of freedom to express the subtle differences and complex relationships between data points in the high-dimensional coordinate data.

[0064] The feature extraction module is used to perform dimensionality increasing processing on the UTM coordinates, so as to convert the original two-dimensional coordinate data into data in a high-dimensional feature space, and map the original coordinates to a higher-dimensional space. In this high-dimensional space, the tower point group that can be difficult to distinguish in a two-dimensional plane can be more clearly distinguished, and the high-dimensional space provides more degrees of freedom to express the subtle differences and complex relationships between data points in the high-dimensional coordinate data.

[0065] Specifically, by further performing feature extraction on the high-dimensional coordinate data, in the preferred embodiment of the present application, the feature extraction module further processes the high-dimensional coordinate data through a sparse attention layer and an intermediate layer. The sparse attention layer uses a sliding window attention and a global attention parallel computing mode to capture local and global dependencies respectively, thereby obtaining the coordinate features of each position. These features not only contain the spatial position information of the tower, but also incorporate the learned environmental and contextual information, providing a more comprehensive data representation for clustering analysis.

[0066] The clustering module maps the coordinate features to determine the probability distribution of the tower in each cluster.

[0067] Specifically, the clustering module, as a key component for generating clustering weights and calculating clustering centers, maps the coordinate features to a clustering space, thereby realizing automatic clustering of the input data. The main functions of the clustering module include generating clustering weights, mapping the extracted features to a clustering space, and generating the probability distribution of each point belonging to different clusters. The clustering module calculates the probability of each tower belonging to each cluster, providing a basis for subsequent clustering center determination.

[0068] According to the probability distribution of the tower in each cluster, the number of clusters of the cluster is adjusted, and the corresponding clustering center of each tower in the tower group is determined, to obtain a UAV nest planning scheme.

[0069] Specifically, the number of clusters is dynamically adjusted according to the probability distribution of each tower in each cluster, and the center position of each cluster is determined. In the preferred embodiment of the present application, the model is trained by initializing the Transformer sparse encoder and cluster head, defining the cluster-related threshold, using the Adam optimizer, setting the learning rate to 0.001, calculating the total loss and updating the model parameters in each training cycle, and optimizing the model parameters according to the distance constraint loss between each tower and its cluster center, as well as the distance constraint loss between cluster centers. Finally, through the optimized model, the number of clusters and the cluster center are dynamically adjusted according to the UTM coordinates of all towers, and the site selection planning scheme of the UAV nest is obtained.

[0070] In summary, in combination with Figure 2 the present application inputs the UTM coordinates into the trained model with feature extraction module and clustering module, and outputs the site selection planning scheme of the UAV nest. Among them, regarding the model training, in the embodiment of the present method, the loss function is used for optimization:

[0071] The loss function of the present method contains four sub-loss functions, namely reconstruction loss , distance constraint loss between cluster centers , distance constraint loss between towers and cluster centers , and cluster center filtering constraint loss . Among them, the reconstruction loss is a loss for measuring the ability of the model to reconstruct the input data, which ensures that the features extracted by the model can effectively represent the input data, while constraining the position of the cluster center.

[0072] The reconstruction loss function is:

[0073] ;

[0074] Among them, N is the total number of tower points, is the input data, is the output of the model reconstruction.

[0075] The distance constraint loss function between cluster centers is:

[0076] ;

[0077] Among them, and are cluster centers, is 5km, and the distance constraint loss between the tower and the cluster center is to ensure that the distance between each tower point and its cluster center does not exceed 5km.

[0078] The distance constraint loss function between the tower and the cluster center is:

[0079] ;

[0080] wherein, is a tower point, is a tower point the cluster center belongs to, is 5km. The cluster center filtering constraint loss is a loss function used in cluster analysis, aiming to ensure that the distribution of cluster centers meets certain constraint conditions, thereby improving the quality of the clustering result.

[0081] The cluster center filtering constraint loss function is the sum of the reconstruction loss, the distance constraint loss between cluster centers, the distance constraint loss between towers and cluster centers, and the kdtree loss:

[0082] ;

[0083] wherein, is a weight coefficient for balancing the contribution of each loss, in the embodiment all are 1.

[0084] The unmanned aerial vehicle nest site planning method of the application first acquires the UTM coordinates of the towers, ensures the unity and accuracy of the coordinate system, and then uses the feature extraction module to perform dimensionality processing on the UTM coordinates to obtain high-dimensional coordinate data, thereby more comprehensively capturing the spatial distribution characteristics of the tower group. Then, by further feature extraction on the high-dimensional coordinate data, the coordinate features of the tower group are obtained, providing rich information for cluster analysis. Through the mapping of the coordinate features by the clustering module, not only the probability distribution of the towers in each cluster is determined, but also the number of clusters can be dynamically adjusted according to these distributions, ensuring that the number of clusters can be updated in real time, so that the corresponding cluster center of each tower can be adaptively determined, making the site selection of the unmanned aerial vehicle nest more scientific and reasonable, not only reducing the operation cost and energy consumption of the unmanned aerial vehicle inspection, but also improving the coverage rate and efficiency of the inspection, especially in complex environments such as uneven distribution of towers and variable terrain areas. The application can cluster the towers to adapt to environmental changes, thereby optimizing the layout of the unmanned aerial vehicle nest and improving the limitations of the unmanned aerial vehicle site selection in complex environments, realizing intelligent site selection of the power distribution network tower group, and significantly improving the efficiency and accuracy of the unmanned aerial vehicle nest site selection in complex environments.

[0085] Optionally, the UTM coordinates of all towers in the tower group are acquired, including:

[0086] The latitude and longitude coordinates of all the towers in the tower group are acquired;

[0087] The UTM coordinates of the towers are determined by projection processing according to the latitude and longitude coordinates.

[0088] Specifically, in the unmanned aerial vehicle nest site planning method of the present application, obtaining the UTM coordinates of all towers in the tower group is a key initial step. The UTM coordinate system is a plane rectangular coordinate system based on the transverse Mercator projection, which divides the earth's surface into multiple 6-degree longitudinal bands, each with its own coordinate system, thereby reducing the influence of earth curvature on measurement accuracy. In this way, the position of each tower is converted into a point in a unified and accurate coordinate system, providing a standardized data basis for subsequent feature extraction and clustering analysis. First, collect the latitude and longitude coordinates of all towers in the tower group, which provide the geographical position information of the towers on the earth's surface. Then, project the latitude and longitude coordinates to determine the UTM coordinates of each tower. In the preferred embodiment of the present application, the PyProj library can be used to perform this conversion, accurately converting latitude and longitude coordinates to UTM coordinates.

[0089] In the embodiment of the present application, by converting latitude and longitude coordinates to UTM coordinates, the influence of earth curvature can be eliminated, making the relative positions between towers more accurate, and the use of UTM coordinates also helps to simplify the subsequent data processing and analysis process, not only improving the scientificity and accuracy of unmanned aerial vehicle nest site selection, but also helping to optimize the flight path of the unmanned aerial vehicle, reduce energy consumption, and improve operation efficiency.

[0090] Optionally, the feature extraction module includes a data embedding layer; the high-dimensional coordinate data of the tower group is obtained by dimensionality processing according to the UTM coordinates through the feature extraction module, including:

[0091] According to the UTM coordinates of all towers in the tower group, a point group coordinate matrix of the tower group is obtained;

[0092] The point group coordinate matrix is input into the data embedding layer;

[0093] The point group coordinate matrix is mapped by the data embedding layer, and the initial dimension of the point group coordinate matrix is mapped to a preset high-order dimension;

[0094] The point group coordinate matrix of the preset high-order dimension is taken as the high-dimensional coordinate data.

[0095] Specifically, in combination with Figure 3As shown, the feature extraction module includes a data embedding layer. First, according to the UTM coordinates of all towers in the tower group, a point group coordinate matrix is constructed, wherein the matrix contains the position information of all towers. Then, the point group coordinate matrix is input into the data embedding layer in the feature extraction module. In the data embedding layer, the initial dimension of the point group coordinate matrix is mapped to a pre-set high-order dimension space through the mapping ability of the deep learning model, wherein a nonlinear transformation of the original coordinate data is involved, so that the model can capture more complex spatial relationships and potential features between towers. In the preferred embodiment of the present application, a feature extraction module with a sparse Transformer encoder can be used to map the tower point group coordinates to a high-dimensional space, for example, to map the 2D UTM coordinates to 64 dimensions. In this way, the feature vector of each tower point is expanded to 64 dimensions, providing more information for subsequent clustering analysis. The mapping can be performed through the mapping formula:

[0096] ;

[0097] wherein N is the total number of towers, X is the point group coordinate matrix, Linear(X) represents a linear transformation operation on the point group coordinate matrix, is the high-dimensional coordinate data.

[0098] In the embodiment of the present application, by mapping the original two-dimensional UTM coordinates to a high-dimensional space, the model not only retains the original geographical position information, but also learns the relative position relationship between towers, distribution density and other environmental factors that may affect the unmanned aerial vehicle operation. This high-dimensional feature representation makes the clustering analysis more accurate and can distinguish different tower groups in more detail, thereby providing a more scientific basis for the site selection of the unmanned aerial vehicle nest.

[0099] Optionally, the feature extraction module includes a sparse attention layer and an intermediate layer, and the feature extraction on the high-dimensional coordinate data to obtain the coordinate feature of the tower group comprises:

[0100] inputting the high-dimensional coordinate data into the sparse attention layer, performing parallel calculation through sliding window attention and global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and then performing normalization processing on the local attention features and the global attention features to obtain a fusion feature representation;

[0101] inputting the fusion feature representation into the intermediate layer, performing feature enhancement on the fusion feature representation through a multi-head attention mechanism to obtain an enhanced feature representation, and then performing residual linking and layer normalization processing on the enhanced feature representation to obtain the coordinate feature.

[0102] Specifically, in combination withFigure 3 As shown, the feature extraction module includes a sparse attention layer and an intermediate layer. The high-dimensional coordinate data is first input into the sparse attention layer, which combines a sliding window attention and a global attention mechanism to calculate local and global attention features in parallel. The sliding window attention focuses on the neighboring region of each data point, capturing local dependencies, while the global attention focuses on the entire data sequence, capturing long-range dependencies. The outputs of these two attention mechanisms are then normalized and fused into a feature representation. This fused feature representation is then fed into the intermediate layer, which contains a multi-head attention mechanism to further enhance the feature representation. In the intermediate layer, through the multi-head attention mechanism, the model can learn the features of the data from multiple subspace perspectives, thereby obtaining a richer information representation. Next, the enhanced feature representation is subjected to residual connection and layer normalization to stabilize the training process and improve the generalization ability of the model.

[0103] In the embodiments of the present application, by combining local and global attention mechanisms, the spatial distribution features of the tower point group can be more comprehensively captured. This feature extraction method not only can identify the local relationship between towers, but also can understand the global structure of the entire tower group. Since the most helpful features for clustering can be automatically learned instead of relying on manually designed feature extraction methods, the adaptability of the model to complex environments is improved.

[0104] Optionally, the inputting the high-dimensional coordinate data into the sparse attention layer includes performing parallel calculation through a sliding window attention and a global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and the method further includes:

[0105] Sliding a window of a preset size on the high-dimensional coordinate data to obtain an attention score between each data point in the window and other data points in the window in the high-dimensional coordinate data;

[0106] Performing a dot product according to the attention score to obtain an attention feature of each data point in the high-dimensional coordinate data;

[0107] According to the attention features of all the data points in the high-dimensional coordinate data, the local attention features of the high-dimensional coordinate data are obtained.

[0108] Specifically, inputting high-dimensional coordinate data into the sparse attention layer is a key step for extracting local and global features. First, by sliding a window of a preset size on the high-dimensional coordinate data, the attention score of each data point and other points in its neighborhood is calculated, quantifying the importance of each data point to determine its influence in the local neighborhood. Next, according to the calculated attention score, a dot product operation is performed to obtain the attention feature of each data point, which reflects the relative importance of the data point in the local window. Finally, the attention features of all data points are integrated to obtain the local attention feature of the entire high-dimensional coordinate data. The sparse attention layer uses a sliding window attention and global attention parallel computing method to capture local and global dependencies, respectively. In the preferred embodiment of the present application, the window size of the sliding window attention is 32, and the attention score between the position and the 16 positions before and after it (i.e. the window size is 32) is calculated for each position:

[0109] ;

[0110] wherein, is the query vector of data point position i, and is the key and value vector of data point position j, is the dimension of the key vector, Attention( ） is the attention score between the data point and other data points in the window, and T is the transpose matrix.

[0111] In the embodiment of the present application, the local attention feature is used to identify the close relationship and local pattern in the tower point group, providing more detailed data features for subsequent clustering analysis, thereby improving the accuracy of the UAV nest site selection.

[0112] Optionally, the input of the high-dimensional coordinate data into the sparse attention layer is parallel calculated by sliding window attention and global attention to obtain the local attention feature and the global attention feature of the high-dimensional coordinate data, further comprising:

[0113] by adding a preset number of global attention points to the sequence of the high-dimensional coordinate data;

[0114] obtaining the attention score between each data point in the high-dimensional coordinate data and the global attention point;

[0115] performing a dot product according to the attention score to obtain the global attention feature of the high-dimensional coordinate data.

[0116] Specifically, the model first adds a preset number of global attention points in the high-dimensional coordinate data sequence, which are key positions for capturing global information. Then, the model calculates the attention scores between each data point and these global attention points, which reflect the importance of the data points in the entire data sequence. Then, according to these attention scores, a dot product operation is performed to obtain the global attention features of the high-dimensional coordinate data. In the preferred embodiment of the present application, global attention is added by adding 4 global attention points in the input sequence, and for each position i, the attention scores of all global attention points are calculated using dot product:

[0117]

[0118] wherein, and are the key vector and value vector of the global attention point, is the query vector of the data point position i, and T is the transpose matrix, is the dimension of the key vector, Attention( ） is the attention score between the data point and the global attention point;

[0119] The local attention features and global attention features are combined and normalized:

[0120]

[0121]

[0122] wherein, WindowAttention( ） and GlobalAttention( ） are the local attention features and global attention features, LayerNorm( ） is the fusion feature representation, and the final fusion feature representation is sent to the intermediate layer, is the fusion feature representation, is the feature representation output after fusion.

[0123] In the embodiment of the present application, by adding global attention points, the model can identify key positions that have important influence on the entire data sequence, so as to capture important features and relationships in the entire region, facilitating better understanding of the distribution and structure of the data. The feature extraction method combining local and global information provides more abundant and detailed data features for subsequent clustering analysis, thereby improving the scientificity and accuracy of the unmanned aerial vehicle nest site selection.

[0124] ​​​Optionally, the intermediate layer comprises an encoder layer, the encoder layer comprises a multi-head attention network containing a preset number of attention heads; the inputting the fusion feature representation into the intermediate layer comprises performing feature enhancement on the fusion feature representation through a multi-head attention mechanism to obtain an enhanced feature representation, which comprises:

[0125] inputting the fusion feature representation into the multi-head attention network, performing linear transformation on the fusion feature representation through the attention heads to obtain an attention matrix;

[0126] obtaining an attention score of each attention head according to the attention matrix;

[0127] obtaining a weighted feature representation of each attention head according to the attention score and the attention matrix;

[0128] splicing the weighted feature representations of all the attention heads to obtain the enhanced feature representation.

[0129] Specifically, first, the fusion feature representation is input into the multi-head attention network, each attention head performs linear transformation on the fusion feature to generate a corresponding attention matrix. Subsequently, the attention scores of each attention head are calculated according to the attention matrices, the importance of different features is reflected by using the attention scores, then, the weighted feature representation of each attention head is obtained by using the attention scores and the corresponding attention matrix. Finally, the weighted feature representations of all the attention heads are spliced to form the enhanced feature representation. In the preferred embodiment of the present application, the intermediate layer is composed of 2 encoder layers, each encoder layer comprises a multi-head attention and a feedforward neural network, the multi-head attention contains 4 independent attention heads, which are 、 、 and , each attention head generates an attention matrix by linear transformation on the input data, including Q matrix , K matrix and V matrix :

[0130] ;

[0131] wherein, Q, V, K are query (Query), key (Key) and value (Value) respectively. These are the three core components in the attention mechanism, which are used to calculate the correlation or attention score between different parts of the input sequence. For the fusion feature representation, the attention scores of each attention head are calculated, and the outputs of all attention heads are spliced and mapped to 64 dimensions through a linear transformation. Finally, the outputs of all multi-head attention mechanisms are spliced and input to the feedforward neural network after layer normalization:

[0132] ;

[0133] ;

[0134] wherein, MultiHead( ） is the attention matrix output by the multi-head attention mechanism, concat( ， ， ） is a linear transformation of the independent attention head, is a linear layer, and X is the fusion feature representation, LayerNorm( MultiHead( ）） is the spliced weighted feature representation of the weighted feature representation of all attention heads and the fusion feature representation, is the spliced weighted feature representation.

[0135] The feedforward neural network is composed of two linear layers and an activation function:

[0136] ;

[0137] wherein, ReLU is an activation function, FFN is a feedforward neural network, , is a learnable weight matrix, is the hidden dimension of the feedforward neural network, and is a bias vector. In this embodiment, . The weighted feature representation processed by the feedforward neural network is connected in residual and normalized by layer, and then the enhanced feature representation is output:

[0138] ;

[0139] wherein, FFN( ） is the attention score obtained by processing the spliced weighted feature representation through the feedforward neural network, LayerNorm( FFN( ）） To obtain by residual connection through the feedforward neural network and normalization operation As an enhanced feature representation.

[0140] In the embodiments of the present application, by splicing the weighted feature representations of different attention heads, the model can integrate information from different attention heads to form a comprehensive feature representation, which helps subsequent clustering analysis and decision making. This method not only improves the scientificity and accuracy of the unmanned aerial vehicle nest site selection, but also enhances the adaptability and robustness of the model to complex environments, so as to realize better site selection planning in practical applications.

[0141] Optionally, the clustering module comprises a fully connected layer and a Softmax layer; the mapping of the coordinate feature by the clustering module to determine the probability distribution of the tower in each cluster comprises:

[0142] The feature dimension of the coordinate feature is mapped to the maximum cluster number of the cluster by the fully connected layer to obtain the cluster weight of the tower;

[0143] The cluster weight of each tower is normalized by the Softmax layer to obtain the probability distribution of the tower in each cluster.

[0144] Specifically, in combination with Figure 4 As shown in the figure, the clustering module comprises a fully connected layer and a Softmax layer, the fully connected layer maps the feature dimension of the coordinate feature to the dimension of the maximum cluster number K to generate the cluster weight, and the cluster weight represents the possibility of each tower belonging to each cluster. Then, the Softmax layer normalizes the cluster weight to convert the cluster weight of each tower into a probability distribution, so that the sum of the probabilities of each tower in all clusters is 1. In the preferred embodiments of the present application, the clustering module is composed of 1 layer of fully connected layer and 1 layer of Softmax layer, the fully connected layer maps the feature dimension to the maximum cluster number K dimension, k can take the value of 20, that is, 20 dimensions, and then the Softmax layer normalizes the original output of the cluster weight to the probability distribution.

[0145] ;

[0146] is a weight matrix, is a bias vector, is the original output of the cluster weight, b is the bias vector, and H is the feature matrix of the input data.

[0147] In the embodiments of the present application, the clustering module provides each tower with a clear probability distribution to indicate the likelihood of the tower belonging to each cluster. Since the output of the Softmax layer can adapt to different cluster numbers, it can not only automatically determine the cluster to which each tower is most likely to belong, but also handle cases where the number of clusters is unknown or changing. In addition, by normalizing the weight matrix into a probability distribution, the model can more intuitively understand the relative position of each tower in different clusters, thereby providing more accurate guidance for the location of the UAV nest, improving the flexibility and accuracy of cluster analysis, and helping to optimize the inspection path of the UAV, reduce flight distance and energy consumption, improve inspection efficiency and quality, and ultimately achieve a more optimal UAV nest location planning.

[0148] Optionally, the adjusting the number of clusters of the cluster according to the probability distribution of the tower in each cluster, determining the cluster center corresponding to each tower in the tower group, and obtaining the UAV nest address comprise:

[0149] According to the probability distribution of the tower in each cluster, the cluster is screened, the number of clusters of the cluster is adjusted, and the corresponding tower is allocated to the screened cluster;

[0150] According to the probability distribution corresponding to the tower in each cluster, the cluster center of the cluster is determined by weighted average;

[0151] The cluster center and the cluster allocation result of each tower are used as the UAV nest planning scheme.

[0152] Specifically, the clusters are screened according to the probability distribution of the tower in each cluster, and those clusters with significant probability distribution are retained, and the corresponding tower is allocated to these screened clusters. Among them, the clusters with low probability distribution are merged or the clusters with high probability distribution are split to ensure that each cluster has representativeness and practicality. Then, for each screened cluster, the system performs weighted average on the probability distribution corresponding to each tower in the cluster to calculate the cluster center of the cluster. Finally, the calculated cluster center and the cluster allocation result of each tower are integrated to form a planning scheme of the UAV nest. Generally, the UAV nest will be planned near the cluster center, because the cluster center represents the concentrated position of the towers in the region, which can maximize the coverage efficiency and reduce the flight distance. The planning scheme will consider the distance between each cluster center and the tower, as well as the possible flight path and obstacles, to determine the optimal nest position.

[0153] In the preferred embodiments of the present application, the original output of the cluster weight is normalized into a probability distribution by the Softmax layer:

[0154] ;

[0155] wherein, represents the probability that the i-th tower belongs to the j-th cluster, k is the number of clusters, exp is the exponential function, is the cluster weight of the i-th tower belonging to the j-th cluster. According to the cluster weight and the input features, the center of each cluster is calculated.

[0156] ;

[0157] wherein, is the center of the j-th cluster, is the high-dimensional feature representation of the i-th tower, and N is the total number of towers.

[0158] In the embodiments of the present application, by calculating the weighted average cluster center for each cluster, it can be ensured that the UAV nest is located at a position that can effectively cover its responsible area, so as to optimize the flight path and inspection efficiency of the UAV. Since it does not depend on the preset number of clusters, it can adapt to different geographical and environmental conditions. Not only does it improve the flexibility and adaptability of the UAV nest site selection, but it also helps to reduce operating costs and improve the efficiency of power grid maintenance, ultimately realizing a more optimal UAV nest planning scheme.

[0159] In combination with Figure 5 shown, the UAV nest site planning system provided by the embodiments of the present application comprises:

[0160] A coordinate acquisition unit is configured to acquire the UTM coordinates of all towers in a tower group.

[0161] A feature extraction unit is configured to perform dimensionality processing on the UTM coordinates by a feature extraction module to obtain high-dimensional coordinate data of the tower group, and perform feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group.

[0162] A clustering unit is configured to map the coordinate features by a clustering module to determine the probability distribution of the towers in each cluster, adjust the cluster number of the clusters according to the probability distribution of the towers in each cluster, determine the cluster center corresponding to each tower in the tower group, and obtain a UAV nest planning scheme.

[0163] The unmanned aerial vehicle nest site planning system of the application firstly acquires the UTM coordinates of the tower, ensures the uniformity and accuracy of the coordinate system, then uses the feature extraction module to process the UTM coordinates to obtain high-dimensional coordinate data, so as to more comprehensively capture the spatial distribution characteristics of the tower group. Then, through further feature extraction of the high-dimensional coordinate data, the coordinate features of the tower group are obtained, which provides rich information for clustering analysis. Through the mapping of the clustering module to the coordinate features, not only the probability distribution of the tower in each cluster is determined, but also the number of clusters can be dynamically adjusted according to the distribution, which ensures that the number of clusters can be updated in real time, so as to adaptively determine the corresponding cluster center of each tower, so that the site selection of the unmanned aerial vehicle nest is more scientific and reasonable, which not only reduces the operation cost and energy consumption of the unmanned aerial vehicle inspection, but also improves the coverage rate and efficiency of the inspection, especially in complex environments, such as uneven distribution of towers and variable terrain areas, the application can cluster the towers, so that the selection of clustering can adapt to environmental changes, thereby optimizing the layout of the unmanned aerial vehicle nest, improving the limitations of the unmanned aerial vehicle site selection in complex environments, realizing intelligent site selection of the distribution network tower group, and significantly improving the efficiency and accuracy of the unmanned aerial vehicle nest site selection in complex environments.

[0164] Although the present application is disclosed as above, the protection scope of the present application is not limited to this. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and these changes and modifications will fall within the protection scope of the present application.

Claims

1. A method for site planning of a drone nest, characterized in that, The method comprises the following steps: obtaining the UTM coordinates of all towers in a tower group; performing dimensionality processing on the UTM coordinates by a feature extraction module to obtain high-dimensional coordinate data of the tower group; wherein the feature extraction module comprises a data embedding layer, a sparse attention layer and an intermediate layer; performing dimensionality processing on the UTM coordinates by the feature extraction module to obtain high-dimensional coordinate data of the tower group, specifically comprising: obtaining a point group coordinate matrix of the tower group according to the UTM coordinates of all towers in the tower group; inputting the point group coordinate matrix into the data embedding layer; mapping the point group coordinate matrix by the data embedding layer to map the initial dimension of the point group coordinate matrix to a preset high-order dimension; taking the point group coordinate matrix of the preset high-order dimension as the high-dimensional coordinate data; performing feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group; wherein, specifically comprising: inputting the high-dimensional coordinate data into the sparse attention layer, performing parallel calculation by sliding window attention and global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and then performing normalization processing on the local attention features and the global attention features to obtain a fusion feature representation; inputting the fusion feature representation into the intermediate layer, performing feature enhancement on the fusion feature representation by a multi-head attention mechanism to obtain an enhanced feature representation, and then performing residual linking and layer normalization processing on the enhanced feature representation to obtain the coordinate features; mapping the coordinate features by a clustering module to determine the probability distribution of the towers in each cluster; wherein the clustering module comprises a full connection layer and a Softmax layer; mapping the coordinate features by the full connection layer to map the feature dimension of the coordinate features according to the maximum cluster number of the clusters to obtain the cluster weight of the towers; performing normalization processing on the cluster weight of each tower by the Softmax layer to obtain the probability distribution of the towers in each cluster; adjusting the cluster number of the clusters according to the probability distribution of the towers in each cluster to determine the cluster center corresponding to each tower in the tower group to obtain a UAV nest planning scheme; wherein, specifically comprising: screening the clusters according to the probability distribution of the towers in each cluster, adjusting the cluster number of the clusters, and assigning the corresponding towers to the screened clusters; performing weighted average on the probability distribution corresponding to the towers in each cluster to determine the cluster center of the cluster; taking the cluster center and the cluster assignment result of each tower as the UAV nest planning scheme. 2.The method of claim 1, wherein, The method comprises the following steps: obtaining the UTM coordinates of all towers in a tower group; performing dimensionality processing on the UTM coordinates by a feature extraction module to obtain high-dimensional coordinate data of the tower group; wherein the feature extraction module comprises a data embedding layer, a sparse attention layer and an intermediate layer; performing dimensionality processing on the UTM coordinates by the feature extraction module to obtain high-dimensional coordinate data of the tower group, specifically comprising: obtaining a point group coordinate matrix of the tower group according to the UTM coordinates of all towers in the tower group; inputting the point group coordinate matrix into the data embedding layer; mapping the point group coordinate matrix by the data embedding layer to map the initial dimension of the point group coordinate matrix to a preset high-order dimension; taking the point group coordinate matrix of the preset high-order dimension as the high-dimensional coordinate data; performing feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group; wherein, specifically comprising: inputting the high-dimensional coordinate data into the sparse attention layer, performing parallel calculation by sliding window attention and global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and then performing normalization processing on the local attention features and the global attention features to obtain a fusion feature representation; inputting the fusion feature representation into the intermediate layer, performing feature enhancement on the fusion feature representation by a multi-head attention mechanism to obtain an enhanced feature representation, and then performing residual linking and layer normalization processing on the enhanced feature representation to obtain the coordinate features; mapping the coordinate features by a clustering module to determine the probability distribution of the towers in each cluster; wherein the clustering module comprises a full connection layer and a Softmax layer; mapping the coordinate features by the full connection layer to map the feature dimension of the coordinate features according to the maximum cluster number of the clusters to obtain the cluster weight of the towers; performing normalization processing on the cluster weight of each tower by the Softmax layer to obtain the probability distribution of the towers in each cluster; adjusting the cluster number of the clusters according to the probability distribution of the towers in each cluster to determine the cluster center corresponding to each tower in the tower group to obtain a UAV nest planning scheme; wherein, specifically comprising: screening the clusters according to the probability distribution of the towers in each cluster, adjusting the cluster number of the clusters, and assigning the corresponding towers to the screened clusters; performing weighted average on the probability distribution corresponding to the towers in each cluster to determine the cluster center of the cluster; taking the cluster center and the cluster assignment result of each tower as the UAV nest planning scheme. The method comprises the following steps: obtaining the UTM coordinates of all towers in a tower group; performing dimensionality processing on the UTM coordinates by a feature extraction module to obtain high-dimensional coordinate data of the tower group; wherein the feature extraction module comprises a data embedding layer, a sparse attention layer and an intermediate layer; performing dimensionality processing on the UTM coordinates by the feature extraction module to obtain high-dimensional coordinate data of the tower group, specifically comprising: obtaining a point group coordinate matrix of the tower group according to the UTM coordinates of all towers in the tower group; inputting the point group coordinate matrix into the data embedding layer; mapping the point group coordinate matrix by the data embedding layer to map the initial dimension of the point group coordinate matrix to a preset high-order dimension; taking the point group coordinate matrix of the preset high-order dimension as the high-dimensional coordinate data; performing feature extraction on the high-dimensional coordinate data to obtain coordinate features of the tower group; wherein, specifically comprising: inputting the high-dimensional coordinate data into the sparse attention layer, performing parallel calculation by sliding window attention and global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and then performing normalization processing on the local attention features and the global attention features to obtain a fusion feature representation; inputting the fusion feature representation into the intermediate layer, performing feature enhancement on the fusion feature representation by a multi-head attention mechanism to obtain an enhanced feature representation, and then performing residual linking and layer normalization processing on the enhanced feature representation to obtain the coordinate features; mapping the coordinate features by a clustering module to determine the probability distribution of the towers in each cluster; wherein the clustering module comprises a full connection layer and a Softmax layer; mapping the coordinate features by the full connection layer to map the feature dimension of the coordinate features according to the maximum cluster number of the clusters to obtain the cluster weight of the towers; performing normalization processing on the cluster weight of each tower by the Softmax layer to obtain the probability distribution of the towers in each cluster; adjusting the cluster number of the clusters according to the probability distribution of the towers in each cluster to determine the cluster center corresponding to each tower in the tower group to obtain a UAV nest planning scheme; wherein, specifically comprising: screening the clusters according to the probability distribution of the towers in each cluster, adjusting the cluster number of the clusters, and assigning the corresponding towers to the screened clusters; performing weighted average on the probability distribution corresponding to the towers in each cluster to determine the cluster center of the cluster; taking the cluster center and the cluster assignment result of each tower as the UAV nest planning scheme. 3.The method of claim 1, wherein, The high-dimensional coordinate data is input into the sparse attention layer, parallel calculation is performed through sliding window attention and global attention, local attention features and global attention features of the high-dimensional coordinate data are obtained, and the method comprises the following steps: Sliding a window of a preset size on the high-dimensional coordinate data to obtain attention scores between each data point in the window and other data points in the window; Dot product is performed according to the attention scores to obtain attention features of each data point in the high-dimensional coordinate data; The local attention features of the high-dimensional coordinate data are obtained according to the attention features of all data points in the high-dimensional coordinate data. 4.The method of claim 3, wherein, The high-dimensional coordinate data is input into the sparse attention layer, parallel calculation is performed through sliding window attention and global attention, local attention features and global attention features of the high-dimensional coordinate data are obtained, and the method further comprises the following steps: A preset number of global attention points are added to the sequence of the high-dimensional coordinate data; Attention scores between each data point in the high-dimensional coordinate data and the global attention points are obtained; Dot product is performed according to the attention scores to obtain the global attention features of the high-dimensional coordinate data. 5.The method of claim 1, wherein, The intermediate layer comprises an encoder layer, the encoder layer comprises a multi-head attention network, and the multi-head attention network comprises a preset number of attention heads; The fusion feature representation is input into the intermediate layer, and the fusion feature representation is subjected to feature enhancement through a multi-head attention mechanism to obtain an enhanced feature representation, which comprises the following steps: The fusion feature representation is input into the multi-head attention network, and the fusion feature representation is subjected to linear transformation through the attention heads to obtain an attention matrix; Attention scores of each attention head are obtained according to the attention matrix; Weighted feature representations of each attention head are obtained according to the attention scores and the attention matrix; The enhanced feature representation is obtained by splicing the weighted feature representations of all attention heads.

6. An unmanned aerial vehicle nest site planning system, comprising: Comprise: A coordinate acquisition unit is configured to acquire UTM coordinates of all towers in a tower group; A feature extraction unit is configured to acquire high-dimensional coordinate data of the tower group by performing dimension lifting on the UTM coordinates through a feature extraction module; wherein the feature extraction module comprises a data embedding layer, a sparse attention layer and an intermediate layer; The feature extraction module is configured to acquire high-dimensional coordinate data of the tower group by performing dimension lifting on the UTM coordinates, and specifically comprises the following steps: acquiring a point group coordinate matrix of the tower group according to the UTM coordinates of all towers in the tower group; inputting the point group coordinate matrix into the data embedding layer; mapping the point group coordinate matrix through the data embedding layer to map an initial dimension of the point group coordinate matrix to a preset high-order dimension; and taking the point group coordinate matrix of the preset high-order dimension as the high-dimensional coordinate data. The high-dimensional coordinate data is feature extracted to obtain a coordinate feature of the tower group; wherein, specifically includes: inputting the high-dimensional coordinate data into the sparse attention layer, performing parallel calculation through sliding window attention and global attention to obtain local attention features and global attention features of the high-dimensional coordinate data, and then normalizing the local attention features and the global attention features to obtain a fusion feature representation; inputting the fusion feature representation into the intermediate layer, performing feature enhancement on the fusion feature representation through a multi-head attention mechanism to obtain an enhanced feature representation, and then performing residual linking and layer normalization processing on the enhanced feature representation to obtain the coordinate feature; A clustering unit is configured to map the coordinate feature through a clustering module to determine a probability distribution of the tower in each cluster; wherein, the clustering module includes a full connection layer and a Softmax layer; the mapping of the coordinate feature through the clustering module to determine the probability distribution of the tower in each cluster specifically includes: mapping the feature dimension of the coordinate feature according to the maximum cluster number of the cluster through the full connection layer to obtain a cluster weight of the tower; and normalizing the cluster weight of each tower through the Softmax layer to obtain the probability distribution of the tower in each cluster; According to the probability distribution of the tower in each cluster, the cluster number of the cluster is adjusted to determine a cluster center corresponding to each tower in the tower group to obtain a UAV nest planning scheme; wherein, specifically includes: screening the cluster according to the probability distribution of the tower in each cluster, adjusting the cluster number of the cluster, and allocating the corresponding tower to the screened cluster; and performing weighted average on the probability distribution corresponding to the tower in each cluster to determine the cluster center of the cluster; and taking the cluster center and the cluster allocation result of each tower as the UAV nest planning scheme.

Images