Multi-source heterogeneous data communication line optimal path planning method and system
By employing a path planning method based on multi-source heterogeneous data, and utilizing graph neural networks and convolutional feature fusion, combined with long short-term memory networks and low-rank attention modules, the computational resource consumption and robustness issues of traditional path planning methods in complex environments are resolved, achieving efficient and coherent optimal path planning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN GREAT POWER GEO INFORMATION TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-08
AI Technical Summary
Traditional path planning methods consume large amounts of computational resources in complex geographical environments, and the search results are unstable, making it difficult to guarantee the robustness and overall optimality of the path. Furthermore, deep learning methods fail to fully consider the topological dependencies and geospatial feature interactions between nodes, resulting in inaccurate path planning.
An optimal path planning method for communication lines using multi-source heterogeneous data is proposed. This method extracts node topological features and fuses them with convolutional features through graph neural networks, combines them with long short-term memory networks for temporal modeling, and introduces a low-rank attention module to reduce computational complexity. Finally, a pre-trained path planning model is used for path decision-making.
It improves the adaptability and rationality of path planning, reduces the probability of generating redundant paths, achieves efficient real-time response and path coherence, and ensures the optimality and robustness of the path.
Smart Images
Figure CN121997507A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for optimal path planning of communication lines with multi-source heterogeneous data, belonging to the field of communication line planning technology. Background Technology
[0002] Traditional power distribution communication path planning methods typically employ random sampling-based strategies. Examples include Probabilistic Route Maps (PRM), Rapid Random Tree Exploration (RRT), and Extended Spatial Tree (EST). These methods explore feasible paths by randomly sampling in a continuous space and constructing a graph or tree structure, avoiding complete discrete modeling of the environment. However, they still require extensive collision detection or probabilistic calculations at the sampling points, resulting in high computational resource consumption. Furthermore, the randomness leads to unstable search results, making it difficult to guarantee path robustness and overall optimality, especially in long-distance or complex geographical environments.
[0003] The development of deep convolutional neural networks has greatly improved path planning results. For example, deep reinforcement learning, in fixed thematic scenarios, learns the scores of different path node selections and selects the node with the highest score as the result. However, this method cannot solve the problem of continuous prediction. Furthermore, although existing deep learning methods have achieved success on specific tasks, they fail to fully consider the topological dependencies between nodes and lack the ability to interact between node attributes and geospatial features. In areas with significantly uneven distribution of building function types and imbalances in energy supply and load, such methods often struggle to accurately capture the dynamic relationship between local structures and global patterns, leading to feasible paths that involve detours or contain redundant nodes, which are not optimal paths.
[0004] Therefore, there is an urgent need for a method and system that can intelligently plan the optimal path for communication lines using multi-source heterogeneous data such as nodes and images. Summary of the Invention
[0005] To address the problems existing in the prior art, this invention proposes a method and system for optimal path planning of communication lines for multi-source heterogeneous data.
[0006] The technical solution of the present invention is as follows: On the one hand, this invention proposes an optimal path planning method for communication lines of multi-source heterogeneous data, including the following steps: Obtain the thematic map of building function classification and spatial data of communication nodes in the target area; preprocess the thematic map of building function classification, construct an undirected graph based on the spatial data of communication nodes, and obtain the source node and target node according to the preset path planning task; The undirected graph and the preprocessed building function classification thematic map are input into the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features. After concatenation, the environment features are output. The source node is used as the initial node to construct a sequence of selected nodes, which is input into the pre-trained path planning model path memory encoder to obtain path features. At the same time, based on the connection relationship of the last node in the selected node sequence in the undirected graph, multiple candidate nodes to be selected are determined. The environment features and path features are input into the feature fusion module to obtain fused features, which are then input into the pre-trained path planning model path decision decoder to output the predicted probability score of each candidate node. Based on the predicted probability scores of candidate nodes, the candidate node with the highest predicted probability score is selected as the next node and added to the selected node sequence. It is then input into the pre-trained path planning model path memory encoder and iterated continuously until the target node is reached, outputting the optimal path of the communication line. The pre-trained path planning model includes an environment-aware encoder, a path memory encoder, and a path decision decoder.
[0007] Preferably, the method inputs the undirected graph and the preprocessed building function classification thematic map into the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features, specifically: A graph attention network is used to calculate the attention weights between the current node and its neighboring nodes based on Euclidean distance and dispersion, and then the neighborhood features are aggregated. Subsequently, a graph convolutional network is used to aggregate second-order neighborhood information and output the node topological features. The building function classification thematic map is processed using a convolutional feature extractor in an environment-aware encoder: high-level semantic information of the image is extracted through multi-layer convolutional operations and downsampling operations, and features at different levels are fused through skip connections to output spatial semantic features; The node topological features and spatial semantic features are concatenated and processed by an activation function to obtain environmental features.
[0008] Preferably, the method uses the source node as the initial node to construct a sequence of selected nodes, which is then input into a pre-trained path planning model, the path memory encoder, to obtain path features, specifically: The path memory encoder employs a long short-term memory network; The node features in the selected node sequence are sequentially input into the memory units of the Long Short-Term Memory (LSTM) network according to the access order. The sequence is then modeled temporally through the gating mechanism of the LSTM network, and the hidden state is updated. The final hidden state output by the LSTM network is used as the path feature representing the path history information.
[0009] Preferably, the feature fusion module is a low-rank attention module.
[0010] Preferably, the method inputs environmental features and path features into a low-rank attention module to obtain fused features, specifically as follows: One-dimensional convolution operations are performed on environmental features and path features respectively to generate query, key, value and output gated feature vectors corresponding to environmental features and path features; The query, key, value, and output gating feature vectors corresponding to environmental features and path features are concatenated to obtain the query, key, value, and output gating feature vectors of the fused features. The query and key feature vectors of the fused features are processed by activation function, a bias term is added, and rotational position encoding is applied to obtain query and key feature vectors with positional information; The key feature vectors with location information are weighted by importance to obtain the weighted key feature vectors of the fused features; The traditional attention matrix calculation is split into two low-rank matrices, and the linear features output by the attention mechanism are calculated, expressed by the formula: ; In the formula, This represents the linear features output after the attention mechanism. This represents the query feature vector after applying positional encoding. This represents the key feature vector after applying positional encoding. This represents the weighted key feature vector after applying positional encoding. Represents the height of the feature vector. Represents the width of the feature vector. This indicates the transpose operation. The feature vector represents the fused features; Linear features output after attention mechanism and the output gated feature vector of the fused features Performing element-wise multiplication yields a fused feature incorporating both environmental and path information, expressed as a formula: ; In the formula, Indicates fusion features, This indicates an operation that changes the shape of the array. This indicates an element-wise multiplication operation. The output gated feature vector represents the fused features. This indicates a fully connected operation.
[0011] Preferably, the method fuses features and inputs them into a pre-trained path planning model path decision decoder, outputting the predicted probability score of each candidate node, specifically as follows: The path decision decoder includes a first fully connected layer and a second fully connected layer; The fused features are input into the first fully connected layer for dimensionality reduction to obtain intermediate features; The intermediate features are concatenated with the environmental features to obtain the fused features; The fused features are input into the second fully connected layer, and the output is the guidance value corresponding to each neighboring node of the current node; The guidance value is normalized to obtain the predicted probability score of each candidate node.
[0012] Preferably, the pre-trained path planning model is obtained by training using the following loss function: The loss function includes binary cross-entropy loss and shortest path ratio loss, wherein: Binary cross-entropy loss is used to measure the difference between the predicted probability distribution of the model output and the manually labeled optimal path node labels; The shortest path ratio loss is used to measure the ratio between the length of the complete predicted path generated by the model and the length of the optimal path labeled by humans, thus constraining path redundancy. The total loss of the path planning model is the weighted sum of the shortest path ratio loss and the shortest path ratio loss.
[0013] On the other hand, the present invention also proposes an optimal path planning system for communication lines of multi-source heterogeneous data, comprising the following modules: Preprocessing module: acquires thematic maps of building function classification and communication node spatial data of the target area; preprocesses the thematic maps of building function classification, constructs an undirected graph based on the communication node spatial data, and obtains source nodes and target nodes according to the preset path planning task; The node prediction module takes the undirected graph and the preprocessed building function classification thematic map as input to the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features, and outputs environmental features after concatenation. It uses the source node as the initial node to construct a selected node sequence, which is then input to the pre-trained path planning model path memory encoder to obtain path features. Simultaneously, based on the connection relationships of the last node in the selected node sequence in the undirected graph, multiple candidate nodes are determined. The environmental features and path features are then input to the feature fusion module to obtain fused features, which are then input to the pre-trained path planning model path decision decoder to output the predicted probability score of each candidate node. Path result search module: Based on the predicted probability scores of candidate nodes, select the candidate node with the highest predicted probability score as the next node and add it to the selected node sequence. Input it into the pre-trained path planning model path memory encoder, and continue to iterate until the target node is reached, and output the optimal path of the communication line.
[0014] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any embodiment of the present invention.
[0015] In another aspect, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.
[0016] The present invention has the following beneficial effects: (1) This invention provides a method and system for optimal path planning of communication lines with multi-source heterogeneous data. It uses an environmentally perceptive encoder to deeply fuse node topological features extracted based on graph neural networks with building spatial semantic features obtained based on convolutional feature extractors. This feature effectively solves the limitation of traditional planning methods that only consider spatial distance or simple adjacency relationships, enabling the model to perceive complex geographical environments, thereby improving the adaptability of communication lines in multi-source heterogeneous environments and the rationality of planning.
[0017] (2) This invention provides a method and system for optimal path planning of communication lines with multi-source heterogeneous data. It utilizes a long short-term memory network to perform time-series modeling on the selected node sequence, transforming historical path information into path features. This feature significantly enhances the model's memory capacity during the decision-making process, effectively identifying and avoiding redundant loops common in path planning. By dynamically sensing the already traversed paths, it ensures that the generated paths have strong directionality and coherence, greatly reducing the probability of generating invalid paths.
[0018] (3) This invention is a method and system for optimal path planning of communication lines with multi-source heterogeneous data. By introducing a low-rank attention module, the traditional calculation of the complete attention matrix is split into the product operation of two low-rank matrices. This feature reduces the computational complexity of the attention mechanism from quadratic to linear. While ensuring full interaction between environmental features and path features, it significantly reduces the computational overhead and memory usage of the model and effectively prevents the gradient explosion problem in deep network training. This enables the system to achieve efficient real-time response to large-scale communication node networks while ensuring planning accuracy. Attached Figure Description
[0019] Figure 1 This is a flowchart of the method proposed in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the method model proposed in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the LRAM module execution flow proposed in Embodiment 1 of the present invention; Figure 4This is a schematic diagram of some path planning results proposed in Embodiment 1 of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.
[0022] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0023] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.
[0024] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.
[0025] Example 1: See Figure 1 This embodiment proposes an optimal path planning method for communication lines of multi-source heterogeneous data, including the following steps: S100. Obtain a thematic map of building function classification in the target area and preprocess it; the preprocessing includes image cropping, image flipping, and contrast enhancement. Simultaneously, spatial data of communication nodes in the target area is acquired, and an undirected graph is constructed based on this data, expressed by the formula: ; In the formula, Represents an undirected graph. Represents a set of points. Represents the set of edges; Based on the constructed undirected graph and combined with the preset path planning task, the source node and target node of the optimal path of the communication line are determined, and the manually labeled optimal path is obtained. S200, see also Figure 2A path planning model based on neural networks is constructed, which includes an environmental perception encoder, a path memory encoder, and a path decision decoder. Specifically: S201, The environment-aware encoder includes an environment feature extractor and a convolutional feature extractor, wherein: A graph convolutional encoder, consisting of a graph neural network (GNN), is used to process node data with a graph structure. By hierarchically aggregating the features of adjacent nodes, it enables each node to perceive global topological information. It should be noted that in the Graph Neural Network (GNN), a Graph Attention Network (GAT) module is designed to learn the neighborhood priority features of the current node based on distance and dispersion, expressed by the formula: ; In the formula, This represents the output after passing through the graph attention network module. Each node's characteristics Indicates the first The node and the first Attention weights between nodes express Activation function Indicates the first 1 node This represents the set of nodes adjacent to the current node. Indicates the first The node Initial features after secondary aggregation; Subsequently, a graph convolutional network (GCN) module is designed to acquire second-order neighborhood information, and its propagation process is expressed by the following formula: ; ; In the formula, This represents the graph convolutional network. The node feature matrix of the layer The degree matrix represents the adjacency matrix. This represents the adjacency matrix with self-loops. Represents the adjacency matrix of an undirected graph. Represents the identity matrix. This represents the graph convolutional network. The node feature matrix of the layer The graph convolutional network represents the first... Layer trainable weight matrix, This represents the layer index of the graph convolutional network.
[0026] In this embodiment, the undirected graph is input into the graph convolution encoder, first processing the current (e.g., the source node) node. its own coordinates and the target node The coordinates are concatenated and then appended to other nodes in the undirected graph. The second aggregation operation yields the current node. The initial characteristics are expressed by the formula: ; ; In the formula, Indicates the first The node The initial features after the second aggregation. Indicates the first 1 node Indicates the target node. Indicates the first The node The initial features after the second aggregation. This indicates a fully connected operation. This indicates a splicing operation. Indicates the first The node The initial features after the second aggregation. This indicates a global average pooling operation. Indicates the first 1 node Represents the set of all nodes. Indicates the first The node The initial features after the second aggregation. Indicates the number of aggregations. and Indicates the node index; The calculated initial features of the current node are input into the graph attention network (GAT) to calculate the current node's features. The attention weights between it and its neighboring nodes are expressed by the formula: ; In the formula, Indicates the first The node and the first Attention weights between nodes Indicates taking Index trading. express Activation function Indicates the first The node The initial features after the second aggregation. Indicates the first 1 node Represents the set of nodes adjacent to the current node; Based on the current node The attention weights between the node and its neighboring nodes are calculated, and the node features output after passing through the graph attention network are expressed by the formula: ; In the formula, This represents the output after passing through the graph attention network. Each node's characteristics; The node features output by all nodes after passing through the graph attention network are obtained, forming the node features of the graph convolutional network. Layer Node Feature Matrix ; The graph convolutional network Layer Node Feature Matrix The input is fed into the Graph Convolutional Network (GCN) module to obtain the second-order Graph Convolutional Network after fusing second-order neighborhood information. Layer node feature matrix As a topological feature of nodes in an undirected graph, it is denoted as... .
[0027] Furthermore, a convolutional feature extractor, BsiNetv2, is constructed and used to extract the spatial semantic feature hidden representation of the preprocessed building function classification thematic map. Specifically: The encoder structure of the convolutional feature extractor is constructed, and a 3×3 convolution operator with a stride of 1 is used to perform convolution processing on the input building function classification thematic map, and the number of feature channels is expanded to 32 to obtain multi-level image features; Perform a 2×2 max pooling operation on the convolutional feature map to reduce the spatial resolution of the feature map and reduce the feature dimension; Multiple ConBlock modules are connected in series in the encoder. Each ConBlock module includes a 3×3 convolutional layer, an activation function, and a downsampling operation based on max pooling, which is used to enhance the high-level semantic expression capability of features layer by layer. To mitigate the loss of feature information during downsampling, a skip connection mechanism is introduced to fuse semantic features from different encoder levels with the upsampling output of the corresponding ConBlock module. Based on the above multi-layer feature extraction and fusion operations, spatial semantic features representing the functional distribution of buildings are obtained, denoted as... .
[0028] Furthermore, spatial semantic features Topological characteristics of undirected graph nodes By fusing these features, an environmental feature with global scene awareness is constructed, which can be expressed as a formula: ; In the formula, Indicates environmental characteristics, Representing spatial semantic features, This represents the topological characteristics of nodes in an undirected graph.
[0029] S202, The path memory encoder includes a Long Short-Term Memory (LSTM) network and a Low-Rank Attention (LRAM) module, wherein: The path memory encoder receives the sequence of nodes selected during the current path planning process as input, and sequentially inputs them into the LSTM memory cells according to the order in which the nodes are visited. By performing temporal modeling on the node sequence, the hidden states are updated step by step, thereby obtaining path features that represent the historical information of the path, denoted as . ; See Figure 3 Environmental characteristics and path features The input is fed into the low-rank attention module (LRAM) to obtain fused features with environmental and path information, specifically: After inputting environmental features and path memory features, a one-dimensional convolution operation is performed to obtain query, key, value, and output gated feature vectors, expressed by the formula: ; ; In the formula, Indicates environmental characteristics, Query feature vectors representing environmental characteristics. Key feature vectors representing environmental characteristics. The feature vector represents the value of environmental characteristics. The output gated feature vector represents the environmental characteristics. Indicate path features, A query feature vector representing path characteristics. Key feature vectors representing path features. The value feature vector represents the path features. The output gated feature vector represents the path characteristics; By concatenating the above feature vectors, we obtain the query, key, value, and output gated feature vectors of the fused features, expressed by the formula: ; In the formula, The query feature vector represents the fused features. The key feature vector represents the fused features. The feature vector represents the value of the fused features. The output gated feature vector represents the fused features; The initially generated query and key feature vectors are subjected to a ReLU activation function and biased to ensure non-negativity and achieve a gating effect, as expressed by the formula: ; ; In the formula, This represents the query feature vector after activation and bias. This represents the key feature vector after activation and paranoia; Apply rotational position encoding to the activated and biased query and key feature vectors, expressed by the formula: ; ; In the formula, This represents the query feature vector after applying positional encoding. This represents the key feature vector after applying positional encoding. This indicates a rotation position encoding operation; To highlight information from important nodes, the key feature vectors after positional encoding are further weighted by importance, as expressed by the formula: ; In the formula, This represents the weighted key feature vector after applying positional encoding. Represents the normalization function. Indicates the weighting factor. Represents the height of the feature vector. Represents the width of the feature vector. Indicates the transpose operation; It should be noted that the height of the feature vector and width The aim is to The probability distribution value is adjusted to a value similar to the original environment and path feature values to avoid feature loss due to the feature values being too small during weighted summation.
[0030] To avoid the quadratic computation overhead of the matrix dot product of the query feature vector and the key feature vector in traditional attention, the traditional complete attention matrix calculation is split into two low-rank matrices, which are calculated separately, as expressed by the formula: ; In the formula, This represents the linear features output after the attention mechanism. Linear features output after attention mechanism and the output gated feature vector of the fused features Performing element-wise multiplication yields a fused feature incorporating both environmental and path information, expressed as a formula: ; In the formula, Indicates fusion features, This indicates an operation that changes the shape of the array. This indicates an element-wise multiplication operation.
[0031] S203. The path decision decoder adopts a fully connected network structure, which integrates the fused features output by the LRAM module. The input is fed into the path decision decoder, specifically: The fused features are processed through the first fully connected layer. Perform dimensionality reduction mapping to obtain intermediate decision features; To mitigate potential information loss during dimensionality reduction, the intermediate decision features are concatenated with the environmental features output by the environmental perception encoder to obtain a fused feature representation. Subsequently, the fused features are input into the second fully connected layer, which outputs a set of node guidance values consistent with the number of neighboring nodes of the current node. This set represents the relative priority of each candidate neighboring node being selected as the next-hop node, expressed by the formula: ; In the formula, Represents the set of node guidance values. Indicates the first guiding value. Indicates the second guiding value. Indicates the third guiding value; It should be noted that the guidance value is the evaluation value of each neighbor of the last vertex of the current search path by the path planning model; the larger the guidance value, the more likely the model believes that the adjacent node is located on the optimal path. The next hop node is selected based on the set of node guidance values. The node with the largest guidance value is taken as the next node of the path, thus realizing the search for a feasible path. S204. Based on the node guidance value set, construct a loss function system for model training, including the Softmax binary cross-entropy loss function and the shortest path ratio loss function, specifically: The set of node guidance values output by the path decision decoder is normalized using the Softmax function to obtain the predicted probability distribution of the corresponding candidate nodes. The predicted probabilities are then compared with the manually labeled optimal path node labels to calculate the binary cross-entropy loss, mitigating the impact of the imbalance between positive and negative samples on model training. This is expressed by the formula: ; In the formula, This represents the binary cross-entropy loss. The path decision decoder represents the first... The node guidance value output by each node. The path decision decoder represents the first... The node guidance value output by each node. Represents the set of candidate nodes; After the model continuously selects nodes and generates a complete predicted path, a shortest path ratio loss is introduced to measure the proportional relationship between the predicted path length and the corresponding shortest path length, in order to constrain redundant nodes generated during path planning. This is expressed by the formula: ; In the formula, This represents the shortest path ratio loss. The path length function is represented. Indicates the complete prediction path, This represents the optimal path marked by the manual annotation. Finally, the binary cross-entropy loss and the shortest path ratio loss are weighted and summed to obtain the model's total loss function, expressed by the formula: ; In the formula, This represents the total loss of the path planning model. Indicates the weighting parameter; In this embodiment, the weight parameter is set to 0.5.
[0032] Furthermore, based on the total loss of the path planning model, the Adam optimizer is used to update and train the path planning model. During training, the batch size is set to N, the initial learning rate is E, and the model is trained on an NVIDIA RTX 3090 GPU for a total of F epochs. The learning rate is dynamically adjusted using a cosine annealing strategy. At the same time, the backpropagation algorithm is used to update the network parameters of the environment-aware encoder, path memory encoder, and path decision decoder in the path planning model. This ensures that in subsequent training iterations, the node guidance values output by the model gradually approach the optimal path node distribution labeled by humans, and generate communication lines with path lengths closer to the shortest path.
[0033] S300. Based on the set of node guidance values, execute the following greedy search strategy: Select the candidate node with the largest indicator value as the next-hop node for the communication line; add this next-hop node to the end of the selected node sequence; determine whether this next-hop node is the target node. If it is the target node, stop the search and output the optimal path of the currently selected node sequence as the communication line; If the selected node is not the target node, the updated sequence of selected nodes is re-inputted into the path memory encoder and iterated until the target node is reached.
[0034] Example 2: This embodiment proposes an optimal path planning system for communication lines of multi-source heterogeneous data, including the following modules: Preprocessing module: acquires thematic maps of building function classification and communication node spatial data of the target area; preprocesses the thematic maps of building function classification, constructs an undirected graph based on the communication node spatial data, and obtains source nodes and target nodes according to the preset path planning task; The node prediction module takes the undirected graph and the preprocessed building function classification thematic map as input to the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features, and outputs environmental features after concatenation. It uses the source node as the initial node to construct a selected node sequence, which is then input to the pre-trained path planning model path memory encoder to obtain path features. Simultaneously, based on the connection relationships of the last node in the selected node sequence in the undirected graph, multiple candidate nodes are determined. The environmental features and path features are then input to the feature fusion module to obtain fused features, which are then input to the pre-trained path planning model path decision decoder to output the predicted probability score of each candidate node. Path result search module: Based on the predicted probability scores of candidate nodes, select the candidate node with the highest predicted probability score as the next node and add it to the selected node sequence. Input it into the pre-trained path planning model path memory encoder, and continue to iterate until the target node is reached, and output the optimal path of the communication line.
[0035] See Figure 4 As can be seen from the partial path planning results on the thematic map of building function types in a certain city in this embodiment, the path planning effect obtained by this method is consistent with the real labels, and the predicted path is smooth and has no node redundancy or looping phenomenon.
[0036] Example 3: This embodiment proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in any embodiment of the present invention.
[0037] Example 4: This embodiment proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.
[0038] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.
[0039] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0040] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0041] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0042] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for multi-source heterogeneous data communication line optimal path planning, characterized in that, Includes the following steps: Acquire the thematic map of building function classification and spatial data of communication nodes in the target area; The thematic map of building function classification is preprocessed, an undirected graph is constructed based on the spatial data of communication nodes, and the source node and target node are obtained according to the preset path planning task. The undirected graph and the preprocessed building function classification thematic map are input into the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features. After concatenation, the environment features are output. The source node is used as the initial node to construct a sequence of selected nodes, which is input into the pre-trained path planning model path memory encoder to obtain path features. At the same time, based on the connection relationship of the last node in the selected node sequence in the undirected graph, multiple candidate nodes to be selected are determined. The environment features and path features are input into the feature fusion module to obtain fused features, which are then input into the pre-trained path planning model path decision decoder to output the predicted probability score of each candidate node. Based on the predicted probability scores of candidate nodes, the candidate node with the highest predicted probability score is selected as the next node and added to the selected node sequence. It is then input into the pre-trained path planning model path memory encoder and iterated continuously until the target node is reached, outputting the optimal path of the communication line. The pre-trained path planning model includes an environment-aware encoder, a path memory encoder, and a path decision decoder. 2.The multi-source heterogeneous data communication line optimal path planning method according to claim 1, wherein, The method inputs an undirected graph and a preprocessed thematic map of building function classification into a pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features, specifically: Using a graph attention network, the attention weights between the current node and its neighboring nodes are calculated based on Euclidean distance and dispersion, and the neighboring nodes are initially aggregated. Using a graph convolutional network, a second convolution is performed on the basis of the initial aggregation to output the node topological features; The convolutional feature extractor in the environment-aware encoder is used to process the thematic map of building function classification: high-level semantic information of the thematic map of building function classification is extracted through multi-layer convolutional operations and downsampling operations, and features at different levels are fused through skip connections to output spatial semantic features; The node topological features and spatial semantic features are concatenated and processed by an activation function to obtain the environmental features.
3. The multi-source heterogeneous data communication line optimal path planning method according to claim 1, characterized in that, The method uses the source node as the initial node to construct a sequence of selected nodes, which is then input into a pre-trained path planning model, the path memory encoder, to obtain path features, specifically: The path memory encoder employs a long short-term memory network; The node features in the selected node sequence are sequentially input into the memory units of the Long Short-Term Memory (LSTM) network according to the access order. The LTM network's gating mechanism is used to perform temporal modeling of the sequence and update the hidden state. The final hidden state output by the LTM network is used as the path feature representing the path's historical information.
4. The optimal path planning method for communication lines of multi-source heterogeneous data according to claim 1, characterized in that, The feature fusion module is a low-rank attention module.
5. The optimal path planning method for communication lines of multi-source heterogeneous data according to claim 4, characterized in that, The method inputs environmental features and path features into a low-rank attention module to obtain fused features, specifically: One-dimensional convolution operations are performed on environmental features and path features respectively to generate query, key, value and output gated feature vectors corresponding to environmental features and path features; The query, key, value, and output gating feature vectors corresponding to environmental features and path features are concatenated to obtain the query, key, value, and output gating feature vectors of the fused features. The query and key feature vectors of the fused features are processed by activation function, a bias term is added, and rotational position encoding is applied to obtain query and key feature vectors with positional information; The key feature vectors with location information are weighted by importance to obtain the weighted key feature vectors of the fused features; The traditional attention matrix calculation is split into two low-rank matrices, and the linear features output by the attention mechanism are calculated, expressed by the formula: ； In the formula, This represents the linear features output after the attention mechanism. This represents the query feature vector after applying positional encoding. This represents the key feature vector after applying positional encoding. This represents the weighted key feature vector after applying positional encoding. Represents the height of the feature vector. Represents the width of the feature vector. This indicates the transpose operation. The feature vector represents the fused features; Linear features output after attention mechanism and the output gated feature vector of the fused features Performing element-wise multiplication yields a fused feature incorporating both environmental and path information, expressed as a formula: ； In the formula, Indicates fusion features, This indicates an operation that changes the shape of the array. This indicates an element-wise multiplication operation. The output gated feature vector represents the fused features. This indicates a fully connected operation.
6. The optimal path planning method for communication lines of multi-source heterogeneous data according to claim 1, characterized in that, The method fuses features, inputs them into a pre-trained path planning model path decision decoder, and outputs the predicted probability score of each candidate node, specifically: The path decision decoder includes a first fully connected layer and a second fully connected layer; The fused features are input into the first fully connected layer for dimensionality reduction to obtain intermediate features; The intermediate features are concatenated with the environmental features to obtain the fused features; The fused features are input into the second fully connected layer, and the output is the guidance value corresponding to each neighboring node of the current node; The guidance value is normalized to obtain the predicted probability score of each candidate node.
7. The optimal path planning method for communication lines of multi-source heterogeneous data according to claim 1, characterized in that, The pre-trained path planning model is obtained by training using the following loss function: The loss function includes binary cross-entropy loss and shortest path ratio loss, wherein: Binary cross-entropy loss is used to measure the difference between the predicted probability distribution of the model output and the manually labeled optimal path node labels; The shortest path ratio loss is used to measure the ratio between the length of the complete predicted path generated by the model and the length of the optimal path labeled by humans, thus constraining path redundancy. The total loss of the path planning model is the weighted sum of the shortest path ratio loss and the shortest path ratio loss.
8. A system for optimal path planning of communication lines for multi-source heterogeneous data, characterized in that, Includes the following modules: Preprocessing module: Acquires thematic maps of building function classification and spatial data of communication nodes in the target area; The thematic map of building function classification is preprocessed, an undirected graph is constructed based on the spatial data of communication nodes, and the source node and target node are obtained according to the preset path planning task. The node prediction module takes the undirected graph and the preprocessed building function classification thematic map as input to the pre-trained path planning model environment-aware encoder to obtain node topological features and spatial semantic features, and outputs environmental features after concatenation. It uses the source node as the initial node to construct a selected node sequence, which is then input to the pre-trained path planning model path memory encoder to obtain path features. Simultaneously, based on the connection relationships of the last node in the selected node sequence in the undirected graph, multiple candidate nodes are determined. The environmental features and path features are then input to the feature fusion module to obtain fused features, which are then input to the pre-trained path planning model path decision decoder to output the predicted probability score of each candidate node. Path result search module: Based on the predicted probability scores of candidate nodes, select the candidate node with the highest predicted probability score as the next node and add it to the selected node sequence. Input it into the pre-trained path planning model path memory encoder, and continue to iterate until the target node is reached, and output the optimal path of the communication line.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.