An AI-driven lightning cartridge replenishment strategy dynamic adjustment method

Abstract

The application discloses an AI-driven lightning bin replenishment strategy dynamic adjustment method, relates to the technical field of warehouse logistics artificial intelligence, and comprises the following steps: collecting multi-source heterogeneous data of a physical supply chain network and constructing a digital twin knowledge graph; extracting a dynamic attribute sequence from the digital twin knowledge graph, performing mode recognition based on a pre-defined business rule library and abnormality detection logic, obtaining a potential risk event, and injecting the potential risk event into the digital twin knowledge graph as a disturbance source node; performing multi-round message transmission and state iteration on the digital twin knowledge graph with the injected disturbance source node by using a graph neural network, simulating a cascading diffusion process of the risk along business edges and logistics edges, and obtaining a risk propagation report. The application solves the fundamental defect of gradual failure due to the lack of dynamic evolution capability, and ensures the prediction accuracy and the reliability of the decision.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology in warehousing and logistics, and in particular to a dynamic adjustment method for AI-driven lightning warehouse replenishment strategy. Background Technology

[0002] In today's globalized and digitalized supply chain environment, warehousing and logistics, as a core link in the supply chain, directly impact a company's core competitiveness through their operational efficiency and robustness. The "flash warehouse" model, in particular, catering to high-frequency, small-batch demands, places extremely high demands on the real-time and adaptive nature of replenishment strategies. Traditional replenishment strategies are often based on statistical analysis of historical sales data (such as time series forecasting and inventory control theory) or simple heuristic rules. These methods often exhibit lag and vulnerability when faced with multi-source concurrent disturbances such as supplier delays, transportation disruptions, and sudden demand fluctuations. With the development of artificial intelligence technology, especially the rise of digital twins and graph neural networks, a new paradigm has emerged for constructing dynamic virtual mappings of supply chain systems and simulating their complex internal interactions. Existing technologies have begun to explore constructing supply chain networks as graph structures, utilizing GNNs for demand forecasting or risk identification, and generating replenishment decisions through optimization algorithms.

[0003] Existing technologies suffer from two main shortcomings: First, at the level of risk perception and modeling, most methods rely on isolated anomaly detection algorithms, failing to place identified risk events within the dynamic propagation simulation of the entire supply chain network topology. This results in localized and static risk assessments, unable to anticipate the cascading effects that a single node's failure might trigger through business and logistics edges, thus severely underestimating potential systemic risks. Second, at the level of strategy generation and evaluation, existing methods typically employ static optimization objectives or conduct strategy testing under limited pre-defined scenarios, making it difficult for their strategies to withstand the complex combinations of uncertainties in real-world environments. Although Monte Carlo simulations are used to evaluate strategy performance, their random sampling mechanism is inefficient and lacks in-depth insights into the interaction mechanism between risk propagation paths and strategies, resulting in generated strategy performance spectra that cannot accurately reveal the vulnerabilities of strategies under different risk scenarios. Summary of the Invention

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

[0005] Therefore, this invention provides an AI-driven method for dynamically adjusting replenishment strategies in lightning warehouses to address the problems of existing technologies being unable to dynamically simulate the cascading propagation of risks in the supply chain network topology and struggling to generate replenishment strategies with deep robustness.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an AI-driven method for dynamically adjusting replenishment strategies in a flash warehouse. The method includes: collecting multi-source heterogeneous data from a physical supply chain network and constructing a digital twin knowledge graph; extracting dynamic attribute sequences from the digital twin knowledge graph and performing pattern recognition based on a predefined business rule base and anomaly detection logic to identify potential risk events; injecting these potential risk events as disturbance source nodes into the digital twin knowledge graph; using a graph neural network to perform multiple rounds of message passing and state iteration on the digital twin knowledge graph containing the disturbance source nodes, simulating the cascading diffusion process of risk along the business and logistics edges, and obtaining a risk propagation report; generating multiple candidate replenishment strategies based on the risk propagation report, and using Monte Carlo simulation to evaluate the robustness of these strategies, generating a strategy performance spectrum; using a multi-objective optimization algorithm to perform a trade-off analysis on the strategy performance spectrum, obtaining the optimal target strategy from the candidate replenishment strategies, and distributing it to the execution terminal; collecting actual execution data from the execution terminal, comparing the actual data with the risk propagation report and the strategy performance spectrum to obtain deviation data, and optimizing the graph neural network and the digital twin knowledge graph based on the deviation data.

[0008] As a preferred embodiment of the AI-driven lightning warehouse replenishment strategy dynamic adjustment method described in this invention, the multi-source heterogeneous data includes inventory level, order flow, in-transit logistics coordinates, and supplier capacity status.

[0009] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategy described in this invention, the specific steps for constructing the digital twin knowledge graph are as follows:

[0010] Collect multi-source heterogeneous data from the physical supply chain network, and perform spatiotemporal alignment and normalization processing to generate a standardized data stream;

[0011] Node-level spatiotemporal features are extracted from a standardized data stream, and dimensionality reduction is performed using an autoencoder to generate a spatiotemporal feature vector.

[0012] Based on spatiotemporal feature vectors, the association strength between any two nodes is calculated, and weighted edges are created for node pairs whose association strength exceeds the connection threshold to obtain the initial knowledge graph.

[0013] Based on the initial knowledge graph, an incremental learning strategy is used to dynamically update node attributes and edge weights, and logical conflicts are eliminated through consistency verification rules to generate a digital twin knowledge graph.

[0014] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategy described in this invention, the specific steps for acquiring potential risk events are as follows:

[0015] Extract the dynamic attribute sequence of each node within a continuous time window from the digital twin knowledge graph to form a spatiotemporal feature sequence;

[0016] Based on the spatiotemporal feature sequence, calculate the real-time anomaly index of each node relative to the historical normal behavior pattern.

[0017] The system performs logical matching between real-time anomaly indicators and a predefined business rule base to identify anomaly nodes that violate at least one business rule.

[0018] By aggregating the changes in the weights of the associated edges of abnormal nodes and the state information of neighboring nodes, potential risk events with risk levels and scope of impact are generated.

[0019] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategy described in this invention, the specific steps for injecting potential risk events as disturbance source nodes into the digital twin knowledge graph are as follows:

[0020] Based on the risk level and impact scope of potential risk events, obtain the distribution of disturbance field generated by disturbance source nodes in the digital twin knowledge graph;

[0021] Based on the perturbation field distribution, the adjustment strategy for node state and edge weight is obtained through a multi-agent collaborative decision-making mechanism;

[0022] The digital twin knowledge graph is updated based on the adjustment strategy of node status and edge weight, and potential risk events are injected into the digital twin knowledge graph as perturbation source nodes.

[0023] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategy described in this invention, the specific steps for obtaining the risk propagation report are as follows:

[0024] Extract the node state features of risk disturbance information from the digital twin knowledge graph of the already injected disturbance source nodes, and construct a heterogeneous graph structure representing the supply chain topology based on the type differences between business edges and logistics edges.

[0025] Based on a heterogeneous graph structure, a graph attention network layer is used to obtain the risk propagation attention weights between adjacent nodes along different edge types.

[0026] Through multiple rounds of message passing, the risk features of adjacent nodes are aggregated along the risk propagation attention weight, and the risk state vector of each node is updated using a gating update mechanism. When the rate of change of the risk state vector of a node is less than the convergence judgment threshold, the simulation process is judged to have converged.

[0027] Based on the risk state vectors of each node after convergence, the set of high-risk nodes, key propagation paths, and affected business scope are identified, and a risk propagation report is generated.

[0028] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategies described in this invention, the specific steps for generating multiple candidate replenishment strategies based on risk propagation reports are as follows:

[0029] Extract key risk parameters and business constraints of affected nodes from the risk propagation report;

[0030] Based on key risk parameters and business constraints of affected nodes, a multi-objective optimization strategy parameter space is constructed.

[0031] In the strategy parameter space, multiple candidate replenishment strategies are generated using the particle swarm optimization algorithm.

[0032] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategies described in this invention, the specific steps for generating the strategy performance spectrum are as follows:

[0033] Based on multiple candidate replenishment strategies, a set of test scenarios with various risk evolution paths is constructed;

[0034] In each scenario within the test scenario set, evaluate the performance of each candidate replenishment strategy in terms of cost, service level, and inventory turnover, and generate a strategy performance spectrum.

[0035] As a preferred embodiment of the AI-driven dynamic adjustment method for lightning warehouse replenishment strategies described in this invention, the steps of using a multi-objective optimization algorithm to perform a trade-off analysis on the strategy performance spectrum, obtaining the optimal target strategy from candidate replenishment strategies, and sending it to the execution terminal are as follows:

[0036] Key performance indicators of each candidate replenishment strategy in terms of cost, service level and inventory turnover were extracted from the strategy performance spectrum.

[0037] Based on key performance indicators, a multi-objective optimization algorithm is applied to perform trade-off analysis and obtain the optimal objective strategy.

[0038] The optimal target strategy is decomposed into specific replenishment instructions and inventory control parameters, and then sent to the corresponding execution terminals.

[0039] As a preferred embodiment of the AI-driven dynamic adjustment method for flash warehouse replenishment strategies described in this invention, the following steps are taken: The actual execution data from the collection and execution terminal is compared with the risk propagation report and strategy performance spectrum to obtain deviation data. Based on the deviation data, the graph neural network and digital twin knowledge graph are then optimized.

[0040] The actual execution data of replenishment instructions and inventory control parameters are collected by the monitoring agent deployed on the execution terminal;

[0041] The actual execution data is compared and analyzed with the node risk status vector and multi-dimensional performance indicators in the strategy performance spectrum in the risk propagation report to generate a deviation dataset.

[0042] Based on the biased dataset, an adaptive gradient optimization algorithm is used to update the parameters of the graph neural network and the node relationship weights of the digital twin knowledge graph.

[0043] The beneficial effects of this invention are as follows: By using graph neural networks to simulate the cascading diffusion of risks, the core problem of being unable to predict global chain reactions due to viewing risk events in isolation, resulting in fragmented and delayed risk assessments, is solved. This enables dynamic tracking of the transmission path of supply chain risks and accurate quantification of the scope of impact, providing key decision-making basis for replenishment strategy formulation. By using actual deviation data for closed-loop optimization, the fundamental defect of gradually failing due to lack of dynamic evolution capability is solved, ensuring the accuracy of prediction and the reliability of decision-making. Attached Figure Description

[0044] 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.

[0045] Figure 1 A flowchart illustrating the dynamic adjustment method for AI-driven flash warehouse replenishment strategies.

[0046] Figure 2 A flowchart for constructing a digital twin knowledge graph.

[0047] Figure 3 A flowchart for risk event identification and injection.

[0048] Figure 4 A flowchart for generating a risk communication report. Detailed Implementation

[0049] 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.

[0050] 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.

[0051] 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.

[0052] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI-driven method for dynamically adjusting a flash warehouse replenishment strategy, including the following steps:

[0053] S1. Collect multi-source heterogeneous data from the physical supply chain network and construct a digital twin knowledge graph.

[0054] Collect multi-source heterogeneous data from the physical supply chain network, and perform spatiotemporal alignment and normalization processing to generate a standardized data stream.

[0055] The specific process includes collecting multi-source heterogeneous data from the physical supply chain network, including inventory levels, order flow, in-transit logistics coordinates, and supplier capacity status. Inventory levels are collected in real time by IoT sensors within the flash warehouse. Order flow data comes from e-commerce platform transaction interfaces and includes order time, product type, quantity, and delivery address. In-transit logistics coordinates are uploaded periodically by GPS devices on transport vehicles or packages. Supplier capacity status is obtained periodically through application programming interfaces (APIs) that interface with supplier enterprise resource planning (ERP) systems, including production line load, raw material inventory, and output capacity. This data is then timestamped and spatially aligned, and normalized to form a standardized data stream.

[0056] Node-level spatiotemporal features are extracted from the standardized data stream, and dimensionality reduction is performed using an autoencoder to generate spatiotemporal feature vectors.

[0057] The specific process includes extracting corresponding inventory levels, order flow, in-transit logistics coordinates, and supplier capacity status from the standardized data stream according to time steps and node positions in the supply chain network. This constitutes multidimensional observations of each node at multiple consecutive time points, forming node-level spatiotemporal features. The node-level spatiotemporal features are then fed into an autoencoder. The autoencoder gradually compresses the input dimension through the neural network layer of the encoder part, learns the inherent spatiotemporal correlation structure of the node-level spatiotemporal features, and outputs a compact representation at the bottleneck layer. This compact representation is the dimensionality-reduced spatiotemporal feature vector.

[0058] It should be noted that an autoencoder refers to an artificial neural network, which consists of two parts: an encoder and a decoder. The encoder maps the high-dimensional input data to a low-dimensional latent space to generate a compact feature representation. The decoder reconstructs the output from the low-dimensional representation to be as close as possible to the original input.

[0059] Based on spatiotemporal feature vectors, the association strength between any two nodes is calculated, and weighted edges are created for node pairs with association strength exceeding a connection threshold to obtain the initial knowledge graph. The expression is as follows:

[0060] ;

[0061] in, Represents a node With nodes The strength of the correlation between them and Identifiers representing any two nodes in a digital twin knowledge graph. Indicates the balancing weight coefficient. This represents the natural exponential function. Represents a node With nodes The combined spatiotemporal distance between them Indicates the scale parameter. Represents a node The spatiotemporal feature vector, Represents a node The spatiotemporal feature vectors.

[0062] It should be noted that the balancing weight coefficient is used to adjust the relative importance of the spatiotemporal integrated distance and the cosine similarity of the spatiotemporal feature vector in the association strength calculation. The balancing weight coefficient is determined through historical operating data and its value ranges from zero to one. The scale parameter is used to control the influence range of the spatiotemporal integrated distance on the association strength, reflecting the sensitivity of distance changes in the exponential decay function. The scale parameter includes a time scale component and a spatial scale component, which correspond to the normalized scale of the time interval and the geographic distance, respectively. The spatiotemporal integrated distance is a comprehensive measure obtained by combining the topological path length (spatial distance) of the node in the digital twin knowledge graph with the dynamic time warping or Euclidean distance (time distance) of its spatiotemporal feature sequence, after normalization and weighted fusion.

[0063] The specific process includes: pairing any two nodes in the digital twin knowledge graph based on spatiotemporal feature vectors, calculating the association strength between each pair of nodes, which comprehensively reflects their similarity and distance relationship in spatiotemporal features; when the association strength of a pair of nodes is higher than a set connection threshold, establishing an edge between the two nodes, and using the association strength as the weight of the edge; by traversing all node pairs and performing the above judgment and edge building operations, an initial knowledge graph is formed; where nodes represent supply chain entities, including warehouses, transportation vehicles, and suppliers, and edges represent the real-time business relationships between supply chain entities, including supply relationships, transportation relationships, and warehousing relationships.

[0064] The connection threshold is preset based on the requirements of the business scenario for the tightness of the connection between supply chain entities. It is usually determined by analyzing the distribution of the connection strength of node pairs in a stable operating state in historical standardized data streams. The exemplary range of the connection threshold is between zero and one.

[0065] Based on the initial knowledge graph, an incremental learning strategy is used to dynamically update node attributes and edge weights, and logical conflicts are eliminated through consistency verification rules to generate a digital twin knowledge graph.

[0066] The specific process includes: based on the initial knowledge graph, whenever a new standardized data stream arrives, extracting the corresponding node-level spatiotemporal features and updating the attribute values ​​of each node, while recalculating the association strength between affected node pairs to adjust edge weights, thus achieving dynamic evolution under the incremental learning strategy; after each update, applying consistency verification rules to check for logical conflicts between node attributes and edge weights, such as mismatches between supplier capacity status and corresponding supply relationship strength, or contradictions between transportation tool location and warehousing relationship, and correcting or deleting conflicting parts; through continuous attribute updates, weight adjustments, and conflict elimination, a digital twin knowledge graph with a consistent structure that reflects the real-time status of the physical supply chain network is finally generated.

[0067] It should be noted that the consistency verification rules are based on the prior constraints of the supply chain business logic and the relationships between entities. By analyzing the reasonable range of node attributes and edge weights and their interdependencies in historical operating data, logical constraints are derived from frequent patterns.

[0068] S2. Extract dynamic attribute sequences from the digital twin knowledge graph, and perform pattern recognition based on a predefined business rule base and anomaly detection logic to obtain potential risk events. Inject these potential risk events as disturbance source nodes into the digital twin knowledge graph.

[0069] The dynamic attributes of each node within a continuous time window are extracted from the digital twin knowledge graph to form a spatiotemporal feature sequence.

[0070] The specific process includes: selecting a continuous time window from the digital twin knowledge graph, which consists of multiple equally spaced time steps; for each node in the digital twin knowledge graph, extracting the dynamic attributes corresponding to each time step in sequence, including inventory level, order flow, in-transit logistics coordinates, and supplier capacity status; concatenating all dynamic attributes of the same node in chronological order within the time window into an ordered multidimensional sequence; repeating the above process for all nodes, and finally generating a spatiotemporal feature sequence for each node that reflects the state evolution within the time window.

[0071] Based on the spatiotemporal feature sequence, the real-time anomaly index of each node relative to its historical normal behavior pattern is calculated, and the expression is:

[0072] ;

[0073] in, Indicates a point in time Real-time abnormal indicators Indicates the current time point in the calculation. Indicates a point in time Actual observations collected from the nodes, Indicates a point in time Exponentially weighted moving average, Indicates a point in time The exponentially weighted moving standard deviation, Represents extremely small positive numbers. This represents the time decay weighting coefficient. Represents the linear rectified function. Indicates a point in time Actual observations collected from the nodes, This represents the standard deviation of the change in adjacent time periods within historical data. This represents the threshold parameter.

[0074] It should be noted that, The time decay weighting coefficient is used to control the influence of historical observations on the current estimate when calculating the exponentially weighted moving average. Observations closer to the current time are given a higher weight. The value is determined by the business's requirements for sensitivity to anomaly response and is usually between zero and one. The threshold parameter is used to determine whether the real-time abnormal indicators have reached the abnormal triggering conditions. When the real-time abnormal indicators exceed the threshold parameter, the corresponding node is considered to have an abnormality. The threshold parameter is obtained by analyzing the distribution of real-time abnormal indicators in the historical spatiotemporal feature sequence under normal operating conditions. Specific methods include calculating the mean and standard deviation and setting it as the mean plus a certain number of times the standard deviation, or using the percentile method to select high quantiles as the initial threshold parameter.

[0075] The specific process includes calculating a real-time anomaly index for each node at the current time point relative to its historical normal behavior pattern, based on the spatiotemporal feature sequence. This real-time anomaly index is dynamically adjusted by comparing the deviation of the current observation value with the exponentially weighted moving average and combining it with the standard deviation of the change in adjacent time points. The deviation part is normalized to eliminate the scale effect. At the same time, a linear rectification function is combined to enhance the response to sudden trends (such as a sudden drop in inventory level, a sudden increase in order flow, a long period of stagnation in the logistics coordinates, or a sharp increase in the supplier's capacity utilization rate). Finally, a real-time anomaly index reflecting the degree of deviation of the node's current state is output.

[0076] Historical normal behavior patterns refer to the typical temporal changes in inventory levels, order flow, in-transit logistics coordinates, and supplier capacity status of each node under normal operating conditions. These patterns are obtained by selecting historical spatiotemporal feature sequences without abnormal events from the digital twin knowledge graph and calculating sliding window statistics or exponentially weighted moving averages.

[0077] The system performs logical matching between real-time anomaly indicators and a predefined business rule base to identify anomaly nodes that violate at least one business rule.

[0078] The specific process includes logically matching the real-time anomaly indicators of each node with each rule in the predefined business rule base, determining whether the real-time anomaly indicators meet the anomaly judgment conditions described by any business rule, and marking the node as an anomaly node if the real-time anomaly indicators of a node violate at least one business rule.

[0079] It should be noted that the predefined business rule base is predefined based on supply chain domain knowledge and experience in handling historical anomalies. By analyzing confirmed anomaly cases and their corresponding real-time anomaly indicator characteristics in past flash warehouse operations (e.g., order fulfillment rate drops sharply beyond the tolerance range, long-term stagnation or deviation from the predetermined path in transit logistics, and continuous over-limit of supplier capacity utilization), formally expressible logical judgment conditions are summarized. Combined with compliance requirements for dimensions such as inventory, orders, logistics, and supplier capacity, these rules are manually compiled into a set of Boolean logic rules.

[0080] By aggregating the changes in the weights of the associated edges of abnormal nodes and the state information of neighboring nodes, potential risk events with risk levels and scope of impact are generated.

[0081] The specific process includes extracting the weight changes of the edges associated with the abnormal nodes in the digital twin knowledge graph for all abnormal nodes, and simultaneously obtaining the current state of the neighboring nodes directly connected to these abnormal nodes. The magnitude of the weight changes and the degree of deviation of the neighboring node states are weighted and fused together, and the overall abnormal impact intensity is assessed according to the preset risk level classification standard, thereby generating potential risk events that include risk level and impact scope.

[0082] It should be noted that the risk level classification criteria are based on the actual impact and consequences of historical potential risk events and the business tolerance. By analyzing the correspondence between the number of abnormal nodes, the magnitude of changes in the weight of related edges, the degree of deviation of the state of neighboring nodes and the final business loss in past events, the threshold range of different risk levels is determined by using the quantile method.

[0083] Based on the risk level and impact range of potential risk events, the distribution of disturbance fields generated by disturbance source nodes in the digital twin knowledge graph is obtained.

[0084] The specific process includes locating the specific position of the disturbance source node in the digital twin knowledge graph based on the risk level and impact range of the potential risk event. The disturbance source node is the abnormal node that triggers the potential risk event. Starting from the disturbance source node, the disturbance impact propagates outward along the edge structure in the digital twin knowledge graph. During the propagation process, the disturbance intensity decreases with the increase of the path length and is adjusted by the edge weight. The larger the edge weight, the tighter the coupling between entities and the stronger the disturbance propagation. At the same time, the feedback effect of the state response degree of neighboring nodes on the disturbance diffusion is considered. By combining the topological distance, edge weight and node state changes, the distribution intensity of the disturbance on each reachable node in the digital twin knowledge graph is obtained, forming a disturbance field distribution that characterizes the spatial range and local intensity of the disturbance impact.

[0085] Based on the perturbation field distribution, the adjustment strategy for node states and edge weights is obtained through a multi-agent collaborative decision-making mechanism.

[0086] The specific process includes assigning an agent to each node in the digital twin knowledge graph based on the perturbation field distribution. Each agent exchanges information and negotiates adjustment intentions with the agents of neighboring nodes based on the current state of its own node and the intensity of the perturbation. Through a multi-agent collaborative decision-making mechanism, consensus is reached under the premise of satisfying global constraints, and the correction direction of each node's state and the update magnitude of the associated edge weights are jointly determined. Finally, an adjustment strategy covering the node state and edge weights of all affected nodes is output.

[0087] It should be noted that the multi-agent collaborative decision-making mechanism configures an agent for each node in the digital twin knowledge graph. Each agent interacts and negotiates based on the local node state, the distribution of the disturbance field, and the information of its neighbors. Through iterative coordination, consensus is reached, and a strategy for adjusting the node state and edge weights is jointly generated.

[0088] The digital twin knowledge graph is updated based on the adjustment strategy of node status and edge weight, and potential risk events are injected into the digital twin knowledge graph as perturbation source nodes.

[0089] The specific process includes modifying the attribute values ​​of each affected node in the digital twin knowledge graph, such as inventory level, order flow, in-transit logistics coordinates, or supplier capacity status, according to the node status and edge weight adjustment strategy, and simultaneously updating the weight values ​​of the edges connecting these nodes; at the same time, the generated potential risk events are clearly identified as disturbance source nodes, and attribute information such as disturbance type, risk level, and impact range are added to the potential risk events in the digital twin knowledge graph, and a mapping relationship between potential risk events and original abnormal nodes is established by means of directed edges or attribute associations, thus completing the complete update of the digital twin knowledge graph.

[0090] S3. Utilize graph neural networks to perform multi-round message passing and state iteration on the digital twin knowledge graph of the injected disturbance source nodes, simulate the cascading diffusion process of risks along the business edge and logistics edge, and obtain risk propagation reports.

[0091] Extract the node state features of risk disturbance information from the digital twin knowledge graph of the injected disturbance source nodes, and construct a heterogeneous graph structure representing the supply chain topology based on the type differences between business edges and logistics edges.

[0092] The specific process includes extracting node state features containing disturbance type, risk level, and impact range from the digital twin knowledge graph of the injected disturbance source nodes, and distinguishing between business edges and logistics edges according to the business relationship type represented by the edges. Business edges reflect supply or warehousing relationships, while logistics edges reflect transportation relationships. Based on the semantic and functional differences between these two types of edges, the nodes and different types of edges are organized into a graph structure with multiple edge types, thus constructing a heterogeneous graph structure that represents the supply chain topology.

[0093] It should be noted that the difference between business edge and logistics edge refers to the fact that business edge represents supply relationship or warehousing relationship, reflecting the ownership of goods or storage responsibility, while logistics edge represents transportation relationship, reflecting the physical movement of goods; this difference is obtained through semantic annotation of edges when constructing the digital twin knowledge graph. Business edge originates from order or inventory records, while logistics edge originates from transportation trajectory or delivery plan.

[0094] Based on a heterogeneous graph structure, a graph attention network layer is used to obtain the risk propagation attention weights between adjacent nodes along different edge types.

[0095] The specific process includes, based on the heterogeneous graph structure, the graph attention network layer processes the neighboring nodes connected to each node through business edges and logistics edges, concatenates or linearly transforms the node state features of the central node and the neighboring nodes corresponding to each type of edge, and then obtains the correlation score between the central node and the neighboring node under a specific edge type through a learnable attention mechanism. The softmax function is applied to all neighbor scores under the same edge type to generate risk propagation attention weights along business edges and risk propagation attention weights along logistics edges, thereby distinguishing the relative importance of risk impact under different semantic relationships.

[0096] Through multiple rounds of message passing, the risk features of adjacent nodes are aggregated along the risk propagation attention weight, and the risk state vector of each node is updated using a gating update mechanism. When the rate of change of the risk state vector of a node is less than the convergence judgment threshold, the simulation process is judged to have converged.

[0097] The specific process includes: through multiple rounds of message passing, each node receives risk features sent by neighboring nodes connected via business and logistics edges in each round, and weights and aggregates these risk features according to the risk propagation attention weight of the corresponding edge type to form the comprehensive risk information received by the node in this round; this comprehensive risk information and the node's current risk state vector are input into a gating update mechanism, which calculates and updates the gating signal to determine how much historical state to retain and how much new aggregated information to integrate, thereby generating the updated risk state vector of the node; synchronous iterative execution on all nodes, and when the rate of change of the risk state vector of any node in all nodes is less than the preset convergence judgment threshold between two consecutive rounds, the risk cascading diffusion simulation process is determined to have converged.

[0098] It should be noted that the convergence judgment threshold is preset based on the stable fluctuation level of the risk state vector in the historical simulation process. By statistically analyzing the distribution of the rate of change between adjacent iterations of the node risk state vector in multiple simulations, the low percentile value or empirical stable interval is selected as the convergence judgment threshold; the exemplary value range is usually between 0.001 and 0.01.

[0099] Based on the risk state vectors of each node after convergence, the set of high-risk nodes, key propagation paths, and affected business scope are identified, and a risk propagation report is generated.

[0100] The specific process includes: based on the risk state vectors of each node after convergence, selecting nodes whose risk state vector values ​​exceed the risk judgment threshold to form a high-risk node set; identifying the critical propagation path from the disturbance source node to the high-risk node by tracing back the edge sequence with the largest risk propagation attention weight; determining the scope of affected business based on the warehouses, transportation vehicles, and suppliers associated with the high-risk nodes and their related supply, transportation, and warehousing relationships; and finally integrating the high-risk node set, critical propagation path, and scope of affected business to generate a risk propagation report.

[0101] It should be noted that the risk assessment threshold is preset based on the actual distribution of node risk state vectors in historical risk events. It is usually selected as the upper bound of the maximum value of the risk state vector under normal operating conditions or determined by adding a certain number of standard deviations to its mean. An exemplary value range is generally between 0.7 and 0.95.

[0102] S4. Generate multiple candidate replenishment strategies based on the risk propagation report, and use Monte Carlo simulation to evaluate the robustness of multiple candidate replenishment strategies and generate strategy performance spectrum.

[0103] Extract key risk parameters and business constraints of affected nodes from the risk propagation report.

[0104] The specific process includes extracting risk state vector values ​​corresponding to the set of high-risk nodes, edge types and attention weights on key propagation paths, inventory level constraints, order fulfillment time limits, transportation capacity limits and supplier capacity limits from the risk propagation report, forming key risk parameters and business constraints on affected nodes.

[0105] Based on key risk parameters and business constraints of affected nodes, a multi-objective optimization strategy parameter space is constructed.

[0106] The specific process includes defining the feasible value range of decision variables such as replenishment quantity, replenishment timing, transportation route selection, and supplier allocation based on key risk parameters and business constraints of affected nodes. Combined with multiple objectives such as cost minimization, service level maximization, and inventory turnover optimization, a multi-objective optimization strategy parameter space is formed, which is jointly defined by these objective functions and constraints.

[0107] In the strategy parameter space, multiple candidate replenishment strategies are generated using the particle swarm optimization algorithm.

[0108] The specific process includes the following steps: In the strategy parameter space of multi-objective optimization, the particle swarm optimization algorithm initializes a group of particles. The position vector of each particle corresponds to a candidate replenishment strategy, which includes the specific values ​​of decision variables such as replenishment quantity, replenishment timing, transportation route selection and supplier allocation. The particles update their speed and position based on their own historical best position and the group's historical best position. After multiple iterations, multiple candidate replenishment strategies distributed in the strategy parameter space are generated.

[0109] It should be noted that the particle swarm optimization algorithm refers to a swarm intelligence optimization method. Each candidate replenishment strategy is represented as a particle in the strategy parameter space of multi-objective optimization. The particle continuously updates its position and velocity by tracking its own historical best position and the group's historical best position. During the iteration process, it explores the strategy parameter space and generates multiple candidate replenishment strategies.

[0110] Based on multiple candidate replenishment strategies, a set of test scenarios with various risk evolution paths is constructed.

[0111] The specific process includes applying each candidate replenishment strategy to the digital twin knowledge graph of the injected disturbance source node based on multiple candidate replenishment strategies, running multiple rounds of message passing and risk state updates under the risk propagation simulation framework, generating the corresponding risk evolution process, and combining the high-risk node changes, key propagation path transfers and affected business scope fluctuations caused by each candidate replenishment strategy into an independent risk evolution path, ultimately forming a set of test scenarios covering the effects of different strategies.

[0112] It should be noted that the risk propagation simulation framework is a simulation method based on digital twin knowledge graphs. It obtains risk propagation attention weights under different edge types through graph attention network layers, aggregates the risk features of adjacent nodes through multi-round message passing, and iteratively updates the risk state vector of each node in combination with a gating update mechanism until the rate of change of the node risk state vector is less than the convergence judgment threshold, thereby simulating the cascading diffusion process of risks caused by disturbance source nodes in the supply chain network.

[0113] In each scenario within the test scenario set, evaluate the performance of each candidate replenishment strategy in terms of cost, service level, and inventory turnover, and generate a strategy performance spectrum.

[0114] The specific process includes, in each scenario in the test scenario set, obtaining the total replenishment and transportation costs generated during the execution of each candidate replenishment strategy, the service level reflected by the order fulfillment rate or stockout rate, and the inventory turnover rate determined by the inventory consumption rate and average inventory level. After normalizing the three indicators, they are combined into a multi-dimensional performance vector, and the multi-dimensional performance vectors of all candidate replenishment strategies in this scenario are aggregated to form a strategy performance spectrum.

[0115] S5. Use a multi-objective optimization algorithm to perform a trade-off analysis on the strategy performance spectrum, obtain the optimal target strategy from the candidate replenishment strategies, and send it to the execution terminal.

[0116] Key performance indicators for each candidate replenishment strategy in terms of cost, service level, and inventory turnover were extracted from the strategy performance spectrum.

[0117] The specific process includes extracting the multi-dimensional performance vector corresponding to each candidate replenishment strategy from the strategy performance spectrum, and obtaining the cost index representing the total replenishment and transportation costs, the service level index reflecting the order fulfillment rate or stockout rate, and the inventory turnover rate index derived from the inventory consumption rate and average inventory level, to form the key performance indicators of each candidate replenishment strategy in terms of cost, service level and inventory turnover rate.

[0118] Based on key performance indicators, a multi-objective optimization algorithm is applied to perform trade-off analysis and obtain the optimal objective strategy.

[0119] The specific process includes, based on the key performance indicators of each candidate replenishment strategy in terms of cost, service level and inventory turnover rate, using a multi-objective optimization algorithm to conduct a trade-off analysis on the conflict and synergy among the three, finding the non-dominated solution set on the Pareto front, and selecting the solution with the best overall performance (e.g., a replenishment strategy that achieves service level improvement within an acceptable cost increase range and inventory turnover rate not lower than the historical average) as the optimal target strategy.

[0120] It should be noted that the multi-objective optimization algorithm is a method for simultaneously optimizing multiple conflicting objectives such as cost, service level, and inventory turnover rate, and supports the selection of the optimal objective strategy by generating a Pareto optimal solution set.

[0121] The preset weights are based on the priority of business objectives and historical decision preferences, and the importance of cost, service level and inventory turnover rate are quantified and assigned using the analytic hierarchy process.

[0122] The optimal target strategy is decomposed into specific replenishment instructions and inventory control parameters, and then sent to the corresponding execution terminals.

[0123] The specific process involves breaking down the optimal target strategy into specific replenishment instructions based on entity dimensions such as warehouses, transportation vehicles, and suppliers. These instructions include the types of goods to be replenished, the quantity to be replenished, and the replenishment time. At the same time, the corresponding inventory control parameters are extracted, including safety stock levels, reorder points, and maximum inventory limits. These replenishment instructions and inventory control parameters are then sent to the corresponding execution terminals.

[0124] S6. Collect actual execution data from the execution terminal, compare the actual data with the risk propagation report and strategy performance spectrum to obtain deviation data, and optimize the graph neural network and digital twin knowledge graph based on the deviation data.

[0125] The actual execution data of replenishment instructions and inventory control parameters are collected by a monitoring agent deployed on the execution terminal.

[0126] The specific process includes collecting real-time data on the actual execution of replenishment instructions through a monitoring agent deployed on the execution terminal, including the actual replenishment time, the actual replenishment product types, and the actual replenishment quantity. At the same time, it collects the actual performance of inventory control parameters during the operation, including the actual inventory level, the actual reorder trigger time, and the actual safety stock status, forming actual execution data.

[0127] The actual execution data is compared and analyzed with the node risk status vector and multi-dimensional performance indicators in the strategy performance spectrum in the risk propagation report to generate a deviation dataset.

[0128] The specific process involves comparing and analyzing the actual replenishment time, actual replenishment product types, actual replenishment quantity, actual inventory level, actual reorder trigger time, and actual safety stock achievement status in the actual execution data with the node risk status vector in the risk propagation report and the cost indicators, service level indicators, and inventory turnover rate indicators in the strategy performance spectrum. The differences between the actual and expected values ​​are obtained, and these differences and the corresponding contextual information are organized into a deviation dataset.

[0129] Based on the biased dataset, an adaptive gradient optimization algorithm is used to update the parameters of the graph neural network and the node relationship weights of the digital twin knowledge graph.

[0130] The specific process includes obtaining the loss value between the predicted output of the graph neural network and the actual execution data based on the deviation dataset, using an adaptive gradient optimization algorithm to iteratively adjust the parameters of the graph neural network according to the loss value, and using the actual interaction changes between nodes reflected in the deviation dataset to correct the node relationship weights of the corresponding edges in the digital twin knowledge graph, so that the prediction ability of the graph neural network and the structural expression of the digital twin knowledge graph continuously approximate the real supply chain operation status.

[0131] It should be noted that the adaptive gradient optimization algorithm is an optimization method that automatically adjusts the learning rate based on the historical gradient of the parameters, which can improve convergence speed and stability. The adaptive gradient optimization algorithm is used to update the parameters of graph neural networks and the node relationship weights of digital twin knowledge graphs.

[0132] In summary, this invention addresses the core problem of fragmented and delayed risk assessment caused by viewing risk events in isolation and simulating cascading risk diffusion through graph neural networks. This enables dynamic tracking of supply chain risk propagation paths and precise quantification of impact range, providing crucial decision-making basis for replenishment strategy formulation. Furthermore, by using actual deviation data for closed-loop optimization, it overcomes the fundamental deficiency of gradually failing due to a lack of dynamic evolution capabilities, ensuring both prediction accuracy and decision reliability.

[0133] 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 dynamically adjusting a flash warehouse replenishment strategy based on AI, characterized in that: include, Collect multi-source heterogeneous data from the physical supply chain network and construct a digital twin knowledge graph; Dynamic attribute sequences are extracted from the digital twin knowledge graph, and pattern recognition is performed based on a predefined business rule base and anomaly detection logic to identify potential risk events. The specific steps for injecting potential risk events as disturbance source nodes into the digital twin knowledge graph are as follows: Based on the risk level and impact scope of potential risk events, obtain the distribution of disturbance field generated by disturbance source nodes in the digital twin knowledge graph; Based on the perturbation field distribution, the adjustment strategy for node state and edge weight is obtained through a multi-agent collaborative decision-making mechanism; The digital twin knowledge graph is updated based on the adjustment strategy of node status and edge weight, and potential risk events are injected into the digital twin knowledge graph as perturbation source nodes. By using graph neural networks to perform multi-round message passing and state iteration on the digital twin knowledge graph of the injected disturbance source node, the cascading diffusion process of risk along the business edge and logistics edge is simulated to obtain a risk propagation report. Multiple candidate replenishment strategies were generated based on the risk communication report; Monte Carlo simulation was used to robustly evaluate multiple candidate replenishment strategies and generate strategy performance spectra. The specific steps are as follows: Based on multiple candidate replenishment strategies, a set of test scenarios with various risk evolution paths is constructed; In each scenario within the test scenario set, evaluate the performance of each candidate replenishment strategy in terms of cost, service level, and inventory turnover, and generate a strategy performance spectrum. A multi-objective optimization algorithm is used to perform a trade-off analysis on the strategy performance spectrum, obtain the optimal target strategy from the candidate replenishment strategies, and send it to the execution terminal. The actual execution data of the execution terminal is collected, and the actual data is compared with the risk propagation report and strategy performance spectrum to obtain deviation data. Based on the deviation data, the graph neural network and digital twin knowledge graph are optimized.

2. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 1, characterized in that: The multi-source heterogeneous data includes inventory levels, order flow, in-transit logistics coordinates, and supplier capacity status.

3. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 2, characterized in that: The specific steps for constructing the digital twin knowledge graph are as follows. Collect multi-source heterogeneous data from the physical supply chain network, and perform spatiotemporal alignment and normalization processing to generate a standardized data stream; Node-level spatiotemporal features are extracted from a standardized data stream, and dimensionality reduction is performed using an autoencoder to generate a spatiotemporal feature vector. Based on spatiotemporal feature vectors, the association strength between any two nodes is calculated, and weighted edges are created for node pairs whose association strength exceeds the connection threshold to obtain the initial knowledge graph. Based on the initial knowledge graph, an incremental learning strategy is used to dynamically update node attributes and edge weights, and logical conflicts are eliminated through consistency verification rules to generate a digital twin knowledge graph.

4. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 3, characterized in that: The specific steps for identifying potential risk events are as follows. Extract the dynamic attribute sequence of each node within a continuous time window from the digital twin knowledge graph to form a spatiotemporal feature sequence; Based on the spatiotemporal feature sequence, calculate the real-time anomaly index of each node relative to the historical normal behavior pattern. The system performs logical matching between real-time anomaly indicators and a predefined business rule base to identify anomaly nodes that violate at least one business rule. By aggregating the changes in the weights of the associated edges of abnormal nodes and the state information of neighboring nodes, potential risk events with risk levels and scope of impact are generated.

5. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 4, characterized in that: The specific steps for obtaining the risk transmission report are as follows. Extract the node state features of risk disturbance information from the digital twin knowledge graph of the already injected disturbance source nodes, and construct a heterogeneous graph structure representing the supply chain topology based on the type differences between business edges and logistics edges. Based on a heterogeneous graph structure, a graph attention network layer is used to obtain the risk propagation attention weights between adjacent nodes along different edge types. Through multiple rounds of message passing, the risk features of adjacent nodes are aggregated along the risk propagation attention weight, and the risk state vector of each node is updated using a gating update mechanism. When the rate of change of the risk state vector of a node is less than the convergence judgment threshold, the simulation process is judged to have converged. Based on the risk state vectors of each node after convergence, the set of high-risk nodes, key propagation paths, and affected business scope are identified, and a risk propagation report is generated.

6. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 5, characterized in that: The specific steps for generating multiple candidate replenishment strategies based on the risk propagation report are as follows: Extract key risk parameters and business constraints of affected nodes from the risk propagation report; Based on key risk parameters and business constraints of affected nodes, a multi-objective optimization strategy parameter space is constructed. In the strategy parameter space, multiple candidate replenishment strategies are generated using the particle swarm optimization algorithm.

7. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 6, characterized in that: The process involves using a multi-objective optimization algorithm to perform a trade-off analysis on the strategy performance spectrum, obtaining the optimal target strategy from the candidate replenishment strategies, and then sending it to the execution terminal. The specific steps are as follows: Key performance indicators of each candidate replenishment strategy in terms of cost, service level and inventory turnover were extracted from the strategy performance spectrum. Based on key performance indicators, a multi-objective optimization algorithm is applied to perform trade-off analysis and obtain the optimal objective strategy. The optimal target strategy is decomposed into specific replenishment instructions and inventory control parameters, and then sent to the corresponding execution terminals.

8. The AI-driven dynamic adjustment method for flash warehouse replenishment strategy as described in claim 7, characterized in that: The actual execution data from the collection and execution terminal is compared with the risk propagation report and strategy performance spectrum to obtain deviation data. Based on the deviation data, the graph neural network and digital twin knowledge graph are optimized. The specific steps are as follows. The actual execution data of replenishment instructions and inventory control parameters are collected by the monitoring agent deployed on the execution terminal; The actual execution data is compared and analyzed with the node risk status vector and multi-dimensional performance indicators in the strategy performance spectrum in the risk propagation report to generate a deviation dataset. Based on the biased dataset, an adaptive gradient optimization algorithm is used to update the parameters of the graph neural network and the node relationship weights of the digital twin knowledge graph.

Images