A neural division method and device for heterogeneous multi-agv scheduling for warehouse logistics

Abstract

The application belongs to the technical field of path optimization and intelligent decision-making, and aims at the problems of unbalanced and low efficiency of multi-agent collaborative heterogeneous scheduling in the prior art, and provides a heterogeneous multi-AGV neural divide-and-conquer scheduling method and device for warehouse logistics, the method constructs an MDMTSP-V problem model with the objective of minimizing the maximum travel time, designs a neural divide-and-conquer framework, decomposes the warehouse logistics problem into two stages of city allocation and sub-traveling salesman planning solution, constructs a neural allocation strategy network, fuses spatial features and heterogeneous speed features, realizes efficient and fair city cluster division, adopts a reinforcement learning mechanism to perform end-to-end training on the network, optimizes the allocation strategy, and solves the sub-TSP problem of each agent by using an efficient and accurate algorithm. The learning ability of the neural network and the accuracy of the classical optimization algorithm are combined, and the fairness and overall efficiency of the multi-agent collaborative heterogeneous scheduling scheme for warehouse logistics are significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer network and information security, and relates to network feature recognition technology, in particular to a heterogeneous multi-AGV neural divide-and-conquer scheduling method and device for warehouse logistics. BACKGROUND

[0002] In a modern warehouse logistics system, multi-automatic guided vehicle (AGV) collaborative scheduling is the core link to improve operation efficiency. In actual scenarios, AGV clusters often have significant speed heterogeneity due to differences in model, load or power consumption, requiring the scheduling system not only to reasonably allocate task sequences, but also to take into account the performance differences of different AGVs to achieve overall efficiency optimization and load balancing. Such a heterogeneous multi-AGV scheduling problem for warehouse logistics is essentially a highly complex and practically significant topic in the field of combinatorial optimization.

[0003] To solve this problem, technicians usually model it as a multiple traveling salesman problem (MTSP) or an extended form. MTSP is an extension of the traveling salesman problem (TSP), aiming to find the optimal path planning for multiple traveling salesmen to visit all task points collaboratively, such as the scheduling problem of multiple AGVs visiting all cities collaboratively. It is a classic model for solving multi-agent (agent) collaborative path planning. On this basis, the multi-depot MTSP (MDMTSP) further introduces multiple warehouses (i.e., start and end points), allowing each traveling salesman to start from different warehouses and return to their respective warehouses, which is more suitable for warehouse logistics scenarios with multiple distribution centers and AGV parking areas.

[0004] Among the many variants of MTSP, the multi-depot MTSP (MDMTSP) introduces multiple warehouses (or multiple start and end points), allowing each traveling salesman to start from different warehouses and return to the corresponding warehouse location, enhancing the practical adaptability of the model and better fitting real-world problems such as multi-vehicle scenarios and multi-robot base deployment.

[0005] The MTSP and its numerous variants are generally concerned with two core optimization objectives: one is to minimize the total travel distance of all traveling salesmen (Min-Sum MTSP) to reduce overall operating costs or energy consumption; the other is to minimize the travel path length of the longest traveling salesman (Min-Max MTSP), focusing on load balancing and synchronous completion of tasks, to avoid efficiency bottlenecks caused by excessively long routes of a traveling salesman (AGV path). Min-Max MTSP is particularly important in heterogeneous AGV scheduling, effectively improving the synchronization and fairness of task completion.

[0006] To address the above problems, existing solution methods mainly include:

[0007] Exact algorithms, such as dynamic programming, mixed integer linear programming (MILP), and its branch-and-bound method and cutting plane algorithm, can obtain optimal solutions, but have high algorithm complexity and are difficult to handle large-scale instances.

[0008] Heuristic algorithms, such as ant colony algorithm, genetic algorithm, and simulated annealing, have strong optimization ability, but rely on parameter tuning and cannot guarantee optimality.

[0009] Learning algorithms, especially neural combinatorial optimization methods based on deep learning, learn the internal structure of the problem through data-driven methods, showing potential in solving speed and generalization ability, and have become a frontier in this field.

[0010] However, existing MTSP and MDMTSP technologies, especially emerging learning methods, are mostly based on the assumption of homogeneous AGVs and fail to effectively integrate heterogeneous speed information of AGVs. This leads to significant limitations in real heterogeneous speed environments: the model cannot accurately depict the impact of speed differences on task allocation and path planning, resulting in scheduling precision and efficiency limitations, imbalance between solution quality and solution time, and insufficient adaptability to complexity and scalability.

[0011] Therefore, developing new scheduling methods and devices that can deeply integrate speed characteristics and spatial task information has become a key challenge to improve the performance of warehouse logistics AGV systems. SUMMARY

[0012] In view of the above technical problems, the present application provides a heterogeneous multi-AGV neural divide-and-conquer scheduling method and device for warehouse logistics. The method is suitable for the heterogeneous multi-AGV planning of the actual heterogeneous scene with different speeds in urban warehouse logistics, and establishes a speed-sensitive model of the Min-Max multi-warehouse multi-traveling salesman problem. By introducing a traveling salesman speed embedding module and an additive attention mechanism, the spatial features and speed information are effectively fused, and a divide-and-conquer strategy is used to process the city allocation and path optimization in stages, significantly improving the fairness and efficiency of scheduling. At the same time, a variance reduction baseline based on speed is designed to optimize the reinforcement learning training process and improve the model convergence speed and solution quality.

[0013] The present application provides a heterogeneous multi-AGV neural divide-and-conquer scheduling method for warehouse logistics, which comprises:

[0014] Step 110, taking the heterogeneous speed of AGV as a key parameter, a heterogeneous multi-warehouse multi-traveling salesman problem model is constructed with the optimization objective of minimizing the maximum travel time;

[0015] Step 120, the heterogeneous multi-warehouse multi-traveling salesman problem is divided into two optimization stages in series: a task allocation stage and a path optimization stage based on a sub-traveling salesman problem;

[0016] Step 130, the optimization objective of the heterogeneous multi-warehouse multi-traveling salesman problem model is reconstructed as a whole double-layer optimization objective;

[0017] Step 140, based on the whole double-layer optimization objective, a neural divide-and-conquer framework is used to solve the heterogeneous multi-warehouse multi-traveling salesman problem model, including:

[0018] In the task allocation stage, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous speed features, and output a city point allocation subset for each agent;

[0019] The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters;

[0020] In the path optimization stage based on the sub-traveling salesman problem:

[0021] The path planning of each city point allocation subset is regarded as an independent sub-traveling salesman problem, and a traveling salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain the path planning scheme of each sub-traveling salesman problem;

[0022] The path planning schemes of all sub-traveling salesman problems are combined to obtain the overall path planning scheme of the heterogeneous multi-warehouse multi-traveling salesman problem.

[0023] On the other hand, the present application provides a heterogeneous multi-AGV neural divide-and-conquer scheduling device for warehouse logistics, comprising:

[0024] The first module is used to construct a model for the heterogeneous speed multi-warehouse multi-traveling salesman problem with the heterogeneous speed of AGVs as the key parameter and the optimization objective of minimizing the maximum travel time.

[0025] The second module is used to decompose the heterogeneous multi-warehouse multi-traveling salesman problem into two sequentially coupled optimization stages: the task allocation stage and the path optimization stage based on the sub-traveling salesman problem.

[0026] The third module is used to reconstruct the optimization objective of the multi-warehouse multi-trading salesman problem model into an overall bi-layer optimization objective;

[0027] The fourth module, based on the overall bi-layer optimization objective, uses a neural divide-and-conquer framework to solve the allometric multi-warehouse multi-trading salesman problem model, including:

[0028] In the task allocation phase, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous velocity features, and output a subset of city point allocations for each agent;

[0029] The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters.

[0030] In the path optimization phase based on the sub-traveling salesman problem:

[0031] The path planning for each city point subset is treated as an independent sub-traveling salesman problem. A travel salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain the path planning scheme for each sub-traveling salesman problem.

[0032] By merging the path planning schemes of all the sub-traveling salesman problems, we obtain the overall path planning scheme for the allometric multi-warehouse multi-traveling salesman problem.

[0033] Compared with the prior art, the technical effects of the present invention include:

[0034] (1) Effectively integrate heterogeneous speed information to improve scheduling fairness and optimization quality. Through the AGV speed embedding module and additive attention mechanism, the deep integration of spatial node features and speed information is achieved. During the allocation stage, the task allocation is dynamically adjusted according to the speed difference of AGVs to avoid the uneven load caused by the homogeneity assumption and significantly improve the balance of the completion time of multiple AGV tasks.

[0035] (2) Use a divide-and-conquer strategy to reduce problem complexity. The problem is divided into two stages: task allocation and sub-path optimization, through a neural divide-and-conquer framework: the neural network quickly generates speed-sensitive task partitions, and each AGV sub-path is optimized by an efficient TSP solver (such as Google OR-Tools), taking into account both scheduling quality and computational efficiency.

[0036] (3) Based on the variance reduction baseline design, the stability of reinforcement learning is ensured. In view of the reward distribution drift and excessive training variance caused by heterogeneous speed, the variance reduction baseline based on the normalization of speed sum is designed, which effectively integrates the speed factor into the reward normalization, significantly reduces the gradient estimation variance, improves the training stability and convergence speed, and finally obtains a better strategy.

[0037] (4) Significantly improve the computing efficiency, suitable for large-scale heterogeneous multi-warehouse multi-traveler scheduling. While ensuring the optimal maximum normalized completion time, the running time is greatly shortened compared with traditional accurate and heuristic algorithms. For example, in the case of thousands of path planning, the reasoning time of the present application is maintained at seconds, while the calculation time of traditional methods increases exponentially and is difficult to complete, effectively meeting the actual industrial multi-agent complex path planning demand.

[0038] The modules of the scheme are clear and easy to integrate, covering state input, graph embedding, speed coding, attention fusion, strategy sampling, path optimization and reinforcement learning modules, supporting zero-training migration of different AGV numbers and task scales, and having good generalization ability and engineering practicability. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the drawings shown.

[0040] Figure 1 The step flow chart of the heterogeneous multi-AGV neural sub-task scheduling method for warehouse logistics in an embodiment of the present application;

[0041] Figure 2 The neural sub-task framework process schematic diagram in an embodiment of the present application;

[0042] Figure 3 The neural allocation strategy network architecture schematic diagram in an embodiment of the present application;

[0043] Figure 4 The result schematic diagram of solving the path AGV agent scheduling based on the AGV number of 10 and the city point number of 50 by using the ant colony algorithm in the experiment of the present application, and the maximum time is 6.6998;

[0044] Figure 5 The result schematic diagram of solving the path AGV agent scheduling based on the AGV number of 10 and the city point number of 50 by using the ant colony algorithm in the experiment of the present application, and the maximum time is 10.4449;

[0045] Figure 6 In the experiment of the present application, the method of the present application is used to solve the path AGV agent scheduling based on 10 AGVs and 50 city points, and a result diagram with a maximum time of 2.5576 is obtained;

[0046] Figure 7 In the experiment of the present application, the method of the present application is used to solve the path AGV agent scheduling based on 10 AGVs and 50 city points, and a result diagram with a maximum time of 2.2760 is obtained;

[0047] Figure 8 In the experiment of the present application, the ant colony algorithm is used to solve the path AGV agent scheduling based on 10 AGVs and 200 city points, and a result diagram with a maximum time of 13.1765 is obtained;

[0048] Figure 9 In the experiment of the present application, the ant colony algorithm is used to solve the path AGV agent scheduling based on 10 AGVs and 200 city points, and a result diagram with a maximum time of 35.9214 is obtained;

[0049] Figure 10 In the experiment of the present application, the method of the present application is used to solve the path AGV agent scheduling based on 10 AGVs and 200 city points, and a result diagram with a maximum time of 20.5488 is obtained;

[0050] Figure 11 In the experiment of the present application, the method of the present application is used to solve the path AGV agent scheduling based on 10 AGVs and 200 city points, and a result diagram with a maximum time of 3.5121 is obtained. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0052] The present application aims to solve the problems of path allocation imbalance, low system efficiency and poor scalability caused by ignoring the speed difference of automatic guided vehicles (AGVs) in the prior art, and provides a high-efficiency, scalable and complex multi-warehouse multi-AGV collaborative scheduling solution suitable for complex logistics scenarios, to meet the urgent demand for path optimization and load balancing in actual complex logistics scenarios.

[0053] To solve the above problems, the application further extends the MDMTSP and defines a Multi-Depot Multiple Traveling Salesman Problem with Heterogeneous Velocities (MDMTSP-V) model. The MDMTSP-V considers the key element of speed heterogeneity in the planning model based on the MDMTSP, and introduces a key parameter - heterogeneous speed for each traveling salesman (for example, AGV). That is, each AGV can have different moving speeds.

[0054] In one embodiment, as shown in FIG. 1, the application provides a heterogeneous speed multi-AGV neural divide-and-conquer scheduling method for warehouse logistics, which comprises the following steps: Figure 1

[0055] Step 110, taking the heterogeneous speed of AGV as a key parameter, a Multi-Depot Multiple Traveling Salesman Problem with Heterogeneous Velocities (MDMTSP-V) model is constructed with the optimization goal of minimizing the maximum travel time;

[0056] Step 120, the MDMTSP-V is decomposed into two optimization stages coupled in series: a task allocation stage and a path optimization stage based on a sub-traveling salesman problem;

[0057] Step 130, the optimization goal of the MDMTSP-V model is restructured into a whole double-layer optimization goal;

[0058] Step 140, based on the whole double-layer optimization goal, a neural divide-and-conquer framework is used to solve the MDMTSP-V model, which comprises the following steps:

[0059] In the task allocation stage, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous speed features, and output a city point allocation subset for each agent;

[0060] The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters;

[0061] In the path optimization stage based on the sub-traveling salesman problem:

[0062] The path planning of each city point allocation subset is regarded as an independent sub-traveling salesman problem, and a traveling salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain a path planning scheme for each sub-traveling salesman problem;

[0063] All the path planning schemes of the sub-traveling salesman problems are merged to obtain the whole path planning scheme of the MDMTSP-V.

[0064] ​Specifically, in step 110, a problem definition and modeling are performed for the collaborative scheduling scenario of multiple AGVs with different speeds in warehousing and logistics planning.

[0065] Traditional MDMTSP typically optimizes the shortest path distance (Min-Sum or Min-Max). However, in MDMTSP-V, due to the different speeds of AGVs, the optimization objective naturally shifts to minimizing the shortest travel time. Specifically, in this invention, it minimizes the maximum travel time among all AGVs, directly addressing the following practical application requirements: all AGVs should be able to complete tasks simultaneously, avoiding waiting and achieving load balancing. The MDMTSP-V model abandons the ideal assumption of homogeneous agents, more realistically depicting the actual scenario in warehousing and logistics where different AGV models and battery capacities lead to speed differences.

[0066] Step 111: To achieve efficient and fair collaborative scheduling of multiple AGVs at different speeds, the first step is to define the problem in the model:

[0067] Mapping the task points that need to be accessed in the warehousing and logistics system (such as picking stations, shelf locations, receiving stations, etc.) to the multi-traveling salesman problem in allometric multi-warehouse operations. A set of coordinates for each city point. This represents the physical location of these task points in the two-dimensional space of the warehouse, describing the spatial characteristics; where, , .

[0068] The automated guided vehicles (AGVs) performing the tasks in Taiwan correspond to the multi-vehicle, multi-traveling salesman problem. Individual agent (travel agent);

[0069] The starting and stopping points / charging stations of AGVs correspond to the warehouses in the multi-warehouse, multi-traveling agent problem; each AGV (agent) departs from its respective warehouse, and the set of warehouses is... This indicates that the docking point / charging station is also located in the task space;

[0070] Each agent speed , indicating the first The heterogeneous speed (actual moving speed) of the AGV. The velocity set of all agents is represented by the velocity vector of the agent, denoted as . This describes the heterogeneous speed parameters of multiple AGVs; this heterogeneous speed characteristic reflects the performance differences caused by differences in AGV model, load, battery status, or drive system in real-world scenarios.

[0071] Step 112, Obtain model constraints and objective definitions:

[0072] All non-warehouse city points need to be assigned to agents, each city point must be assigned and visited only once.

[0073] Agent path (the complete journey of AGV path) is defined as:

[0074] (1)

[0075] Where agent (the th AGV) starts from its own warehouse and finally returns to ; represents the number of city points assigned to agent , is the city point assignment subset of agent , which satisfies , and when ; .

[0076] Agent path length (Euclidean distance) is defined as:

[0077] (2)

[0078] However, under the condition of different speeds, the key indicator to measure the efficiency of AGV (agent) operation is not the path length, but the actual travel time, so it is necessary to normalize the path length into travel time, and define the travel time of agent : .

[0079] Step 113, construct MDMTSP-V model, the optimization target (objective function) of the MDMTSP-V model is to minimize the maximum value of the travel time of all agents:

[0080] (3)

[0081] The MDMTSP-V model constructed by the above process focuses on the task fairness and synchronous completion efficiency of multiple agents under the condition of different speeds in the objective function, which directly corresponds to the load balancing and operation synchronization requirements in AGV scheduling, to ensure that no AGV becomes a system bottleneck due to slow speed or long task path, thereby improving the throughput and completion efficiency of the warehouse operation line as a whole.

[0082] Further, in step 120, to solve the multi-AGV cooperative scheduling problem described in the above MDMTSP-V model, the present application adopts the Divide-and-Conquer idea to divide the complex heterogeneous multi-depot multi-traveler problem into two serial and coupled optimization stages.

[0083] The first stage is the task allocation stage (Divide):

[0084] Design a neural allocation strategy ( is the allocation strategy parameter), distribute the non-depot city point set to agents according to probability, and output the city point allocation subset of each agent , aiming to achieve load balancing and optimal scheduling according to speed and spatial characteristics. The optimization core of this stage is to explicitly consider the speed

[0085] of each agent and the spatial distribution of task points in the allocation strategy, in order to achieve load balancing in terms of travel time and lay the foundation for subsequent path optimization.

[0086] The second stage is the path optimization stage based on the sub-traveler problem (TSP) (Conquer):

[0087] For each agent's city point allocation subset , together with the warehouse (warehouse node) , use an efficient Traveler Problem (TSP) solver to obtain the optimal path of the agent, aiming to minimize the path length .

[0088] In one embodiment, the TSP solver uses Google OR-Tools.

[0089] Google OR-Tools is an open source, cross-platform software suite developed by Google, used to solve the search problem of the optimal solution in combinatorial optimization problems, such as vehicle routing, scheduling, packing, linear programming, etc.

[0090] Further, in step 130, based on the two optimization stages, the optimization objective of the MDMTSP-V model is reconstructed as the overall bi-level optimization objective in the following form:

[0091] (4)

[0092] where denotes the TSP solver,​ are control parameters (hyperparameters) of the TSP solver.

[0093] The model solving of the present application mainly focuses on optimizing the allocation strategy parameters in the task allocation stage , while the TSP solver in the path optimization stage is regarded as a stable performance and fixed parameter component, that is, the control parameters of the TSP solver are pre-configured.

[0094] In step 140, the model of the variable-speed multi-bin multi-traveling salesman problem is solved by using the neural divide-and-conquer framework.

[0095] The whole process of model solving by using the neural divide-and-conquer framework is shown in Figure 2 . Mainly includes:

[0096] (1) Task allocation stage:

[0097] By using the graph embedding module, the city points and warehouses are taken as nodes of the graph, and the connection relationship of the nodes is taken as the edge in the graph; the coordinate (position) information of the city points or warehouses is converted into a vector representation, and the graph data is constructed by using the node features and edge connection relationship of the graph .

[0098] The graph data and the speed vector of the agent are taken as inputs and provided to the neural allocation strategy. The neural allocation strategy outputs the task allocation result, that is, the city point allocation subset (node subset) corresponding to each agent (AGV).

[0099] Figure 2 The allocation result is shown in the figure, forming a set composed of three subsets. Each subset contains the warehouse node corresponding to an agent (represented by the node marked with the corresponding agent serial number) and the city points allocated to the agent (represented by the ordinary node in the figure).

[0100] (2) Sub-TSP solving stage: Figure 2 In the figure, the three subsets output by the neural allocation strategy are regarded as independent TSP instances, and are input into the TSP solver to generate three sub-TSP solutions, each corresponding to an agent. Finally, these sub-TSP solutions are combined to obtain the complete solution of the MDMTSP-V problem.

[0101] It can be seen that the neural divide-and-conquer framework explicitly incorporates the speed of heterogeneous agents in the task assignment stage, and utilizes an efficient TSP solver in the path optimization stage. The key of the neural divide-and-conquer framework is to design a neural assignment strategy network to integrate spatial and speed information through the interconnected modules in the network to assign each city point in MDMTSP-V to an agent with heterogeneous speed. The detailed architecture of the proposed neural assignment strategy model is shown in Figure 3

[0102] Specifically, in step 141, in the task assignment stage, a neural assignment strategy network is constructed to integrate spatial features and heterogeneous speed features, and output the city point assignment subset of each agent.

[0103] First, a neural assignment strategy network is constructed, which at least includes a graph embedding module, a speed embedding module, an additive attention module, an agent embedding module and an assignment decision module.

[0104] First, the graph embedding module uses a composite message passing neural network (CMPNN) to extract spatial features from node coordinates, capturing geometric relationships without speed input. At the same time, the speed embedding module is used to encode heterogeneous speed features, specifically by using a DeepSet architecture to encode the speed of the agent, ensuring permutation invariance and equivariance.

[0105] In order to integrate spatial features with the speed of a specific agent and improve the quality of assignment, the additive attention module applies an additive attention mechanism based on the speed embedding to dynamically weight the node embedding according to the corresponding speed embedding, obtaining the agent embedding, which can capture the interaction between city location and agent capability. Using these integrated representations, the agent embedding module uses a multi-step attention mechanism to aggregate global graph context and speed-enhanced node information, thereby generating a comprehensive embedding to represent each agent.

[0106] Finally, the assignment decision module uses a dot-product attention mechanism to calculate the probability of node assignment to an agent, thereby forming a random assignment strategy, and outputting the city point assignment subset assigned to each agent.

[0107] The modular design of the neural assignment strategy network enables spatial features and speed features to jointly guide decision-making to achieve collaborative learning. By treating MDMTSP-V as a probabilistic assignment problem, the model can efficiently solve the sub-TSP problem while balancing the workload between heterogeneous agents.

[0108] The specific implementation of each module in the neural assignment strategy network is described in detail below.

[0109] (1) Graph Embedding Module ​

[0110] Input: Two-dimensional coordinates of all nodes (city points);

[0111] Structure: Employs a Composite Message Passing Neural Network (CMPNN), through... The message passing is iterated in rounds to spread spatial information, allowing each node to gradually aggregate neighbor information and finally form the embedding vector of the node describing the spatial characteristics.

[0112] Message calculation formula between nodes:

[0113] (5)

[0114] in, , is a node The generated embedding vector, This is a node The generated embedding vector, It is the preset dimension of the embedding vector (e.g., 64, 128, 256, etc.). ; Indicates from node Send to node The message vector is used to encode the node. For nodes The impact; Represents a node and nodes Relationship information; It is a message function, which is a learnable neural network.

[0115] In one embodiment, the learnable neural network employs a multilayer perceptron (MLP).

[0116] Perform neighbor aggregation and updates:

[0117] (6)

[0118] (7)

[0119] in, Represents a node Aggregate all neighboring nodes The total message obtained after the message, Represents a node The set of neighboring nodes; This refers to aggregate functions, such as summation functions. Average function Maximum value function Or attention weighting, etc.; Represents a node The current round ( embedding vectors of the last round (R) of the node embedding vectors of the last round (R) of the node embedding vectors of the next round (R+1) of the node embedding vectors of the next round (R+1) of the node is an update function, such as Gated Recurrent Unit (GRU) or MLP. The dimension of the embedding vector remains unchanged during the process of updating the embedding vector through iteration.

[0120] embedding vectors of the final node embedding vectors of the final node , which is used to represent the spatial features, where embedding vectors of the last round (R) of the node embedding vectors of the last round (R) of the node embedding vectors of the last round (R) of the node

[0121] The CMPNN (Compound Message Passing Neural Network) is an advanced graph neural network architecture designed specifically for processing graph data with complex structures and multiple relationship types. Through multiple message function parallel methods, adaptive weighted aggregation strategies, and attention mechanism integration, it supports more complex and detailed feature extraction compared to traditional message passing neural networks (MPNN).

[0122] (2) Velocity Embedding Module

[0123] Input: the speed of each agent (AGV), represents the speed of the agent , ;

[0124] Action: encode the speed features of each agent (AGV) to ensure the invariance and equivariance of the permutation of agent order;

[0125] Structure: DeepSet is used. First, individual feature mapping is used to convert the speed of each agent into a high-dimensional embedded individual speed feature representation. Then, high-dimensional embedding aggregation is performed to obtain global speed features. Finally, individual speed features and global speed features are combined (fused) to obtain the final speed embedding (speed feature) of each agent, which contains both individual speed information and global speed information.

[0126] The specific steps are as follows:

[0127] First, individual feature mapping is performed through a trainable linear mapping to convert the speed of each agent into a high-dimensional embedded individual speed feature representation:

[0128] (8)

[0129] where, is a trainable linear mapping, mapping results in is the agent 's individual velocity feature, is a trainable weight matrix, is a trainable bias.

[0130] In one embodiment, the trainable linear mapping is a multi-layer perceptron, is a multi-layer perceptron (MLP): .

[0131] Then aggregate the individual velocity features of all agents (i.e. combine the velocity information of all agents) to obtain the aggregated global velocity feature:

[0132] (9)

[0133] Here we use average aggregation (instead of summation) to keep the feature scale independent of the number of agents, and enhance the generalization ability.

[0134] Finally, fuse the individual velocity feature and the global velocity feature to obtain and output the velocity embedding of the agent :

[0135] (10)

[0136] where, are both another trainable mapping (e.g. multi-layer perceptron); the velocity embedding of the agent : .

[0137] The DeepSet is a neural network architecture for processing set data, whose core property is permutation invariance. Specifically, it extracts features of each element in the set (through a shared neural network), and aggregates the element features into a set-level representation through a permutation-invariant aggregation function (e.g. summation, average, max, etc.), and finally processes the representation through a neural network to obtain the output. The key property of the permutation-invariant aggregation function is that its output is the same regardless of the order of the elements in the set, which is very suitable for processing tasks such as point cloud, multi-agent / proxy system, set prediction, etc.

[0138] (3) Additive Attention module (Additive Attention, or Attention Fusion Module)

[0139] Purpose: Based on the additive attention mechanism, dynamically fuse the embedding vector of the spatial node and the velocity embedding of the agent to improve the expression and allocation quality.

[0140] Input: embedding vector of a node , ; velocity embedding of an agent , ;

[0141] For each node-agent pair , the dynamic fusion process includes:

[0142] First, feature alignment and fusion are performed using a multi-layer perceptron (MLP) to project the spatial and velocity features of different dimensions into the same hidden space, obtaining the fused feature vector:

[0143] (11)

[0144] wherein, is the output of processing the embedding vector of a node using a multi-layer perceptron (MLP), is the output of processing the velocity embedding using a multi-layer perceptron (MLP); is the spatial feature transformation weight, is the velocity feature transformation weight, is the dimension of the attention hidden space in the multi-layer perceptron (e.g., 64, 128, 256, etc.); the fused feature vector .

[0145] In one embodiment, the dimension of the attention hidden space in the multi-layer perceptron satisfies: .

[0146] The attention score is calculated as:

[0147] (12)

[0148] wherein, is the hyperbolic tangent activation function used to introduce a nonlinear transformation to compress the features into the range [-1, 1]; denotes the transpose of the attention weight vector ; is the unnormalized attention score, which is a scalar.

[0149] The attention score is normalized using the activation function to obtain the normalized attention weight:

[0150] (13)

[0151] Obviously, the above normalized attention weight satisfies the probability distribution condition, that is, for the same node , the attention weights of all agents satisfy the normalization condition: . The greater the value of , the more suitable the agent for the task of the node .

[0152] Finally, the normalized attention weights are used to weight the fusion of the node embedding vectors to obtain the attention-weighted node embedding:

[0153] (14)

[0154] The dimension of is consistent with .

[0155] (4) Agent Embedding Module (Agent Embedding or Agent Embedding Module)

[0156] Input:

[0157] Global context is composed of two parts: global spatial features and warehouse features ; the global spatial features are obtained by aggregating all node features, such as taking the sum average of embedding vectors;

[0158] Velocity embedding ;

[0159] Node embedding , ;

[0160] Purpose: aggregate global context, node embedding, velocity embedding, and form a rich agent vector;

[0161] First, by concatenating the global context and the velocity embedding, the initial query vector of the agent is constructed:

[0162] (15)

[0163] wherein the dimension of , is the dimension of .

[0164] Then, a nonlinear transformation is introduced using the activation function to obtain the query vector of the agent :

[0165] (16)

[0166] wherein is the query transformation weight matrix, is the transformation bias, is the query vector dimension, the query vector .

[0167] Compute the Key / Value vector:

[0168] (17)

[0169] where, is the Key vector, is the Key vector weight matrix, is the Key vector bias; is the Value vector, is the Value vector weight matrix, is the Value vector bias.

[0170] Next, perform attention allocation, recompute attention scores and normalized attention weights:

[0171] (18)

[0172] (19)

[0173] Finally, utilize the recomputed normalized attention weights to weight-sum the Value vector, obtain and output the Agent Embedding:

[0174] (20)

[0175] (5) Allocation Module

[0176] Input: Node Embedding , ; Agent Embedding ;

[0177] Utilize the Agent Embedding and Node Embedding to perform feature mapping, remove bias effect to regenerate Key vector and query vector:

[0178] (21)

[0179] Compute similarity:

[0180] (22)

[0181] Compute importance score:

[0182] (23)

[0183] where, is a preset scaling factor.

[0184] The distribution probability is calculated by scaling dot-product attention:

[0185] (24)

[0186] The distribution probability matrix .

[0187] The sampling distribution is performed to obtain the agent distribution result:

[0188] (25)

[0189] wherein, is a random distribution of a discrete random variable, indicates that for each node , an agent index is sampled according to the probability distribution to obtain the agent distribution result of the node , ; and:

[0190] .

[0191] The action vector of the agent distribution result is obtained for each node .

[0192] Secondly, the neural distribution strategy network is trained by using a reinforcement learning optimization and training mechanism to optimize the distribution strategy parameters, including:

[0193] (1) a policy gradient algorithm (Policy Gradient) is designed, including:

[0194] An optimization method similar to the REINFORCE algorithm is used to optimize the distribution strategy parameters , and the actual maximum normalized completion time is used as a reward signal:

[0195] (26)

[0196] In each round of training, the sampling distribution, reward value (Reward) calculation, and policy update are performed.

[0197] (2) a variance reduction baseline is designed, including:

[0198] To solve the reward distribution drift caused by heterogeneous speeds, a normalized baseline based on the heterogeneous speed parameter (the speed set of the agent) is designed:

[0199] (27)​

[0200] (28)

[0201] in, It is the average total path length of the batch. For the first In the batch instance, the first The actual path length of each proxy. This refers to the batch size.

[0202] Advantage function:

[0203] (29)

[0204] Design the strategy loss function:

[0205] (30)

[0206] in, These are the allocation strategy parameters to be optimized. It is a probability distribution, used as a parameterized policy function. Indicates the state Select (assign) action under conditions The probability (conditional probability); It is a logarithmic function; This indicates the calculation of the expected value.

[0207] Based on the above strategy loss function Stochastic gradient descent optimization is employed. The baseline is dynamically updated based on the actual path statistics of the current batch.

[0208] (3) Design and optimize the process, including:

[0209] Initialize all weight parameters and optimizer parameters of the neural assignment policy network;

[0210] The speed of batch sampling of city points and agents in each round of training is input into the neural allocation strategy network, and the allocation probability and sampling allocation scheme are output.

[0211] For each agent, assign a subset of city points and call the TSP solver to obtain the agent's path. and path length ;

[0212] Calculate the normalized maximum completion time Update the baseline To obtain the advantage function ;

[0213] according to Backpropagate gradients and update parameters;

[0214] Loop iteration until convergence or preset stop condition is met.

[0215] Thirdly, the trained neural assignment strategy network is utilized to fuse the spatial features and heterogeneous speed features, and each agent's city point assignment subset is re-outputted.

[0216] Next, in step 142, in the path optimization stage based on the sub-traveling salesman problem: the path planning of each city point assignment subset is taken as an independent sub-traveling salesman problem, and a traveling salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain the path planning scheme of each sub-traveling salesman problem.

[0217] Specifically, for the agent along with the finally obtained city point assignment subset: , the corresponding sub-TSP problem is solved by using a classic TSP solver such as Google OR-Tools to obtain the solution of the sub-TSP problem, that is, the optimal path of the agent , and the path optimization objective is to minimize the path length

[0218] Speed only affects the assignment decision, and does not affect the model structure and solver of the sub-TSP.

[0219] In step 143, the path planning schemes of all sub-traveling salesman problems are combined to obtain the final path set , and the overall path planning scheme of the heterogeneous speed multi-warehouse multi-traveling salesman problem is obtained.

[0220] In addition, the method steps of the present application further include step 150: calculating the strategy loss function using the overall double-layer optimization objective to evaluate the overall path planning scheme.

[0221] In summary, the present application proposes a heterogeneous speed multi-AGV neural divide-and-conquer scheduling method for warehouse logistics, which mainly obtains the following beneficial effects:

[0222] (1) For the first time in the cooperative scheduling scenario of heterogeneous speed multi-AGV for warehouse logistics, the heterogeneous agent feature is systematically introduced, and the DeepSet speed embedding and additive attention fusion mechanism are designed.

[0223] (2) The neural divide-and-conquer strategy divides the complex scheduling problem into stages, significantly reducing the optimization difficulty.

[0224] (3) The assignment strategy network structure is flexible and has strong expression, and can be directly generalized to different agent warehouse locations and different city scales without the need for further training.

[0225] (4) Variance reduction baseline innovative solution to reward distribution drift under allometric, improve training stability;

[0226] (5) The whole process can be modularized, easy to implement and industrialize.

[0227] In fact, the method of the application is also applicable to other engineering applications that can be refined into allometric multi-warehouse multi-traveler problems, and the specific implementation architecture can be summarized as:

[0228] (1) Modular system architecture:

[0229] Data input module: receive city point coordinates, warehouse information and agent speed parameters;

[0230] Neural allocation strategy module: including spatial feature extraction, speed feature embedding, attention fusion, agent embedding, allocation decision and other sub-modules;

[0231] Path optimization module: call efficient TSP solver for each agent allocation result;

[0232] Optimization training module: including reward calculation, baseline estimation, gradient update, etc.

[0233] Visualization and evaluation module: output allocation results and optimized paths, support performance statistics and effect analysis.

[0234] (2) Model parameter and hyperparameter setting:

[0235] Neural network parameter initialization of each layer;

[0236] Model iteration rounds , embedding dimension , attention depth , etc.

[0237] Optimizer type (such as Adam, SGD), learning rate, batch size ;

[0238] Baseline estimation constant Reference theoretical or empirical value setting or estimate according to the current batch;

[0239] The hyperparameters of the TSP solver, such as branch and bound depth and search strategy, can be flexibly adjusted.

[0240] In one embodiment, the specific implementation architecture process of a typical application includes:

[0241] (1) Input: 100 city point coordinates, 3 warehouses, and agent speeds ;

[0242] (2) Data into neural allocation strategy, get allocation probability matrix , sampling distribution scheme ;

[0243] (3) Call OR-Tools for each proxy subset allocation to obtain the optimal path of the agent and the corresponding path length ;

[0244] (4) Calculate and the baseline to obtain the advantage and update the strategy parameters;

[0245] (5) Iterative training for several rounds, output the optimal allocation and path, and visualize the synchronization completion time.

[0246] It can be seen that the application discloses a detailed, directly implementable neural divide-and-conquer and attention mechanism heterogeneous multi-warehouse multi-traveling salesman problem solving method, which has good scalability, scheduling fairness and industrial application value.

[0247] In addition, the application also verifies the effectiveness and superiority of the heterogeneous multi-AGV neural divide-and-conquer scheduling method for warehouse logistics through experiments, and designs and implements a series of systematic experiments. The experiments aim to evaluate the rationality of the method for modeling the heterogeneous speed of AGV (corresponding to the speed of the agent), the contribution of speed coding and attention mechanism, the influence of the speed-based variance reduction baseline on reinforcement learning optimization, and the performance and generalization ability of the method under different scales and various speed distributions. The following specifically introduces the experimental settings, comparison baseline, experimental results and analysis, and the advantages of the application in actual path planning are shown through the visualization results.

[0248] (I) Experimental setup

[0249] (1) Data generation

[0250] To simulate the heterogeneous multi-warehouse multi-traveling salesman problem scenario corresponding to the multi-AGV cooperative operation in warehouse logistics, the city nodes (representing shelves, workstations, etc. Task points) and warehouse coordinates are uniformly randomly sampled in a unit square region , ensuring no spatial bias and meeting typical combinatorial optimization path planning benchmarks. Node coordinates are input in the form of a tensor with shape , where represents the batch size, represents the total number of nodes in the instance.

[0251] To reflect the heterogeneous speed characteristics in actual vehicle fleets due to model, load, or performance degradation, the agent (i.e., AGV) speed vector is sampled in two stages:

[0252] First, the lowest speed threshold of the instance is uniformly sampled from the interval :

[0253] (31)

[0254] Subsequently, for each AGV corresponding to the speed of the agent:

[0255] (32)

[0256] This hierarchical sampling ensures that the speed of the agent within the instance shares a common lower bound, and the speed distribution between different instances is diverse, enhancing the robustness of the model. After sampling, the speed vector is normalized:

[0257] (33)

[0258] Ensure that the speed value range is unified in , which facilitates stable model training.

[0259] (2) Evaluate the baseline

[0260] To verify the performance advantage of the method in AGV scheduling, three types of classical algorithms are selected for comparison:

[0261] ORTools: Google's open-source combinatorial optimization library, which uses heuristic strategies such as guided local search, supports constraint path planning, and realizes it by adjusting the cost function for heterogeneous AGV scheduling problems;

[0262] Ant Colony Optimization (ACO): a probabilistic heuristic algorithm based on the foraging behavior of ants, which simulates pheromone to guide path search and supports speed-weighted path cost, commonly used for static AGV path planning;

[0263] Genetic Algorithm (GA): simulates the natural evolution mechanism, explores the solution space through selection, crossover and mutation, and adjusts the fitness function by embedding the speed factor, suitable for large-scale AGV scheduling combination.

[0264] The above baseline algorithms do not rely on the training phase, optimize each instance independently, and cover the typical paradigm of optimal solution search and approximate solution in the AGV scheduling field.

[0265] (3) Experiment scale and parameters

[0266] The experiment covers 11 node sizes to simulate different task quantities in the warehouse scenario:

[0267] ;

[0268] And two kinds of AGV fleet (agent) quantities:

[0269] ;

[0270] A total of 11 2=22 configurations, all with 512 instances sampled for average performance statistics. Model training is only performed on fixed agent numbers, and the generalization ability is evaluated across node scales in testing. Running time and maximum normalized completion time are used as key indicators to measure the efficiency and real-time performance of AGV systems.

[0271] (II) Experimental results and analysis

[0272] (1) Comparison of training and testing performance of matching scale

[0273] Table 1 Comparison of average maximum relative completion time and computation time (seconds) of different algorithms MDMTSP-V under different experimental settings:

[0274]

[0275] The average maximum normalized completion time (Max Cost) and computation time of the present method compared with three baselines are summarized in Table 1 under the settings of node number agent (AGV) number

[0276] The results show that the present system achieves the lowest Max Cost in all configurations, and the advantage becomes more significant as the problem scale increases. For example, the present method achieves an average maximum normalized completion time of 2.195, which is about 11% lower than the 2.465 of the suboptimal ACO, while the computation time is only about 0.04 seconds, which is much better than the tens of seconds of ACO and GA and the seconds of ORTools.

[0277] This phenomenon reflects that the present method efficiently captures the characteristics of AGVs with different speeds through speed perception embedding and attention mechanisms, improves the load balancing and task allocation of the vehicle fleet, and benefits from the computational efficiency brought by end-to-end training and divide-and-conquer structure, which better meets the real-time requirements of scheduling in dynamic warehouse environments.

[0278] (2) Large-scale generalization ability test

[0279] Table 2 Comparison of average maximum relative completion time and computation time (seconds) of each algorithm (M=5, N=200~1000)

[0280]

[0281] The above table shows the generalization performance of the model trained on nodes and tested on to nodes, with agent number .

[0282] ​Table 3 Comparison of average maximum relative completion time and calculation time (seconds) of each algorithm (M = 10, N = 200-1000)

[0283]

[0284] The above table shows that the model only in training and to the generalization performance of the node, the number of agents 10.

[0285] Even far beyond the training scale, the model still maintains the lowest Max Cost, and the performance degradation trend of ACO and GA is obviously weak. Especially in the largest scale , the Max Cost of the present application is 6.036, which is significantly better than 10.153 of ACO and 88.615 of GA. The traditional algorithm of ortools fails to complete the solution within the limited time.

[0286] In addition, the reasoning time of the present application slowly rises with the increase of the scale, still maintains the order of seconds, and is significantly better than the sharp increase of the traditional heuristic method, which shows good scalability and practical value when dealing with large warehouse mass task scheduling.

[0287] (3) Path planning visualization

[0288] As shown in Figures 4-7 , the path solving AGV agent scheduling result comparison of a small-scale instance is shown; as shown in Figures 8-11 , the path solving AGV agent scheduling result comparison of a medium-scale instance is shown. Figures 4-11 The horizontal coordinate and the vertical coordinate respectively represent the X coordinate and the Y coordinate of the position point in the two-dimensional plane.

[0289] Specifically, Figure 4 is the AGV agent scheduling result diagram solved by the ant colony algorithm under the condition of a small-scale setting of 10 AGVs and 50 city points, wherein the maximum time is the scheduling time 6.6998 of AGV9; Figure 5 is the AGV agent scheduling result diagram solved by the ant colony algorithm under the condition of a small-scale setting of 10 AGVs and 50 city points in the experiment of the present application, wherein the maximum time is the scheduling time 10.4449 of AGV0; Figure 6 is the AGV agent scheduling result diagram solved by the method of the present application under the condition of a small-scale setting of 10 AGVs and 50 city points in the experiment of the present application, wherein the maximum time is the scheduling time 2.5576 of AGV1; Figure 7For the small-scale setting condition of 10 AGVs and 50 city points in the experiment of the present application, the AGV agent scheduling result diagram solved by the method of the present application is shown in the figure, wherein the maximum time is the scheduling time of AGV1 numbered 2.2760.

[0290] Further, Figure 8 For the medium-scale setting condition of 10 AGVs and 200 city points in the experiment of the present application, the AGV agent scheduling result diagram solved by the ant colony algorithm is shown in the figure, wherein the maximum time is the scheduling time of AGV9 numbered 13.1765. Figure 9 For the medium-scale setting condition of 10 AGVs and 200 city points in the experiment of the present application, the AGV agent scheduling result diagram solved by the ant colony algorithm is shown in the figure, wherein the maximum time is the scheduling time of AGV9 numbered 35.9214. Figure 10 For the medium-scale setting condition of 10 AGVs and 200 city points in the experiment of the present application, the AGV agent scheduling result diagram solved by the method of the present application is shown in the figure, wherein the maximum time is the scheduling time of AGV0 numbered 20.5488. Figure 11 For the medium-scale setting condition of 10 AGVs and 200 city points in the experiment of the present application, the AGV agent scheduling result diagram solved by the method of the present application is shown in the figure, wherein the maximum time is the scheduling time of AGV5 numbered 3.5121.

[0291] As can be seen from the example diagram, the AGV travel path generated by the method of the present application is more evenly distributed, reasonably utilizes the speed advantage of different AGVs, and maximizes the acceleration of the overall operation completion time. Compared with the baseline method, part of the AGV load is often too heavy, the path is intertwined, resulting in an increase in the longest completion time. The visualization directly verifies the effectiveness of the speed embedding and attention mechanism in improving the speed AGV collaborative scheduling.

[0292] The visualization directly verifies the effectiveness of the speed embedding and attention mechanism in improving the speed AGV collaborative scheduling.

[0293] (4) Result summary

[0294] Based on the experimental results, the method of the present application significantly improves the calculation efficiency while maintaining a low maximum normalized completion time in multi-AGV path planning and exhibits good generalization ability. The embedding and attention mechanism based on speed enable the model to have stronger expression and scheduling optimization ability, effectively addressing the heterogeneous speed challenge of multi-warehouse AGV systems.

[0295] The solution of the foregoing embodiment focuses on simultaneously implementing an efficient task point allocation strategy, accurate heterogeneous AGV speed information fusion, and stable and high-quality training and optimization process, and finally aims to minimize the maximum normalized travel time to improve the fairness and actual effect of multi-AGV collaborative scheduling. The specific technical solutions include problem modeling, neural divide-and-conquer framework, allocation strategy network design, sub-TSP problem solving, reinforcement learning optimization mechanism, etc.

[0296] In view of the problems of ignoring the traveling salesman speed heterogeneity, unbalanced scheduling and low solving efficiency in the existing AGV scheduling technology, a heterogeneous speed multi-bin multi-traveling salesman problem solving method and device based on neural divide-and-conquer and attention mechanism are proposed. Relying on the technical solutions described in the application, the following remarkable beneficial effects are achieved:

[0297] (1) effectively fuse heterogeneous AGV speed information, improve scheduling fairness and optimization quality: by introducing AGV speed embedding module and combining additive attention mechanism, the application realizes the deep fusion of spatial node features and heterogeneous speed information. In the allocation stage, the model can dynamically adjust the city allocation strategy according to the speed difference between AGVs, avoiding the waste of resources and uneven load caused by the homogenization of speed assumption in traditional methods, greatly improving the balance of the maximum job completion time of each AGV. It can be better implemented so that slower AGVs are not over-allocated tasks, and faster AGVs reasonably bear more load, significantly improving the synchronous completion efficiency of the entire warehouse operation.

[0298] (2) Divide-and-conquer strategy reduces problem complexity and improves solving efficiency and adaptability: the application adopts a neural divide-and-conquer framework, which divides the complex heterogeneous multi-AGV scheduling problem into two stages of city allocation and sub-TSP path optimization. The allocation strategy is based on neural network learning, which quickly generates a reasonable city division, and the sub-problem is solved by an efficient TSP solver (such as Google OR-Tools) to obtain the optimal solution. This strategy takes into account the expansion of the warehouse scheduling problem, ensuring the speed sensitivity of the allocation, and fully utilizing existing efficient path optimization algorithms, achieving high-quality solutions while significantly shortening the calculation time, meeting the real-time rescheduling needs of the AGV system.

[0299] (3) variance reduction baseline design based on speed to ensure stable convergence of reinforcement learning

[0300] In view of the problems of reward distribution drift and excessive training variance caused by heterogeneous AGV speed, the application designs a variance reduction baseline based on the normalization of the sum of speeds, effectively incorporating the speed factor into the reward normalization, significantly reducing the gradient estimation variance, improving the training stability and convergence speed, and finally obtaining a better AGV scheduling strategy.

[0301] (4) Significantly improve the computing efficiency, suitable for large-scale warehouse AGV scheduling: Experiments show that the method of the application can greatly shorten the running time compared with traditional precise and heuristic algorithms while ensuring better maximum normalized completion time. For example, in the case of AGV path planning with thousands of task points, the reasoning time of the application is maintained at seconds, while the calculation time of traditional methods increases exponentially and is difficult to complete, effectively meeting the real-time planning needs of complex paths of multiple AGVs in actual industry.

[0302] (5) Modular design, good scalability and engineering implementation value: The technical solution of the application has clear structure and clear module division, covering data input, graph embedding, AGV speed coding, attention fusion, strategy sampling, path optimization and reinforcement learning optimization modules, which are easy to develop, debug and upgrade independently. The network structure supports migration to different AGV (agent) quantities and warehouse sizes without retraining, showing good generalization ability and practicality, and is easy to integrate into existing warehouse management systems (WMS).

[0303] In summary, the application solves the scheduling imbalance and solution efficiency bottleneck problem caused by heterogeneous speed in multi-warehouse AGV scheduling by deeply integrating heterogeneous AGV information and combining neural division and attention mechanism, providing a new efficient, stable and scalable heterogeneous multi-AGV path planning solution with high industrial application value and promotion prospects.

[0304] In one embodiment, the application also provides a heterogeneous multi-AGV neural division scheduling device for warehouse logistics, comprising:

[0305] The first module is used to construct a heterogeneous multi-warehouse multi-traveler problem model with the minimum maximum travel time as the optimization objective, taking the heterogeneous speed of AGV as the key parameter;

[0306] The second module is used to decompose the heterogeneous multi-warehouse multi-traveler problem into two optimization stages in series: the task allocation stage and the path optimization stage based on the sub-traveler problem;

[0307] The third module is used to reconstruct the optimization objective of the heterogeneous multi-warehouse multi-traveler problem model into a whole double-layer optimization objective;

[0308] The fourth module is used to solve the heterogeneous multi-warehouse multi-traveler problem model based on the whole double-layer optimization objective using a neural division framework, comprising:

[0309] In the task allocation stage, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous speed features, and output the city point allocation subset of each agent;

[0310] The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters;

[0311] In the path optimization phase based on the sub-tourist problem:

[0312] The path planning of each city point allocation subset is regarded as an independent sub-tourist problem, and a tourist problem solver is used to solve each sub-tourist problem in parallel to obtain the path planning scheme of each sub-tourist problem;

[0313] The path planning schemes of all sub-tourist problems are combined to obtain the overall path planning scheme of the heterogeneous multi-warehouse multi-tourist problem.

[0314] On the other hand, the present application provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to realize the steps of the heterogeneous multi-AGV neural divide-and-conquer scheduling method for warehouse logistics provided in any of the above embodiments. The computer device can be a server. The computer device includes a processor, a memory, a network interface and a database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control ability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with the external terminal through the network connection.

[0315] On the other hand, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the heterogeneous multi-AGV neural divide-and-conquer scheduling method for warehouse logistics provided in any of the above embodiments.

[0316] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0317] The details of the present application are as follows.

[0318] The technical features of the above embodiments can be combined in any way. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0319] The above embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application.

[0320] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A neural divide-and-conquer scheduling method for multiple AGVs with varying speeds in warehousing and logistics, characterized in that, The method includes: Step 110: Using the heterogeneous speeds of AGVs as key parameters, construct a model for the heterogeneous speed multi-warehouse multi-traveling salesman problem with the optimization objective of minimizing the maximum travel time; including: Step 111: Define the problem of collaborative scheduling of multiple AGVs with different speeds in warehouse logistics planning: Map the task points that need to be accessed in the warehousing and logistics system to the multi-warehouse multi-trading salesman problem. A set of coordinates of each city point. Describe spatial features; The AGVs performing the tasks in Taiwan correspond to the multi-warehouse, multi-traveling salesman problem. One agent; the AGV refers to an automated guided vehicle; Map the starting docking point / charging station of the AGV to the warehouse in the multi-warehouse multi-traveling salesman problem; Each AGV departs from its respective warehouse, and the set of warehouses is represented as follows: ; Each agent speed , indicating the first The heterogeneous speed of the AGVs ; Let the velocity vector of each agent represent the set of velocities of all agents, denoted as . Describes the heterogeneous speed parameters of multiple AGVs; Step 112, Obtain model constraints and objective definitions: Define proxy Path: ； Among them, agents From its own warehouse Departure, and eventual return ; Indicates assignment to the agent The number of city points as an agent A subset of city points is allocated, satisfying: ,and hour, ; ; Define proxy Path length: ； Step 113: Construct a multi-agency, multi-traveling salesman problem model. The optimization objective of this model is to minimize the maximum value of the travel time for all agents: ； Step 120 decomposes the multi-warehouse multi-traveling salesman problem into two sequentially coupled optimization stages: the task allocation stage and the path optimization stage based on the sub-traveling salesman problem; Step 130: Reconstruct the optimization objective of the multi-warehouse multi-trading salesman problem model into an overall bi-layer optimization objective; Step 140: Based on the overall bilayer optimization objective, a neural divide-and-conquer framework is used to solve the allometric multi-warehouse multi-trading salesman problem model, including: In the task allocation phase, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous velocity features, and output a subset of city point allocations for each agent; The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters. In the path optimization phase based on the sub-traveling salesman problem: The path planning for each city point subset is treated as an independent sub-traveling salesman problem. A travel salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain the path planning scheme for each sub-traveling salesman problem. By merging the path planning schemes of all the sub-traveling salesman problems, we obtain the overall path planning scheme for the allometric multi-warehouse multi-traveling salesman problem.

2. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 1, characterized in that, In step 120, the task allocation phase includes: designing a neural allocation strategy. Set up city points that are not warehouses Assigned according to probability Each agent outputs a value. City point allocation subset ;in, These are the allocation strategy parameters; The path optimization phase based on the sub-traveling salesman problem includes: assigning a subset of city points to each agent. Along with the warehouse The optimal path for the agent is obtained using a traveling salesman problem solver. The goal is to minimize the path length. .

3. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 2, characterized in that, The Traveling Salesman Problem solver uses Google OR-Tools.

4. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 1, characterized in that, Step 130 includes reconstructing the optimization objective of the allochthonous multi-warehouse multi-trading salesman problem model into a holistic two-layer optimization objective in the following form: ； in, This represents a solver for the traveling salesman problem. These are the hyperparameters pre-configured by the Traveling Salesman Problem solver.

5. The neural divide-and-conquer scheduling method for multiple AGVs with varying speeds for warehousing and logistics according to claim 1, characterized in that, In step 140, during the task allocation phase, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous velocity features, outputting a subset of city point allocations for each agent, including: A neural allocation strategy network is constructed, which includes at least: a graph embedding module, a velocity embedding module, an additive attention module, a surrogate embedding module, and an allocation decision module; Spatial features are extracted using a graph embedding module: a composite message-passing neural network is employed, through... The iterative message passing propagates spatial information, allowing each node to gradually aggregate neighbor information and generate embedding vectors that describe the spatial characteristics of the nodes. Speed ​​features are encoded using a speed embedding module: DeepSet is used to convert the speed of each agent into a high-dimensional embedded individual speed feature through individual feature mapping, and then the high-dimensional embedding is aggregated to obtain the global speed feature. Finally, the individual speed features and the global speed features are fused to obtain the final speed feature of each agent. Using an additive attention module to fuse spatial and velocity features: Based on the additive attention mechanism, the embedding vectors of spatial nodes and the velocity features of the agent are dynamically fused to obtain attention-weighted node embeddings; The proxy embedding module is used to obtain the proxy embedding: the global context, node embedding, and velocity features are input into the proxy embedding module, and the global context, node embedding, and velocity features are aggregated by the nonlinear transformation of the proxy embedding module to generate the proxy embedding. The allocation decision module outputs a subset of city point allocations for each agent: the node embedding and agent embedding are input into the allocation decision module, and the allocation probability is calculated through feature mapping; based on the allocation probability, sampling allocation is performed to obtain the agent allocation result; The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters; By utilizing a pre-trained neural allocation strategy network and fusing spatial features with heterogeneous velocity features, a new subset of city point allocations for each agent is output.

6. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 5, characterized in that, The step of training the neural allocation strategy network using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters includes: Initialize the parameters and optimizer of the neural allocation strategy network; The speed of batch sampling of city points and agents in each round of training is input into the neural allocation strategy network, and the allocation probability and sampling allocation scheme are output. For each agent, assign a subset of city points and call the Traveling Salesman Problem solver to obtain the agent's path. and the path length of the proxy ; Calculate the normalized maximum completion time ; Dynamically update baseline ; Obtain the advantage function ; Stochastic gradient descent algorithm is used to optimize the allocation strategy parameters. Including policy-based loss functions Backpropagate the gradient and update the parameters of the neural assignment policy network; The process is repeated until convergence or a preset stopping condition is met.

7. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 6, characterized in that, The stochastic gradient descent algorithm has the following maximum normalized completion time. As a reward signal: ； By normalization, a baseline based on heterogeneous velocity parameters is designed: ； in, It is the average total path length of the batch. For the first Agent in a batch of instances The actual path length, Batch size; The advantage function It is given by the following formula: ； The policy loss function is given by the following formula: ； in, It is a probability distribution. Indicates the state Selecting and assigning actions under certain conditions The conditional probability; It is a logarithmic function; This indicates the calculation of the expected value.

8. The neural divide-and-conquer scheduling method for multi-AGVs with varying speeds for warehousing and logistics according to claim 1, characterized in that, The method further includes step 150, which uses the overall two-layer optimization objective to calculate the strategy loss function and evaluate the overall path planning scheme.

9. A neural-based multi-AGV (Automated Guided Vehicle) scheduling device for warehousing and logistics, characterized in that, include: The first module is used to construct a model for the heterogeneous speed multi-warehouse multi-traveling salesman problem, using the heterogeneous speed of AGVs as the key parameter and minimizing the maximum travel time as the optimization objective; it includes: Submodule one is used to define the problem of coordinated scheduling of multiple AGVs with different speeds in warehouse logistics planning: Map the task points that need to be accessed in the warehousing and logistics system to the multi-warehouse multi-trading salesman problem. A set of coordinates of each city point. Describe spatial features; The AGVs performing the tasks in Taiwan correspond to the multi-warehouse, multi-traveling salesman problem. One agent; the AGV refers to an automated guided vehicle; Map the starting docking point / charging station of the AGV to the warehouse in the multi-warehouse multi-traveling salesman problem; Each AGV departs from its respective warehouse, and the set of warehouses is represented as follows: ; Each agent speed , indicating the first The heterogeneous speed of the AGVs ; Let the velocity vector of each agent represent the set of velocities of all agents, denoted as . Describes the heterogeneous speed parameters of multiple AGVs; Submodule 2 is used to obtain model constraints and objective definitions: Define proxy Path: ； Among them, agents From its own warehouse Departure, and eventual return ; Indicates assignment to the agent The number of city points as an agent A subset of city points is allocated, satisfying: ,and hour, ; ; Define proxy Path length: ； Submodule 3 is used to construct a model for the multi-warehouse multi-traveling salesman problem (MTS). The optimization objective of this model is to minimize the maximum value of the travel time for all agents: ； The second module is used to decompose the heterogeneous multi-warehouse multi-traveling salesman problem into two sequentially coupled optimization stages: the task allocation stage and the path optimization stage based on the sub-traveling salesman problem. The third module is used to reconstruct the optimization objective of the multi-warehouse multi-trading salesman problem model into an overall bi-layer optimization objective; The fourth module, based on the overall bi-layer optimization objective, uses a neural divide-and-conquer framework to solve the allometric multi-warehouse multi-trading salesman problem model, including: In the task allocation phase, a neural allocation strategy network is constructed to fuse spatial features and heterogeneous velocity features, and output a subset of city point allocations for each agent; The neural allocation strategy network is trained using a reinforcement learning optimization and training mechanism to optimize the allocation strategy parameters. In the path optimization phase based on the sub-traveling salesman problem: The path planning for each city point subset is treated as an independent sub-traveling salesman problem. A travel salesman problem solver is used to solve each sub-traveling salesman problem in parallel to obtain the path planning scheme for each sub-traveling salesman problem. By merging the path planning schemes of all the sub-traveling salesman problems, we obtain the overall path planning scheme for the allometric multi-warehouse multi-traveling salesman problem.

Images