Search

CN-121995915-A - AGV sampling vehicle optimal path planning method based on graph neural network

CN121995915ACN 121995915 ACN121995915 ACN 121995915ACN-121995915-A

Abstract

The invention discloses an AGV sampling vehicle optimal path planning method based on a graph neural network, which comprises the following steps of firstly constructing an environment graph for path planning, secondly calling an improved GRAPHSAGE model, thirdly executing comparison and fusion to obtain a bidirectional relationship attention weight of a target node to a neighbor node, fourthly constructing a neighbor profit model to obtain a neighbor competition aggregation weight, fifth generating updated node representation, sixth repeatedly executing the second to fifth steps on all nodes in the environment graph, seventh inputting a sampling point feature sequence into a Seq2Seq network based on an attention mechanism, and eighth generating an optimal path planning result of the AGV sampling vehicle. The invention has higher decision accuracy and path execution efficiency, and is suitable for diversified and complicated industrial sampling application scenes.

Inventors

  • ZHANG CHEN
  • LI ZHIPENG

Assignees

  • 滨州魏桥国科高等技术研究院

Dates

Publication Date
20260508
Application Date
20260128

Claims (9)

  1. 1. The AGV sampling vehicle optimal path planning method based on the graph neural network is characterized by comprising the following steps of: Acquiring map data and sampling point data of an AGV sampling vehicle operation area, and constructing an environment map for path planning; calling a pointer type neighbor selection unit of the improved GRAPHSAGE model, generating pointer scores for neighbor nodes of the target node, and selecting a preset number of neighbor nodes as a target neighbor set according to the pointer scores; the improved GRAPHSAGE model comprises a pointer type neighbor selection unit, a bidirectional relationship attention calculation unit, a game type neighbor competition aggregation unit and a node updating unit; In a bidirectional relationship attention calculating unit, respectively generating forward relationship expression and reverse relationship expression for each neighbor node in the target node and the target neighbor set, and executing comparison and fusion to obtain bidirectional relationship attention weight of the target node to the neighbor node; In a game neighbor competition aggregation unit, taking the node characteristics of the attention weight and the target neighbor set in a two-way relationship as input, constructing a neighborhood profit model, and iteratively updating the aggregation weight of each neighbor node according to the neighborhood profit model to obtain the neighbor competition aggregation weight; Step five, carrying out weighted aggregation on node characteristics of a target neighbor set according to neighbor competition aggregation weights, splicing characteristic vectors of target nodes with the aggregation characteristics, inputting the spliced characteristic vectors and the aggregation characteristics into a node updating unit, and generating updated node representations; step six, repeatedly executing the second to the fifth steps on all nodes in the environment map, and completing the propagation of the multilayer improved GRAPHSAGE model according to the preset layer number to obtain the final representation of all nodes; the nodes corresponding to the sampling points are finally expressed to form a sampling point characteristic sequence, the sampling point characteristic sequence is input into a Seq2Seq network based on an attention mechanism, and a sampling point access sequence is generated; and step eight, generating an optimal path planning result of the AGV sampling vehicle from the initial node to each sampling point and then to the final node according to the communication relation between the sampling point access sequence and the nodes in the environment map.
  2. 2. The method for planning an optimal path of an AGV sampling vehicle based on a neural network according to claim 1, wherein the first step specifically comprises: Acquiring electronic map data and corresponding coordinate system information of an AGV sampling vehicle operation area, extracting roads, channels and operation area boundaries in the operation area from the electronic map data, and determining a passable area and an obstacle area; in the passable area, discretizing the operation area according to a preset space division rule, and marking passable positions obtained by discretizing and positions of sampling points as graph nodes; establishing graph edges for paired graph nodes with drivable connection according to the road communication relationship recorded in the electronic map data and the communication relationship between the discretized adjacent passable positions so as to form a topological structure of the environment graph; and writing the position information of the nodes in the operation area, the mark information of whether the nodes are sampling points or not and the attribute information related to the node passing state into feature vectors according to a preset node initial feature rule aiming at each graph node in the environment graph, and completing the construction of the environment graph for path planning.
  3. 3. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the second step specifically comprises: selecting a current graph node to be processed as a target node in the environment graph according to a preset node traversing rule; reading a neighbor node set of a target node from a topological structure of an environment map, acquiring a feature vector of the target node and feature vectors of all neighbor nodes, and inputting the feature vector of the target node and the feature vector of the neighbor node into a pointer type neighbor selection unit of the improved GRAPHSAGE model; In the pointer type neighbor selection unit, the relation between the target node and each neighbor node is coded based on a pointer network, and corresponding pointer scoring is generated for each neighbor node in the neighbor node set; And sorting the neighbor node sets from high to low according to the pointer scoring, selecting the neighbor node with the pointer scoring front from the sorting result according to a preset neighbor node quantity threshold, and determining the selected neighbor node set as a target neighbor set.
  4. 4. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the third step specifically comprises: combining the feature vector of the target node with the feature vector of the neighbor node in the target neighbor set according to a preset forward relation coding rule to generate a forward relation representation of the target node pointing to the neighbor node; Combining the characteristic vectors of the neighbor nodes in the target neighbor set with the characteristic vectors of the target nodes according to a preset reverse relation coding rule to generate reverse relation representation representing that the neighbor nodes point to the target nodes; Performing contrast fusion operation on the forward relation representation and the reverse relation representation corresponding to the same neighbor node to obtain a bidirectional relation association score corresponding to the current neighbor node; and carrying out normalization processing on the bi-directional relation association scores of all the neighbor nodes in the target neighbor set, and generating the bi-directional relation attention weight of the target node to all the neighbor nodes.
  5. 5. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the fourth step specifically comprises: storing the bidirectional relationship attention weight and the neighbor node feature vector in an associated manner; initializing aggregation weights for each neighbor node in a target neighbor set, and establishing a neighbor profit model for each neighbor node in a neighbor; The neighborhood profit model comprises self profit values generated by the feature vectors of the current neighbor nodes and the attention weights of the bidirectional relations according to preset profit composition rules, and neighbor interaction influence values generated by the feature vectors of the current neighbor nodes, the feature vectors of other neighbor nodes and the attention weights of the bidirectional relations according to preset interaction composition rules; In the iterative updating process, updating the aggregation weight of each neighbor node in the target neighbor set according to the neighbor profit model; on the basis of initializing aggregation weight, calculating self-income value and neighbor interaction influence value of each neighbor node based on a neighbor income model; Forming comprehensive benefits by the self benefit value and the neighbor interaction influence value; Executing increasing processing or decreasing processing on the aggregation weights of the corresponding neighbor nodes according to the change direction of the comprehensive benefits between the current iteration round and the last iteration round, wherein if the comprehensive benefits of the current iteration round show increasing trend relative to the last iteration round, the increasing processing is executed on the aggregation weights of the corresponding neighbor nodes; After each round of updating is completed, normalization processing is carried out on the aggregation weights of all the neighbor nodes; and stopping updating and outputting neighbor competition aggregation weights when the number of iteration rounds reaches a preset iteration upper limit or the variation of all neighbor node aggregation weights in two adjacent rounds of iterations is smaller than a preset threshold.
  6. 6. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the fifth step specifically comprises the following steps: After the game type neighbor competition aggregation unit completes iterative updating, neighbor competition aggregation weights corresponding to all neighbor nodes in the target neighbor set and feature vectors of all neighbor nodes are obtained, and the neighbor competition aggregation weights are associated with the corresponding neighbor node feature vectors one by one; performing weighted operation on the feature vectors of all neighbor nodes in the target neighbor set according to the neighbor competition aggregation weight to generate a neighborhood aggregation feature vector used for representing the neighborhood information of the target node; Acquiring a feature vector of a target node, and splicing the feature vector of the target node and the neighborhood aggregation feature vector according to a preset feature splicing sequence to form a combined feature vector of the target node; The combined feature vector of the target node is input into a node updating unit, linear transformation processing is carried out on the combined feature vector of the target node in the node updating unit, activation processing is carried out on the combined feature vector of the target node, and updated node representation of the target node is generated.
  7. 7. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the sixth step specifically comprises: writing the updated node representation generated by each graph node in the node updating unit into a corresponding graph node in the environment graph, and taking the updated node representation as a feature vector of the current graph node in the next round of processing; In each round of processing, sequentially taking a graph node in an environment graph as a target node, calling a pointer type neighbor selection unit, a bidirectional relationship attention calculation unit, a game type neighbor competition aggregation unit and a node updating unit of the improved GRAPHSAGE model, and processing feature vectors of the target node and neighbor node feature vectors in a target neighbor set to generate updated node representation of the target node; The updated node representation of all the graph nodes of the current round is used as the characteristic vector of the graph nodes in the next round of processing, and the multi-round processing is repeatedly executed according to the preset network layer number until the multi-layer improved GRAPHSAGE model propagation corresponding to the preset layer number is completed; After the multi-layer propagation is completed, the updated node representation of each graph node in the last round of processing is determined as the final representation of each graph node in the environmental graph.
  8. 8. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the seventh step specifically comprises: Reading the final representation of each sampling point corresponding graph node from the environment graph, and sequentially arranging the final representations corresponding to each sampling point according to the serial number sequence of the sampling points to form a sampling point characteristic sequence; inputting the sampling point feature sequence into an encoder end of a Seq2Seq network based on an attention mechanism, sequentially receiving each sampling point feature vector in the sampling point feature sequence at the encoder end according to a time step sequence, and carrying out encoding processing on the sampling point feature vector of each time step to generate a corresponding encoder hiding state sequence; Setting an initial state of a decoder end according to the hidden state sequence of the encoder, inputting the initial state of the decoder and a start identifier for indicating the start of decoding into the decoder end of the Seq2Seq network, and starting a decoding process; In the current time step of the decoder side, receiving a sampling point identifier output by the previous time step and a decoder hiding state of the previous time step, taking the decoder hiding state of the current time step and an encoder hiding state sequence as attention calculation input, and calculating attention weight distribution corresponding to the current time step; Weighting the encoder hidden state sequence according to the attention weight distribution of the current time step to generate a context vector of the current time step, and combining the context vector of the current time step with the decoder hidden state of the current time step to obtain a decoding output vector of the current time step; Generating a next sampling point identifier to be accessed according to the decoding output vector of the current time step, writing the sampling point identifier into a sampling point access sequence as a decoding output result of the current time step, and taking the current sampling point identifier as one of inputs of a decoder end of the next time step; And repeatedly executing the attention calculation at the decoder end, the context vector generation, the decoding output vector generation and the sampling point identification output according to the time step sequence until the sampling point access sequence contains all the identifications corresponding to the sampling points, so as to obtain the sampling point access sequence output by the Seq2Seq network based on the attention mechanism.
  9. 9. The optimal path planning method for the AGV sampling vehicle based on the graphic neural network according to claim 1, wherein the eighth step specifically comprises: determining a starting node and a terminating node of the AGV sampling vehicle in an environment map, and respectively corresponding the starting node and the terminating node with a starting position and a terminating position of the AGV sampling vehicle; According to a sampling point access sequence output by a Seq2Seq network based on an attention mechanism, sequentially acquiring graph nodes corresponding to each sampling point identification in the sampling point access sequence from an environmental graph to form a sampling point node sequence arranged according to an access sequence; Inserting a starting node into the head of the sampling point node sequence, and adding a termination node into the tail of the sampling point node sequence to obtain a target node sequence comprising the starting node, the graph nodes corresponding to the sampling points and the termination node; Aiming at the adjacent previous node and the next node in the target node sequence, calling a Dijkstra shortest path algorithm to execute path search processing based on a communication relation between the nodes of the graph in the environment graph; Determining a graph node driving path from a previous node to a next node, sequentially splicing the graph node driving paths corresponding to each adjacent node pair according to the sequence of the target node sequence, and generating a graph node sequence which covers all sampling points and is arranged according to the driving sequence as an AGV sampling vehicle integral driving path; and determining a graph node sequence and a corresponding graph edge connection relation contained in the whole running path of the AGV sampling vehicle as an optimal path planning result of the AGV sampling vehicle from the initial node to each sampling point and then to the final node.

Description

AGV sampling vehicle optimal path planning method based on graph neural network Technical Field The invention relates to the technical field of industrial automation and path planning, in particular to an AGV sampling vehicle optimal path planning method based on a graph neural network. Background With the development of intelligent manufacturing and intelligent logistics, automatic guided vehicles (Automated Guided Vehicle, AGVs) are widely used in industrial production, warehouse logistics, medical distribution and other scenes. AGV sampling car is as an intelligent mobile platform that possesses route autonomous decision-making and task execution ability, extensively is arranged in medical examination, biological product sampling etc. fine scene. The quality of the path planning, which is a key component of the AGV system, directly determines the efficiency of the job and the consumption of resources. The traditional path planning method is based on classical graph search algorithms such as Dijkstra and A, and the like, and has good performance under a fixed graph structure and a single target, but when a plurality of target sampling points, a changeable operation environment and dynamic interference factors are faced, the traditional method has difficulty in considering the overall path optimality and the execution efficiency. In order to improve the overall performance of multi-objective path planning, research attempts have been made in recent years to introduce intelligent technologies such as deep learning, reinforcement learning and the like, and particularly, the graph neural network (Graph Neural Network, GNN) has a strong capability in the aspect of modeling graph structure data. GRAPHSAGE is used as a mainstream graph neural network variant, and efficient learning of the graph structure is realized through neighbor sampling and feature aggregation. However, the standard GRAPHSAGE model still has limitations on neighbor node selection, relationship modeling and node aggregation strategies, and particularly has defects in relationship difference modeling among nodes, game analysis of local optimal paths and global path optimization when an AGV path planning scene with complex space constraint and task sequence requirements is processed. The existing path planning method based on the graph neural network often fails to fully utilize the two-way dependency relationship between adjacent nodes, and lacks an effective dynamic weight adjustment mechanism in the aggregation process of the adjacent nodes, so that local information is easy to lead the whole nodes to be updated, and the global path quality is further influenced. Meanwhile, in the multi-point path decision, a cooperative mechanism capable of uniformly processing the sampling point ordering and the path segment minimization is lacking, so that the coupling degree between the path access sequence and the driving path is insufficient, and the rationality and the optimality of the final path are affected. Therefore, how to provide an optimal path planning method for an AGV sampling vehicle based on a graph neural network is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an AGV sampling vehicle optimal path planning method based on a graph neural network. The invention fully utilizes the improved GRAPHSAGE model and the Seq2Seq network based on the attention mechanism, carries out joint modeling on the access sequence decision of the sampling points and the shortest path solving in the environmental map, realizes the deep expression and the refined distinction of the association information between the nodes of the map by introducing a pointer type neighbor selection mechanism, a bidirectional relationship attention calculation mode and a game type neighbor competition aggregation strategy, and completes the optimal search of the driving path between the adjacent target nodes by combining Dijkstra shortest path algorithm. The method can generate the whole optimal driving path covering all the sampling points through double optimization of the access sequence of the sampling points and the shortest path among the nodes, and has the advantages of high rationality of path planning, fine modeling of node relation, low global path cost, suitability for complex sampling operation scenes and the like. According to the embodiment of the invention, the AGV sampling vehicle optimal path planning method based on the graph neural network comprises the following steps of: Acquiring map data and sampling point data of an AGV sampling vehicle operation area, and constructing an environment map for path planning; calling a pointer type neighbor selection unit of the improved GRAPHSAGE model, generating pointer scores for neighbor nodes of the target node, and selecting a preset number of neighbor nodes as a target neighbor set according to the pointer scores; the improved GRAPHSAGE model comp