E-commerce storage order real-time picking path planning method and system

Abstract

The invention discloses an e-commerce storage order real-time picking path planning method and system, and relates to the technical field of intelligent storage and logistics automation, and the method comprises the following steps: training a behavior prediction model based on historical picking behavior data, and outputting a path candidate region of a picker; constructing a path game model according to the path candidate area in combination with the current order task; based on a Stackelberg mechanism, solving a game equilibrium path distribution result; according to a solving result, evaluating the candidate path scheme through a multi-objective optimization algorithm, and outputting an optimal path scheduling strategy; according to the method, by introducing path candidate region extraction based on behavior prediction, fusion graph construction, a Stackelberg hierarchical game solving mechanism and multi-target optimization strategy output, real-time performance and global collaborative optimization of order picking path selection are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing and logistics automation technology, and more specifically, to a method and system for real-time picking route planning for e-commerce warehouse orders. Background Technology

[0002] In modern e-commerce warehousing systems, order fulfillment speed and picking efficiency have become key indicators affecting consumer experience and platform competitiveness. With the continuous growth in e-commerce order volume and the increasing complexity of SKUs, warehousing operations face the challenges of "multi-tasking, high concurrency, and short lead times." Especially in the order picking process, how to quickly generate efficient and low-conflict picking routes based on real-time order tasks has become one of the core issues for improving warehousing operational efficiency.

[0003] For example, the invention patent with publication number CN119130324A discloses a two-stage scheduling method for intelligent warehousing material picking based on a genetic algorithm with an embedded neighborhood search strategy. This method relates to the field of intelligent logistics scheduling and is used to solve the warehousing material picking scheduling problem in enterprise production logistics. It includes: considering the heterogeneity of the production line, establishing a two-stage optimization model for order task batching and grouping path optimization with the objective function of minimizing the total cost of picking order delays; invoking a tunnel-style warehousing picking path strategy; proposing four neighborhood search strategies; and improving the traditional genetic algorithm for problem solving by embedding a neighborhood search strategy. This automates the batching and grouping path planning of orders to be picked, minimizing the total cost of order picking delays. It can solve the two-stage scheduling problem of "order batching-path planning" in the warehousing picking process of enterprise production material supply, thereby improving the overall efficiency of production logistics and achieving the goal of cost reduction and efficiency improvement.

[0004] For example, the invention patent with publication number CN109447317A discloses a warehouse shortest path optimization method based on multiple people picking simultaneously. Taking the shortest picking path based on multiple people picking simultaneously as the optimization objective, it makes assumptions about the warehouse operating environment and establishes a mathematical model, determining the required number of pickers based on outbound product information; it introduces the concept of a separator into the genetic operation of the genetic algorithm, combining it with a selection operator improved based on cumulative probability to improve the algorithm and solve the objective function, obtaining the optimal picking path set for multiple people picking simultaneously; it uses a picking path conflict elimination strategy based on time windows to plan the obtained path set, adopts a priority-based waiting strategy to eliminate path conflicts, and obtains the picking time for each path; finally, it compares the picking times of each path and selects the path with the shortest time as the optimization result. This can improve warehouse picking efficiency.

[0005] The above-disclosed technical solutions have at least the following technical problems: Traditional systems cannot adapt to frequently changing order tasks, resulting in frequent path conflicts, low overall operational efficiency, and crude path conflict resolution, leading to frequent bottleneck congestion. Furthermore, the lack of global coordination capabilities makes it difficult to achieve orderly collaboration among multiple agents.

[0006] To address the above problems, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for real-time picking path planning for e-commerce warehouse orders. By introducing path candidate region extraction based on behavior prediction, fusion graph construction, Stackelberg hierarchical game solving mechanism, and multi-objective optimization strategy output, the method achieves real-time and global collaborative optimization of picking path selection, solves problems such as low task adaptability, frequent path conflicts, and unstable scheduling results in traditional path planning, and significantly improves the intelligence level and operating efficiency of e-commerce warehouse picking system.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for real-time picking route planning for e-commerce warehouse orders includes: training a behavior prediction model based on historical picking behavior data and outputting candidate route areas for pickers; constructing a path game model based on the candidate route areas and the current order task; solving the game equilibrium path allocation result based on the Stackelberg mechanism; evaluating the candidate route schemes based on the solution result and outputting the optimal route scheduling strategy through a multi-objective optimization algorithm.

[0009] In a preferred embodiment, the step of training a behavior prediction model based on historical picking behavior data and outputting candidate path regions for pickers is as follows: Acquiring warehouse environment data within the target area to construct a warehouse space structure map; acquiring historical order data, picking trajectory data, and warehouse operation procedures to construct a picking behavior knowledge graph; fusing the warehouse space structure map and the picking behavior knowledge graph using a graph embedding alignment mechanism to form a fused graph; selecting candidate path nodes based on the fused graph and extracting prediction input data from the candidate path nodes; inputting the prediction input data into a relational graph convolutional network model and outputting the selection probability of each candidate path node as the picker's confidence level in selecting that path region; sorting all candidate nodes according to their selection probabilities and, based on a set threshold, selecting the K regions with the highest confidence levels as candidate path regions.

[0010] In a preferred embodiment, the fusion of the warehouse space structure map and the picking behavior knowledge graph through a graph embedding alignment mechanism is as follows: Based on the warehouse space structure map, structural feature representations of each node are extracted to form a structural embedding vector; representation learning of entities and relationships in the picking behavior graph is performed using a graph embedding algorithm to obtain a knowledge embedding vector for each knowledge entity; the structural embedding vector and the knowledge embedding vector are mapped using a multilayer perceptron to form a mapped embedding vector; the mapped embedding vector is scored using an attention network mechanism, and the scores are normalized to output modality weights; the structural embedding vector and the knowledge embedding vector are weighted and fused according to the modality weights to obtain the final unified node representation vector; the unified node representation vector is used as the feature of each node to construct a fusion graph.

[0011] In a preferred embodiment, the step of selecting candidate path nodes based on the fusion graph is as follows: Obtain the current location node of the picker, and perform a multi-hop reachability search in the fusion graph starting from the current location node of the picker to construct an initial reachable node set; based on the initial reachable node set, analyze the picker's historical trajectory data, count the access frequency under similar task conditions, and output the frequency weight; sort the nodes in the initial reachable node set according to the frequency weight, retain the top N nodes with the highest weights to form a frequent access candidate set; obtain the semantic features of the current picking task, construct a task vector, and combine it with the fusion representation vector to calculate the semantic similarity with the task vector using cosine similarity; select nodes with high semantic relevance based on semantic similarity to form a semantic matching candidate set; and form the final candidate path node set by using the intersection method based on the frequent access candidate set and the semantic matching candidate set.

[0012] In a preferred embodiment, the step of constructing a path game model based on path candidate regions and the current order task is as follows: Based on the output of the behavior prediction model, obtain the path candidate region set and its behavior confidence for each picker, as the initial policy set for the agent; extract task features from the current order task and match the task features with the path candidate set to determine whether each path meets the requirements of the corresponding task; after matching, treat each picker as an agent, and its path candidate path node set constitutes the policy space, and construct a composite payoff function for each path; output the expected value of the path policy by linearly weighting the composite payoff function; construct a path game model with all pickers as agents, their respective path candidate sets as policy spaces, and the expected value as the utility function.

[0013] In a preferred embodiment, the process of solving the game equilibrium path allocation result based on the Stackelberg mechanism is as follows: All pickers are divided into several priority levels, forming multi-level agent groups. A leader-responder structure of the Stackelberg game is adopted between levels, with the upper-level group acting as the leader and the lower-level group as the responders. Based on the path candidate set, a path conflict graph is constructed and the edge weights are dynamically adjusted. The leader agent of each priority level selects the path with the highest expected value as the optimal path strategy, while recording the path's occupied nodes and influence range. This path and its occupied node information are considered pre-occupancy information and passed to the responders of the next priority level. After receiving the path occupancy information from the leader level, the responder agents dynamically adjust the conflict penalty weight and task reward weight in the payoff function according to the path conflict graph structure. After all levels complete their strategy selection, a set of global path strategies converges, forming a multi-level Stackelberg equilibrium solution.

[0014] In a preferred embodiment, the dynamic adjustment of edge weights is specifically as follows: For any two candidate path schemes, extract the spatial coordinate sequence of their respective path segments and their expected arrival time intervals, determine their overlapping segments in spatial location, and calculate the intersection duration of the overlapping segments in the time dimension; for path segments that do not directly overlap but are spatially adjacent, calculate their minimum spatial distance in the warehouse coordinate system using the Manhattan distance method; detect whether the current path pair overlaps with the upper-level picker path segment. If there is an overlap, it is considered that the node has been pre-occupied, and the conflict penalty value of the current edge is increased; normalize the overlap time, spatial distance, and pre-occupancy status, and introduce weight parameters for weighted fusion to form the final edge weight; the edge weight is updated before each round of path strategy solution.

[0015] In a preferred embodiment, the step of evaluating candidate path schemes based on the solution results and using a multi-objective optimization algorithm to output the optimal path scheduling strategy is as follows: Obtain a path set generated based on the path game solution results; define a multi-objective optimization function according to the operational objectives of the warehousing system; normalize the objective function and set weight coefficients for the objective function according to different application scenarios to form a weighted scoring model; solve the objective function and path data using a multi-objective evolutionary decomposition algorithm to output a Pareto optimal solution set for a set of path allocation solutions; select the most suitable path scheduling scheme from the optimal solution set based on the current warehousing resource status as the final path allocation result.

[0016] A real-time picking route planning system for e-commerce warehouse orders includes a training module, a game theory module, a solution module, and an output module, with connections between the modules. The training module trains a behavior prediction model based on historical picking behavior data and outputs candidate route areas for pickers. The game theory module constructs a route game model based on the candidate route areas and the current order task. The solution module solves for the equilibrium route allocation result based on the Stackelberg mechanism. The output module evaluates the candidate route schemes based on the solution results and uses a multi-objective optimization algorithm to output the optimal route scheduling strategy.

[0017] The technical effects and advantages of the real-time picking route planning method and system for e-commerce warehouse orders of this invention are as follows: 1. This invention introduces a behavior prediction mechanism to construct a picking behavior knowledge graph based on historical picking behavior data, order trajectories, and the warehousing environment. This graph is then combined with a warehousing spatial structure graph and fused into a unified spatial-behavioral representation graph through a graph embedding alignment mechanism. A graph neural network (such as a relational graph convolutional network) is used to predict the selection probability of path nodes, accurately outputting candidate path regions. This mechanism significantly improves the accuracy of path prediction, providing reliable prior knowledge for subsequent path game theory.

[0018] 2. This invention solves the path game model using the Stackelberg mechanism, stratifying pickers according to indicators such as task urgency to construct a multi-level leader-responder structure. Upper-level leaders, the pickers, prioritize path strategies and occupy key nodes, while lower-level respondents adjust their payoff functions to make optimal responses based on already occupied information and the dynamic edge weights of the path conflict graph. Through this hierarchical solution and strategy-response mechanism, the system can converge to a multi-level Stackelberg equilibrium solution that satisfies both global coordination and local optima, minimizing path conflicts and efficiently utilizing path resources. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the real-time picking path planning method for e-commerce warehouse orders according to the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a real-time picking path planning system for e-commerce warehouse orders according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, Figure 1 This invention provides a real-time picking path planning method for e-commerce warehouse orders, comprising: S1, based on historical picking behavior data, trains a behavior prediction model and outputs candidate areas for the picker's path; In this embodiment, a behavior prediction model is trained based on historical picking behavior data, and the candidate areas for the picker's path are output, as follows: Acquire warehousing environment data within the target area and construct a warehousing space structure map, wherein... It is a set of nodes, including location nodes, aisle nodes, and picking point nodes. It is a set of edges, including accessible path edges, physical adjacency edges, and historical migration edges, used to describe the accessible relationships and historical movement trends of pickers in the warehouse; By acquiring historical order data, picking trajectory data, and warehouse operation procedures, a knowledge graph of picking behavior can be constructed. ,in, It is a set of nodes, including pickers, task types, goods categories, operation time periods, and path segments. The set of edges includes frequent selection relationships, path preference relationships, time-active relationships, and path co-occurrence relationships; The warehouse space structure map and the picking behavior knowledge graph are merged through a graph embedding alignment mechanism to form a fused graph; Based on the fusion graph, candidate path nodes are selected, and the predicted input data from the candidate path nodes is extracted. The predicted input data is fed into the relational graph convolutional network model, and the selection probability of each candidate path node is output as the confidence level of the picker in selecting that path area. The confidence scores of all candidate nodes are sorted according to the selection probability, and the regions with the highest K confidence scores are selected as path candidate regions based on the set threshold.

[0023] In this embodiment, the warehouse space structure diagram and the picking behavior knowledge graph are fused together using a graph embedding alignment mechanism, as follows: Based on the warehouse space structure diagram, extract the spatial coordinates, cargo storage density, historical popularity and other structural features of each node (including storage locations, aisle nodes, picking stations, etc.) to form a structural embedding vector; Representation learning is performed on entities and relationships in the picking behavior graph using graph embedding algorithms (such as TransE or R-GCN) to obtain the knowledge embedding vector of each knowledge entity, which is used to express the semantic preferences of the picker's behavior and the potential structural relationships between tasks. By mapping the structure embedding vector and the knowledge embedding vector using a multilayer perceptron, their dimensions are made consistent and aligned to a unified semantic space, forming a mapped embedding vector; The mapping embedding vector is scored using an attention network mechanism, and the scores are normalized to output the modality weights. The structural embedding vector and the knowledge embedding vector are weighted and fused according to the modality weights to obtain the final unified node representation vector; Using the unified node representation vector as the feature of each node, a fusion graph is constructed, including: The node set of the fusion graph is consistent with the node set of the warehouse space structure graph. It is still composed of various storage locations, aisle nodes, etc. in the warehouse space, and the representation vector of each node adopts a unified node representation vector. Regarding the construction of the edge set, in addition to retaining the physical connection edges in the warehouse space structure graph, semantic association edges derived from the picking knowledge graph are introduced. Specifically, if there is a significant historical co-occurrence frequency between two nodes, continuous access behavior on the picking path, or sharing similar product categories and operational strategies, such semantic relationships are represented as additional edges and incorporated into the edge set of the fusion graph. The completed fusion graph uses unified node vectors as vertex features and edge sets composed of multi-source relationships as the connection structure, forming graph structure data with high semantic and spatial coupling capabilities.

[0024] The scoring of the mapping embedding vector using an attention network mechanism is as follows:

[0025]

[0026] The normalization is as follows:

[0027] In the formula: It is the attention score of the warehouse space structure diagram modality. This represents the embedding vector of node v in the warehouse space structure graph after mapping. This represents the attention scoring function used to calculate the importance of each modality embedding. This is the weight matrix in the attention network, used to perform a linear transformation on the input embedding. It is the bias vector in the attention network, used to introduce non-linear offsets. It is the transpose of the weight vector output by the attention layer, used to map the activated representation to a scalar score. It is the attention score of the picking behavior graph modality. This represents the embedding vector of node v in the picking behavior graph after mapping. These are the normalized weights of the warehouse space structure diagram modes. It is the normalized weight of the picking behavior graph modality.

[0028] The process of selecting candidate path nodes based on the fusion graph is as follows: Obtain the current location node of the picker, and perform a multi-hop reachability search in the fusion graph starting from the current location node of the picker to construct an initial set of reachable nodes; Based on the initial set of reachable nodes, analyze the historical trajectory data of pickers, count their access frequency under similar task conditions, and output the frequency weight. The nodes in the initial reachable node set are sorted according to frequency weight, and the top N nodes with the highest weight are retained to form a frequently accessed candidate set. Obtain the semantic features of the current picking task (composed of task type, goods category, operation time period, etc.), construct the task vector, and combine it with the fused representation vector to calculate the semantic similarity with the task vector through cosine similarity. Nodes with high semantic relevance are selected based on semantic similarity to form a semantic matching candidate set; Based on the frequently accessed candidate set and the semantically matched candidate set, the final candidate path node set is formed by the intersection method.

[0029] The frequency weight calculation formula is as follows:

[0030] In the formula: Represents a node The frequency weight value, with a value range of [0,1], is used to measure the relative frequency of the node's appearance in similar historical picking tasks. Represents a node In historical trajectory data, the cumulative number of times a node is accessed by pickers under contexts similar to the current task conditions is called the historical access frequency of a node. Represents the set of spatially reachable nodes. Any node in the list represents a path area node that the picker may move to in the current task scenario.

[0031] S2, construct a path game model based on the candidate path regions and the current order task; In this embodiment, a path game model is constructed based on the candidate path regions and the current order task, as follows: Based on the output of the behavior prediction model, obtain the set of candidate paths and their behavior confidence for each picker, which serves as the initial policy set for the agent. Extract task features from the current order task. These features include order priority, fulfillment time limit, and spatial location of the corresponding picking point. These features will be used to construct the objective function and constraint mechanism in the game model. The task features are matched with the path candidate set to determine whether each path meets the requirements of the corresponding task, such as whether it covers the required storage location and whether it can be completed within the expected time window, thereby generating a path-task fit score. After matching, each picker is regarded as an intelligent agent, and the set of candidate path nodes constitutes the policy space. A composite benefit function is constructed for each path. The composite benefit function includes task completion benefit (reflecting the path's coverage and completion capability for high-priority and urgent orders), path execution cost (including path length, estimated time, energy consumption, and warehouse congestion level), and conflict penalty cost (measuring the degree of spatial and temporal overlap between the path and other pickers' paths, used to quantify the cost of interference). The expected value of the path strategy is output by combining the composite payoff functions through linear weighting. A path game model is constructed, with all pickers as agents, their respective path candidate sets as policy spaces, and expected value as the utility function.

[0032] S3, based on the Stackelberg mechanism, solves the game equilibrium path allocation result; In this embodiment, the equilibrium path allocation result of the game is solved based on the Stackelberg mechanism, as follows: Based on indicators such as the urgency, value, and storage space distribution of current orders, all pickers are divided into several priority levels, forming multi-level agent groups. Pickers in each priority level participate in path strategy selection simultaneously. The leader-responder structure of Stackelberg game is adopted between levels. The upper level group acts as the leader and makes path selection decisions first, while the lower level group acts as the responder and adjusts its own path strategy based on the leader's path decision. Based on the path candidate set, a real-time updated path conflict graph is constructed. Each node in the graph corresponds to a path candidate, and the edges represent spatial or temporal overlapping conflicts between paths. The edge weights are dynamically adjusted. Each priority layer leader agent selects the path with the highest return as the optimal path strategy based on the expected value. At the same time, it records the path's occupied nodes and the scope of influence. The path and its occupied node information are regarded as pre-occupancy information and passed to the responders of the next priority layer as constraints for subsequent path selection. After receiving the path occupancy information from the leadership layer, the response layer agent dynamically adjusts the conflict penalty weight and task reward weight in the reward function according to the path conflict graph structure to reflect the degree of constraint of the strategy space. The reward function is obtained by linearly weighting a composite reward function. After all levels have completed their strategy selections, a set of global path strategies converges, forming a multi-level Stackelberg equilibrium solution. This multi-level Stackelberg equilibrium solution satisfies the following conditions: Top leaders prioritize stability in their strategies; Lower-level responders achieve local optimality under the constraints of upper-level policies; Minimize overall path conflicts and maximize task benefits.

[0033] In this embodiment, the edge weights are dynamically adjusted, as follows: For any two candidate paths, extract the spatial coordinate sequence of each path segment and its expected arrival time interval, determine the overlapping segment in spatial location, and calculate the intersection duration of the overlapping segment in time dimension. For path segments that do not directly overlap but are spatially adjacent, the minimum spatial distance in the warehouse coordinate system is calculated using the Manhattan distance method. The smaller the distance, the greater the potential interference, and the conflict weight of the corresponding edge should also be increased. Check if the current path pair overlaps with the path segment of the upper picker (i.e., the game leader whose strategy has been selected). If there is an overlap, it is considered that the node has been pre-occupied, and the conflict penalty value of the current edge is increased. The overlapping time, spatial spacing and pre-occupancy state are normalized and weighted by introducing weight parameters to form the final edge weights. The edge weights are updated before each round of path strategy solving.

[0034] The specific steps for dynamically adjusting the conflict penalty weight and task reward weight in the reward function based on the path conflict graph structure are as follows: Increase the penalty weight for candidate paths that conflict with the leader's pre-occupied path, thereby reducing the reward of that path; The reward weight is dynamically adjusted based on the current task priority and time window to ensure timely task completion. Introduce a path flexibility assessment to prioritize paths with strong alternatives and low conflict.

[0035] S4. Based on the solution results, the candidate path schemes are evaluated through a multi-objective optimization algorithm, and the optimal path scheduling strategy is output. In this embodiment, based on the solution results, candidate path schemes are evaluated using a multi-objective optimization algorithm, and the optimal path scheduling strategy is output as follows: Obtain a path set generated based on the path game solution results. The path set contains multiple candidate paths for pickers, and each path is accompanied by evaluation indicators such as task matching degree, path cost, and conflict risk. Based on the operational objectives of the warehousing system, a multi-objective optimization function is defined, which includes maximizing the overall order task coverage, minimizing the total path length, and reducing the spatial overlap between paths. The objective function is normalized, and weight coefficients are set according to different application scenarios to form a weighted scoring model. Based on the constructed objective function and path data, the multi-objective evolutionary decomposition algorithm (MOEA / D) is used to solve the problem, and a set of Pareto optimal solutions for path allocation is output. Based on the current state of warehouse resources, the most suitable set of path scheduling schemes is selected from the optimal solution set as the final path allocation result. The scheduling result includes the matching and allocation of pickers and paths, the specific path travel order and estimated time for each picker, and early warning prompts for conflict nodes.

[0036] Example 2, Figure 2 The present invention provides a real-time picking path planning system for e-commerce warehouse orders, including a training module, a game theory module, a solution module, and an output module, with connections between the modules; The training module is used to train a behavior prediction model based on historical picking behavior data and output candidate areas for the picker's path. The game theory module is used to construct a path game model based on the candidate path regions and the current order task. The solution module is used to solve for the equilibrium path allocation result in the game based on the Stackelberg mechanism; The output module is used to evaluate candidate path schemes based on the solution results and through a multi-objective optimization algorithm, and output the optimal path scheduling strategy.

[0037] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0038] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0039] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0040] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0041] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0042] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time picking path planning for e-commerce warehouse orders, characterized in that, include: Based on historical picking behavior data, a behavior prediction model is trained and candidate areas for picking personnel's paths are output. Construct a path game model based on the candidate path regions and the current order task; Based on the Stackelberg mechanism, the equilibrium path allocation result of the game is obtained; Based on the solution results, the candidate path schemes are evaluated using a multi-objective optimization algorithm, and the optimal path scheduling strategy is output.

2. The method for real-time picking path planning of e-commerce warehouse orders according to claim 1, characterized in that, The behavior prediction model is trained based on historical picking behavior data, and the candidate areas for the picker's path are output as follows: Acquire warehousing environment data within the target area and construct a warehousing space structure map; Acquire historical order data, picking trajectory data, and warehouse operation procedures to construct a knowledge graph of picking behavior; The warehouse space structure map and the picking behavior knowledge graph are merged through a graph embedding alignment mechanism to form a fused graph; Based on the fusion graph, candidate path nodes are selected, and the predicted input data from the candidate path nodes is extracted. The predicted input data is fed into the relational graph convolutional network model, and the selection probability of each candidate path node is output as the confidence level of the picker in selecting that path area. The confidence scores of all candidate nodes are sorted according to the selection probability, and the regions with the highest K confidence scores are selected as path candidate regions based on the set threshold.

3. The method for real-time picking path planning of e-commerce warehouse orders according to claim 2, characterized in that, The process of fusing the warehouse space structure map with the picking behavior knowledge graph through a graph embedding alignment mechanism is as follows: Based on the warehouse space structure diagram, extract the structural feature representation of each node to form a structural embedding vector; The entity and relationship representations in the picking behavior graph are learned by using a graph embedding algorithm to obtain the knowledge embedding vector of each knowledge entity; The structure embedding vector and the knowledge embedding vector are mapped separately by a multilayer perceptron to form a mapped embedding vector; The mapping embedding vector is scored using an attention network mechanism, and the scores are normalized to output the modality weights. The structural embedding vector and the knowledge embedding vector are weighted and fused according to the modality weights to obtain the final unified node representation vector; A unified node representation vector is used as the feature of each node to construct a fusion graph.

4. The method for real-time picking path planning of e-commerce warehouse orders according to claim 3, characterized in that, The process of selecting candidate path nodes based on the fusion graph is as follows: Obtain the current location node of the picker, and perform a multi-hop reachability search in the fusion graph starting from the current location node of the picker to construct an initial set of reachable nodes; Based on the initial set of reachable nodes, analyze the historical trajectory data of pickers, count the access frequency under similar task conditions, and output the frequency weight. The nodes in the initial reachable node set are sorted according to frequency weight, and the top N nodes with the highest weight are retained to form a frequently accessed candidate set. Obtain the semantic features of the current picking task, construct the task vector, and combine it with the fused representation vector to calculate the semantic similarity with the task vector through cosine similarity. Nodes with high semantic relevance are selected based on semantic similarity to form a semantic matching candidate set; Based on the frequently accessed candidate set and the semantically matched candidate set, the final candidate path node set is formed by the intersection method.

5. The method for real-time picking path planning of e-commerce warehouse orders according to claim 4, characterized in that, The path game model is constructed based on the candidate path regions and the current order task, as follows: Based on the output of the behavior prediction model, obtain the set of candidate paths and their behavior confidence for each picker, which serves as the initial policy set for the agent. Extract task features from the current order task and match the task features with the path candidate set to determine whether each path meets the requirements of the corresponding task; After matching is completed, each picker is regarded as an intelligent agent, and the set of candidate path nodes constitutes the policy space. A composite reward function is constructed for each path. The expected value of the path strategy is output by combining the composite payoff functions through linear weighting. A path game model is constructed, with all pickers as agents, their respective path candidate sets as policy spaces, and expected value as the utility function.

6. The method for real-time picking path planning of e-commerce warehouse orders according to claim 5, characterized in that, The solution to the game equilibrium path allocation result based on the Stackelberg mechanism is as follows: All pickers are divided into several priority levels to form multi-level agent groups. The leader-responder structure of Stackelberg game is used between the layers, with the upper group as the leader and the lower group as the responder. Based on the path candidate set, a path conflict graph is constructed in real time, and the edge weights are dynamically adjusted. Each priority layer's leader agent selects the path with the highest expected value as the optimal path strategy, while recording the path's occupied nodes and influence range. This path and its occupied node information are considered pre-occupancy information and are passed on to the responders of the next priority layer. After receiving path occupancy information from the leadership layer, the response layer agent dynamically adjusts the conflict penalty weight and task reward weight in the reward function according to the path conflict graph structure. After all levels have completed their strategy selections, a set of global path strategies converges, forming a multi-level Stackelberg equilibrium solution.

7. The method for real-time picking path planning of e-commerce warehouse orders according to claim 6, characterized in that, The dynamic adjustment of edge weights is as follows: For any two candidate paths, extract the spatial coordinate sequence of each path segment and its expected arrival time interval, determine the overlapping segment in spatial location, and calculate the intersection duration of the overlapping segment in time dimension. For path segments that do not directly overlap but are spatially adjacent, the minimum spatial distance in the warehouse coordinate system is calculated using the Manhattan distance method; Check if the current path pair overlaps with the upper-level picker's path segment. If there is an overlap, it is considered that the node has been pre-occupied, and the conflict penalty value of the current edge is increased. The overlapping time, spatial spacing and pre-occupancy state are normalized and weighted by introducing weight parameters to form the final edge weights. The edge weights are updated before each round of path strategy solving.

8. The method for real-time picking path planning of e-commerce warehouse orders according to claim 7, characterized in that, Based on the solution results, the candidate path schemes are evaluated using a multi-objective optimization algorithm to output the optimal path scheduling strategy, as detailed below: Obtain the set of paths generated based on the path game solution results; Define a multi-objective optimization function based on the operational objectives of the warehousing system; The objective function is normalized, and weight coefficients are set according to different application scenarios to form a weighted scoring model. Based on the constructed objective function and path data, a multi-objective evolutionary decomposition algorithm is used to solve the problem, and a Pareto optimal solution set of path allocation solutions is output. Based on the current state of warehouse resources, the most suitable set of path scheduling schemes is selected from the optimal solution set as the final path allocation result.

9. A real-time picking route planning system for e-commerce warehouse orders as described in any one of claims 1-8, characterized in that, It includes a training module, a game-playing module, a solution module, and an output module, and there are connections between the modules; The training module is used to train a behavior prediction model based on historical picking behavior data and output candidate areas for the picker's path. The game theory module is used to construct a path game model based on the candidate path regions and the current order task. The solution module is used to solve for the equilibrium path allocation result in the game based on the Stackelberg mechanism; The output module is used to evaluate candidate path schemes based on the solution results and through a multi-objective optimization algorithm, and output the optimal path scheduling strategy.

Images