A multi-channel order-oriented intelligent order distribution and logistics matching method

Abstract

The application discloses a kind of multi-channel order-oriented intelligent order distribution and logistics matching method, it is related to intelligent decision-making technical field, including the construction multi-channel order's performance knowledge graph, continuously obtains multi-channel order execution performance knowledge graph update;Based on the updated performance knowledge graph, the state transition graph containing directed transition edge is constructed, the state transition graph is embedded learning by graph neural network, after generating node embedding vector, first intrinsic reward, second intrinsic reward are constructed and aggregated, generate comprehensive intrinsic reward, then obtain total reward by accessing decay mechanism;After splicing total reward with embedding vector, action value is calculated by action value network, and the maximum action value is used as current order distribution and logistics matching decision;The application can improve multi-channel order distribution accuracy, logistics matching executability and order performance stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to an intelligent order allocation and logistics matching method for multi-channel orders. Background Technology

[0002] With the continuous integration of e-commerce, instant retail, omnichannel stores, wholesale distribution, and third-party platform businesses, order fulfillment systems have gradually evolved from the traditional single-warehouse delivery model to complex decision-making systems involving multi-channel order acceptance, multi-entity collaboration, multi-node fulfillment, and multi-capacity linkage. Existing order fulfillment technologies typically achieve functions such as order access, inventory query, warehouse allocation, carrier selection, and delivery route planning through data interaction between order management systems, warehouse management systems, transportation management systems, supplier management systems, and store systems. As business scales up, some systems are beginning to introduce heuristic algorithms, operations research optimization models, machine learning ranking models, or reinforcement learning methods to improve order allocation efficiency and logistics resource utilization. Furthermore, technologies such as knowledge graphs, graph neural networks, and state transition modeling are increasingly being applied to supply chain networks, warehousing and distribution collaboration, and fulfillment risk prediction to characterize the multi-dimensional relationships between orders, goods, warehouses, suppliers, vehicles, routes, and regions, thereby enabling fulfillment decisions to evolve from single-point rule matching to a structured, dynamic, and intelligent approach.

[0003] Existing technologies are mostly based on static rules or locally optimal indicators, such as inventory satisfaction, warehouse distance, or transportation costs. They struggle to simultaneously consider the dynamic coupling relationships between warehouse operational capacity, supplier fulfillment stability, real-time transportation capacity status, route reliability, and order priority. Furthermore, existing technologies typically treat order splitting and logistics matching as a one-off choice problem, lacking explicit modeling of fulfillment status transitions such as "single warehouse delivery, split order delivery, supplier replenishment, direct supplier delivery, transportation capacity reassignment, and route switching." This makes it difficult to make interpretable dynamic corrections in the event of inventory fluctuations, warehouse congestion, supplier delays, or transportation capacity unavailability. In addition, existing technologies often lack constrained exploration mechanisms when exploring new warehouses, new suppliers, new transportation capacity, or low-frequency delivery routes, easily leading to path dependence on mature resources. On the other hand, blindly introducing exploration may increase the risks of stockouts, delays, and abnormal fulfillment. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent order allocation and logistics matching method for multi-channel orders, solving the problems of existing technologies that are difficult to perform interpretable dynamic corrections and lack constrained exploration mechanisms.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent order allocation and logistics matching method for multi-channel orders, comprising, Construct a knowledge graph for the fulfillment of orders from multiple channels and continuously obtain updates to the knowledge graph for the execution and fulfillment of orders from multiple channels; Based on the updated performance knowledge graph, a state transition graph containing directed transition edges is constructed. By embedding the state transition graph into a graph neural network, node embedding vectors are generated. Then, the first intrinsic reward and the second intrinsic reward are constructed and aggregated to generate a comprehensive intrinsic reward. Finally, the total reward is obtained through an access decay mechanism. After concatenating the total reward with the embedded vector, the action value is calculated through the action value network, and the maximum action value is used as the current order and logistics matching decision. Perform hard constraint checks on the current order allocation and logistics matching decisions, and generate check results to determine whether it is necessary to call the particle swarm optimization algorithm to optimize the order allocation and logistics matching decisions.

[0007] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the construction of a multi-channel order fulfillment knowledge graph and the continuous acquisition and updating of the multi-channel order execution fulfillment knowledge graph are as follows: Construct order fulfillment units and generate unified internal order identifiers; Acquire multi-source data and perform preprocessing; After mapping standardized entities and generating nodes, edges are established between nodes to form a knowledge graph of contract fulfillment.

[0008] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the construction of a state transition graph containing directed transition edges is as follows: Retrieve candidate performance plans from the updated performance knowledge graph as performance status nodes; Establish directed transition edges between nodes in the performance status and construct decision actions; After calculating the edge weights based on the directed transition edges, a state transition graph is constructed.

[0009] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the step of generating node embedding vectors by embedding the state transition graph through a graph neural network includes: In each layer of the graph neural network, the information of the neighboring performance status nodes is aggregated according to the edge weights to obtain the embedding vector of each node.

[0010] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the step of constructing a first intrinsic reward and a second intrinsic reward and aggregating them to generate a comprehensive intrinsic reward includes: The first intrinsic reward is constructed based on the centrality of each fulfilling state node in the state transition graph; then, the second intrinsic reward is generated by calculating the inverse degree of the fulfilling state node. The first intrinsic reward and the second intrinsic reward are combined to obtain the comprehensive intrinsic reward.

[0011] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the step of obtaining the total reward through an access decay mechanism includes: The number of visits to the fulfillment status node is counted, and an access decay mechanism is constructed to decay and correct the comprehensive intrinsic reward, generating the decayed comprehensive intrinsic reward. Develop external rewards based on actual business objectives; The decayed intrinsic reward and extrinsic reward are weighted and combined to generate the total reward.

[0012] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the step of concatenating the total reward with the embedded vector includes: The total reward is normalized and scaled to a numerical range that matches the feature scale of the embedded vector; then the normalized total reward is used as a one-dimensional feature and concatenated with the embedded vector of the performance status node to form a fused feature vector.

[0013] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the step of performing item-by-item hard constraint checks on the current order allocation and logistics matching decisions, and generating check results to determine whether it is necessary to call the particle swarm optimization algorithm to optimize the order allocation and logistics matching decisions, includes: Based on the current order allocation and logistics matching decisions, check whether the available inventory of the goods in each warehouse in the current plan can meet the order demand; check whether the remaining processing capacity of the selected warehouse in the current decision cycle can support the outbound shipment of the order or its sub-orders; check whether the supplier can provide the corresponding goods, and check whether the estimated delivery time meets the order fulfillment time requirements; check whether the selected transportation capacity is available, whether it can reach the pickup point within the required time, and whether the load capacity meets the order transportation requirements. If all the check results are satisfied, the current order and logistics matching decision is determined to be the final result; if any check result is not satisfied, the current order and logistics matching decision is determined to be the result to be optimized, and then optimized using the particle swarm optimization algorithm.

[0014] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the optimization using the particle swarm optimization algorithm includes: The current order allocation and logistics matching decision is treated as an individual particle. By perturbing the individual particles, a population is generated for initialization. Define a particle fitness function; update the particle velocity and position using standard velocity and position update formulas; when the maximum number of iterations is reached, output the optimized optimal particle; extract the order allocation and logistics matching decision corresponding to the optimal particle for inspection; if all inspection results are satisfied, the order allocation and logistics matching decision corresponding to the optimal particle is determined as the final result; otherwise, the particle swarm optimization algorithm is executed again.

[0015] As a preferred embodiment of the intelligent order allocation and logistics matching method for multi-channel orders described in this invention, the optimization using particle swarm optimization further includes: During optimization, if a continuous loop occurs, the decision action will be immediately sent to the management personnel as a detection result of the current order and logistics matching decision not being met, for review.

[0016] The beneficial effects of this invention are as follows: By constructing a fulfillment knowledge graph that includes orders, goods, warehouses, suppliers, and transportation capacity, and forming a state transition diagram based on candidate fulfillment schemes, this invention can structurally express the decision-making processes in multi-channel order fulfillment, such as single-warehouse delivery, split-order delivery, supplier replenishment, direct supplier delivery, transportation capacity matching, and route switching. This improves the comprehensive perception of inventory, timeliness, cost, risk, and resource availability in the order allocation and logistics matching process. Furthermore, by learning fulfillment state characteristics through graph neural networks and combining intrinsic rewards, extrinsic rewards, and action value networks to select the current order allocation and logistics matching decision, multi-channel orders can generate better order allocation and logistics matching schemes based on real-time fulfillment resource status. At the same time, before the scheme is executed, warehouse inventory, warehouse processing capacity, supplier supply capacity, and transportation capacity feasibility are verified, and if they are not met, particle swarm optimization is used to optimize, thereby improving the accuracy of multi-channel order allocation, the feasibility of logistics matching, and the stability of order fulfillment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for intelligent order allocation and logistics matching for multi-channel orders.

[0019] Figure 2 A flowchart for generating the total reward.

[0020] Figure 3 A flowchart for obtaining the current order allocation and logistics matching decision. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-3 This is the first embodiment of the present invention, which provides an intelligent order allocation and logistics matching method for multi-channel orders, including the following steps: S1. Construct a knowledge graph for the fulfillment of orders from multiple channels and continuously obtain updates to the knowledge graph for the execution and fulfillment of orders from multiple channels; S1.1 Construct an order fulfillment unit and generate a unified internal order identifier; Specifically, an API gateway connects to e-commerce platforms, offline store systems, wholesale order systems, and third-party online mall systems. Each channel is connected using a unified interface authentication method to form an order fulfillment unit. When the order fulfillment unit is running, it pulls order data from multiple channels at preset time intervals. Then, the order number, order source channel, product code, product quantity, shipping address, expected delivery time, order priority, order creation time, payment status, and order service level fields are written into the order access cache, and a unified internal order identifier is generated for each order record. It should be noted that each type of channel is accessed using a unified interface authentication method, such as OAuth2.0, Access Token, or signature-based interface authentication.

[0025] S1.2. Acquire multi-source data and perform preprocessing; It should be noted that: multi-source data includes warehouse information, supplier information, and transportation resource information; preprocessing includes, but is not limited to, time alignment and outlier removal; Specifically, when accessing order data, the order fulfillment unit obtains warehouse information from the warehouse management system, supplier information from the supplier management system, and transportation resource information from the transportation management system or transportation capacity scheduling system through API interfaces; subsequently, time alignment and outlier removal are performed on various types of data.

[0026] It should be noted that warehouse information includes, but is not limited to, warehouse number, warehouse type, warehouse location, available storage locations, current inventory, frozen inventory, salable inventory, processing capacity, current workload, delivery coverage, historical fulfillment performance, historical on-time delivery rate, and warehouse order acceptance status. Supplier information includes, but is not limited to, supplier number, available goods, delivery time, historical delivery records, historical fulfillment success rate, delay rate, probability of abnormal fulfillment, and current availability status; Transportation capacity information includes, but is not limited to, capacity number, capacity type, load capacity, real-time location, current status, estimated availability time, historical operating trajectory, historical on-time rate, historical anomaly rate, and service area.

[0027] S1.3 After mapping standardized entities and generating nodes, establish edges between nodes to form a performance knowledge graph.

[0028] Specifically, orders, goods, warehouses, suppliers, and transportation capacity are uniformly mapped as standardized entities and used as nodes in the knowledge graph, forming order nodes, goods nodes, warehouse nodes, supplier nodes, and transportation capacity nodes. Edges are established between nodes based on the fulfillment relationship; the fields of inventory quantity, processing capacity, fulfillment success rate, service level, anomaly rate, current location, available time, historical on-time rate, and delivery timeliness are encapsulated as node attributes; Encapsulate the fields of transportation cost, delivery time, route stability, historical fulfillment success rate, replenishment time, inventory matching degree, distance, estimated arrival time, and anomaly probability as edge attributes; Based on the nodes and edges, a fulfillment knowledge graph is formed that includes orders, products, warehouses, suppliers, and transportation capacity.

[0029] It should be noted that: the order node is used to describe the demand characteristics of orders to be fulfilled; the product node is used to describe the product category, volume, weight, and inventory requirements; the warehouse node is used to describe the warehouse's inventory capacity, processing capacity, and coverage capacity; the supplier node is used to describe the supplier's supply capacity and supply stability; and the transportation capacity node is used to describe the transportation capacity and real-time availability of vehicles, riders, trunk carriers, or third-party delivery resources. Examples of establishing edges based on fulfillment relationships include: an "inclusion" edge between an order and a product, an "in stock" edge between a product and a warehouse, an "available" edge between a product and a supplier, an "loadable" edge between a warehouse and transportation capacity, a "deliverable" edge between a warehouse and a receiving area, a "replenishable" edge between a supplier and a warehouse, a "serviceable" edge between transportation capacity and a delivery area, an "allocable" edge between an order and a warehouse, and a "deliverable" edge between an order and transportation capacity. Secondly, construct node attributes, which are attribute vectors. For example, the attribute vector of a warehouse node may include salable inventory, frozen inventory, current workload, processing capacity per unit time, delivery coverage radius, historical on-time rate, and historical anomaly rate; the attribute vector of a supplier node may include the quantity of available goods, average delivery time, historical delay rate, probability of abnormal fulfillment, and current availability status; the attribute vector of a transportation capacity node may include load capacity, current location, real-time status, estimated availability time, historical on-time rate, and historical trajectory stability; the attribute vector of an order node may include order priority, quantity of goods, remaining expected delivery time, order source channel, and service level. In addition, edge attributes are used to describe the relationships between nodes. For example, the edge attributes between an order node and a warehouse node may include product inventory fulfillment rate, warehouse distance, warehouse coverage matching degree, estimated outbound time, and historical order fulfillment success rate; the edge attributes between a warehouse node and a transportation capacity node may include loadability, vehicle arrival time, transportation cost, historical on-time delivery rate, and current availability; and the edge attributes between a supplier node and a warehouse node may include delivery timeliness, replenishment cost, supply stability, and delay probability.

[0030] Furthermore, we continuously acquire order data from multiple channels and update the fulfillment knowledge graph accordingly.

[0031] It should be noted that: when the fulfillment knowledge graph is updated, the real-time inventory is updated to the corresponding warehouse node and the product-warehouse edge attribute when warehouse inventory changes, such as when a product is locked by a new order, released by a canceled order, replenished by an inbound order, or adjusted during inventory count; when warehouse operational capacity changes, such as when a warehouse experiences an increase in outbound backlog or a temporary shutdown, the current operational load and available processing capacity are updated to the warehouse node; when the location or status of transportation resources changes, such as when a vehicle changes from idle to en route, a rider accepts an order, a vehicle arrives at the warehouse, or a carrier becomes unavailable, the real-time location, current status, and estimated availability time are updated to the transportation node; when a supplier generates new fulfillment records, such as delivery delays, on-time delivery, or abnormal cancellations, the supplier node's historical fulfillment records, delay rate, and abnormal fulfillment probability are updated.

[0032] S2. Based on the updated performance knowledge graph, construct a state transition graph containing directed transition edges. After embedding the state transition graph into a graph neural network and generating node embedding vectors, construct the first intrinsic reward and the second intrinsic reward and aggregate them to generate a comprehensive intrinsic reward. Then, obtain the total reward through the access decay mechanism. S2.1 Retrieve candidate performance plans from the updated performance knowledge graph as performance status nodes; Specifically, for orders to be fulfilled, candidate warehouses, candidate suppliers, and candidate transportation capacity that can meet the requirements of the orders to be fulfilled are retrieved from the updated fulfillment knowledge graph, and candidate fulfillment plans are generated for the orders to be fulfilled; for any candidate fulfillment plan, it is converted into a fulfillment status node, which constitutes the candidate status space of the orders to be fulfilled. It should be noted that the retrieval of candidate warehouses, candidate suppliers, and candidate transportation capacity that can meet the requirements of orders to be fulfilled from the updated fulfillment knowledge graph is carried out according to the following rules: If a single warehouse can meet the total quantity of goods, a single warehouse fulfillment plan is generated; if a single warehouse cannot meet the total quantity of goods, but multiple warehouses can meet it together, an order splitting fulfillment plan is generated; if the current warehouse inventory is insufficient but the supplier can replenish the goods within the allowed timeframe, a fulfillment plan is generated after the supplier replenishes the goods; if the supplier can ship directly to the customer, a supplier direct shipping plan is generated. The fulfillment status node does not simply represent a warehouse or a transportation capacity, but rather the "complete status of an order under a certain candidate fulfillment combination". Therefore, the content of the fulfillment status node should include the order summary (such as product category combination hash, receiving area grid ID, expected delivery time), candidate warehouse set, candidate supplier set, candidate transportation capacity set, inventory fulfillment ratio, estimated outbound time, estimated delivery time, transportation cost, fulfillment risk, whether the order has been split, and whether replenishment has been required.

[0033] S2.2 Establish directed transition edges between performance status nodes and construct decision actions; Specifically, based on the candidate state space, directed transition edges are established between the performance state nodes, and corresponding decision actions are constructed for the directed transition edges.

[0034] It should be noted that when establishing directed transition edges, if a fulfillment plan can be transformed into another fulfillment plan through a decision action, such as changing from single-warehouse delivery to split-order delivery, changing from regular transportation capacity to expedited transportation capacity, changing from warehouse delivery to direct delivery from supplier, changing from the original warehouse to a forward warehouse, or changing from the original delivery route to an alternative route, then a state transition edge is established between the two fulfillment state nodes; and each edge corresponds to an executable decision action, which includes warehouse selection, order splitting, inventory locking, supplier replenishment, transportation capacity matching, route selection, and abnormal re-delivery. Suppose the current fulfillment status is "order shipped from warehouse W1 and matched with regular delivery capacity R1". If it is detected that warehouse W1's workload is increasing or its inventory is occupied by other high-priority orders, the "reassign warehouse" action can be executed, switching the order to the next fulfillment status of "order shipped from warehouse W2 and matched with delivery capacity R2". In this case, a directed transition edge is established between the previous and next status nodes. Similarly, if the current fulfillment status is "warehouse W2 has insufficient inventory but can wait for supplier S1 to replenish stock", and the "trigger supplier replenishment" action is executed, switching the fulfillment status to "supplier S1 replenishes stock and then warehouse W2 ships", a corresponding directed transition edge can also be established. Furthermore, if the current status is "order shipped from warehouse W3", and the "split order" action is executed, shipping product A from warehouse W3 and product B from warehouse W4, a directed transition edge is established from the single-warehouse status to the split order status.

[0035] S2.3 After calculating the edge weights based on the directed transition edges, construct the state transition diagram; Specifically, based on the directed transition edges, the edge weights are calculated, and a state transition graph is constructed according to the performance state nodes and edge weights. The expression for calculating edge weights is as follows: ; In the formula, Indicates the status of performance Transfer to performance status The right to the border, This represents the probability of fulfilling the performance conditions. Indicates the next performance status. Indicates the current performance status. This indicates the currently executing order splitting or logistics matching action. Indicates the moment of decision-making regarding contract fulfillment; It should be noted that in actual implementation, multiple candidate fulfillment states are first generated for each order to be fulfilled. Each fulfillment state represents the business status of the order under a specific fulfillment plan. For example, an order contains two items of product A and one item of product B, with the delivery address located in the eastern part of a city, and is expected to be delivered before 8 PM on the same day. After querying warehouse inventory and delivery coverage, the following fulfillment states may be generated: The order is shipped from warehouse W1 and matched with delivery capacity R1; the order is shipped from warehouse W2 and matched with delivery capacity R2; the order is split between warehouses W1 and W3 and shipped separately with different delivery capacities; the order is replenished by supplier S1 to warehouse W2 and then shipped by W2; the order is shipped directly by supplier S2 and matched with a third-party carrier. Each executable fulfillment plan can be considered a fulfillment state node. When calculating the weight of a directed transfer edge, the current fulfillment status, the next fulfillment status, and the currently executed order allocation or logistics matching action are taken as input. The output is a value describing the probability of successfully entering the next fulfillment status and meeting the fulfillment conditions after executing the action in the current fulfillment status. This probability can be between zero and one. The closer the value is to one, the more reliable and easier the transfer is to succeed; the closer the value is to zero, the higher the risk or the less likely the transfer is to be executed. To ensure reproducibility, the edge weight can be determined jointly through historical statistics, real-time rule scoring, probability calibration, and normalization. For example, in one implementation, samples similar to the current order, current state, next state, and current action are retrieved from historical order fulfillment records. Similarity criteria may include similar or identical order origin channels, similar or identical order priorities, similar or identical product categories, similar or identical delivery areas, similar candidate warehouses, similar candidate suppliers, similar candidate transport capacity types, similar or identical delivery routes, and similar or identical fulfillment periods. The number of times the same action was successfully executed and fulfillment conditions were met in these similar samples, along with the total number of times the action was executed, is counted. The proportion of successful executions to the total number of executions is used as the base probability. For example, in the past thirty days, among orders similar to the current order, there were one hundred instances of the action "switch to warehouse W2 and match transport capacity R2" being executed from "warehouse W1 normal delivery status." Of these, eighty-seven were fulfilled on time, and thirteen experienced delays, stockouts, or transport capacity anomalies. In this case, the historical base probability of this directed transfer edge can be set to 0.87.

[0036] Furthermore, in each layer of the graph neural network, the information of the neighboring performance status nodes is aggregated according to the edge weights to obtain the embedding vector of each node; The embedding vector of the current node is obtained by the following expression: ; In the formula, Indicates the first Nodes output by a layered graph neural network Embedded vector, Represents the ReLU activation function. Indicates the first The weight matrix of a layered graph neural network. Represents a node , Indicates belonging to, Represents a node The set of neighboring nodes, Represents a node The degree, Represents a node The degree, Indicates the first Nodes output by a layered graph neural network Embedded vector, Indicates the first Bias terms in layered graphical neural networks; It should be noted that when embedding the performance state nodes in the state transition graph using a graph neural network, the input features of the performance state nodes should first be uniformly encoded; the weight matrix and bias terms can be generated through random initialization. The input to a graph neural network includes a set of nodes, a set of edges, an initial feature vector for each node, and the weight of each edge. A graph neural network can employ a multi-layer message-passing structure. Each layer performs neighbor information aggregation and feature transformation for each node in the current graph's fulfillment state. For a given current node, its set of neighboring nodes is first determined. Neighboring nodes can include predecessor fulfillment state nodes pointing to the current node, or successor fulfillment state nodes pointed to by the current node. For example, in a preferred embodiment, both the predecessor edge and the successor edge are retained and treated as a directional message passing relationship, enabling the model to distinguish between the different business meanings of "switching out of the current scheme" and "switching in from other schemes". It should also be noted that for each neighboring node of the current node, the embedding vector of that neighboring node in the previous layer of the graph neural network is read, along with the edge weight connecting that neighboring node to the current node. Then, the embedding vector of the neighboring node is multiplied by the corresponding edge weight, ensuring that high-feasibility, high-success-rate, or high-stability fulfillment transitions have a greater impact on the current node, while low-feasibility, low-stability, or high-risk fulfillment transitions have a smaller impact. To avoid nodes with a particularly large number of connections having an excessively strong influence on other nodes, normalization is performed based on the degree of the neighboring nodes and the degree of the current node. The degree can be understood as the number of edges connected to a node in the state transition graph.

[0037] For example, in one implementation, if the state transition graph is a directed graph, the in-degree and out-degree can be counted separately. The in-degree represents how many fulfilling states can transition to the current fulfilling state, and the out-degree represents how many subsequent fulfilling states the current fulfilling state can transition to. When aggregating predecessor neighbors, the in-degree can be used for normalization; when aggregating successor neighbors, the out-degree can be used for normalization; when using both predecessor and successor neighbors, the sum of the in-degree and out-degree can be used as the node degree, or the predecessor and successor neighbors can be normalized separately before merging.

[0038] Preferably, the present invention configures a three-layer graph neural network. The first layer is used to map the original performance state features into hidden representations, the second layer is used to further aggregate performance state transition information within a two-hop range, and the third layer is used to output the final embedding vector; Assuming the initial feature dimension of a node is 128, the output dimension of the first layer can be set to 64, the output dimension of the second layer to 64, and the output dimension of the third layer to 32. Ultimately, each fulfillment status node receives a 32-dimensional embedding vector. During the training phase, samples are first constructed using historical order fulfillment data. Each historical order record can reconstruct the order information, candidate warehouse information, candidate supplier information, candidate transportation capacity information, final selected solution, and actual fulfillment result at that time. Fulfillment state nodes and state transition edges are generated based on historical records, and edge weights are calculated. For solutions that have been successfully fulfilled and delivered on time, a higher supervision signal can be given to their corresponding state node or transition edge; for solutions that have experienced delays, stockouts, transportation capacity anomalies, supplier anomalies, or customer cancellations, a lower supervision signal or a higher risk label can be given. The graph neural network learns through training how to map states with similar fulfillment effects to similar embedding spaces and states with significantly different fulfillment effects to different regions. Therefore, when a new order arrives, even if some candidate combinations do not have completely identical historical records, the graph neural network can infer the potential fulfillment effect of the combination based on the graph structure and neighbor information of similar fulfillment states. When training a graph neural network, it is essential to ensure that the training data and the online inference data use the same feature processing procedures. Specifically, the field dictionary, normalization parameters, missing value handling rules, node type encoding rules, and edge weight calculation rules used during the training phase should all be saved as part of the model configuration; the same configuration must be loaded during the online inference phase. If the system adds new order channels, warehouse types, or transportation capacity types, corresponding values ​​need to be added to the field dictionary, and the model should learn the meaning of the new values ​​through incremental training or retraining.

[0039] Furthermore, a first intrinsic reward is constructed based on the centrality of each fulfilling state node in the state transition graph; subsequently, the inverse degree of the fulfilling state node is calculated to generate a second intrinsic reward. The first intrinsic reward and the second intrinsic reward are combined to obtain the comprehensive intrinsic reward; Centrality, expressed as: ; In the formula, Indicates the performance status node centrality, Indicates the performance status node The set of neighboring nodes, Indicates the performance status node To the node The right to the border, This represents the sum of the weights of all directed transition edges in the state transition graph; The first intrinsic reward, expressed as: ; In the formula, Indicates the performance status node The first intrinsic reward, Indicates the centrality reward weight; The second intrinsic reward is generated by calculating the inverse degree of the fulfillment state node, expressed as:

[0040] ; In the formula, Indicates the performance status node inverse degree, Indicates the performance status node The number of neighboring nodes, Represents extremely small positive numbers (e.g., ), Indicates the performance status node The second intrinsic reward, This indicates the reward weight based on the inverse degree; It should be noted that the first intrinsic reward and the second intrinsic reward are aggregated by simply adding the sum of the first intrinsic reward and the second intrinsic reward. Before performing centrality calculations, the set of neighboring nodes for each fulfillment state node needs to be determined. This embodiment preferably uses outgoing neighbors as the set of neighboring nodes. That is, for any fulfillment state node, all subsequent fulfillment state nodes to which that node can be directly transferred through an executable fulfillment action are considered its neighboring nodes. In this embodiment, when calculating the centrality of a node in a performance state, the directed transition edge weights from the node to all its neighboring nodes are first obtained, and these edge weights are accumulated to obtain the total local transition weight of the node. This total local transition weight reflects the ability and stability of the current performance state to transition to other feasible performance states. If a performance state can stably transition to multiple high-quality successor states, or if its transition edge weights are high, then its total local transition weight is large. Then, the sum of the edge weights of all directed transition edges in the entire state transition graph is counted as the total global transition weight. Finally, the ratio of the total local transition weight of the node to the total global transition weight is calculated to obtain the centrality of the performance state node. Thus, centrality is not simply a matter of counting the number of neighbors, but considers both the neighbor connection relationships and the size of the directed edge weights. Even if a node has few neighbors, if the weights of its connected transition edges are high, it may still have high centrality; conversely, even if a node has many neighbors, if the feasibility of each transition edge is low, then its centrality will not be excessively amplified. The centrality reward weight of the first intrinsic reward is used to control the contribution strength of centrality in the overall intrinsic reward. The value range of the centrality reward weight can be set, for example, to 0.5 to 0.9; The inverse degree reward weight in the second intrinsic reward is used to control the contribution of low-connectivity state exploration to the overall intrinsic reward. The inverse degree reward weight can be set to an example value of 0.1 to 0.5. Ideally, the centrality reward weight is set at 0.7, and the inverse degree reward weight is set at 0.3, with a sum of 1. This value ensures that the overall intrinsic reward prioritizes fulfillment stability and secondarily explores low-connectivity states. The reason for setting the centrality reward weight at 0.7 is that fulfillment states with higher centrality typically correspond to order fulfillment schemes with high inventory satisfaction, stable transfer paths, numerous alternative fulfillment solutions, and high historical fulfillment success rates. Prioritizing such states can improve the stability of order allocation and logistics matching. The reason for setting the inverse degree reward weight at 0.3 is that low-connectivity fulfillment states typically correspond to new warehouses, new stores, new suppliers, new transportation capacity, or low-frequency delivery routes. Moderately exploring such states can increase the system's adaptability to new resources and low-frequency scenarios. However, there are fewer historical samples of such states, so their weight is lower than the centrality reward weight to avoid excessive exploration leading to increased fulfillment risk.

[0041] Furthermore, the number of visits to the fulfillment status nodes is counted, and an access decay mechanism is constructed to decay and correct the comprehensive intrinsic reward, generating the decayed comprehensive intrinsic reward. Develop external rewards based on actual business objectives; The diminished intrinsic reward and extrinsic reward are weighted and combined to generate the total reward; The process involves constructing an access decay mechanism and generating a decayed comprehensive intrinsic reward, expressed as follows: ; In the formula, Indicates the performance status node Decaying overall intrinsic reward This represents the intrinsic reward decay coefficient. Indicates the performance status node Number of visits, Indicates the performance status node Comprehensive intrinsic rewards; The extrinsic reward is constructed using the following expression: ; In the formula, Indicates external reward, This indicates that the reward will be delivered on time. Full performance reward This indicates a penalty for logistics costs. This indicates a delayed punishment. This indicates a penalty for stockouts. This indicates a penalty for splitting orders. This indicates penalties for abnormal performance of obligations; The weighted sum of the decayed intrinsic and extrinsic rewards is expressed as follows: ; In the formula, Indicates the total reward. Indicates the weight of external rewards. This represents the weight of the overall intrinsic reward after decay; It should be noted that the number of visits should be counted based on the fulfillment status node, not simply by warehouse, supplier, or transportation capacity. For example, the fulfillment status of warehouse A for ordinary next-day delivery orders has different time constraints and risk characteristics than the fulfillment status of warehouse A for high-priority instant delivery orders; the fulfillment status of the same vehicle R will also differ under different receiving areas, different load utilization rates, and different route conditions. Therefore, the number of visits needs to be linked to a specific fulfillment status node. The fulfillment status node can be determined by order type, product combination, candidate warehouse, candidate supplier, candidate transportation capacity, delivery route, whether the order is split, whether replenishment is required, estimated delivery time, and service level. For fulfillment statuses that are completely identical or have equivalent business functions, the same node can be used to identify and accumulate the number of visits; for fulfillment statuses with key business differences, they should be split into different nodes and counted separately. The access decay mechanism controls the decay rate through both the intrinsic reward decay coefficient and the number of accesses. The intrinsic reward decay coefficient can be set to a value greater than 0 and less than 1. Since multi-channel order fulfillment systems need to allow the model to perform a small number of repeated verifications of the same fulfillment status to avoid model judgment bias caused by accidental success or failure of a single fulfillment result, this embodiment sets the default value of the intrinsic reward decay coefficient to 0.95. With this value, if the number of accesses to a certain fulfillment status node is zero, the decayed comprehensive intrinsic reward is equal to the original comprehensive intrinsic reward. It should also be noted that external rewards consist of both positive incentives and penalties. Positive incentives include on-time delivery rewards and full fulfillment rewards; penalties include logistics cost penalties, delay penalties, stockout penalties, order splitting penalties, and abnormal fulfillment penalties. The external reward is obtained by adding the on-time delivery rewards and full fulfillment rewards, and then subtracting the logistics cost penalties, delay penalties, stockout penalties, order splitting penalties, and abnormal fulfillment penalties. This design ensures that external rewards simultaneously reflect service levels, cost control, and risk control. A solution that delivers on time, fully meets product demand, has low costs, and is free of abnormalities will have a higher external reward; conversely, a solution with delays, stockouts, excessive order splitting, or capacity anomalies will have a lower external reward. The rules for assigning values ​​to positive rewards and penalties are as follows: On-time delivery rewards measure whether the estimated or actual delivery time of an order is no later than the expected delivery time. If an order is delivered before or exactly at the expected delivery time, the on-time delivery reward is set to one; if an order is delivered later than the expected delivery time, the on-time delivery reward is set to zero.

[0042] The full fulfillment bonus measures whether all items in an order are fulfilled. If all items and quantities in an order can be fulfilled by the candidate fulfillment plan, the full fulfillment bonus is 1; if any item cannot be fulfilled or is insufficient in quantity, the bonus is calculated as a percentage of the fulfilled items relative to the total order demand. For example, if an order requires ten items and the current plan can fulfill eight, the full fulfillment bonus is 0.8. Logistics cost penalty reflects the basic cost components of transportation, warehousing, packaging, and order splitting costs for candidate fulfillment plans. For all candidate fulfillment plans generated for the same order, the estimated logistics cost for each plan is first calculated, then the lowest and highest costs among the candidate plans for that order are taken. If the highest cost is greater than the lowest cost, the logistics cost penalty for a particular plan is equal to the difference between that plan's cost and the lowest cost divided by the difference between the highest and lowest costs, multiplied by a cost penalty coefficient. The default cost penalty coefficient is 0.5. If the highest cost equals the lowest cost, the logistics cost penalty for all candidate plans is zero, indicating that the costs cannot distinguish between the plans. The delay penalty measures the extent to which the estimated or actual delivery time exceeds the expected delivery time. If the estimated delivery time of the candidate fulfillment plan is no later than the expected delivery time, the delay penalty is zero; if the estimated delivery time is later than the expected delivery time, the delay penalty is one. The stockout penalty measures the extent to which candidate fulfillment options cannot meet the quantity of goods required for an order. If all goods in an order can be fulfilled, the stockout penalty is zero; if some goods cannot be fulfilled, the stockout penalty is equal to the proportion of unfulfilled goods to the total order demand. The order splitting penalty measures the cost and experience impact of splitting an order into multiple fulfillment packages, multiple warehouses, or multiple suppliers. If the order is not split, the splitting penalty is zero; if the order is split into two packages, the penalty is 0.1%; if it's split into three packages, the penalty is 0.2%, and so on, with an additional 0.1% penalty for each additional package, capped at 0.5%. An additional package refers to one less package than the total number of packages. For example, if an order is split into three packages, there are two additional packages, resulting in a splitting penalty of 0.2%. Abnormal fulfillment penalties are used to reflect the abnormal risks of warehouses, suppliers, transportation capacity, and routes in candidate fulfillment plans. The comprehensive abnormal risk can be calculated based on the historical abnormality rate of warehouses, the probability of abnormal fulfillment by suppliers, the historical abnormality rate of transportation capacity, the probability of route instability, and the current real-time status. The comprehensive abnormal risk is a weighted average. If the candidate plan does not involve supplier replenishment, the supplier abnormality probability can be zero or not included in the average; if the candidate plan does not involve multiple routes, only the abnormality probability of the actual routes involved is used. The default weights can be set to 30% for warehouse abnormality rate, 20% for supplier abnormality probability, 30% for transportation capacity abnormality rate, and 20% for route abnormality probability. If a certain item is missing, its weight is proportionally allocated to the remaining items, and the value of the abnormal fulfillment penalty should be limited to between zero and one. Secondly, when calculating external rewards, the scales for each reward and penalty item should be standardized. On-time delivery rewards and complete fulfillment rewards are defaulted to between zero and one; logistics cost penalties, delay penalties, out-of-stock penalties, order splitting penalties, and abnormal fulfillment penalties are also defaulted to between zero and one. Therefore, the calculated external rewards are typically between negative and positive numbers. To ensure scale stability when subsequently integrated with internal rewards, external rewards can be normalized or truncated. In this implementation, external rewards are truncated to between -1 and 2; if the calculated result is less than -1, it is taken as -1; if it is greater than 2, it is taken as 2. It should also be noted that: the weight of external rewards The weight of the decayed overall intrinsic reward It can be set to a value greater than 0 and less than 1, and the sum of the two is preferably 1; Superior extrinsic reward weighting Take a weight of 0.7 for the overall intrinsic reward after decay. We set it to 0.3 to ensure a balance between performance stability and exploratory nature.

[0043] Furthermore, after obtaining the embedding vector and total reward corresponding to each fulfillment status node, the total reward is normalized and scaled to a numerical range that matches the feature scale of the embedding vector; then, the normalized total reward is used as a one-dimensional feature and concatenated with the embedding vector of the fulfillment status node output by the graph neural network to form a fused feature vector; and the action value of the decision action is calculated, and the decision action with the maximum action value is selected as the current order and logistics matching decision.

[0044] It should be noted that: after obtaining the embedding vector and total reward corresponding to each performance state node, the total reward is normalized and scaled to a numerical range that matches the feature scale of the embedding vector; the normalized total reward is used as a one-dimensional feature and concatenated with the performance state node embedding vector output by the graph neural network to form a fused feature vector; while the action value is the network, which can be implemented using a multi-layer fully connected neural network, for example, using a two-layer fully connected neural network as the action value network; Assume the current order originates from an e-commerce platform, containing two items of product A and one item of product B, with a delivery expectation of four hours. After generating the current pending fulfillment status node, the graph neural network outputs the embedding vector of this node. Subsequently, four candidate actions are generated: the first action is to allocate the entire order to the main warehouse W1 and match it with vehicle R1; the second action is to allocate the entire order to the forward warehouse W2 and match it with local delivery capacity R2; the third action is to split the order into W1 and the store warehouse W3, and match them with R1 and rider R4 respectively; the fourth action is for supplier S1 to replenish the goods to W2, and then R2 will deliver the goods. Action features are constructed for each of these four actions, including inventory fulfillment rate, estimated outbound time, estimated delivery time, cost, whether to split the order, replenishment time, delivery capacity availability, and anomaly risk. The action value network calculates the action value of each of the four actions. If the second action has the highest action value and satisfies the hard constraints of inventory, timeliness, delivery capacity, and route, then the second action is selected as the order splitting and logistics matching decision for the current order, i.e., the order is allocated to the forward warehouse W2 and matched with local delivery capacity R2.

[0045] It should also be noted that the loss function of the action value network can be the error between the target action value and the predicted action value, and its expression is: ; In the formula, The loss function value represents the value of the action. This indicates the number of samples used in each training session. Indicates the first The target action value of each sample This represents the predicted action value under the network parameters of the action value. Indicates the first The performance status of each sample Indicates the first The process of matching individual samples with orders or logistics; Graph neural networks can be jointly trained based on the performance results. For example, if two performance states have similar performance results in the past, such as similar delivery time, similar costs, and similar anomaly probabilities, their embedding distance should be closer; if the performance results of two performance states differ greatly, their embedding distance should be farther. Preferably, the present invention employs an embedding loss based on result similarity, expressed as: ; In the formula, This represents the embedding loss value of the graph neural network. Indicates belonging to, This represents the set of fulfillment state node pairs used for training. The nodes representing a pair of nodes With nodes , Represents a node With nodes The similarity is calculated based on the performance results. This indicates the transpose operation. Represents norm operations. Represents a node Embedded vector, Represents a node The embedding vector; Among them, similarity Generated based on a preset similarity table, as shown in the figure: On time and complete 1.00 0.85 0.75 0.60 0.40 0.10 Orders split on time 0.85 1.00 0.65 0.75 0.45 0.10 Slight delay in completion 0.75 0.65 1.00 0.85 0.50 0.15 Slight delay in order splitting 0.60 0.75 0.85 1.00 0.55 0.15 Abnormal but completed 0.40 0.45 0.50 0.55 1.00 0.25 Failure to fulfill contract 0.10 0.10 0.15 0.15 0.25 1.00

[0046] In the preset similarity table, the diagonal line has a value of 1, indicating that two fulfillment status nodes corresponding to the same fulfillment result type are completely similar; the off-diagonal line values ​​are preset according to the business similarity between the fulfillment result types. For example, both on-time complete fulfillment and on-time split fulfillment meet the timeliness requirements and ultimately complete the order, so the similarity is high and set to 0.85; both on-time complete fulfillment and slightly delayed complete fulfillment fully meet the product requirements, but their timeliness performance is different, so the similarity is set to 0.75; both slightly delayed complete fulfillment and slightly delayed split fulfillment have slight delays and ultimately complete the fulfillment, so the similarity is set to 0.85; both abnormal but completed fulfillment and fulfillment failure reflect high fulfillment risk, so the similarity can be set to 0.25; on-time complete fulfillment and fulfillment failure have the greatest difference in fulfillment results, so the similarity is set to 0.10.

[0047] It should be further noted that the action value network and the graph neural network can be trained synchronously; Ideally, during synchronous training, the overall loss is a weighted fusion of the action value loss and the embedding loss, expressed as: ; In the formula, This represents the overall loss function value. The weights represent the embedding loss; The parameters are updated iteratively using the Adam optimizer. During the iteration process, the iteration stops when the overall loss value no longer decreases, and the trained action value network and graph neural network are output. Secondly, the weights of the embedding loss are used to control the contribution of the graph neural network embedding learning target to the overall training target. Therefore, the weights of the embedding loss can be set in the range of 0 to 1. Ideally, the weight of the embedding loss should be set to 0.2. This allows the action value loss to be the main optimization objective and the embedding loss to be the auxiliary constraint during synchronous training. This enables the graph neural network to learn the similarity of results between performance states, but does not allow the embedding space constraint to excessively affect the value judgment of the action value network on specific order and logistics matching actions.

[0048] S4. Perform hard constraint checks on the current order allocation and logistics matching decisions, generate check results to determine whether it is necessary to call the particle swarm optimization algorithm to optimize the order allocation and logistics matching decisions; Specifically, based on the current order allocation and logistics matching decision, check whether the available inventory of the goods in each warehouse in the current plan can meet the order demand; check whether the remaining processing capacity of the selected warehouse in the current decision cycle can support the outbound shipment of the order or the sub-orders split from the order; check whether the supplier can provide the corresponding goods, and check whether the estimated delivery time meets the order fulfillment time requirements; check whether the selected transportation capacity is available, whether it can reach the pickup point within the required time, and whether the load capacity meets the order transportation requirements; If all the check results are satisfied, the current order and logistics matching decision is determined to be the final result; if any check result is not satisfied, the current order and logistics matching decision is determined to be the result to be optimized, and then optimized using the particle swarm optimization algorithm.

[0049] It should be noted that when checking whether the available inventory of each warehouse in the current plan can meet the order demand, it is necessary to read the details of each warehouse involved in the current order splitting and logistics matching decision, as well as the corresponding goods handled by each warehouse. If the order is not split, it is only necessary to verify whether the selected warehouse can meet all the goods and quantities in the order; if the order is split into multiple sub-orders, it is necessary to verify whether the warehouse corresponding to each sub-order can meet the goods and quantities in that sub-order. When checking whether the remaining processing capacity of the selected warehouse within the current decision-making cycle can support the outbound shipment of the order or its sub-orders, it is necessary to read the remaining processing capacity of the selected warehouse within the current decision-making cycle. Warehouse processing capacity can include picking capacity, verification capacity, packing capacity, handover capacity, outbound platform capacity, and wave processing capacity. For a standard implementation, warehouse processing capacity can be simplified to the number of orders or items that can be processed in the current cycle; for a refined implementation, it can be calculated separately by product category, temperature zone, packaging type, wave, and outbound channel. The remaining warehouse processing capacity should equal the total processing capacity of the current cycle minus the capacity occupied by scheduled orders, the capacity occupied by locked pending orders, the capacity occupied by abnormal orders, and the capacity reserved by staff. If the processing demand of the current order or sub-order is less than or equal to the corresponding warehouse's remaining processing capacity, the warehouse processing capacity is satisfied; otherwise, it is deemed unsatisfactory. When checking whether a supplier can provide the corresponding goods and whether the estimated delivery time meets the order fulfillment time requirements, the first step is to determine whether the current plan involves a supplier. If the current plan involves warehouse stock delivery and does not involve supplier replenishment or direct delivery, the supplier verification can be recorded as passed. If the current plan involves supplier replenishment to the warehouse, direct delivery to the user, supplier allocation to stores, or supplier provision of out-of-stock goods, a supplier capability verification must be performed. Supplier capability verification includes goods availability verification, supply quantity verification, supplier current status verification, delivery time verification, service area verification, and anomaly risk verification. Subsequently, the supplier's available goods list is read to confirm whether the supplier can provide the corresponding goods in the order or sub-order; the supplier's current available quantity is read to confirm whether the available quantity is greater than or equal to the required quantity; the supplier's current fulfillment status is read to confirm whether it is in a state of accepting orders, being able to ship, or being able to replenish; the supplier's estimated preparation time and transportation time to the warehouse or user are read to confirm whether the estimated delivery time meets the order fulfillment time requirements. When checking whether the selected transport capacity is available, can reach the pickup point within the required time, and whether the load capacity meets the order's transportation needs, the system reads the selected transport capacity resources in the current plan and checks the capacity status, arrival time at the pickup point, load capacity, loading restrictions, service area, and historical trajectory reliability. It checks whether the current status of the selected transport capacity is available. If the transport capacity status is: accepted order, en route and unable to accept new orders, under maintenance, offline, resting, abnormal, timed out of location, or unavailable, the transport capacity verification is deemed to have failed. If the transport capacity status is available or can be booked, the system continues to check the estimated arrival time between its current location and the pickup point. Based on real-time location, road distance, traffic conditions, and transport capacity type, the estimated arrival time is calculated, and it is determined whether it can arrive before the order's required pickup time or outbound completion time. The transport capacity should be determined based on the total weight, volume, number of items, and special transport requirements of the order or sub-order. For example, this involves reading the maximum weight, maximum volume, transportable product categories, temperature control capabilities, embargoes, and currently occupied capacity of the transport resources. If the total weight of the order exceeds the remaining transport capacity, or the total volume exceeds the remaining transport capacity, or the goods require cold chain logistics but the transport capacity lacks cold chain capabilities, or the goods belong to a category that the transport capacity cannot transport, then the transport capacity verification fails. For order splitting schemes, the transport capacity selected for each sub-order should be verified separately; if multiple sub-orders share the same transport capacity, then the combined total weight, total volume, and delivery sequence should be verified to ensure feasibility. Capacity time verification also includes the estimated transit time from the pickup point to the delivery address, loading or handover time, route congestion time, and necessary buffer time. If the estimated arrival time at the pickup point is later than the warehouse's available handover window, or the estimated delivery time is later than the expected delivery time of the order, or the delivery service area does not cover the delivery address, the capacity verification will fail. For guaranteed delivery or instant delivery orders, stricter buffer times can be set, such as requiring the estimated delivery time to be at least ten or fifteen minutes earlier than the expected delivery time, to prevent defaults due to traffic fluctuations. For regular orders, more lenient buffer times can be set.

[0050] Furthermore, when optimizing using the particle swarm optimization algorithm, the current order allocation and logistics matching decision is treated as individual particles. By perturbing these individual particles, a population is generated for initialization. Define the particle fitness function; update the particle velocity and position using standard velocity and position update formulas; when the maximum number of iterations is reached, output the optimized optimal particle; extract the order allocation and logistics matching decision corresponding to the optimal particle for inspection; if the inspection results are all satisfied, the order allocation and logistics matching decision corresponding to the optimal particle is determined to be the final result; otherwise, the particle swarm optimization algorithm is executed again. The particle fitness function is expressed as follows: ; In the formula, Indicates the first The fitness function value of each particle. Indicates the first The estimated logistics cost per particle. This represents the delay penalty coefficient. Indicates the first The expected delivery delay for each particle.

[0051] It should be noted that: for estimated logistics costs, the candidate order matching schemes for each particle must first be retrieved; if the scheme is single-warehouse delivery, only the warehouse operation cost, transportation cost from the warehouse to the delivery address, and last-mile delivery cost are calculated; if the scheme is multi-warehouse split delivery, the warehouse operation cost, transportation cost, and last-mile delivery cost for each sub-order are calculated separately, and the costs of each sub-order are summed; if the scheme involves supplier replenishment, the supplier replenishment cost to the warehouse also needs to be calculated; if the scheme involves direct delivery from the supplier, the supplier direct delivery cost needs to be calculated; after determining the cost calculation scope, the operation billing rules of the candidate warehouses are retrieved from the warehouse management system. Warehouse operation costs can be determined based on the number of items in the order, the volume of the items, the weight of the items, the picking complexity, the packaging complexity, and whether expedited delivery is required. In specific operations, first, the number of different types of goods and the total number of items for the current order to be shipped from this warehouse are counted, and then the basic outbound fee, single-item picking fee, single-item packaging fee, special packaging fee, and expedited operation fee corresponding to this warehouse are retrieved. If the order consists of general merchandise, the standard picking and packing rate applies. If the order consists of cold chain goods, bulky items, fragile items, or items requiring special packaging, the corresponding special handling rate is applied based on the product label. If the order has a high priority or requires expedited processing, an expedited processing fee is added to the basic handling cost. This yields the warehouse outbound and handling costs for this candidate solution. Next, it is determined whether the solution involves suppliers. If the candidate solution does not involve supplier replenishment or direct delivery from suppliers, the supplier-related costs are recorded as zero. If the candidate solution involves supplier replenishment, the supplier's replenishment quote, replenishment transportation costs, minimum order quantity, emergency replenishment surcharge, and potential cross-warehouse transfer costs are obtained from the supplier. The replenishment cost is determined based on the quantity of out-of-stock items, the distance from the supplier to the target warehouse, the delivery method, and the delivery time requirements. If the supplier charges based on the quantity of goods, the replenishment cost is obtained by multiplying the quantity of replenished goods by the corresponding unit price; if the supplier charges based on the number of replenishment batches, the replenishment cost is determined based on the number of replenishment batches, the minimum order quantity, and transportation costs; if the current order requires the supplier to complete replenishment within a short period of time, the emergency replenishment cost is added. If the candidate solution involves direct delivery from suppliers, the supplier's direct delivery service fee, supplier's side packaging fee, supplier's delivery fee to the delivery address, and the direct delivery settlement rules agreed upon between the platform and the supplier are retrieved to obtain the supplier's direct delivery cost. After obtaining warehouse operation costs and supplier-related costs, the next step is to obtain transportation costs. Transportation costs mainly consist of trunk transportation costs, intra-city transportation costs, or city distribution costs between the warehouse or supplier's shipping point and the target delivery area. First, the shipping origin and destination are determined based on the candidate options. If the shipment is from the warehouse, the origin is the selected warehouse location; if it's a direct shipment from the supplier, the origin is the supplier's shipping address; if the supplier replenishes stock and then the warehouse ships the goods, the replenishment segment originates at the supplier's location, and the replenishment segment ends at the warehouse location; the order delivery segment originates at the warehouse location, and the destination is the order's delivery address. Map services are then used to obtain the estimated distance from the origin to the destination, estimated travel time, transportation route type, and corresponding freight rate. If the transportation capacity uses fixed-route pricing, the cost is obtained based on the route quote; if the transportation capacity uses distance pricing, the cost is calculated by multiplying the distance between the origin and destination by the unit distance rate; if the transportation capacity uses weight or volume pricing, the cost is calculated by multiplying the order weight, volume, or chargeable weight by the corresponding rate; if the transportation capacity is for instant delivery, express delivery, or dedicated vehicle delivery, the express surcharge, base price, mileage fee, and waiting fee are retrieved according to the corresponding service type. For multi-warehouse order splitting schemes, the transportation cost of each sub-order is calculated separately and then added together to form the transportation cost of the scheme. Finally, the last-mile delivery cost is obtained. The last-mile delivery cost represents the expense required to deliver goods from the last-mile station, store warehouse, forward warehouse, delivery point, or carrier's last-mile node to the delivery address. The last-mile delivery cost is determined based on the order's delivery address area, delivery distance, delivery service level, delivery time, number of packages, package weight, and last-mile delivery capacity type. If the order is completed by same-city instant delivery, the last-mile delivery cost can be directly obtained from the instant delivery platform's quotation interface; if the order is delivered by in-house capacity, it is calculated according to the internal last-mile delivery rate table; if the order is delivered by a third-party courier or carrier, it is calculated based on the carrier's quotation table, initial weight and subsequent weight rules, regional surcharges, and service level surcharges. If the order is split into multiple packages, the last-mile delivery cost is calculated separately for each package and then added together; if multiple packages can be delivered together, the last-mile cost is calculated once according to the combined delivery rules, and additional package handling fees are added. After obtaining the warehouse operation cost, supplier-related costs, transportation costs, and last-mile delivery cost, all costs are added together to obtain the estimated logistics cost for the corresponding candidate solution for each particle. For the anticipated delivery delay, the expected delivery time of the order is first obtained. This expected delivery time can come from the original order fields, such as the user-selected scheduled delivery time, the latest delivery time promised by the channel, the agreed arrival time for wholesale orders, or the store's required delivery time for offline store orders. Then, the expected delivery time is calculated based on the candidate solutions corresponding to the particles. The expected delivery time is not given directly by a single field, but is obtained by summing the estimated time consumption of multiple fulfillment stages. For the warehouse delivery solution, the following are calculated sequentially: order allocation confirmation time, inventory locking time, warehouse picking time, verification time, packaging time, outbound waiting time, capacity arrival time, loading or handover time, transportation time, last-mile delivery time, and receipt buffer time. For the supplier replenishment-followed warehouse delivery solution, the supplier's preparation time, the supplier's transportation time to the warehouse, and the warehouse receiving and shelving time or cross-docking time also need to be added. For the supplier direct delivery solution, the supplier's order confirmation time, supplier's picking and packaging time, supplier's outbound time, supplier's transportation time to the delivery address, and last-mile delivery time need to be calculated. There are three main ways to obtain the time spent at each stage of fulfillment. The first is to directly read real-time estimated values ​​from the business system. For example, a warehouse management system returns the estimated picking completion time for the current warehouse, a transportation management system returns the estimated vehicle arrival time, a map service returns the estimated travel time under current traffic conditions, and an instant delivery platform returns the estimated delivery time. The second is to calculate averages or quantiles based on historical fulfillment data. For example, warehouse operation time can be calculated based on historical picking times for the same warehouse, the same product type, and the same order quantity range; last-mile delivery time can be calculated based on historical delivery times for the same region, the same delivery time period, and the same transportation type. The third is to estimate based on rule tables. For example, regular orders are calculated using standard operating time, expedited orders using expedited operating time, large item orders using large item processing time, and cold chain orders using cold chain handover time. In practice, the fulfillment chain of the candidate solutions is first determined. If the solution is single-warehouse delivery, the fulfillment chain includes the complete link from the selected warehouse to the delivery address; if the solution involves multi-warehouse order splitting, each sub-order forms a separate fulfillment chain; if the solution involves replenishment, the replenishment chain and the order delivery chain together constitute a complete fulfillment chain. The estimated completion time for each fulfillment chain is calculated separately. For single-warehouse single-parcel orders, the estimated delivery time is the estimated completion and receipt time for that fulfillment chain. For multi-warehouse split orders, if the business requires that all parcels in an order be delivered before the order is considered complete, the estimated delivery time is the latest time among the estimated delivery times of all sub-orders; if the business allows some parcels to be delivered first, the estimated delivery time of each sub-order can be recorded separately, but the order-level estimated delivery time used for the fitness function is still preferably the latest delivery time to ensure that the overall order fulfillment commitment is not underestimated. Secondly, if an order is assigned to a warehouse after its daily cut-off time, the outbound time needs to be postponed to the next available processing wave; if the selected carrier's last pickup for the day has ended, the transportation start time needs to be postponed to the next pickup period; if the order's delivery address has nighttime delivery restrictions, the last-mile delivery time needs to be postponed to the available delivery window. These business time rules are used to adjust the estimated delivery time to avoid overly optimistic delivery times due to relying solely on theoretical transportation time calculations. After obtaining the estimated delivery time of the candidate solution corresponding to the particle, the estimated delivery time is compared with the expected delivery time of the order. If the estimated delivery time is not later than the expected delivery time of the order, the estimated delivery delay of the candidate solution is zero; if the estimated delivery time is later than the expected delivery time of the order, the system calculates the time difference between the two and uses this time difference as the estimated delivery delay. It should also be noted that: delay penalty coefficient A tiered configuration approach can be used to determine this; For example, in one implementation, a fee of 10 yuan per hour is charged for regular orders, 20 yuan per hour for same-day delivery orders, and 50 yuan per hour for instant delivery orders or high-priority orders; if a delay in delivery is expected... By using minutes as the unit, hourly values ​​can be converted to minute-level values. For example, a regular order costs approximately one-sixth of a yuan per minute, a same-day delivery order costs approximately one-third of a yuan per minute, and an instant delivery order costs approximately five-sixths of a yuan per minute. This tiered configuration allows the particle swarm optimization process to reduce projected logistics costs while fully considering business losses caused by order delays.

[0052] Furthermore, during optimization, if a continuous loop occurs, the decision action is immediately sent to the management personnel as a detection result of the current order and logistics matching decision not being met, for review.

[0053] It should be noted that the continuous loop case can be regarded as the particle swarm optimization algorithm performing more than 3 optimizations.

[0054] In summary, this invention constructs a fulfillment knowledge graph encompassing orders, goods, warehouses, suppliers, and transportation capacity, and forms a state transition diagram based on candidate fulfillment schemes. This allows for a structured representation of decision-making processes in multi-channel order fulfillment, including single-warehouse delivery, split-order delivery, supplier replenishment, direct supplier delivery, transportation capacity matching, and route switching. This improves the comprehensive perception of inventory, timeliness, cost, risk, and resource availability in the order allocation and logistics matching process. Furthermore, by learning fulfillment state characteristics through graph neural networks and combining intrinsic rewards, extrinsic rewards, and action value networks to select the current order allocation and logistics matching decision, multi-channel orders can generate better order allocation and logistics matching schemes based on real-time fulfillment resource status. Simultaneously, before scheme execution, warehouse inventory, warehouse processing capacity, supplier supply capacity, and transportation capacity feasibility are verified, and optimization is performed using particle swarm optimization when conditions are not met. This improves the accuracy of multi-channel order allocation, the feasibility of logistics matching, and the stability of order fulfillment.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent order allocation and logistics matching for multi-channel orders, characterized in that: include, Construct a knowledge graph for the fulfillment of orders from multiple channels and continuously obtain updates to the knowledge graph for the execution and fulfillment of orders from multiple channels; Based on the updated performance knowledge graph, a state transition graph containing directed transition edges is constructed. The state transition graph is embedded and learned through a graph neural network. After generating node embedding vectors, the first intrinsic reward and the second intrinsic reward are constructed and aggregated to generate a comprehensive intrinsic reward. Finally, the total reward is obtained through an access decay mechanism. After concatenating the total reward with the embedded vector, the action value is calculated through the action value network, and the maximum action value is used as the current order and logistics matching decision. Perform hard constraint checks on the current order allocation and logistics matching decisions, and generate check results to determine whether it is necessary to call the particle swarm optimization algorithm to optimize the order allocation and logistics matching decisions.

2. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The construction of a multi-channel order fulfillment knowledge graph, and the continuous acquisition of updates to the multi-channel order execution fulfillment knowledge graph, are detailed as follows: Construct order fulfillment units and generate unified internal order identifiers; Acquire multi-source data and perform preprocessing; After mapping standardized entities and generating nodes, edges are established between nodes to form a knowledge graph of contract fulfillment.

3. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The construction includes a state transition graph with directed transition edges, as detailed below: Retrieve candidate performance plans from the updated performance knowledge graph as performance status nodes; Establish directed transition edges between nodes in the performance status and construct decision actions; After calculating the edge weights based on the directed transition edges, a state transition graph is constructed.

4. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The step of embedding the state transition graph using a graph neural network to generate node embedding vectors includes: In each layer of the graph neural network, the information of the neighboring performance status nodes is aggregated according to the edge weights to obtain the embedding vector of each node.

5. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The process of constructing the first intrinsic reward and the second intrinsic reward and aggregating them to generate a comprehensive intrinsic reward includes: Construct the first intrinsic reward based on the centrality of each fulfilling state node in the state transition graph; calculate the inverse degree of the fulfilling state node to generate the second intrinsic reward; The first intrinsic reward and the second intrinsic reward are combined to obtain the comprehensive intrinsic reward.

6. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The method of obtaining the total reward through the access decay mechanism includes: The number of visits to the fulfillment status node is counted, and an access decay mechanism is constructed to decay and correct the comprehensive intrinsic reward, generating the decayed comprehensive intrinsic reward. Develop external rewards based on actual business objectives; The decayed intrinsic reward and extrinsic reward are weighted and combined to generate the total reward.

7. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The concatenation of the total reward with the embedded vector includes: The total reward is normalized and scaled to a numerical range that matches the feature scale of the embedded vector; then the normalized total reward is used as a one-dimensional feature and concatenated with the embedded vector of the performance status node to form a fused feature vector.

8. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 1, characterized in that: The step of performing a step-by-step hard constraint check on the current order allocation and logistics matching decision, and generating check results to determine whether the particle swarm optimization algorithm needs to be called to optimize the order allocation and logistics matching decision, includes: Based on the current order allocation and logistics matching decisions, check whether the available inventory of the goods in each warehouse in the current plan can meet the order demand; check whether the remaining processing capacity of the selected warehouse in the current decision cycle can support the outbound shipment of the order or its sub-orders; check whether the supplier can provide the corresponding goods, and check whether the estimated delivery time meets the order fulfillment time requirements; check whether the selected transportation capacity is available, whether it can reach the pickup point within the required time, and whether the load capacity meets the order transportation requirements. If all the check results are satisfied, the current order and logistics matching decision is determined to be the final result; if any check result is not satisfied, the current order and logistics matching decision is determined to be the result to be optimized, and then optimized using the particle swarm optimization algorithm.

9. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 8, characterized in that: The optimization using the particle swarm optimization algorithm includes: The current order allocation and logistics matching decision is treated as an individual particle. By perturbing the individual particles, a population is generated for initialization. Define a particle fitness function; update the particle velocity and position using standard velocity and position update formulas; when the maximum number of iterations is reached, output the optimized optimal particle; extract the order allocation and logistics matching decision corresponding to the optimal particle for inspection; if all inspection results are satisfied, the order allocation and logistics matching decision corresponding to the optimal particle is determined as the final result; otherwise, the particle swarm optimization algorithm is executed again.

10. The intelligent order allocation and logistics matching method for multi-channel orders as described in claim 9, characterized in that: The optimization using the particle swarm optimization algorithm also includes: During optimization, if a continuous loop occurs, the decision action will be immediately sent to the management personnel as a detection result of the current order and logistics matching decision not being met, for review.

Images