A full-link collaborative optimization method for flexible supply chain resilience evaluation

Abstract

The application discloses a kind of flexible supply chain resilience evaluation full-link collaborative optimization method, it is related to supply chain intelligent management technical field, including obtaining supply chain multi-source data and pre-processing, obtain structured time series dataset;Supply chain heterogeneous graph is constructed and is fused with domain knowledge graph, and supply chain semantic graph structure is obtained;Through the way that rule calculation and large language model inference are combined, generate resilience index set and identify weak link of supply chain;Order data is input into time series prediction model to obtain future demand prediction value, calculate supply chain risk prediction result;Generation candidate scheme set, and resilience index set and candidate scheme set are input into multi-objective optimization model and are solved, and collaborative instruction set is generated;Collaborative instruction set is executed and feedback data is collected, and model parameters are updated according to feedback data.The application realizes the intelligent evaluation of supply chain resilience and the accurate early warning of risk, and significantly improves the anti-risk ability and response efficiency of supply chain through collaborative optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain management technology, and in particular to a full-chain collaborative optimization method for assessing the resilience of flexible supply chains. Background Technology

[0002] With the deepening of globalization and the increasing complexity of the market environment, supply chains face multiple uncertainties and risks, including demand fluctuations, supply disruptions, and logistical delays. Supply chain resilience, as a crucial indicator of a supply chain's ability to withstand risks and recover quickly, has received widespread attention from businesses and academia. Traditional supply chain management methods primarily rely on statistical analysis of historical data and human experience for decision-making, which often proves insufficient for timely responses to unforeseen events or rapidly changing market demands.

[0003] Existing supply chain risk assessment methods primarily rely on single-dimensional data analysis to perceive the state of the supply chain, lacking a comprehensive understanding of the supply chain network structure and inter-entity relationships, resulting in incomplete risk identification. In terms of optimization decision-making, existing technologies typically focus on localized optimization of single links such as procurement, production, and warehousing, with insufficient coordination between links, easily leading to local optima but overall suboptimal results. Furthermore, existing methods often use fixed weights in setting optimization objectives, making it difficult to dynamically adjust optimization strategies based on the current state of the supply chain, reducing the flexibility and adaptability of decision-making. Summary of the Invention

[0004] This invention provides a full-link collaborative optimization method for assessing the resilience of flexible supply chains, which addresses the technical problems of inaccurate supply chain resilience assessment, insufficient risk warning capabilities, and lack of full-link collaborative optimization in existing technologies.

[0005] The first aspect of this invention provides a full-chain collaborative optimization method for assessing the resilience of a flexible supply chain, comprising:

[0006] Obtain multi-source data from the supply chain and preprocess it to obtain a structured time-series dataset;

[0007] Features are extracted from structured time-series datasets and a supply chain heterogeneous graph is constructed. The supply chain heterogeneous graph is then fused with a knowledge graph in the supply chain domain to obtain a supply chain semantic graph structure.

[0008] Based on the semantic graph structure of the supply chain, a set of resilience indicators is generated and weak links in the supply chain are identified by combining rule calculation and large language model reasoning.

[0009] Order data is input into a time-series forecasting model to obtain future demand forecasts, and the supply chain risk forecasting results are calculated by combining the current supply chain status.

[0010] A set of candidate solutions is generated based on weak links in the supply chain, future demand forecasts, and supply chain risk predictions.

[0011] The set of resilience indicators and the set of candidate solutions are input into a multi-objective optimization model for solving, generating a set of collaborative instructions.

[0012] Execute the collaborative instruction set and collect feedback data, then update the model parameters based on the feedback data.

[0013] Optionally, the supply chain semantic graph structure is obtained, including:

[0014] Extracting statistical and event features from structured time-series datasets;

[0015] Based on structured time-series datasets, identify supply chain entities and interaction relationships, and construct a supply chain heterogeneity graph;

[0016] Align the heterogeneous supply chain graph with the supply chain domain knowledge graph to extract the associated domain knowledge;

[0017] By attaching statistical features, event features, and domain knowledge as attributes to the corresponding nodes and edges of the supply chain heterogeneous graph, a supply chain semantic graph structure is obtained.

[0018] Optionally, a set of resilience indicators is generated and weak links in the supply chain are identified, including:

[0019] Convert the supply chain semantic graph structure into a text description to generate a supply chain status description text.

[0020] Based on the semantic graph structure of the supply chain, redundancy indicators, response speed indicators, and visibility indicators are calculated, and abnormal nodes are identified according to indicator thresholds and event characteristics.

[0021] Fill the resilience assessment prompt template with the supply chain status description text, abnormal node details and calculated indicator values ​​to generate a complete prompt;

[0022] The complete prompt words are input into a large language model for inference, generating the raw results of the resilience assessment.

[0023] The flexibility and synergy indicators are extracted from the original resilience assessment results and merged with the calculated indicators to generate a set of resilience indicators.

[0024] Bottlenecks, root causes, and risk transmission paths are extracted from the original resilience assessment results to identify weak links in the supply chain.

[0025] Optionally, the supply chain risk forecast results are calculated based on the current supply chain status, including:

[0026] Historical order data is extracted from structured time-series datasets to construct order time-series sequences;

[0027] Extract features from the order time series sequence to generate time series features;

[0028] Input the time series features into the time series prediction model to generate demand prediction distribution parameters for each future time node;

[0029] Using the Monte Carlo simulation method, based on the current supply chain status and demand forecast distribution parameters, the inventory distribution at each future time point is calculated;

[0030] Based on inventory distribution, risk indicators are calculated for each future time point to generate supply chain risk prediction results.

[0031] Optionally, a set of candidate solutions is generated based on weak links in the supply chain, future demand forecasts, and supply chain risk predictions, including:

[0032] Based on the types of bottleneck nodes and optimization needs in the weak links of the supply chain, the corresponding candidate solution types are determined;

[0033] Based on the stockout risk probability and forecast time points in the supply chain risk forecast results, identify the key time windows in which the stockout risk probability exceeds the risk threshold.

[0034] For key time windows and candidate solution types, available resource nodes are retrieved from the supply chain semantic graph structure, and candidate solution sets of various types are generated based on the resource attributes of the resource nodes.

[0035] Calculate the resource consumption, risk mitigation effect, and feasibility score of each candidate solution in the candidate solution set;

[0036] The candidate solution set is filtered based on the feasibility score, and the candidate solutions with feasibility scores higher than the feasibility threshold are retained to obtain a candidate solution set containing solution attribute information.

[0037] Optionally, a cooperative instruction set is generated, including:

[0038] Based on the set of resilience indicators, the weight coefficients of each optimization objective in the multi-objective optimization model are determined;

[0039] The candidate solution set is encoded as decision variables to construct a multi-objective optimization model;

[0040] A multi-objective optimization algorithm is used to solve the multi-objective optimization model to obtain the Pareto optimal solution set;

[0041] The optimal solution is selected from the Pareto optimal solution set based on the weight coefficients, and the candidate solution combination corresponding to the optimal solution is converted into a cooperative instruction set.

[0042] Optionally, the model parameters are updated based on the feedback data, including:

[0043] Execute the collaborative instruction set and collect feedback data generated during instruction execution;

[0044] Calculate prediction bias and execution bias based on feedback data;

[0045] Update the parameters of the time series prediction model based on the prediction bias, and update the parameters of the large language model and the multi-objective optimization model based on the execution bias.

[0046] Beneficial effects: This invention achieves comprehensive perception and deep understanding of the supply chain status by constructing a supply chain semantic graph structure and integrating domain knowledge; it improves the accuracy and interpretability of resilience assessment by combining rule calculation and large language model reasoning to generate a set of resilience indicators; it achieves quantitative assessment and forward-looking early warning of supply chain risks by adopting a probabilistic risk prediction method; it generates targeted candidate solutions based on weak links and risk prediction results, and performs global optimization through a multi-objective optimization model with dynamic weight adjustment, achieving a balance between risk control, response speed, and cost-effectiveness; and it continuously updates model parameters through execution feedback data, forming a closed-loop mechanism of "assessment-prediction-optimization-execution-feedback", enabling the quality of decision-making to continuously improve as the process progresses. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0048] Figure 1 This is a flowchart illustrating a full-link collaborative optimization method for assessing the resilience of a flexible supply chain, as provided in an embodiment of the present invention.

[0049] Figure 2 This is a flowchart illustrating the generation process of a supply chain semantic graph structure for a full-link collaborative optimization method for assessing the resilience of a flexible supply chain, as provided in an embodiment of the present invention.

[0050] Figure 3 A flowchart illustrating the resilience assessment process of a full-link collaborative optimization method for assessing the resilience of a flexible supply chain, as provided in this embodiment of the invention. Detailed Implementation

[0051] This invention provides a full-link collaborative optimization method for assessing the resilience of flexible supply chains, which addresses the technical problems of inaccurate supply chain resilience assessment, insufficient risk warning capabilities, and lack of full-link collaborative optimization in existing technologies.

[0052] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] like Figure 1 As shown, the present invention provides a full-chain collaborative optimization method for assessing the resilience of a flexible supply chain, comprising:

[0054] S1: Acquire multi-source data from the supply chain and preprocess it to obtain a structured time-series dataset.

[0055] In one embodiment, the multi-source supply chain data includes order data, inventory data, replenishment and allocation data, fulfillment timeliness data, supplier delivery data, logistics anomaly data, price fluctuation data, capacity data, and material availability data, collected from systems at each node of the supply chain. Capacity data includes information such as equipment identification, rated capacity, equipment utilization rate, operating status, and changeover time.

[0056] Furthermore, for time-series data (order data, supplier delivery data, logistics anomaly data, price fluctuation data), historical data and the latest data within a preset time range are obtained. The preset time range is set according to business needs, such as the past 3 months or the past 12 months. For status-based data (inventory data, in-transit replenishment data, production capacity data), current status data and recent change data are obtained.

[0057] In addition, preprocessing includes: data cleaning to remove duplicate records and format errors; outlier detection and correction, using detection methods based on business rules or statistical distributions; time granularity unification, unifying data from different time granularities to a preset time window; data association, linking multi-source data through key identification fields (order identifier, product identifier, warehouse identifier, supplier identifier) ​​to form a unified end-to-end data view; and numerical feature standardization to eliminate differences in units.

[0058] S2: Extract features based on structured time-series datasets and construct a supply chain heterogeneous graph. Integrate the supply chain heterogeneous graph with the supply chain domain knowledge graph to obtain a supply chain semantic graph structure.

[0059] Specifically, the flowchart for generating the supply chain semantic graph structure is as follows: Figure 2 As shown, it includes:

[0060] S2.1: Extract statistical features and event features from structured time-series datasets.

[0061] In another embodiment, statistical features are calculated from the structured time-series dataset based on a sliding time window. These features include demand fluctuation coefficient, inventory turnover rate, supplier on-time delivery rate, capacity utilization rate, lead time, etc. Supply chain anomaly events, such as stockout events, supply disruption events, logistical anomaly events, and demand surge events, are identified from the structured time-series dataset. Each event includes its occurrence time and characteristic attributes corresponding to its type. These characteristic attributes include one or more of the following: duration, delay duration, surge magnitude, and scope of impact.

[0062] S2.2: Identify supply chain entities and interaction relationships based on structured time-series datasets, and construct a supply chain heterogeneous graph.

[0063] In one embodiment, supply chain entities are identified from a structured time-series dataset and graph nodes are generated. Supply chain entities include suppliers, factories, warehouses, and SKUs. Interaction relationships between entities are extracted from the structured time-series dataset and graph edges are generated. Interaction relationships include supply relationships, production relationships, allocation relationships, and substitution relationships. Graph nodes and graph edges are assembled to form a supply chain heterogeneous graph.

[0064] S2.3: Align the supply chain heterogeneous graph with the supply chain domain knowledge graph to extract the associated domain knowledge.

[0065] Preferably, based on the type attribute of the node, candidate entities of the same type are retrieved from the supply chain domain knowledge graph to form a candidate entity set; the attribute similarity between the node and the candidate entities is calculated, and candidate entities with attribute similarity exceeding the attribute similarity threshold are taken as the alignment result. The attribute similarity threshold is determined based on the verification accuracy of historical alignment samples; when there are multiple candidate entities that meet the conditions, the one with the highest similarity is selected as the alignment result. The attribute similarity is calculated based on the comparable attributes of the node and the candidate entity, which include one or more of name, geographical location, and business category.

[0066] Furthermore, for successfully aligned nodes and edges, the associated business rules and constraints are extracted from the supply chain domain knowledge graph. The business rules include safety stock calculation rules, replenishment trigger rules, etc., and the constraints include supplier capacity limits, warehouse capacity limits, minimum order quantities, etc.

[0067] It should be noted that the knowledge graph in the supply chain domain is built from standard documents and industry best practices in supply chain management. It includes supply chain entity types and their typical attributes, as well as supply chain relationship types and their business constraints.

[0068] The entity alignment method, which combines type constraints and attribute similarity, first narrows down the candidate range through type constraints and then performs precise matching through attribute similarity, avoiding cross-type misalignment and improving the accuracy and efficiency of entity alignment.

[0069] S2.4: Statistical features, event features, and domain knowledge are added as attributes to the corresponding nodes and edges of the supply chain heterogeneous graph to obtain the supply chain semantic graph structure.

[0070] Furthermore, the parameter values ​​of the constraints and the type identifiers of the business rules are attached as knowledge attributes to the corresponding nodes and edges; statistical features are attached to the corresponding types of nodes and edges; and event features are attached to the corresponding nodes and edges in the form of an event record list.

[0071] By adopting the above technical solution, the fusion representation of temporal dynamic information and static domain knowledge is realized, enabling the supply chain semantic graph structure to have topological relevance, temporal dynamism and business semantics, providing multi-dimensional information support for subsequent resilience assessment and weak link identification.

[0072] S3: Based on the supply chain semantic graph structure, it generates a set of resilience indicators and identifies weak links in the supply chain by combining rule calculation and large language model reasoning.

[0073] Specifically, the resilience assessment flowchart is as follows: Figure 3 As shown, it includes:

[0074] S3.1: Convert the supply chain semantic graph structure into a text description to generate a supply chain status description text.

[0075] In this step, by traversing the nodes and edges in the supply chain semantic graph structure, the types, identifiers, statistical features and event features of the nodes are extracted, the relationship types and connection information of the edges are extracted, and the node information and relationship information are organized into natural language descriptions according to the text template to form the supply chain status description text.

[0076] S3.2: Calculate redundancy, response speed and visibility indicators based on the supply chain semantic graph structure, and identify abnormal nodes based on indicator thresholds and event characteristics.

[0077] Furthermore, the calculation of the redundancy index includes: counting the number of alternative supply sources for each key node in the supply chain semantic graph structure, extracting the capacity and geographical location information of alternative supply sources, and calculating the supply source redundancy based on the number, capacity scale and geographical distribution of alternative supply sources; extracting the actual inventory and statistical characteristics of nodes, combining the safety stock calculation rules obtained from the supply chain domain knowledge graph, and calculating the inventory redundancy based on the ratio of actual inventory to safety stock.

[0078] The calculation of the response speed index includes: extracting the actual lead time characteristics of the node, obtaining the standard lead time benchmark value from the supply chain domain knowledge graph, and calculating the internal response speed based on the ratio of the actual lead time to the standard lead time; extracting the actual delivery cycle and promised delivery cycle of the supplier node, and calculating the supplier response speed based on the ratio of the actual delivery cycle to the promised delivery cycle.

[0079] The calculation of visibility metrics includes: extracting the occurrence time and discovery time of abnormal events from event features, and calculating the abnormal event discovery lag; extracting the timestamp information of node data, and calculating the data update delay; and calculating the visibility metrics based on the weighted average of the abnormal event discovery lag and the data update delay.

[0080] Furthermore, based on the statistical distribution of various indicators during historical normal operation, and combined with industry benchmark values ​​obtained from the knowledge graph of the supply chain domain, the judgment thresholds for redundancy indicators, response speed indicators, and visibility indicators are determined. Threshold judgments are performed on each indicator: for redundancy indicators, an indicator value below the corresponding threshold is judged as abnormal; for response speed indicators, an indicator value above the corresponding threshold is judged as abnormal; and for visibility indicators, an indicator value above the corresponding threshold is judged as abnormal. Event features attached to nodes and edges are extracted, and event feature judgments are performed in combination with anomaly frequency, fluctuation amplitude, and resource utilization. Nodes that meet the conditions of exceeding the indicator threshold or event feature abnormality are identified and marked as abnormal nodes.

[0081] S3.3: Fill the resilience assessment prompt template with the supply chain status description text, abnormal node details and calculated indicator values ​​to generate complete prompt words.

[0082] Specifically, the node identifier, anomaly type, indicator deviation degree, and associated events are extracted from the identified abnormal nodes to generate abnormal node details; the calculated redundancy indicators, response speed indicators, and visibility indicators are summarized to generate an indicator value list; the generated supply chain status description text, abnormal node details, and indicator value list are filled into the resilience assessment prompt template to generate a complete prompt word.

[0083] In one embodiment, the resilience assessment prompt template is a pre-designed structured text template used to organize supply chain data into an input format that is understandable by a large language model.

[0084] S3.4: Input the complete prompt words into the large language model for inference and generate the raw results of the resilience assessment.

[0085] Further, the generated complete prompt words are input into the large language model; the large language model infers based on the supply chain status, abnormal node information, and calculated indicators in the complete prompt words, and performs flexibility assessment, coordination assessment, and weak link attribution according to the assessment task instructions; the text results output by the large language model are obtained to obtain the original resilience assessment results, including flexibility score, flexibility constraint list, coordination score, coordination problem list, bottleneck node list, root cause list, and risk transmission path list; the format completeness of the original resilience assessment results is verified. If the format is incomplete, the complete prompt words are re-entered for inference; if the format is complete, the original resilience assessment results are output.

[0086] In one embodiment, the reasoning process of the large language model includes: in flexibility assessment, analyzing the availability of alternative resources, switching costs, capacity adjustment space, and inventory buffer capacity to comprehensively judge the supply chain's ability to adjust to emergencies and output a flexibility score; in coordination assessment, analyzing information visibility, the matching degree of response speed between nodes, and the impact of event linkage to comprehensively judge the coordination efficiency between supply chain nodes and output a coordination score; in weak link attribution, identifying bottleneck nodes based on abnormal nodes and indicator deviations, analyzing the structural causes, strategic causes, and resource constraints that lead to the bottleneck, and describing the transmission path of the impact of a single point of failure on the entire chain.

[0087] In another embodiment, the large language model is a pre-trained language model with natural language understanding and complex reasoning capabilities, such as Claude, GPT series, GLM series, etc.

[0088] S3.5: Extract flexibility and synergy indicators from the raw results of the resilience assessment and merge them with the calculated indicators to generate a set of resilience indicators.

[0089] Furthermore, flexibility score and synergy score fields are extracted from the original resilience assessment results, and the extracted flexibility score and synergy score are used as flexibility index and synergy index, respectively. The flexibility index and synergy index are then merged with the redundancy index, response speed index and visibility index calculated in S3.2 to generate a resilience index set.

[0090] S3.6: Extract bottleneck nodes, root causes, and risk transmission paths from the original resilience assessment results to identify weak links in the supply chain.

[0091] Furthermore, bottleneck node list, root cause list, and risk transmission path list fields are extracted from the original resilience assessment results. The extracted bottleneck node list is used as the key weak nodes in the supply chain, the extracted root cause list is used as the root cause of the weak links, and the extracted risk transmission path list is used as the risk propagation mechanism in the supply chain. A comprehensive description of the weak links in the supply chain is determined based on the bottleneck nodes, root causes, and risk transmission paths.

[0092] By combining rule-based calculation with large language model reasoning, this invention achieves multi-dimensional quantitative assessment and in-depth root cause analysis of supply chain resilience. Specifically, the redundancy, response speed, and visibility indicators obtained through rule-based calculation objectively reflect the structural characteristics and operational status of the supply chain. The flexibility and coordination indicators obtained through large language model reasoning assess the supply chain's dynamic adjustment capability in the face of uncertainty and the efficiency of inter-node coordination. The combination of these two indicators forms a set of resilience indicators that combines objective quantification with comprehensive judgment. At the same time, the large language model is used to perform root cause analysis and risk transmission path reasoning on bottleneck nodes, identifying deep structural problems and complex causal relationships, and providing targeted improvement directions for collaborative optimization.

[0093] S4: Input order data into the time series forecasting model to obtain future demand forecasts, and combine this with the current supply chain status to calculate the supply chain risk forecast results.

[0094] Specifically, S4 includes:

[0095] S4.1: Extract historical order data from the structured time-series dataset and construct an order time-series sequence.

[0096] S4.2: Extract features from the order time series to generate time series features.

[0097] Among them, time-series characteristics include, but are not limited to: historical demand values, time period characteristics (week, month, season, etc.), trend characteristics, lag characteristics, sliding window statistical characteristics (mean, standard deviation, quantiles), and external factor characteristics (promotional activities, holiday markings, etc.).

[0098] S4.3: Input the time series features into the time series prediction model to generate the demand prediction distribution parameters for each future time node.

[0099] Furthermore, the time-series forecasting model processes the time-series features and outputs the predicted distribution parameters for each future time node. When the model outputs parameterized distribution parameters, the mean of the distribution is extracted as the demand forecast value, and the standard deviation is extracted as the uncertainty measure. The forecast interval is calculated based on the confidence level and the distribution parameters. When the model outputs a set of quantiles, the median is extracted as the demand forecast value, and the quantile interval is extracted as the uncertainty measure. The corresponding quantile position is calculated based on the confidence level, and the corresponding quantile is selected as the upper and lower bounds of the forecast interval.

[0100] In one embodiment, the time series prediction model employs a deep learning-based probabilistic prediction architecture, which can output a prediction distribution containing uncertainty information; the prediction distribution can be represented by a parametric distribution (such as a normal distribution or a log-normal distribution) or a non-parametric quantile representation.

[0101] S4.4: Using the Monte Carlo simulation method, based on the current supply chain status and demand forecast distribution parameters, calculate the inventory distribution at each future time point.

[0102] Furthermore, current supply chain status data is acquired, including current inventory levels, replenishment quantities of in-transit replenishment orders, and estimated arrival times. Multiple random samples are performed based on demand forecast distribution parameters to generate multiple future demand fulfillment paths. For each demand fulfillment path, using the current inventory level as the initial value, the inventory status is projected node by node in chronological order. The inventory level at each time node is equal to the inventory level at the previous time node plus the arrival quantity at this node minus the demand sampling value at this node. For each time node, statistical characteristics are calculated based on the inventory results of all simulated paths, including the expected value, standard deviation, and key quantile values ​​of the predicted inventory level, forming a description of the inventory distribution at that time node.

[0103] S4.5: Based on inventory distribution, calculate risk indicators for each future time point and generate supply chain risk prediction results.

[0104] Furthermore, for each forecast time point, the proportion of simulated paths with negative inventory levels is statistically analyzed to obtain the stockout risk probability; for each forecast time point, the expected value of the stockout quantity in all simulated paths is calculated to obtain the expected stockout quantity; the forecast inventory level, inventory distribution characteristics, stockout risk probability, and expected stockout quantity for each forecast time point are integrated to generate the supply chain risk forecast result.

[0105] By quantifying the uncertainty of demand into a probability distribution and transmitting it to inventory risk assessment through Monte Carlo simulation, the system achieves accurate quantification and forward-looking early warning of stockout risk, improves the accuracy and reliability of risk prediction, and enhances the robustness and resilience of the supply chain in uncertain environments.

[0106] S5: Generate a set of candidate solutions based on weak links in the supply chain, future demand forecasts, and supply chain risk forecasts.

[0107] Specifically, S5 includes:

[0108] S5.1: Determine the corresponding candidate solution type based on the type of bottleneck node in the weak link of the supply chain.

[0109] In one embodiment, replenishment and alternative supply plans are generated for supplier nodes, capacity adjustment and production scheduling plans are generated for production facility nodes, and allocation plans are generated for warehouse nodes.

[0110] S5.2: Based on the stockout risk probability and prediction time points in the supply chain risk forecast results, identify the key time windows in which the stockout risk probability exceeds the risk threshold.

[0111] Specifically, each predicted time node and its corresponding stockout risk probability are extracted from the supply chain risk forecast results; the stockout risk probability of each predicted time node is compared with the risk threshold to identify the time nodes where the stockout risk probability exceeds the risk threshold; the identified time nodes are arranged in chronological order, and consecutive time nodes are merged to form a time period; the time period is determined as the critical time window, and the start and end times and the highest risk probability value of the critical time window are recorded.

[0112] In one embodiment, the risk threshold is set as follows: based on the enterprise's target service level and historical stockout data, statistical analysis is used to determine the frequency of actual stockouts when the stockout risk probability exceeds a certain level, and the minimum risk probability that can meet the target service level constraint is selected as the threshold; this threshold is usually in the range of 40% to 60%.

[0113] In other embodiments, the risk threshold can be dynamically adjusted according to the enterprise's risk tolerance and business scenario. For example, a lower threshold (e.g., 30%) can be used during important promotional periods, while a higher threshold (e.g., 60%) can be used during regular sales periods.

[0114] S5.3: For key time windows and candidate solution types, retrieve available resource nodes from the supply chain semantic graph structure, and generate candidate solution sets of various types based on the resource attributes of the resource nodes.

[0115] Furthermore, the types of resource nodes to be retrieved are determined based on the types of candidate solutions; resource nodes of the corresponding types are retrieved from the supply chain semantic graph structure, and the resource attributes of the resource nodes are extracted; the availability of resource nodes is determined based on the time constraints of the key time window and the resource attributes of the resource nodes; for available resource nodes, the solution parameters of the candidate solutions are determined based on their resource attributes; the solution parameters include the number of resources, execution time, and estimated cost; and candidate solutions of various types are generated based on the solution parameters to form a set of candidate solutions.

[0116] S5.4: Calculate the resource consumption, risk mitigation effect, and feasibility score of each candidate solution in the candidate solution set.

[0117] Furthermore, based on the candidate solution's parameters and resource node attributes, the resource consumption of the candidate solution is calculated, including procurement costs, production costs, inventory costs, and logistics costs. Based on the candidate solution's resource quantity and execution time, combined with the current inventory status and future demand forecasts, the expected inventory level at each predicted time node after the candidate solution's execution is calculated. Based on the expected inventory level and future demand forecasts, the stockout risk probability after the candidate solution's execution is calculated and compared with the stockout risk probability before execution to obtain the risk mitigation effect. Based on the resource availability of resource nodes and the execution time constraints of the candidate solution, the feasibility score of the candidate solution is calculated.

[0118] S5.5: Filter the candidate solution set based on the feasibility score, retain the candidate solutions with feasibility scores higher than the feasibility threshold, and obtain the candidate solution set containing solution attribute information.

[0119] It should be noted that the feasibility threshold in this invention can be flexibly set according to the actual application scenario. In a preferred embodiment, the feasibility threshold is set to 0.7. In other embodiments, the feasibility threshold can be dynamically determined based on the statistical characteristics of the candidate solution set.

[0120] The candidate solution set obtained after screening includes solution attribute information for each solution, including solution type, target node, execution time window, resource consumption, risk mitigation effect and feasibility score.

[0121] By linking weak links in the supply chain, risk prediction results, and available resources, targeted generation and precise screening of candidate solutions were achieved. Identifying key time windows focused solution generation on high-risk periods, improving the timeliness of countermeasures; retrieving available resources and calculating risk mitigation effects based on the supply chain semantic graph structure ensured the feasibility of the generated solutions; and quantitative evaluation screened high-quality candidate solutions provided an effective solution pool for subsequent optimization decisions, improving the accuracy and efficiency of supply chain risk response.

[0122] S6: Input the set of resilience indices and candidate solutions into the multi-objective optimization model for solution to generate a set of cooperative instructions.

[0123] Specifically, S6 includes:

[0124] S6.1: Based on the set of resilience indicators, determine the weight coefficients of each optimization objective in the multi-objective optimization model.

[0125] Preferably, the overall resilience score of the supply chain is calculated based on the values ​​of each resilience index in the resilience index set, and the weight coefficients of each optimization objective are determined based on the overall resilience score.

[0126] Specifically, the overall resilience score of the supply chain is calculated based on the redundancy, response speed, visibility, flexibility and coordination indicators in the resilience index set. The weight coefficients of each optimization objective are calculated based on the mapping relationship between the overall resilience score and the weight coefficients of each optimization objective. The mapping relationship is as follows: the lower the resilience score, the higher the weight coefficient of the target of minimizing stockout risk and the lower the weight coefficient of the target of minimizing total cost.

[0127] It should be noted that the overall resilience score can be calculated using different methods depending on the actual application scenario. In a preferred embodiment, a simple averaging method is used to calculate the overall resilience score. In other embodiments, a weighted average method or a minimum value method can be used.

[0128] S6.2: Encode the candidate solution set as decision variables and construct a multi-objective optimization model.

[0129] Specifically, a binary decision variable is assigned to each candidate solution in the candidate solution set; a multi-objective optimization function is constructed with the optimization objectives of minimizing stockout risk, minimizing recovery time, and minimizing total cost. The stockout risk objective function is calculated based on the risk mitigation effect of each candidate solution, the recovery time objective function is calculated based on the execution time, and the total cost objective function is calculated based on resource consumption. Constraints are set, including the upper limit of supplier capacity, the upper limit of warehouse capacity, the capacity of logistics network, and the lower limit of service level.

[0130] S6.3: Use a multi-objective optimization algorithm to solve the multi-objective optimization model and obtain the Pareto optimal solution set.

[0131] S6.4: Select the optimal solution from the Pareto optimal solution set according to the weight coefficients, and convert the candidate solution combination corresponding to the optimal solution into a cooperative instruction set.

[0132] Further, the weighted objective function value of each solution in the Pareto optimal solution set is calculated according to the weighting coefficients determined in S6.1, and the solution with the smallest weighted objective function value is selected as the optimal solution; the selected candidate solution combination is determined according to the values ​​of each decision variable in the optimal solution; and the instructions are converted according to the type of the candidate solution: the replenishment plan and the alternative supply plan are converted into procurement instructions, the capacity adjustment plan and the production scheduling plan are converted into production instructions, and the allocation plan is converted into warehousing and distribution instructions, thereby generating a collaborative instruction set.

[0133] The optimization target weights are dynamically adjusted based on the resilience index set, enabling the optimization strategy to adapt to the current resilience status of the supply chain and avoiding decision-making rigidity caused by fixed weights. Multi-objective optimization is adopted to solve the problem and select the best solution from the Pareto optimal solution set, achieving a balance between stockout risk, recovery time and total cost. The optimization results are converted into an executable set of collaborative instructions, realizing closed-loop management from decision-making to execution.

[0134] S7: Executes the collaborative instruction set and collects feedback data, then updates the model parameters based on the feedback data.

[0135] Specifically, the system executes a set of collaborative instructions; collects feedback data generated during instruction execution, including arrival data, output data, fulfillment data, and sales data; calculates prediction and execution deviations based on the feedback data; updates the parameters of the time-series prediction model based on the prediction deviations; and updates the parameters of the large language model and the multi-objective optimization model based on the execution deviations.

[0136] In summary, this invention achieves comprehensive perception and deep understanding of the supply chain status by constructing a supply chain semantic graph structure and integrating domain knowledge; it improves the accuracy and interpretability of resilience assessment by combining rule calculation and large language model reasoning to generate a set of resilience indicators; it achieves quantitative assessment and forward-looking early warning of supply chain risks by adopting a probabilistic risk prediction method; it generates targeted candidate solutions based on weak links and risk prediction results, and performs global optimization through a multi-objective optimization model with dynamic weight adjustment, achieving a balance between risk control, response speed, and cost-effectiveness; and it continuously updates model parameters through execution feedback data, forming a closed-loop mechanism of "assessment-prediction-optimization-execution-feedback," enabling the quality of decision-making to continuously improve during operation.

[0137] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for end-to-end collaborative optimization of flexible supply chain resilience assessment, characterized in that, include: Obtain multi-source data from the supply chain and preprocess it to obtain a structured time-series dataset; The multi-source supply chain data includes order data, inventory data, replenishment and allocation data, fulfillment timeliness data, supplier delivery data, logistics anomaly data, price fluctuation data, production capacity data, and material availability data. Features are extracted based on the structured time-series dataset and a supply chain heterogeneous graph is constructed. The supply chain heterogeneous graph is then fused with a supply chain domain knowledge graph to obtain a supply chain semantic graph structure. Based on the aforementioned supply chain semantic graph structure, a set of resilience indicators is generated and weak links in the supply chain are identified through a combination of rule calculation and large language model reasoning. Order data is input into a time-series forecasting model to obtain future demand forecasts, and the supply chain risk forecasting results are calculated by combining the current supply chain status. A set of candidate solutions is generated based on the weak links in the supply chain, future demand forecasts, and supply chain risk forecasts. The set of resilience indices and the set of candidate solutions are input into a multi-objective optimization model for solution, generating a set of cooperative instructions. Execute the collaborative instruction set and collect feedback data, then update the model parameters based on the feedback data; The generation of resilience indicator sets and identification of weak links in the supply chain includes: Convert the supply chain semantic graph structure into a text description to generate a supply chain status description text. Based on the semantic graph structure of the supply chain, redundancy indicators, response speed indicators, and visibility indicators are calculated, and abnormal nodes are identified according to indicator thresholds and event characteristics. Fill the resilience assessment prompt template with the supply chain status description text, abnormal node details and calculated indicator values ​​to generate a complete prompt; The complete prompt words are input into a large language model for inference to generate the raw results of the resilience assessment. The flexibility and synergy indices are parsed from the original resilience assessment results and merged with the calculated indices to generate a set of resilience indices. Bottleneck nodes, root causes, and risk transmission paths are extracted from the original resilience assessment results to identify weak links in the supply chain.

2. The end-to-end collaborative optimization method for assessing the resilience of flexible supply chains according to claim 1, characterized in that, The obtained supply chain semantic graph structure includes: Extracting statistical and event features from structured time-series datasets; Based on structured time-series datasets, identify supply chain entities and interaction relationships, and construct a supply chain heterogeneity graph; Align the heterogeneous supply chain graph with the supply chain domain knowledge graph to extract the associated domain knowledge; The statistical features, event features, and domain knowledge are attached as attributes to the corresponding nodes and edges of the heterogeneous supply chain graph to obtain the supply chain semantic graph structure.

3. The end-to-end collaborative optimization method for assessing the resilience of flexible supply chains according to claim 1, characterized in that, The calculation of supply chain risk prediction results based on the current supply chain status includes: Historical order data is extracted from structured time-series datasets to construct order time-series sequences; Feature extraction is performed on the order time series to generate time series features; The time-series features are input into the time-series prediction model to generate demand prediction distribution parameters for each future time node. Using the Monte Carlo simulation method, based on the current supply chain status and demand forecast distribution parameters, the inventory distribution at each future time point is calculated; Based on the inventory distribution, risk indicators for each future time point are calculated to generate supply chain risk prediction results.

4. The end-to-end collaborative optimization method for assessing the resilience of flexible supply chains according to claim 1, characterized in that, The process of generating a candidate solution set based on the weak links in the supply chain, future demand forecasts, and supply chain risk predictions includes: Based on the types of bottleneck nodes and optimization needs in the weak links of the supply chain, the corresponding candidate solution types are determined; Based on the stockout risk probability and forecast time points in the supply chain risk forecast results, identify the key time windows in which the stockout risk probability exceeds the risk threshold. For the key time window and candidate solution type, available resource nodes are retrieved from the supply chain semantic graph structure, and candidate solution sets of each type are generated based on the resource attributes of the resource nodes. Calculate the resource consumption, risk mitigation effect, and feasibility score of each candidate solution in the candidate solution set; Based on the feasibility score, a set of candidate solutions is selected, and candidate solutions with feasibility scores higher than the feasibility threshold are retained to obtain a set of candidate solutions containing solution attribute information.

5. The end-to-end collaborative optimization method for assessing the resilience of flexible supply chains according to claim 1, characterized in that, The generation of the cooperative instruction set includes: Based on the set of resilience indicators, the weight coefficients of each optimization objective in the multi-objective optimization model are determined; The candidate solution set is encoded as decision variables to construct a multi-objective optimization model; A multi-objective optimization algorithm is used to solve the multi-objective optimization model to obtain the Pareto optimal solution set; The optimal solution is selected from the Pareto optimal solution set according to the weight coefficients, and the candidate scheme combination corresponding to the optimal solution is converted into a cooperative instruction set.

6. The end-to-end collaborative optimization method for assessing the resilience of flexible supply chains according to claim 1, characterized in that, The step of updating the model parameters based on the feedback data includes: Execute the collaborative instruction set and collect feedback data generated during instruction execution; Calculate the prediction deviation and execution deviation based on the feedback data; The time series prediction model parameters are updated based on the prediction bias, and the large language model parameters and multi-objective optimization model parameters are updated based on the execution bias.

Images