Cable engineering path intelligent planning method and system based on deep reinforcement learning

Abstract

This invention discloses an intelligent cable engineering path planning method and system based on deep reinforcement learning, belonging to the field of artificial intelligence technology. The method includes: acquiring multi-source engineering data of the planning area; constructing and partitioning the regional data source dependency graph to obtain multiple planning sub-regions; establishing the state space of each sub-region; inputting a deep reinforcement learning policy network and outputting search policy parameters for each region; synchronously switching the corresponding parameters and performing depth-first search when switching regions until the target cable engineering planning path is obtained. This invention solves the technical problem in existing cable engineering path planning methods where the single search algorithm has fixed parameters and lacks regional adaptability, resulting in low search efficiency and planning accuracy in large-scale heterogeneous terrain planning scenarios. It achieves efficient and adaptive dynamic search in large-scale cable engineering path planning, improving path planning efficiency and the quality of planning schemes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for intelligent planning of cable engineering routes based on deep reinforcement learning. Background Technology

[0002] In power engineering construction, cable route planning for 220kV and below urban transmission lines is a typical complex decision-making problem involving large scale, multiple constraints, and multiple objectives. With the accelerating pace of urbanization, factors such as dense underground pipelines, varied topography, and numerous environmentally sensitive areas make traditional planning methods relying on human experience and single algorithms insufficient to meet the ever-increasing demands for efficiency and accuracy. Existing path search methods often employ fixed-parameter depth-first search, A* algorithms, or genetic algorithms. These methods often suffer from low search efficiency, slow convergence speed, and susceptibility to local optima in large-scale planning areas. More importantly, geological conditions, underground pipeline distribution, and urban development plans vary significantly across different regions. Traditional methods typically use uniform search parameters, lacking dynamic adaptability to the characteristics of regional data sources. This leads to decreased search accuracy in areas with complex data sources and wasted computational resources in areas with sparse data sources. Therefore, how to adaptively adjust search strategy parameters based on heterogeneous engineering data from different regions to significantly improve the efficiency and robustness of large-scale cable route planning while maintaining search accuracy has become an urgent technical challenge. Summary of the Invention

[0003] This application provides a method and system for intelligent planning of cable engineering routes based on deep reinforcement learning. The key is that, in the context of large-scale cable engineering route planning scenarios, traditional single search algorithms suffer from low efficiency and lack of regional adaptability due to heterogeneous terrain and significant differences in the types of engineering data from multiple sources. The method uses a deep reinforcement learning policy network to adaptively generate search depth thresholds, branch expansion thresholds, and backtracking trigger thresholds for each planning sub-region. This is combined with a data processing flow that dynamically updates the search strategy parameters when dividing regions based on data source dependency graphs and switching sub-regions. Together, these methods significantly improve the path search efficiency, planning accuracy, and regional adaptability of depth-first search in large-scale cable engineering planning.

[0004] The first aspect of this application provides an intelligent cable engineering path planning method based on deep reinforcement learning, the method comprising: Acquire multi-source engineering data for the cable engineering path planning area; construct a regional data source dependency graph based on the multi-source engineering data, partition the planning area according to the data source dependency graph, and output multiple planning sub-regions; establish multiple regional state spaces corresponding to the multiple planning sub-regions; input the multiple regional state spaces into a deep reinforcement learning policy network, and output search strategy parameters for each planning sub-region to control the depth-first search process, wherein the search strategy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold; if it is detected that the current planning sub-region is switched to the next planning sub-region, switch the search strategy parameters corresponding to the next planning sub-region to perform a depth-first search until the target cable engineering planning path traversing the cable engineering path planning area is obtained.

[0005] A second aspect of this application provides an intelligent cable engineering path planning system based on deep reinforcement learning, the system comprising: Data Acquisition Module: Acquires multi-source engineering data for the cable engineering path planning area; Region Division Module: Constructs a regional data source dependency graph based on the multi-source engineering data, partitions the planning area according to the data source dependency graph, and outputs multiple planning sub-regions; State Space Establishment Module: Establishes multiple regional state spaces corresponding to the multiple planning sub-regions; Policy Generation Module: Inputs the multiple regional state spaces into a deep reinforcement learning policy network, outputs search policy parameters for each planning sub-region to control the depth-first search process, wherein the search policy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold; Depth Search Module: If it is detected that the current planning sub-region is switched to the next planning sub-region, the search policy parameters corresponding to the next planning sub-region are switched to execute a depth-first search until the target cable engineering planning path traversing the cable engineering path planning area is obtained.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multi-source engineering data is collected within the cable engineering planning area. A regional data source dependency graph is constructed based on the data source dependencies of different areas, thus dividing the large planning area into multiple planning sub-regions with different data characteristics. Subsequently, a corresponding regional state space is established for each planning sub-region, and a deep reinforcement learning network is used to learn the appropriate depth-first search control parameters for each sub-region, including search depth, branch expansion, and backtracking triggering. During the actual path search process, the planning sub-region to which the current search position belongs is identified in real time. When the search process moves from one sub-region to another, the search strategy parameters corresponding to the next sub-region are automatically switched. This allows the depth-first search to dynamically adjust the search method according to regional environmental differences, ultimately obtaining a target planning path covering the entire cable engineering path planning area. This enables the adoption of differentiated search strategies for different regions, improving the efficiency and adaptability of global path planning. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the intelligent cable engineering path planning method based on deep reinforcement learning provided in the embodiments of this application.

[0009] Figure 2 A schematic diagram of the structure of a cable engineering path intelligent planning system based on deep reinforcement learning provided in an embodiment of this application.

[0010] Figure 3 This is a schematic diagram illustrating the adaptive changes in search strategy parameters for different planning sub-regions provided in the embodiments of this application.

[0011] Figure labeling: Data acquisition module 11, region division module 12, state space establishment module 13, strategy generation module 14, depth search module 15. Detailed Implementation

[0012] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0013] Example 1, as Figure 1 As shown, this application provides an intelligent cable engineering path planning method based on deep reinforcement learning, wherein the method includes: Obtain multi-source engineering data for the cable engineering route planning area.

[0014] In this embodiment, the starting coordinates and ending coordinates of the cable project to be planned are used as a reference. The starting coordinates and ending coordinates are connected to obtain a reference line. A preset distance, such as 500m, is used on both sides of the reference line as a buffer range to generate the cable project path planning area. The cable project path planning area is then converted to a unified geographic coordinate system. The cable project path planning area is then divided into multiple spatial grid units according to a preset grid scale, such as 10m×10m. For each spatial grid cell, multi-source engineering data corresponding to that spatial grid cell is read from a unified cable engineering spatial data platform according to the spatial grid cell number. This includes topographic elevation data, geological survey data, hydrological data, land use data, underground pipeline data, road traffic data, and ecological protection data. Specifically, the ground elevation, slope, and topographic relief within the spatial grid cell are read as topographic elevation data; soil layer type, rock and soil bearing capacity, distribution of soft soil, and geological risk level within the spatial grid cell are read as geological survey data; groundwater depth, surface water body boundaries, drainage ditch locations, and water accumulation risk level within the spatial grid cell are read as hydrological data; and construction land within the spatial grid cell is read as hydrological data. Land use data includes land use type, green space type, agricultural land type, land use restriction level, and non-construction land boundary. Underground pipeline data includes the location, burial depth, and diameter of water supply and drainage pipelines, gas pipelines, communication pipelines, and existing cable pipelines within the spatial grid unit. Road traffic data includes the road centerline, road red line, road width, road grade, and traffic restrictions within the spatial grid unit. Ecological protection data includes the ecological protection red line, greening protection area, water source protection area, ecological restriction level, and environmentally sensitive area boundary within the spatial grid unit. These data are then used for field association and coordinate registration according to the spatial grid unit number, forming a multi-source engineering data record with the spatial grid unit as the smallest data carrier. Subsequently, coordinate consistency verification, spatial overlap verification, and field integrity verification are performed on each multi-source engineering data record. When coordinate deviations exist within the same spatial grid unit, the grid center point in the geographic coordinate system is used as a unified positioning benchmark for coordinate correction. When fields are missing, they are supplemented based on the average value of the same type of fields in eight adjacent spatial grid units. By uniformly collecting data from different sources, formats, and spatial scales, we can provide basic data support for subsequent regional division, state space establishment, and path search strategy generation.

[0015] Based on the multi-source engineering data, a regional data source dependency graph is constructed. The planning area is then partitioned according to the data source dependency graph, and multiple planning sub-regions are output.

[0016] In one embodiment, the cable engineering route planning area is divided into multiple initial spatial units, and multi-source engineering data related to cable route planning are extracted from each initial spatial unit. Since different spatial units are located in different engineering environments, their dependence on various data sources also varies. Therefore, the data source contribution of each initial spatial unit can be determined based on the influence of various engineering data on route feasibility assessment, obstacle avoidance, construction cost evaluation, and route risk identification. After obtaining the data source contribution, the dependence of each initial spatial unit on different data sources is converted into a data source dependency vector. By comparing the data source dependency vectors of different initial spatial units, the data source dependency similarity between spatial units is determined. Thus, the initial spatial units are used as graph nodes, and the data source dependency similarity relationships between different spatial units are used as graph edges to construct a regional data source dependency graph reflecting the data source dependencies within the planning area. Subsequently, the regional data source dependency graph is aggregated and analyzed, and initial spatial units with similar data source dependencies are grouped into the same type of region, ensuring that spatial units within the same planning sub-region have a relatively consistent data source system and engineering environment characteristics. Finally, the aggregated spatial units are used as corresponding planning sub-regions, resulting in multiple planning sub-regions. The above methods enable the division of planning areas to no longer rely solely on geographical location or human experience, but to adaptively partition the areas based on the actual differences in the dependence of different regions on multi-source engineering data, thus providing a foundation for generating differentiated search strategy parameters for different planning sub-regions.

[0017] Furthermore, the method for constructing a regional data source dependency map based on the multi-source engineering data includes: Obtain the initial spatial unit set of the planned area, and extract multi-source engineering data for each initial spatial unit based on the dependent data source type; analyze the data source contribution set of each initial spatial unit based on the multi-source engineering data; calculate the data source dependency vector of each initial spatial unit based on the data source contribution set; calculate the data source dependency similarity based on the data source dependency vectors of any two initial spatial units, and output the data source dependency similarity set; construct a regional data source dependency graph using the initial spatial unit set as the node set and the data source dependency similarity set as the dependency edge set.

[0018] Preferably, the cable engineering route planning area is divided into multiple initial spatial units according to a preset grid scale, denoted as the initial spatial unit set U. For each initial spatial unit Ui, corresponding multi-source engineering data is extracted in order of preset data source types, such as topographic elevation data, geological survey data, hydrological data, land use data, underground pipeline data, road traffic data, and ecological protection data. For the i-th initial spatial unit Ui, the contribution index of each type of data source to the cable route planning result within the initial spatial unit is calculated, resulting in the data source contribution set Ci, where cij represents the contribution of the j-th data source to the route planning result of the i-th initial spatial unit. During the calculation, the comprehensive route cost Pi of the initial spatial unit is first calculated based on all multi-source engineering data within the i-th initial spatial unit. This comprehensive route cost Pi is obtained by normalizing and weighting the construction difficulty value, route avoidance risk value, accessibility value, ecological restriction value, and engineering cost value of the spatial unit. Subsequently, the j-th data source is masked sequentially. That is, the data field corresponding to the j-th data source is replaced with the average value of the same type of data source in the planning area. Under the condition that the other data sources remain unchanged, the masked path cost Pij of the initial spatial unit is recalculated. The absolute value of the difference between the comprehensive path cost Pi and the masked path cost Pij is taken as the original contribution value aij of the j-th data source. Then, the original contribution values ​​of all data sources in the i-th initial spatial unit are normalized to obtain cij=aij / (ai1+ai2+...+aim), thus forming the data source contribution set Ci of the i-th initial spatial unit.

[0019] Next, based on the data source contribution set Ci, the data source dependency vector Di for each initial spatial unit is calculated. This data source dependency vector Di represents the dependency weight of the initial spatial unit on different data sources when planning cable engineering routes. Specifically, Di = (di1, di2, ..., dim), where Di represents the data source dependency vector of the i-th spatial unit, m represents the number of data sources, and dij represents the dependency weight of the i-th spatial unit on the j-th data source. A normalized contribution index can be selected as this dependency weight. Then, for any two initial spatial units Ui and Uk, the cosine similarity between their data source dependency vectors Di and Dk is calculated. This cosine similarity is taken as the data source dependency similarity sik. Then, any two initial spatial units in the initial spatial unit set are traversed to obtain the data source dependency similarity set S = {sik | i ≠ k}. Finally, a regional data source dependency graph G=(V,E) is constructed using the initial spatial unit set U as the node set and the data source dependency similarity set S as the dependency edge set. Here, V represents the node set, with each node corresponding to an initial spatial unit, and E represents the dependency edge set, with sik used as the edge weight. This yields a regional data source dependency graph that characterizes the data source dependencies between different initial spatial units, providing a graph structure foundation for subsequent partitioning of the planned area according to data source dependencies, thereby improving the matching between the partitioning results and the cable path search strategy.

[0020] Furthermore, the planning area is partitioned according to the data source dependency graph, and multiple planning sub-regions are output. The method includes: Graph clustering is performed on the data source dependency graph, and the initial spatial units with data source dependency similarity greater than a preset similarity threshold are aggregated to obtain a data source dependency aggregate graph. The aggregated spatial unit corresponding to each aggregated node in the data source dependency aggregate graph is used as the corresponding planning sub-region, and multiple planning sub-regions are output.

[0021] Optionally, the system first obtains the previously constructed regional data source dependency graph and a pre-set similarity threshold, which can be set to 0.85. The system traverses all dependency edges in the regional data source dependency graph. When the data source dependency similarity corresponding to a dependency edge is less than the similarity threshold, the dependency edge is deleted from the regional data source dependency graph. When the data source dependency similarity corresponding to a dependency edge is greater than or equal to 0.85, the dependency edge is retained, resulting in a threshold-filtered regional data source dependency graph. Subsequently, according to the spatial arrangement order of the initialized spatial units within the planning area, starting from the initialized spatial unit in the lower left corner, the system sequentially checks whether each node has been assigned a cluster number in the order from left to right and from bottom to top. If the current node has not been assigned a cluster number, the node is determined as the seed node of a new aggregation node. Starting from the seed node, a connectivity search is performed along the retained dependency edges. All nodes that can be connected to the seed node through the retained dependency edges are grouped into the same node set, and the node set is assigned the same cluster number, thus completing one node aggregation. The process continues iterating through nodes that have not yet been assigned cluster numbers until all nodes in the regional data source dependency graph have been assigned cluster numbers, resulting in multiple node aggregation sets. For any node aggregation set, the number of initial spatial units it contains is counted. If this number is less than the preset minimum number of aggregations, the average dependency edge weight between this node aggregation set and its adjacent node aggregation sets is calculated. This node aggregation set is then merged into the adjacent node aggregation set with the largest average dependency edge weight to avoid forming scattered planning areas that are too small and have insufficient engineering significance. After merging these small areas, each node aggregation set is treated as an aggregation node in the data source dependency aggregation graph, and the spatial extent of all initial spatial units within this node aggregation set is merged into one aggregation spatial unit. For two aggregation nodes, if their corresponding aggregation spatial units are adjacent on the spatial boundary, an aggregation dependency edge is established between the two aggregation nodes. The average data source dependency similarity between the initial spatial units corresponding to the adjacent boundaries of the two aggregation spatial units is used as the edge weight of the aggregation dependency edge, thereby generating the data source dependency aggregation graph. Then, each aggregation spatial unit corresponding to each aggregation node in the data source dependency aggregation graph is used as a planning sub-region. For each planning sub-region, a region number, region boundary coordinates, initialization spatial unit number, average data source dependency vector, dominant data source type, and adjacent planning sub-region number are generated. All planning sub-regions are then combined according to their spatial boundaries to output multiple planning sub-regions covering the cable engineering route planning area.By using the above method, initial spatial units with similar data source dependencies and spatial continuity can be aggregated into the same planning sub-region. This allows the planning region division results to reflect the differences in the dependencies of different regions on data sources such as topographic elevation, geological survey, hydrology, land use, underground pipelines, road traffic, and ecological protection. This provides an accurate regional basis for subsequently establishing differentiated regional state spaces and switching search strategy parameters for different planning sub-regions.

[0022] Establish multiple regional state spaces corresponding to the multiple planning sub-regions.

[0023] In one embodiment, multiple planning sub-regions from the aforementioned output are obtained, and the initial spatial unit number, region boundary coordinates, average data source dependency vector, and dominant data source type contained in each planning sub-region are read. For the r-th planning sub-region, the multi-source engineering data corresponding to all initial spatial units within the planning sub-region are extracted, and the data are normalized and quantified according to a preset state field order. In specific processing, firstly, all values ​​are normalized by maximizing and minimizing them to ensure they are all under the same dimension. Then, the average elevation and average slope of all initialized spatial units within the planning sub-region are calculated to obtain the topographic elevation status value. The weighted average of soil layer risk level, soil and rock bearing capacity level, and soft soil proportion is calculated to obtain the geological survey status value. The weighted average of groundwater depth risk level, surface water proportion, and water accumulation risk level is calculated to obtain the hydrological status value. The weighted average of construction land proportion, non-construction land proportion, and land use restriction level is calculated to obtain the land use status value. The normalized results of the total length of underground pipelines, the number of underground pipeline intersections, and the length occupied by existing cable ducts are calculated to obtain the underground pipeline status value. The weighted average of road width, road grade, and traffic restriction level is calculated to obtain the road traffic status value. The weighted average of the ecological protection red line proportion, environmentally sensitive area proportion, and ecological restriction level is calculated to obtain the ecological protection status value. Subsequently, these state values ​​are arranged in a preset order to form the environmental feature vector of the planning sub-region. Then, the data source dependency vectors of all initialized spatial units in the planning sub-region are weighted and averaged according to the area of ​​the spatial units to obtain the data source dependency feature vector of the planning sub-region. This vector is used to represent the dependency weight of the planning sub-region on topographic elevation data, geological survey data, hydrological data, land use data, underground pipeline data, road traffic data, and ecological protection data during the path planning process.

[0024] Next, the search history information of the planned sub-region is read, including the average number of search steps, average number of branch expansions, average number of backtrackings, average path cost, and average number of invalid searches during the historical training process. The search history information is then normalized to obtain a search history feature vector. If the planned sub-region has not yet undergone historical training, the search history feature vector is initialized to a zero vector. Finally, the environmental feature vector, data source dependency feature vector, and search history feature vector are concatenated in field order to generate the region state vector of the planned sub-region. This region state vector, along with the region number, region boundary coordinates, and adjacent planned sub-region numbers, is then encapsulated into the region state space of the planned sub-region. All planned sub-regions are processed sequentially in the above manner to obtain multiple region state spaces corresponding one-to-one with the multiple planned sub-regions. Through this method, the differences in engineering environment, data source dependency, and search history among different planned sub-regions can be uniformly expressed as state inputs that the deep reinforcement learning policy network can recognize. This provides a state basis for subsequently generating search policy parameters corresponding to each planned sub-region, thereby improving the targeting and adaptability of subsequent path search.

[0025] The state spaces of the multiple regions are input into a deep reinforcement learning policy network, and the search policy parameters for each planning sub-region are output to control the depth-first search process. The search policy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold.

[0026] In one embodiment, the state space corresponding to each planning sub-region is used as input to a deep reinforcement learning policy network. This network identifies the search environment and makes parameter decisions for different planning sub-regions. The deep reinforcement learning policy network encodes the input state space through a state encoding layer, converting environmental features, data source dependency features, and search history features into state-encoded features of a unified dimension. A policy feature extraction layer then analyzes these state-encoded features to extract key policy features influencing the depth-first search process, forming a regional search policy representation vector. Subsequently, a policy parameter output layer generates search policy parameters for each planning sub-region based on the regional search policy representation vector. These parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold. The search depth threshold limits the maximum number of layers that the depth-first search can continuously expand downwards within the current planning sub-region; the branch expansion threshold limits the number of candidate neighboring nodes that each search node is allowed to continue expanding to; and the backtracking trigger threshold determines whether to trigger backtracking when the cumulative cost, number of invalid expansions, or number of consecutive failures of the current path reach the threshold. By using the above method, each planning sub-region can obtain depth-first search control parameters that match its state characteristics, thereby improving the relevance and regional adaptability of the search parameter generation.

[0027] Furthermore, the multiple region state spaces are input into a deep reinforcement learning policy network, which includes a state encoding layer, a policy feature extraction layer, and a policy parameter output layer. The state encoding layer encodes features in each input region state space to obtain state-encoded features, where each region state space includes an environmental feature vector, a data source dependency feature vector, and a search history feature vector. The policy feature extraction layer performs deep feature extraction on the state-encoded features to obtain a region search policy representation vector. The policy parameter output layer outputs search policy parameters for each planning sub-region to control the depth-first search process based on the region search policy representation vector.

[0028] Preferably, the regional state space corresponding to each planning sub-region is read sequentially according to the region number of the planning sub-region. For the r-th planning sub-region, its regional state space includes environmental feature vector, data source dependency feature vector, and search history feature vector. The system first uses the state encoding layer of the deep reinforcement learning policy network to normalize the environmental feature vector, data source dependency feature vector, and search history feature vector respectively, so that each feature value is mapped to the numerical range of 0 to 1. The state encoding layer inputs the normalized environmental feature vector into the first fully connected encoding unit, the normalized data source dependency feature vector into the second fully connected encoding unit, and the normalized search history feature vector into the third fully connected encoding unit. The environmental encoding features, data source dependency encoding features, and search history encoding features are obtained by linear mapping and ReLU activation function respectively. Then, the environmental encoding features, data source dependency encoding features, and search history encoding features are concatenated in a preset order to obtain the state encoding features of the r-th planning sub-region. Subsequently, the state-encoded features are passed to the policy feature extraction layer. This layer consists of an attention-based hidden layer and a feedforward neural network. The state-encoded features are mapped to query vectors, key vectors, and value vectors, respectively, and then weighted and fused to obtain a multi-feature fusion vector. This multi-feature fusion vector is then input into the feedforward neural network to extract deep features influencing depth-first search control, thus obtaining a region search policy representation vector. This vector represents the search control tendency that the planned sub-region should adopt under the current engineering environment, data source dependencies, and historical search performance. Next, the region search policy representation vector is passed to the policy parameter output layer. This layer has three parameter output nodes, corresponding to the search depth threshold, branch expansion threshold, and backtracking trigger threshold, respectively. The layer performs non-linear activation and reinforcement learning on the original output values ​​of the three parameter output nodes to obtain a set of search policy parameters corresponding to the r-th planned sub-region. All planning sub-regions are processed sequentially in the manner described above, and the search strategy parameters for each planning sub-region are output to control the depth-first search process. This allows the search depth, the number of branch expansions, and the backtracking trigger conditions to adapt to different planning sub-regions, thereby improving the search efficiency and regional adaptability of large-scale cable engineering path planning.

[0029] Furthermore, the strategy feature extraction layer is used to perform deep feature extraction on the state encoding features, and the method for obtaining the region search strategy representation vector includes: The strategy feature extraction layer includes a hidden layer based on an attention mechanism and a connected feedforward neural network. The attention mechanism maps the environmental feature vector, data source dependency feature vector, and search history feature vector to a query matrix, a key matrix, and a value matrix, respectively, to obtain query vectors, key vectors, and value vectors. Feature association scores are calculated based on the correlation between the query vectors and key vectors. The feature association scores are weighted and fused based on the value vectors to output a multi-feature fusion vector. This multi-feature fusion vector is then input into the feedforward neural network to extract nonlinear features affecting the region search, thereby obtaining a region search strategy representation vector.

[0030] Optionally, the strategy feature extraction layer includes an attention-based hidden layer and a feedforward neural network connected to the hidden layer, used to extract deep features from the state-encoded features that influence the depth-first search strategy parameters of the current planning sub-region. For the r-th planning sub-region, firstly, the environment-encoded features, data source dependency-encoded features, and search history-encoded features output by its state-encoded layer are obtained, and the three types of encoded features are arranged into a feature sequence according to a preset order. Subsequently, the strategy feature extraction layer performs linear mapping on the feature sequence through an attention mechanism to obtain the query vector, key vector, and value vector. Specifically, environmental encoding features, data source dependency encoding features, and search history encoding features are input into shared query mapping matrices, key mapping matrices, and value mapping matrices to obtain query vector Q, key vector K, and value vector V, respectively. Then, feature association scores are calculated based on the dot product correlation between query vector Q and key vector K, and these scores are normalized to obtain attention weights among different features. Features more relevant to the current regional search strategy receive greater weights. For example, when a planning sub-region has a high underground pipeline density and a large number of historical backtrackings, the proportion of underground pipeline-related environmental features and backtracking-related search history features in the attention weights increases. Next, the value vector V is weighted and fused according to the attention weights to output a multi-feature fusion vector. This multi-feature fusion vector represents the comprehensive influence relationship between environmental conditions, data source dependency, and search history performance. The multi-feature fusion vector is then input into a feedforward neural network for nonlinear feature extraction to obtain a regional search strategy representation vector. This regional search strategy representation vector is then passed to the strategy parameter output layer for subsequent generation of search depth thresholds, branch expansion thresholds, and backtracking trigger thresholds.

[0031] For example, a feedforward neural network can employ a three-layer fully connected structure, including an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer's dimension is consistent with the dimension of the multi-feature fusion vector, such as 64 dimensions. The first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the output layer has 32 neurons, outputting a 32-dimensional region search strategy representation vector. ReLU activation functions are connected after the first and second hidden layers, and a Dropout ratio of 0.1 is set during training to reduce the risk of overfitting. The training process of the feedforward neural network is synchronized with the overall training process of the deep reinforcement learning policy network. During training, the region search policy representation vector is input into the policy parameter output layer to obtain the candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold for the current planning sub-region. This set of candidate search policy parameters is then used to control the depth-first search process, and the reward value is calculated based on the path cost, search time, number of backtracking attempts, and number of invalid expansion attempts. When the reward value increases, it indicates that the region search policy representation extracted by the current feedforward neural network is more beneficial to the search process; when the reward value decreases, the attention mechanism and weight parameters in the feedforward neural network are adjusted through backpropagation. This training process is repeated until the average reward value change over multiple training rounds is less than a preset convergence threshold, resulting in the trained policy feature extraction layer. Through this process, the policy feature extraction layer can extract the nonlinear representations most relevant to search control from multiple state features of different planning sub-regions, improving the accuracy and regional adaptability of subsequent search policy parameter outputs.

[0032] Furthermore, the strategy parameter output layer outputs search strategy parameters for each planning sub-region to control the depth-first search process based on the region search strategy representation vector, the method including: The region search strategy representation vector is input into the strategy parameter output layer for nonlinear activation to obtain candidate search strategy parameters, including candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold. The search performance metrics for each planning sub-region performing depth-first search under the corresponding candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold are trained, including path cost, search time, and backtracking error. The path cost, search time, and backtracking error are used to perform reinforcement learning on the strategy parameter output layer to output optimized search strategy parameters.

[0033] Optionally, for the r-th planning sub-region, the region search strategy representation vector output by the strategy feature extraction layer is obtained. This region search strategy representation vector has been fused with environmental features, data source dependency features, and search history features through a self-attention mechanism, and can represent the dominant factors that have the greatest impact on depth-first search in this planning sub-region. Specifically, when the attention weights corresponding to terrain slope and topographic relief are high, the planning sub-region is determined to be a terrain-constrained dominant region; when the attention weights corresponding to underground pipeline density and pipeline intersection number are high, the planning sub-region is determined to be an underground exploration-constrained dominant region; and when the attention weights corresponding to the proportion of ecological protection red lines and the proportion of environmentally sensitive areas are high, the planning sub-region is determined to be an ecologically restricted dominant region. Subsequently, the region search strategy representation vector is input into the strategy parameter output layer. This strategy parameter output layer adopts a three-way parallel output structure. The first path outputs the candidate search depth threshold, the second path outputs the candidate branch expansion threshold, and the third path outputs the candidate backtracking trigger threshold. Each output is first processed through a fully connected mapping, then normalized by a Sigmoid nonlinear activation function, and finally inversely normalized and rounded according to a preset parameter range to obtain the candidate search strategy parameters. For example, the candidate search depth threshold is limited to between 3 and 50, the candidate branch expansion threshold is limited to between 1 and 8, and the candidate backtracking trigger threshold is limited to between 2 and 20. The candidate search depth threshold is used to limit the maximum number of layers that the depth-first search can continuously delve into within the current planned sub-region; the candidate branch expansion threshold is used to limit the number of candidate neighboring nodes that each search node can expand to; and the candidate backtracking trigger threshold is used to limit the number of consecutive invalid expansions, the number of times the path cost exceeds the limit, or the number of search failures to reach the corresponding conditions to trigger backtracking.

[0034] Subsequently, the candidate search strategy parameters corresponding to the r-th planning sub-region are written into the depth-first search controller, and a simulated depth-first search is performed within this planning sub-region, using the entrance boundary node as the search starting point and the exit boundary node as the search target. During the search process, the depth-first search controller limits the maximum search level of the search tree according to the candidate search depth threshold, selects several nodes with lower path cost values ​​from the adjacent candidate nodes of the current node for expansion according to the candidate branch expansion threshold, and determines whether to backtrack from the current search path to the previous search node according to the candidate backtracking trigger threshold. After completing one simulated search, the path cost, search time, and backtracking error within the planning sub-region are recorded. The path cost is obtained by summing the construction difficulty value, underground pipeline avoidance value, ecological restriction value, and passage cost value corresponding to the initial spatial unit traversed by the search path; the search time is obtained by the computation time consumed from the start of the search to obtaining a feasible sub-path; and the backtracking error is calculated by the difference between the actual number of backtracking attempts and the historical best number of backtracking attempts for this planning sub-region. For the same planning sub-region, multiple rounds of simulation search are repeatedly performed to obtain the average path cost, average search time, and average backtracking error of the planning sub-region under the current candidate search strategy parameters, and these are used as search performance indicators.

[0035] Then, action reward values ​​are constructed based on search performance metrics. The lower the path cost, the shorter the search time, and the smaller the backtracking error, the larger the action reward value. Candidate search depth thresholds, candidate branch expansion thresholds, and candidate backtracking trigger thresholds are then used as actions in deep reinforcement learning. The regional state space of the planned sub-region is used as the state, and the action reward value is used as the reward signal for backpropagation updates of the policy parameter output layer. Through this reinforcement learning process, the policy parameter output layer can utilize the dominant constraint information extracted by the self-attention mechanism to suppress redundant information interference from heterogeneous data sources, highlighting the key data source characteristics that are decisive for the current planned sub-region. This results in the output of optimized search depth thresholds, optimized branch expansion thresholds, and optimized backtracking trigger thresholds that match the planned sub-region. This resolves the contradiction between information redundancy and key information overload in heterogeneous data sources in large-scale cable planning across different regions, making depth-first search parameters regionally specific and improving path search efficiency, search accuracy, and adaptability to complex regions.

[0036] like Figure 3 As shown, the adaptive changes of search strategy parameters for different planning sub-regions are illustrated. The search depth threshold, branch expansion threshold, and backtracking trigger threshold for each planning sub-region show differentiated adjustment trends as the region number changes. This indicates that the proposed method can dynamically configure depth-first search parameters according to the environmental status and data source dependency characteristics of different planning sub-regions, thereby improving the regional adaptability of the search strategy.

[0037] Furthermore, the strategy parameter output layer is subjected to reinforcement learning based on the path cost, search time, and backtracking error to output optimized search strategy parameters. The method includes: Obtain an input action space, which includes an action space based on candidate search strategy parameters; calculate the action reward value for each candidate search strategy parameter based on the path cost, search time, and backtracking error in the input action space; perform backpropagation to update the network weights of the strategy parameter output layer with the goal of maximizing the action reward value, and output optimized search strategy parameters.

[0038] Optionally, for the r-th planning sub-region, the candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold initially output by the strategy parameter output layer are obtained, and these three candidate search strategy parameters are used as the central parameters of the action space. Subsequently, a preset depth step is extended before and after the candidate search depth threshold, a preset branch step is extended before and after the candidate branch expansion threshold, and a preset backtracking step is extended before and after the candidate backtracking trigger threshold, forming an input action space composed of multiple sets of search strategy parameters. For example, if the candidate search depth threshold is 20, the candidate branch expansion threshold is 4, and the candidate backtracking trigger threshold is 6, then the search depth thresholds {18, 20, 22}, the branch expansion thresholds {3, 4, 5}, and the backtracking trigger thresholds {5, 6, 7} are combined to obtain 27 sets of candidate search strategy parameters, each set of candidate search strategy parameters being considered as an action. Meanwhile, the search depth threshold is limited to between 3 and 50, the branch expansion threshold is limited to between 1 and 8, and the backtracking trigger threshold is limited to between 2 and 20. Action parameters that exceed the boundary are directly truncated to the corresponding boundary value, thereby obtaining the input action space corresponding to the r-th planning sub-region.

[0039] Next, each set of candidate search strategy parameters in the input action space is selected sequentially, and this set of candidate search strategy parameters is written into the depth-first search controller. The depth-first search controller is then controlled to perform simulated path search in the r-th planning sub-region. During the search process, the maximum number of downward expansion layers of the search tree is limited according to the search depth threshold in this set of parameters, the number of candidate adjacent nodes that each search node can continue to expand to is limited according to the branch expansion threshold, and the backtracking trigger threshold is used to determine whether the current search path triggers backtracking. After completing a search, record the path cost, search time, and backtracking error corresponding to the parameters of the candidate search strategy. Then, normalize the path cost, search time, and backtracking error respectively, and calculate the action reward value according to the reward function. The action with the lower path cost, shorter search time, and smaller backtracking error will receive a larger reward value. The action reward value R can be set as R=1-αC-βT-γE, where C represents the normalized path cost, T represents the normalized search time, E represents the normalized backtracking error, and α, β, and γ represent the weights of path cost, search time, and backtracking error respectively, and α+β+γ=1.

[0040] Then, the network weights of the policy parameter output layer are updated via backpropagation with the objective of maximizing the action reward value. Specifically, the region search policy representation vector of the r-th planning sub-region is used as the input to the policy parameter output layer, each candidate search policy parameter in the input action space is used as an optional action, and the action reward value corresponding to each optional action is used as the reinforcement learning feedback signal. The policy parameter output layer outputs the selection probability of each optional action based on the current network weights, and selects the candidate search policy parameter with the largest action reward value as the positive optimization objective. During training, the difference between the action reward value and the average reward value of the current planning sub-region is calculated as the advantage value. When the advantage value of a candidate search policy parameter is positive, the output probability of the policy parameter output layer for that candidate search policy parameter is increased; when the advantage value of a candidate search policy parameter is negative, the output probability of the policy parameter output layer for that candidate search policy parameter is decreased. Backpropagation is then used to update the weight parameters in the policy parameter output layer corresponding to the search depth threshold, branch expansion threshold, and backtracking trigger threshold. The process of action space construction, simulated search, reward value calculation, and network weight update is repeated until the average change in action reward value over multiple training rounds is less than a preset convergence threshold, or the preset maximum number of training rounds is reached. The optimal action for the current planning sub-region is then determined, and the search depth threshold, branch expansion threshold, and backtracking trigger threshold of this optimal action are output as optimized search strategy parameters. This approach allows the strategy parameter output layer to no longer rely solely on fixed empirical parameters for search control values. Instead, it optimizes the search strategy parameters based on the actual search performance of candidate search strategy parameters within each planning sub-region through reinforcement learning, thereby obtaining search depth, branch expansion, and backtracking control parameters that are more suitable for the current regional engineering environment and data source dependency characteristics.

[0041] Furthermore, after outputting the optimized search strategy parameters, the method also includes: Obtain the search strategy parameters of the first planning sub-region optimized from the plurality of planning sub-regions; initialize the candidate search strategy parameters of adjacent planning sub-regions based on the search strategy parameters.

[0042] Optionally, the planning sub-region where the search starting point is located is determined based on the starting point coordinates of the target cable engineering planning path, and this planning sub-region is taken as the first planning sub-region. If the search starting point is located on the boundary between two planning sub-regions, the planning sub-region containing the search starting point and having the smallest angle with the endpoint direction is determined as the first planning sub-region. Subsequently, the region state space corresponding to the first planning sub-region is read, and the region state space is input into the trained deep reinforcement learning policy network to obtain the candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold for the first planning sub-region. The path cost, search time, and backtracking error are calculated by simulating depth-first search, and then the candidate search policy parameters are optimized by reinforcement learning based on the principle of maximizing action reward value to obtain the optimized search policy parameters for the first planning sub-region, which are denoted as the first search policy parameters. Subsequently, based on the data source dependency aggregation graph, adjacent planning sub-regions with aggregated dependency edges to the first planning sub-region are read, and the aggregated dependency edge weights between the first planning sub-region and each adjacent planning sub-region are obtained. When the aggregated dependency edge weight between an adjacent planning sub-region and the first planning sub-region is greater than a preset migration threshold, the first search strategy parameter optimized from the first planning sub-region is used as the initial value of the candidate search strategy parameter for that adjacent planning sub-region. When the aggregated dependency edge weight is not greater than the preset migration threshold, the region state space of the adjacent planning sub-region itself is read first, and the first search strategy parameter is corrected based on the difference in data source dependency vector between the adjacent planning sub-region and the first planning sub-region, and then used as the candidate search strategy parameter for that adjacent planning sub-region. In specific adjustments, if the underground pipeline dependency weight of an adjacent planning sub-region is higher than that of the first planning sub-region, the candidate branch expansion threshold is increased and the candidate backtracking trigger threshold is decreased; if the terrain elevation dependency weight of an adjacent planning sub-region is higher than that of the first planning sub-region, the candidate search depth threshold is increased and the candidate branch expansion threshold is decreased; if the ecological protection dependency weight of an adjacent planning sub-region is higher than that of the first planning sub-region, the candidate backtracking trigger threshold is decreased, so that the search process triggers backtracking earlier when approaching confined spatial units. After initialization, the candidate search strategy parameters corresponding to each adjacent planning sub-region are written into the strategy parameter cache table, and the candidate search strategy parameters are used as the initial actions for subsequent reinforcement learning optimization of that adjacent planning sub-region. Through the above process, the search experience that has been optimized in the first planning sub-region can be transferred to adjacent planning sub-regions with similar data source dependencies, reducing the number of times adjacent planning sub-regions need to start searching and training from scratch, improving the efficiency of search strategy parameter generation, and maintaining the continuity of search strategy switching between different planning sub-regions.

[0043] If it is detected that the current planning sub-region has switched to the next planning sub-region, the search strategy parameters corresponding to the next planning sub-region are switched to perform a depth-first search until the target cable engineering planning path that traverses the cable engineering path planning area is obtained.

[0044] In one embodiment, the starting coordinates of the cable engineering path planning area are used as the search starting point. The planning sub-region where the search starting point is located is determined, and the search depth threshold, branch expansion threshold, and backtracking trigger threshold corresponding to this planning sub-region are read and written into the depth-first search controller. During the depth-first search process, the spatial coordinates of the current search node are obtained in real time, and the planning sub-region to which the current search node belongs is determined based on the spatial coordinates. If the current search node is still within the original planning sub-region, node expansion, depth restriction, and backtracking judgment continue according to the original search strategy parameters. If the current search node enters the next planning sub-region, the search strategy parameters corresponding to the next planning sub-region are read, and the original search depth threshold, branch expansion threshold, and backtracking trigger threshold in the depth-first search controller are overwritten, while the already formed search path and accumulated path value are retained. Subsequently, the current search node that has entered the next planning sub-region is used as the continuing search node, and the depth-first search continues to be executed according to the updated search strategy parameters until the search node reaches the endpoint coordinates, obtaining the target cable engineering planning path. In this way, the search parameters can be dynamically switched as the planning sub-region changes, improving the efficiency and regional adaptability of large-scale cable engineering path search.

[0045] Furthermore, switching the search strategy parameters for the next planned sub-region to perform a depth-first search can be achieved through methods including: The search starting point of the cable engineering path planning area is initialized; the initial planning sub-region where the search starting point is located is identified, and the candidate group search strategy parameters corresponding to the initial planning sub-region are written into the boundary register of the depth-first search controller; the search space coordinates are detected in real time, and if the search space coordinates are not in the initial planning sub-region, the next planning sub-region where the search space coordinates are located and the corresponding second set of search strategy parameters are extracted; the boundary register of the depth-first search controller is overwritten with the candidate group search strategy parameters by the second set of search strategy parameters.

[0046] Preferably, the coordinates of the cable access starting point, predetermined in the cable engineering route planning task, are read and mapped to the aforementioned set of initialized spatial units. The initialized spatial unit closest to the cable access starting point coordinates and meeting the passability conditions is determined and used as the search starting point for depth-first search. Simultaneously, search spatial coordinates, a starting node number, an initial value of the cumulative path cost, and a visited identifier are generated for the search starting point. The cumulative path cost, search depth count, and consecutive invalid expansion count are initialized to 0. Subsequently, the search spatial coordinates of the search starting point are spatially matched with the boundary coordinates of multiple planning sub-regions to determine the planning sub-region containing the search starting point. This planning sub-region is recorded as the initial planning sub-region. If the search starting point is located on the common boundary of two planning sub-regions, the distances from the search starting point to the exit boundary nodes within each candidate planning sub-region are compared, and the planning sub-region with the shorter distance is determined as the initial planning sub-region.

[0047] After determining the initial planning sub-region, the candidate group search strategy parameters corresponding to the initial planning sub-region are read from the strategy parameter cache table, and the candidate group search strategy parameters are written into the boundary register of the depth-first search controller. The boundary register is used to store the search control parameters that are effective for the current planning sub-region in real time during the depth-first search process. This allows the depth-first search controller to control the maximum continuous expansion level of the search tree according to the currently written search depth threshold, control the number of candidate adjacent nodes that each search node is allowed to expand according to the currently written branch expansion threshold, and control the backtracking operation when continuous invalid expansion reaches the threshold according to the currently written backtracking trigger threshold. Subsequently, during the depth-first search execution, whenever a search node moves from the current initialization spatial unit to the next initialization spatial unit, the search space coordinates corresponding to that next initialization spatial unit are read in real time. These search space coordinates are then matched with the boundary coordinates of the currently recorded initial planning sub-region. If the search space coordinates are still within the initial planning sub-region, the candidate group search strategy parameters in the boundary register remain unchanged, and the depth-first search continues. If the search space coordinates are not within the initial planning sub-region, the planning sub-region where the search space coordinates are located is re-queried based on the search space coordinates. The queried planning sub-region is determined as the next planning sub-region, and the second set of search strategy parameters corresponding to the next planning sub-region is retrieved from the strategy parameter cache table. Then, the depth-first search controller pauses new branch expansion operations, retains the currently formed search path, the set of visited nodes, the cumulative path cost, and the current search node number. The candidate group search strategy parameters previously stored in the boundary register are cleared, and the second set of search strategy parameters is written into the boundary register, overwriting the candidate group search strategy parameters. After overwriting, the initialization spatial unit corresponding to the current search space coordinates is used as the local search starting point within the next planning sub-region, and the depth-first search continues. By using the above method, when the search path crosses different planning sub-regions, only the boundary register parameters of the depth-first search controller are updated, without resetting the searched paths and accumulated costs. This ensures the continuity of path search and allows the search depth, branch expansion, and backtracking trigger conditions to dynamically switch according to the data source dependency characteristics of the planning sub-regions, thereby improving the search efficiency and regional adaptability of large-scale cable engineering path planning.

[0048] In summary, the embodiments of this application have at least the following technical effects: First, multi-source engineering data for the cable engineering route planning area is acquired. Then, a regional data source dependency graph is constructed based on the multi-source engineering data. The planning area is partitioned according to the data source dependency graph, outputting multiple planning sub-regions. Next, multiple regional state spaces corresponding to the multiple planning sub-regions are established. Then, the multiple regional state spaces are input into a deep reinforcement learning policy network, outputting search policy parameters for each planning sub-region to control the depth-first search process. These search policy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold. Finally, if a switch to the next planning sub-region is detected, the search policy parameters corresponding to the next planning sub-region are switched, and a depth-first search is performed until the target cable engineering planning path traversing the cable engineering route planning area is obtained. This invention addresses the technical problem in existing cable engineering route planning methods where fixed parameters of a single search algorithm and a lack of regional adaptability lead to low search efficiency and planning accuracy in large-scale heterogeneous terrain planning scenarios. It achieves the technical effect of improving route planning efficiency and planning scheme quality by adaptively partitioning the planning area based on multi-source engineering data and dynamically switching the search strategy parameters corresponding to each sub-region during the depth-first search process, thus realizing efficient and adaptive dynamic search in large-scale cable engineering route planning.

[0049] Example 2, based on the same inventive concept as the cable engineering path intelligent planning method based on deep reinforcement learning in the previous examples, such as... Figure 2 As shown, this application provides an intelligent cable engineering path planning system based on deep reinforcement learning, wherein the system includes: Data acquisition module 11: Acquires multi-source engineering data for the cable engineering path planning area; Region division module 12: Constructs a regional data source dependency graph based on the multi-source engineering data, partitions the planning area according to the data source dependency graph, and outputs multiple planning sub-regions; State space establishment module 13: Establishes multiple regional state spaces corresponding to the multiple planning sub-regions; Policy generation module 14: Inputs the multiple regional state spaces into a deep reinforcement learning policy network, outputs search policy parameters for each planning sub-region to control the depth-first search process, wherein the search policy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold; Depth search module 15: If it is detected that the current planning sub-region is switched to the next planning sub-region, the search policy parameters corresponding to the next planning sub-region are switched to execute a depth-first search until the target cable engineering planning path that traverses the cable engineering path planning area is obtained.

[0050] Furthermore, the region division module 12 is used to perform the following method: Obtain the initial spatial unit set of the planned area, and extract multi-source engineering data for each initial spatial unit based on the dependent data source type; analyze the data source contribution set of each initial spatial unit based on the multi-source engineering data; calculate the data source dependency vector of each initial spatial unit based on the data source contribution set; calculate the data source dependency similarity based on the data source dependency vectors of any two initial spatial units, and output the data source dependency similarity set; construct a regional data source dependency graph using the initial spatial unit set as the node set and the data source dependency similarity set as the dependency edge set.

[0051] Furthermore, the region division module 12 is used to perform the following method: Graph clustering is performed on the data source dependency graph, and the initial spatial units with data source dependency similarity greater than a preset similarity threshold are aggregated to obtain a data source dependency aggregate graph. The aggregated spatial unit corresponding to each aggregated node in the data source dependency aggregate graph is used as the corresponding planning sub-region, and multiple planning sub-regions are output.

[0052] Furthermore, the strategy generation module 14 is used to perform the following method: The multiple region state spaces are input into a deep reinforcement learning policy network, which includes a state encoding layer, a policy feature extraction layer, and a policy parameter output layer. The state encoding layer encodes features in each input region state space to obtain state-encoded features, where each region state space includes an environmental feature vector, a data source dependency feature vector, and a search history feature vector. The policy feature extraction layer performs deep feature extraction on the state-encoded features to obtain a region search policy representation vector. The policy parameter output layer outputs search policy parameters for each planning sub-region to control the depth-first search process based on the region search policy representation vector.

[0053] Furthermore, the strategy generation module 14 is used to perform the following method: The strategy feature extraction layer includes a hidden layer based on an attention mechanism and a connected feedforward neural network. The attention mechanism maps the environmental feature vector, data source dependency feature vector, and search history feature vector to a query matrix, a key matrix, and a value matrix, respectively, to obtain query vectors, key vectors, and value vectors. Feature association scores are calculated based on the correlation between the query vectors and key vectors. The feature association scores are weighted and fused based on the value vectors to output a multi-feature fusion vector. This multi-feature fusion vector is then input into the feedforward neural network to extract nonlinear features affecting the region search, thereby obtaining a region search strategy representation vector.

[0054] Furthermore, the strategy generation module 14 is used to perform the following method: The region search strategy representation vector is input into the strategy parameter output layer for nonlinear activation to obtain candidate search strategy parameters, including candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold. The search performance metrics for each planning sub-region performing depth-first search under the corresponding candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold are trained, including path cost, search time, and backtracking error. The path cost, search time, and backtracking error are used to perform reinforcement learning on the strategy parameter output layer to output optimized search strategy parameters.

[0055] Furthermore, the strategy generation module 14 is used to perform the following method: Obtain an input action space, which includes an action space based on candidate search strategy parameters; calculate the action reward value for each candidate search strategy parameter based on the path cost, search time, and backtracking error in the input action space; perform backpropagation to update the network weights of the strategy parameter output layer with the goal of maximizing the action reward value, and output optimized search strategy parameters.

[0056] Furthermore, the strategy generation module 14 is used to perform the following method: Obtain the search strategy parameters of the first planning sub-region optimized from the plurality of planning sub-regions; initialize the candidate search strategy parameters of adjacent planning sub-regions based on the search strategy parameters.

[0057] Furthermore, the depth search module 15 is used to perform the following methods: The search starting point of the cable engineering path planning area is initialized; the initial planning sub-region where the search starting point is located is identified, and the candidate group search strategy parameters corresponding to the initial planning sub-region are written into the boundary register of the depth-first search controller; the search space coordinates are detected in real time, and if the search space coordinates are not in the initial planning sub-region, the next planning sub-region where the search space coordinates are located and the corresponding second set of search strategy parameters are extracted; the boundary register of the depth-first search controller is overwritten with the candidate group search strategy parameters by the second set of search strategy parameters.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent cable engineering path planning based on deep reinforcement learning, characterized in that, The method includes: Obtain multi-source engineering data for the cable engineering route planning area; A regional data source dependency graph is constructed based on the multi-source engineering data. The planning area is then partitioned according to the data source dependency graph, and multiple planning sub-regions are output. Establish multiple regional state spaces corresponding to the multiple planning sub-regions; The state spaces of the multiple regions are input into a deep reinforcement learning policy network, and the search policy parameters for each planning sub-region are output to control the depth-first search process. The search policy parameters include at least a search depth threshold, a branch expansion threshold, and a backtracking trigger threshold. If it is detected that the current planning sub-region has switched to the next planning sub-region, the search strategy parameters corresponding to the next planning sub-region are switched to perform a depth-first search until the target cable engineering planning path that traverses the cable engineering path planning region is obtained.

2. The method as described in claim 1, characterized in that, Switch the search strategy parameters for the next planning sub-region and perform a depth-first search. Methods include: Initialize the search starting point for the cable engineering path planning area; Identify the initial planning sub-region where the search starting point is located, and write the candidate group search strategy parameters corresponding to the initial planning sub-region into the boundary register of the depth-first search controller; Real-time detection of search space coordinates; if the search space coordinates are not in the initial planning sub-region, extract the next planning sub-region where the search space coordinates are located and the corresponding second set of search strategy parameters. The boundary register of the depth-first search controller overrides the candidate group search strategy parameters with the second set of search strategy parameters.

3. The method as described in claim 1, characterized in that, The method for constructing a regional data source dependency graph based on the multi-source engineering data includes: Obtain the initial spatial unit set of the planned area, and extract multi-source engineering data for each initial spatial unit based on the dependent data source type; The data source contribution set of each initial spatial unit is analyzed based on the multi-source engineering data. Calculate the data source dependency vector of each initial spatial unit based on the data source contribution set, calculate the data source dependency similarity based on the data source dependency vectors of any two initial spatial units, and output the data source dependency similarity set. A regional data source dependency graph is constructed using the initial spatial unit set as the node set and the data source dependency similarity set as the dependency edge set.

4. The method as described in claim 3, characterized in that, The planning area is partitioned according to the data source dependency graph, and multiple planning sub-regions are output. The method includes: Graph clustering is performed on the data source dependency graph, and the initial spatial units with data source dependency similarity greater than a preset similarity threshold are aggregated into nodes to obtain a data source dependency aggregated graph. Using the aggregation spatial unit corresponding to each aggregation node in the data source dependency aggregation graph as the corresponding planning sub-region, multiple planning sub-regions are output.

5. The method as described in claim 1, characterized in that, The state spaces of the multiple regions are input into a deep reinforcement learning policy network, which includes a state encoding layer, a policy feature extraction layer, and a policy parameter output layer. The state coding layer is used to encode the features of each region's state space to obtain state coding features. Each region's state space includes an environmental feature vector, a data source dependency feature vector, and a search history feature vector. The strategy feature extraction layer is used to perform deep feature extraction on the state encoding features to obtain the region search strategy representation vector; The strategy parameter output layer outputs search strategy parameters for each planning sub-region to control the depth-first search process based on the region search strategy representation vector.

6. The method as described in claim 5, characterized in that, The strategy feature extraction layer is used to perform deep feature extraction on the state encoding features, and the method for obtaining the region search strategy representation vector includes: The strategy feature extraction layer includes an attention-based hidden layer and a connected feedforward neural network; The attention mechanism is used to map the environmental feature vector, the data source dependency feature vector, and the search history feature vector to the query matrix, the key matrix, and the value matrix, respectively, to obtain the query vector, the key vector, and the value vector. Calculate the feature association score based on the correlation between the query vector and the key vector; The feature association scores are weighted and fused based on the value vector to output a multi-feature fusion vector. The multi-feature fusion vector is input into the feedforward neural network to extract nonlinear features affecting the region search, thereby obtaining the region search strategy representation vector.

7. The method as described in claim 5, characterized in that, The strategy parameter output layer outputs search strategy parameters for each planning sub-region to control the depth-first search process based on the region search strategy representation vector, and the method includes: The region search strategy representation vector is input into the strategy parameter output layer for nonlinear activation to obtain candidate search strategy parameters, including candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold. Train the search performance metrics for each planning sub-region to perform depth-first search under the corresponding candidate search depth threshold, candidate branch expansion threshold, and candidate backtracking trigger threshold, including path cost, search time, and backtracking error; The path cost, search time, and backtracking error are used to perform reinforcement learning on the policy parameter output layer to output optimized search policy parameters.

8. The method as described in claim 7, characterized in that, The strategy parameter output layer is subjected to reinforcement learning based on the path cost, search time, and backtracking error to output optimized search strategy parameters. The method includes: Obtain the input action space, which includes the action space based on candidate search strategy parameters; Calculate the action reward value for each candidate search strategy parameter in the input action space based on the path cost, search time, and backtracking error; The network weights of the policy parameter output layer are updated by backpropagation with the goal of maximizing the reward value of the action, and the optimized search policy parameters are output.

9. The method as described in claim 7, characterized in that, After outputting the optimized search strategy parameters, the method also includes: Obtain the search strategy parameters optimized from the first planning sub-region among the plurality of planning sub-regions; Initialize the candidate search strategy parameters for adjacent planning sub-regions based on the search strategy parameters.

10. A cable engineering path intelligent planning system based on deep reinforcement learning, characterized in that, The system is used to implement the intelligent cable engineering path planning method based on deep reinforcement learning as described in any one of claims 1-9, the system comprising: Data acquisition module: Acquires multi-source engineering data for the cable engineering route planning area; Regional division module: Constructs a regional data source dependency graph based on the multi-source engineering data, divides the planning region into sub-regions according to the data source dependency graph, and outputs multiple planning sub-regions; State space establishment module: Establishes multiple region state spaces corresponding to the multiple planning sub-regions; Policy generation module: Input the state spaces of the multiple regions into the deep reinforcement learning policy network, and output the search policy parameters for each planning sub-region to control the depth-first search process, wherein the search policy parameters include at least the search depth threshold, the branch expansion threshold, and the backtracking trigger threshold; Depth Search Module: If it is detected that the current planning sub-region has switched to the next planning sub-region, the search strategy parameters corresponding to the next planning sub-region are switched to perform a depth-first search until the target cable engineering planning path that traverses the cable engineering path planning area is obtained.

Images