A Medical Device Supply Chain Recommendation Method and System Based on Large Language Model

Abstract

This invention provides a medical device supply chain recommendation method and system based on a large language model, belonging to the field of model-assisted technology. The method includes acquiring demand descriptions and converting them into structured vectors, constructing a domain knowledge graph to perform multi-hop reasoning to identify collaborative paths, generating adversarial disturbance analysis to reconstruct failure modes and the supply chain, establishing a risk propagation graph to identify trigger points and generate hierarchical prevention and control strategies. This method can effectively improve the accuracy of intelligent recommendations for the medical device supply chain, enhance the supply chain's robustness and risk response capabilities.

Description

Technical Field

[0001] This invention relates to model-assisted technology, and more particularly to a medical device supply chain recommendation method and system based on large language model assistance. Background Technology

[0002] Medical device supply chain management is a crucial link in ensuring the normal operation of the healthcare system, encompassing the entire process from raw material procurement and manufacturing to final distribution and use. With the increasing demand for healthcare services and the strengthening of globalization, the medical device supply chain has become increasingly complex. Traditional medical device supply chain management relies primarily on human experience and simple information systems for decision-making, making it difficult to cope with the changing demands and unforeseen events in the modern healthcare environment. In recent years, the development of artificial intelligence, especially large language models, has provided new technological support for medical device supply chain management, enabling more intelligent and personalized supply chain recommendations.

[0003] Existing technologies lack the ability to effectively process unstructured medical needs descriptions. Most of them use predefined needs classification methods, which make it difficult to capture the complex and ever-changing needs expressions and implicit information in clinical practice, resulting in discrepancies between supply chain recommendations and actual needs.

[0004] Traditional supply chain recommendation methods are usually based on simple matching algorithms or rule systems. They lack the understanding and reasoning ability of deep causal relationships in the medical device supply chain, and cannot identify potential supply chain collaboration paths and risk points. The recommendation results are often too simplistic and difficult to adapt to the high reliability requirements of the medical environment.

[0005] Existing supply chain recommendation systems lack robust assessment and dynamic adjustment mechanisms when facing emergencies such as supply disruptions, quality anomalies, and sudden changes in demand. They are unable to simulate possible failure scenarios in advance and formulate corresponding prevention and control strategies, resulting in limited ability to cope with risks in practical applications and difficulty in ensuring the continuity and stability of medical device supply. Summary of the Invention

[0006] This invention provides a medical device supply chain recommendation method and system based on a large language model, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a medical device supply chain recommendation method based on a large language model, comprising:

[0008] Obtain the medical device requirement description, and use a large language model to convert the medical device requirement description into a structured requirement vector;

[0009] A domain knowledge graph of the medical device supply chain is constructed. A large language model is used to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links. The supply chain collaboration paths are transformed into embedded representations of candidate supply chain solutions through graph representation learning.

[0010] For the embedded representation of the candidate supply chain solution, adversarial perturbations are generated by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applied to the embedded representation. The semantic deviation between the perturbed embedded representation and the structured demand vector is analyzed using a large language model to identify failure modes and generate failure cause descriptions. Based on the failure modes, alternative paths are re-searched on the domain knowledge graph or redundant nodes are introduced for robust reconstruction.

[0011] A risk propagation graph is established based on the failure modes. Risk trigger points are identified by tracing the diffusion path of risks in the risk propagation graph. The temporal evolution characteristics of the risk trigger points are extracted as risk feature vectors. The large language model is used to generate hierarchical prevention and control strategies for the reconstructed supply chain solution based on the risk feature vectors. The hierarchical prevention and control strategies are transformed into new constraint relationship edges to enhance the risk resistance capability of the solution.

[0012] Constructing a domain knowledge graph for the medical device supply chain, and using a large language model to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths, includes:

[0013] The structured demand vector is used to extract the implicit functional dependencies and scenario constraints in the demand through a hierarchical semantic decomposition algorithm. Based on the functional dependencies, the instrument classification node that matches the semantics of the demand is located in the domain knowledge graph as the starting point for inference. The scenario constraints are then converted into path filtering rules for graph traversal.

[0014] The large language model is used to expand and traverse outward from the inference starting point along the supply and demand matching relationship edge. During the traversal, the supplier capability nodes connected by each relationship edge are evaluated in real time according to the path filtering rules to see if they meet the scenario constraints. Supplier capability nodes that meet the conditions are dynamically aggregated into the current hop reachable node set.

[0015] For each supplier capability node in the set of reachable nodes, the large language model is used to extend downstream logistics service nodes along the quality traceability relationship edge and the timeliness guarantee relationship edge. During the extension process, the timeliness requirements and quality standards in the structured demand vector are combined to perform semantic similarity calculation on the reachability of each downstream node, and the logistics service nodes that can form a complete supply chain closed loop are selected. The complete path from the inference starting point through the supplier capability node to the logistics service node is identified as the supply chain collaborative path.

[0016] The method involves using a large language model to analyze the semantic deviation between the perturbed embedded representation and the structured requirement vector to identify failure modes and generate failure cause descriptions. Based on these failure modes, a robust reconstruction is then performed by re-searching for alternative paths or introducing redundant nodes on the domain knowledge graph.

[0017] The embedded representation of the supply chain solution after applying adversarial perturbation is semantically aligned with the structured demand vector. The degree of semantic deviation is quantified by calculating the semantic distance metric between the embedded representation after perturbation and the structured demand vector. Embedded representations with a semantic deviation exceeding a preset deviation threshold are marked as failed solutions.

[0018] The failure mode of the failed scheme is identified using a large language model. The perturbed embedded representation is mapped back to the corresponding path in the domain knowledge graph. By analyzing the magnitude of the representation changes of each node and relation edge in the path before and after the perturbation, the failed nodes and failed relation edges that cause semantic deviation are located.

[0019] Based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the adjacent nodes of the failed node to search for alternative paths that can bypass the failed node. When there are no alternative paths that meet the conditions, intermediate nodes with redundancy guarantee capabilities are inserted between the upstream and downstream nodes of the failed relationship edge to form a redundant path.

[0020] The semantic distance metric is calculated between the redundant path and the structured requirement vector. When the semantic distance metric is less than the preset deviation threshold, the reconstruction is confirmed to be successful.

[0021] Based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the neighboring nodes of the failed node to search for alternative paths that can bypass the failed node, including:

[0022] In the domain knowledge graph, locate all incoming and outgoing nodes of the failed node, analyze the semantic association strength between the incoming nodes and the structured demand vector using a large language model, and determine the incoming node with the highest semantic association strength as the starting node for alternative path search.

[0023] The process involves expanding the traversal along the relation edges in the domain knowledge graph from the starting node, performing functional completeness verification on each candidate path encountered, and marking the verified candidate paths as potential alternative paths.

[0024] For the potential alternative path, calculate the semantic matching degree between its endpoint node and the outgoing edge node. When the semantic matching degree meets the connection condition, establish a connection edge from the endpoint node to the outgoing edge node to form a complete alternative path that bypasses the failed node.

[0025] A risk propagation graph is established based on the aforementioned failure modes. By tracing the diffusion path of risk in the risk propagation graph, risk trigger points are identified. The temporal evolution characteristics of the risk trigger points are extracted into a risk feature vector, including:

[0026] The failure node in the failure mode is set as the initial risk source of the risk propagation graph. The upstream and downstream related nodes of the failure node are extracted from the domain knowledge graph. A directed risk propagation edge is established according to the business dependency relationship between the nodes. The direction of the directed risk propagation edge points to the downstream node affected by the failure node, and a risk propagation graph is constructed.

[0027] The risk propagation diagram simulates the risk diffusion process, propagating risk signals from the initial risk source to downstream nodes along the directed risk propagation edge, calculating the cumulative risk intensity of each node, and identifying nodes whose cumulative risk intensity exceeds their carrying capacity as secondary failure nodes.

[0028] Using a large language model, trace the node sequence and propagation edge sequence from the initial risk source to the secondary failure node, determine the complete path composed of the node sequence and the propagation edge sequence as the risk diffusion path, and locate the risk trigger point in the risk diffusion path;

[0029] Based on the temporal dimensionality reduction algorithm, the state change data of the risk trigger point within multiple historical time windows are reduced in dimensionality. The dimensionality-reduced state change data is organized into temporal evolution features in chronological order. The temporal evolution features are vectorized and encoded using the large language model to generate a risk feature vector that captures the dynamic evolution law of the risk trigger point.

[0030] Based on a time-series dimensionality reduction algorithm, the state change data of the risk trigger point within multiple historical time windows are reduced in dimensionality. The dimensionality-reduced state change data is then organized into time-series evolution features in chronological order, including:

[0031] Obtain the state change data of the risk trigger point within multiple historical time windows, construct a high-dimensional state vector from the multi-dimensional values ​​corresponding to each historical time window, and form a high-dimensional state vector sequence indexed by historical time windows.

[0032] The high-dimensional state vector sequence is subjected to time-series correlation analysis to calculate the variance of each monitoring dimension in the time series, and to identify the variance dimensions with a change magnitude lower than a preset magnitude threshold and the redundant dimensions with linear correlation in the multiple historical time windows.

[0033] The retention weights of each monitoring dimension are obtained by nonlinear mapping of the variance based on the time series dimensionality reduction algorithm. Basic retention weights are assigned to the variance dimension and the redundant dimension, and enhanced retention weights are assigned to the dominant dimension that contributes significantly to the time series change pattern. The high-dimensional state vector sequence is then projected using the basic retention weights and the enhanced retention weights to generate a low-dimensional state vector with a reduced number of dimensions.

[0034] After associating the low-dimensional state vector with the corresponding historical time window identifier, time series mapping is performed on the low-dimensional state vector to form time series evolution features.

[0035] The method of generating a hierarchical risk prevention and control strategy for the reconstructed supply chain solution using the large language model based on the risk feature vector, and transforming the hierarchical risk prevention and control strategy into newly added constraint relationship edges to enhance the risk resistance capability of the solution includes:

[0036] The numerical and temporal components in the risk feature vector are analyzed. Based on the difference in the amplitude of the numerical components and the rate of change of the temporal components, the risk is divided into multiple risk levels. For each risk level, the corresponding response time requirements and prevention and control priorities are determined.

[0037] The risk sensitivity of each node in the reconstructed supply chain scheme is assessed by using the prevention and control priority assessment, and important nodes with failure probabilities exceeding preset probability thresholds and weak edges with insufficient risk propagation blocking capabilities are identified under different risk levels. The important nodes and the weak edges are then taken as the key targets for prevention and control.

[0038] Construct a hierarchical prevention and control strategy for the key prevention and control targets, and refine the hierarchical prevention and control strategy into enhanced constraint rules for the important nodes and backup path constraint rules for the weak edges according to the response time requirements;

[0039] The strengthening constraint rules are transformed into new constraint relationship edges connecting the important nodes and their limited collaborative objects. The backup path constraint rules are transformed into new constraint relationship edges connecting the starting point of the weak edge and the ending point of the redundant path. New constraint relationship edges are added to the knowledge graph of the reconstructed supply chain solution to enhance the risk resistance capability of the solution.

[0040] A second aspect of this invention provides a medical device supply chain recommendation system based on a large language model, comprising:

[0041] The first unit is used to obtain the medical device requirement description and convert the medical device requirement description into a structured requirement vector using a large language model.

[0042] The second unit is used to construct a domain knowledge graph of the medical device supply chain. It uses a large language model to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links, and the supply chain collaboration paths are converted into embedded representations of candidate supply chain solutions through graph representation learning.

[0043] The third unit is used for the embedded representation of the candidate supply chain solutions. It generates adversarial disturbances by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applies them to the embedded representation. It uses a large language model to analyze the degree of semantic deviation between the embedded representation after the disturbance and the structured demand vector to identify failure modes and generate failure cause descriptions. Based on the failure modes, it re-searches alternative paths on the domain knowledge graph or introduces redundant nodes for robust reconstruction.

[0044] The fourth unit is used to establish a risk propagation graph based on the failure modes, identify risk trigger points by tracing the diffusion path of risks in the risk propagation graph, extract the temporal evolution characteristics of the risk trigger points as risk feature vectors, use the large language model to generate hierarchical prevention and control strategies for the reconstructed supply chain solution based on the risk feature vectors, and transform the hierarchical prevention and control strategies into newly added constraint relationship edges to enhance the risk resistance capability of the solution.

[0045] A third aspect of the present invention provides an electronic device, comprising:

[0046] processor;

[0047] Memory used to store processor-executable instructions;

[0048] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] The beneficial effects of this application are as follows:

[0051] This invention combines the semantic analysis capabilities of a large language model with the structured representation of a knowledge graph to achieve intelligent recommendation in the medical device supply chain. It effectively solves the problem that traditional supply chain recommendation methods struggle to handle complex semantic requirements and implicit relationships, making the recommendation results more aligned with actual business needs.

[0052] This invention uses adversarial perturbation to simulate various risk scenarios that the supply chain may face. By analyzing the degree of semantic deviation after the perturbation, it can identify potential failure modes in advance and perform robust reconstruction, which significantly improves the risk resistance and reliability of the recommended supply chain solution and reduces the risk of supply interruption in actual operation.

[0053] Based on risk propagation diagrams and time-series characteristic analysis, this invention generates hierarchical prevention and control strategies and transforms them into constraint relationships, forming a complete closed-loop mechanism for supply chain risk management. This mechanism can not only cope with known risks, but also adapt to the uncertainties unique to the medical device field, such as regulatory changes and quality fluctuations, providing a safer and more efficient solution for medical device supply chain management. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the medical device supply chain recommendation method based on a large language model, as described in an embodiment of the present invention.

[0055] Figure 2 This is a flowchart of the graph path reconstruction and semantic distance verification and correction process in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1This is a flowchart illustrating the medical device supply chain recommendation method based on a large language model, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Obtain the medical device requirement description, and use a large language model to convert the medical device requirement description into a structured requirement vector;

[0060] A domain knowledge graph of the medical device supply chain is constructed. A large language model is used to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links. The supply chain collaboration paths are transformed into embedded representations of candidate supply chain solutions through graph representation learning.

[0061] For the embedded representation of the candidate supply chain solution, adversarial perturbations are generated by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applied to the embedded representation. The semantic deviation between the perturbed embedded representation and the structured demand vector is analyzed using a large language model to identify failure modes and generate failure cause descriptions. Based on the failure modes, alternative paths are re-searched on the domain knowledge graph or redundant nodes are introduced for robust reconstruction.

[0062] A risk propagation graph is established based on the failure modes. Risk trigger points are identified by tracing the diffusion path of risks in the risk propagation graph. The temporal evolution characteristics of the risk trigger points are extracted as risk feature vectors. The large language model is used to generate hierarchical prevention and control strategies for the reconstructed supply chain solution based on the risk feature vectors. The hierarchical prevention and control strategies are transformed into new constraint relationship edges to enhance the risk resistance capability of the solution.

[0063] In one alternative implementation, a domain knowledge graph of the medical device supply chain is constructed, and multi-hop reasoning is performed on the domain knowledge graph based on the structured demand vector using a large language model to identify implicit supply chain collaboration paths, including:

[0064] The structured demand vector is used to extract the implicit functional dependencies and scenario constraints in the demand through a hierarchical semantic decomposition algorithm. Based on the functional dependencies, the instrument classification node that matches the semantics of the demand is located in the domain knowledge graph as the starting point for inference. The scenario constraints are then converted into path filtering rules for graph traversal.

[0065] The large language model is used to expand and traverse outward from the inference starting point along the supply and demand matching relationship edge. During the traversal, the supplier capability nodes connected by each relationship edge are evaluated in real time according to the path filtering rules to see if they meet the scenario constraints. Supplier capability nodes that meet the conditions are dynamically aggregated into the current hop reachable node set.

[0066] For each supplier capability node in the set of reachable nodes, the large language model is used to extend downstream logistics service nodes along the quality traceability relationship edge and the timeliness guarantee relationship edge. During the extension process, the timeliness requirements and quality standards in the structured demand vector are combined to perform semantic similarity calculation on the reachability of each downstream node, and the logistics service nodes that can form a complete supply chain closed loop are selected. The complete path from the inference starting point through the supplier capability node to the logistics service node is identified as the supply chain collaborative path.

[0067] The knowledge graph construction in the medical device supply chain adopts a three-layer entity relationship architecture. The bottom layer consists of medical device classification nodes, storing structured fields such as device type codes, functional attribute vectors, technical specifications, and regulatory classification identifiers. Each node contains 50 to 120 attribute fields, and attribute values ​​support string, numeric, and boolean data types. The middle layer consists of supplier capability nodes, recording core attributes such as unique supplier identifiers, production qualification levels, production capacity, quality system certification status, and geographic coordinates. The number of attribute fields for each node is controlled within the range of 80 to 150. The top layer consists of logistics service nodes, containing key features such as service provider identifiers, transportation mode types, delivery time commitments, temperature-controlled storage capabilities, and coverage areas. The number of attribute fields is 60 to 100. Nodes are connected through supply and demand matching edges between device classifications and supplier capabilities, quality traceability edges between suppliers and historical quality records, and timeliness assurance edges between suppliers and logistics services. Each edge contains metadata such as weight coefficients, creation timestamps, validity periods, and constraint labels. The graph uses a distributed graph database for storage, supporting tens of millions of nodes. The response time for a single query is controlled within 200 milliseconds, and data updates use an incremental synchronization mechanism with an update delay of no more than 1 hour.

[0068] The structured requirement vector is parsed using a hierarchical semantic decomposition algorithm. This algorithm segments the input 768-dimensional vector into semantic blocks: functional requirements occupy the first 200 dimensions, quality constraints occupy dimensions 201 to 400, timeliness requirements occupy dimensions 401 to 600, and cost constraints occupy dimensions 601 to 768. The decomposition process employs a sliding window technique with a window size of 32 dimensions and a sliding step of 16 dimensions to ensure overlap between semantic segments. The algorithm identifies key semantic features through an attention weighting mechanism, using a multi-head attention calculation with 8 heads, each with 96 dimensions. Functional dependencies are extracted through semantic relevance analysis, calculating the Pearson correlation coefficient between vector segments. A dependency is established when the absolute value of the correlation coefficient exceeds 0.6, and the dependency strength is equal to the absolute value of the correlation coefficient. Scene constraint extraction uses a threshold detection method, statistically analyzing the values ​​of the quality constraint dimensions. Dimensions with a mean deviation exceeding 1.5 times the standard deviation are identified as hard constraints, while dimensions with a deviation between 0.5 and 1.5 times the standard deviation are identified as soft constraints. The decomposition algorithm runs in batch mode with a batch size of 16 vectors, keeping peak memory usage below 4GB and each decomposition taking 300 to 800 milliseconds.

[0069] The reasoning starting point localization employs a multi-stage semantic matching strategy. The first stage involves coarse-grained screening, performing batch cosine similarity calculations between the functional dependency matrix and the functional vectors of device classification nodes in the knowledge graph. Similarity calculations utilize vectorized operations, processing up to 1000 candidate nodes per iteration while maintaining 32-bit floating-point precision. When the similarity value exceeds 0.75, the node enters the fine-grained matching stage. Fine-grained matching combines scene constraints with a comprehensive score, using a weighted linear combination algorithm with a 0.6 weight for functional matching and a 0.4 weight for scene adaptation. Candidate reasoning starting points are sorted in descending order of score, with the top 5 to 15 nodes selected as the final starting point set. The number of nodes is dynamically adjusted based on matching quality; when the highest score is below 0.8, the set expands to 15 nodes, and when it's above 0.9, it shrinks to 5 nodes. When scene constraints are converted into graph traversal rules, each constraint generates a corresponding filtering function. These functions use conditional expressions and support operations such as numerical comparison, string matching, and set inclusion. The time constraint is transformed into a composite function of geographical distance calculation and transportation time prediction. The geographical distance is calculated using the Havesing formula, and the transportation time is predicted based on a regression model built on historical data.

[0070] The large language model employs an improved breadth-first search algorithm for multi-hop inference, limiting the search depth to 5 hops to avoid performance issues caused by excessively long paths. The model uses a pre-trained transformer architecture with 13 billion parameters, employing mixed-precision computation during inference. Weights are 16-bit floating-point numbers, and activation values ​​are 32-bit floating-point numbers. In each inference step, the model reads the adjacency relationships of the current node and filters for candidate expansion nodes based on relationship type, with a maximum of 100 candidate nodes. Semantic understanding of supplier capability nodes utilizes a multi-task learning framework, simultaneously predicting three objectives: capability matching degree, qualification compliance degree, and historical performance score. Predicted results are continuous values ​​between 0 and 1, with computational precision maintained to three decimal places. Path selection rules are deployed as real-time filters. These filters perform constraint checks on each candidate node, including geographical location restrictions, qualification level requirements, and minimum production capacity limits. Nodes that do not meet any hard constraints are immediately pruned; if soft constraints are not met, the node's weight is reduced but it remains in the search space. The set of reachable nodes is managed using a priority queue data structure. The queue is sorted by the comprehensive score of the nodes, and the queue capacity is dynamically expanded, with an initial capacity of 1000 positions and an expansion step of 500 positions.

[0071] The downstream expansion process is executed in parallel along the quality traceability and timeliness assurance relationships. The two expansion branches are implemented asynchronously and concurrently, with computational resources managed by a thread pool, the size of which is set to twice the number of CPU cores. The quality traceability branch focuses on evaluating the supplier's quality management capabilities, calculating a quality reliability score by analyzing historical quality records, certification validity, and the frequency of quality incidents. The timeliness assurance branch focuses on logistics and delivery capabilities, assessing the degree of timeliness fulfillment based on factors such as delivery network coverage, transportation configuration, and warehouse distribution density. The semantic similarity calculation of logistics service nodes employs a multi-dimensional feature fusion method. The feature vector includes dimensions such as service type encoding, timeliness commitment value, quality assurance level, and cost rate range, for a total of 128 dimensions. Similarity calculation uses weighted Manhattan distance, with the weights of each dimension dynamically adjusted according to the urgency of the demand. For urgent demands, the timeliness weight increases to 0.5, while for regular demands, the quality weight remains at 0.4. During the calculation, missing values ​​are filled using the mean, and outliers are detected and replaced using the 3σ criterion.

[0072] The supply chain closed-loop formation assessment employs a combination of path integrity verification and business logic checks. Integrity verification ensures the path includes three essential nodes: equipment classification, supplier capabilities, and logistics services. Business logic checks verify the rationality of relationships between nodes. Path quality assessment uses a multi-factor scoring model, with factors including path length penalty, node quality reward, relationship strength adjustment, and constraint fulfillment incentive. The length penalty uses an exponential decay function; for every additional hop in the path length, the score is multiplied by a decay coefficient of 0.9. The node quality reward is calculated based on the historical performance data of each node, with high-performing nodes receiving a score bonus of 0.1 to 0.3. The relationship strength adjustment fine-tunes the score based on edge weights, providing positive adjustment for strong relationships and negative adjustment for weak relationships, with the adjustment range controlled within 5% of the score. The constraint fulfillment incentive awards extra points to paths that fully meet hard constraints, with the bonus ranging from 10% to 20% of the base score. The final path score is calculated through a weighted summation; a path is marked as a valid collaborative path when its score exceeds a preset threshold of 70 points.

[0073] In the specific implementation case, the input demand vector is a 768-dimensional array with values ​​ranging from -1 to +1, and a vector sparsity of approximately 40%. Hierarchical decomposition extracts a 15×15-dimensional functional dependency matrix containing 42 non-zero elements and 7 scenario constraints, including 2 hard constraints on timeliness, 3 soft constraints on quality, and 2 cost constraints. The inference starting point location identifies 4 equipment classification nodes, numbered MED_A001, MED_B003, MED_C005, and MED_D008, with corresponding matching scores of 0.87, 0.83, 0.79, and 0.76, respectively. Multi-hop inference performs 3 rounds of expansion. The first round expands from the 4 starting points to obtain 35 candidate supplier nodes, from which 28 valid nodes are retained after path filtering. The second round further expands to obtain 67 candidate logistics service nodes, from which 52 nodes are retained after semantic similarity filtering. The third round verifies path integrity, ultimately generating 18 valid supply chain collaboration paths with path lengths ranging from 3 to 4 hops and path scores ranging from 72 to 91. The entire inference process took 1.8 seconds to compute, with a peak memory usage of 5.2 GB and a GPU resource utilization rate of 85%. The successfully identified cooperative paths covered 85% of the original requirement constraints.

[0074] In one optional implementation, a large language model is used to analyze the semantic deviation between the perturbed embedded representation and the structured demand vector to identify failure modes and generate failure cause descriptions. Based on the failure modes, alternative paths are re-searched on the domain knowledge graph or redundant nodes are introduced for robust reconstruction, including:

[0075] The embedded representation of the supply chain solution after applying adversarial perturbation is semantically aligned with the structured demand vector. The degree of semantic deviation is quantified by calculating the semantic distance metric between the embedded representation after perturbation and the structured demand vector. Embedded representations with a semantic deviation exceeding a preset deviation threshold are marked as failed solutions.

[0076] The failure mode of the failed scheme is identified using a large language model. The perturbed embedded representation is mapped back to the corresponding path in the domain knowledge graph. By analyzing the magnitude of the representation changes of each node and relation edge in the path before and after the perturbation, the failed nodes and failed relation edges that cause semantic deviation are located.

[0077] Based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the adjacent nodes of the failed node to search for alternative paths that can bypass the failed node. When there are no alternative paths that meet the conditions, intermediate nodes with redundancy guarantee capabilities are inserted between the upstream and downstream nodes of the failed relationship edge to form a redundant path.

[0078] The semantic distance metric is calculated between the redundant path and the structured requirement vector. When the semantic distance metric is less than the preset deviation threshold, the reconstruction is confirmed to be successful.

[0079] like Figure 2 As shown, the method includes:

[0080] The semantic alignment analysis of the embedded representation of the supply chain scheme and the structured demand vector after adversarial perturbation employs a multi-dimensional semantic mapping algorithm. This algorithm maps the 512-dimensional embedded representation and the 768-dimensional demand vector to a unified 256-dimensional semantic space for comparison. The semantic alignment process is implemented through a linear transformation layer. The transformation matrix parameters are obtained through pre-training, and the weights are initialized using a Xavier normal distribution with a learning rate of 0.001 and 10,000 training iterations. The aligned vectors are normalized using the L2 norm to ensure that the vector magnitude is 1, avoiding the impact of dimensional bias on distance calculation. The semantic distance metric uses a weighted cosine distance method. The weight vector is dynamically adjusted according to the importance of each dimension in the demand vector. The importance score is calculated through analysis of variance: dimensions with a variance greater than 0.1 have a weight of 1.0, dimensions with a variance between 0.05 and 0.1 have a weight of 0.7, and dimensions with a variance less than 0.05 have a weight of 0.3. The distance calculation precision is maintained to 5 decimal places, and the computational complexity is O(n), where n is the vector dimension. The preset deviation threshold is determined based on historical data statistics. The 95th percentile of the semantic distance distribution of historical normal solutions is taken as the threshold benchmark. The default value is set to 0.25, which can be adjusted within the range of 0.15 to 0.35 according to the business scenario.

[0081] Semantic deviation quantification is achieved through hierarchical calculation, comprising two levels: global deviation and local deviation. Global deviation calculates the semantic distance between the overall embedded representation and the demand vector, reflecting the overall matching degree of the solution. Local deviation is calculated segmented by functional module, dividing the 256-dimensional semantic space into 8 functional regions, each with 32 dimensions, corresponding to functional modules such as supplier selection, logistics configuration, quality assurance, cost control, timeliness management, risk control, compliance requirements, and service support. The deviation of each functional module is calculated independently, using the Euclidean distance of vector segments within the module as the metric. When the global deviation exceeds a preset threshold or any local deviation exceeds the module threshold, the corresponding supply chain solution is marked as failed. The module threshold is set to 0.8 times the global threshold to ensure timely detection of local anomalies. Failure markers are stored using a bitmask method, with each solution corresponding to an 8-bit binary code. Each bit represents the failure status of the corresponding functional module: 1 indicates failure, and 0 indicates normal operation.

[0082] The large language model employs a causal analysis framework for failure mode identification of failure scenarios. This framework is built on a transformer architecture, with 7 billion model parameters, 16-bit mixed-precision inference, and an inference batch size of 32. Model input includes multimodal features such as comparative embedding representations before and after perturbation, failure marker bitmasks, and corresponding knowledge graph path information. Failure mode identification is achieved through a multi-classification task, with 12 predefined failure mode types, including supplier capacity shortages, logistics delays, quality degradation, cost overruns, compliance risks, technical failures, demand changes, intensified competition, policy adjustments, natural disasters, staff turnover, and system failures. The model output is the probability distribution of each failure mode, with probability values ​​ranging from 0 to 1, and the sum of the probabilities of all categories equals 1. A failure mode with a probability exceeding 0.6 is identified as a primary failure cause, those with probabilities between 0.3 and 0.6 are marked as secondary failure causes, and those with probabilities below 0.3 are ignored. Model inference time is controlled within 500 milliseconds, and peak GPU memory usage does not exceed 8GB.

[0083] The mapping of embedded representations back to the domain knowledge graph path employs a reverse tracing algorithm, which maintains a bidirectional mapping relationship between embedded representations and graph paths. The mapping relationship is stored in a hash table, where the key is the hash value of the embedded representation and the value is the corresponding graph path identifier. The hash calculation uses the first 64 bits of the SHA256 algorithm to ensure a collision probability of less than one in a million. The path identifier uses hierarchical encoding, with the format: start node ID - path length - end node ID - timestamp, supporting unique path identification and fast retrieval. The reverse tracing process iterates through each node and relation edge in the path, calculating the magnitude of representation change for each element before and after perturbation. Node representation changes are measured by the cosine similarity of the node embedding vector; nodes with a similarity below 0.9 are marked as suspected invalid nodes. Relationship edge representation changes are calculated by the combined change rate of the edge weight and edge attribute vector; relation edges with a change rate exceeding 20% ​​are marked as suspected invalid relation edges. The location of invalid nodes and invalid relation edges uses a combination of threshold filtering and expert rules. Expert rules are designed based on business logic and include conditions such as node type constraints, relation transitivity constraints, and temporal consistency constraints.

[0084] Alternative path search re-executes multi-hop inference from the neighboring nodes of the failed node. The search algorithm employs the A* heuristic search, with the heuristic function combining goal orientation and path cost. Goal orientation is evaluated through semantic similarity between the target node and the current node. Similarity calculation uses a pre-trained graph neural network model, with node feature vectors as input and similarity scores between nodes as output. Path cost comprehensively considers multiple dimensions such as path length, node quality, and relationship strength, and is calculated using a weighted summation method. The weights are configured as follows: path length 0.3, node quality 0.4, and relationship strength 0.3. The search depth is limited to 6 hops to avoid performance degradation due to excessively long paths. During the search, an open list and a closed list are maintained. The open list stores candidate nodes to be expanded, while the closed list stores visited nodes to prevent repeated visits and infinite loops. Candidate nodes are sorted by heuristic function values, with priority given to expanding the node with the highest score. The search terminates when a valid alternative path is found or the open list is empty. A valid path must meet business conditions such as functional integrity, quality requirements, and time constraints.

[0085] The redundant node insertion mechanism is triggered when no suitable alternative path can be found. Insertion occurs between upstream and downstream nodes of a failed relationship edge. Redundant nodes are selected from a pre-built backup node library containing various types of high-reliability nodes. Node reliability scores are calculated based on historical performance data, ranging from 0 to 100; nodes with scores above 85 are included in the backup library. Node selection employs a multi-objective optimization algorithm, with optimization objectives including maximizing reliability, minimizing cost, and minimizing response time. The optimization algorithm uses a genetic algorithm with a population size of 100, 50 generations, a crossover probability of 0.8, and a mutation probability of 0.1. Inserted redundant nodes need to establish new relationship edges with upstream and downstream nodes. The relationship edge type is automatically inferred based on business logic, and the edge weight is determined through similarity calculation. After the redundant path is formed, integrity verification is required to ensure that the path meets the requirements of the supply chain business process.

[0086] The semantic distance metric between redundant paths and structured demand vectors uses the same calculation method as the original solution to ensure consistency in evaluation criteria. Distance calculation is performed in a 256-dimensional semantic space using a weighted cosine distance formula, with the weight vector remaining consistent with the original calculation. The criterion for successful reconstruction is that the semantic distance metric is less than a preset deviation threshold, and the overall score of the redundant path is required to be no less than 90% of the original path's score. The overall score includes three dimensions: reliability score, cost-efficiency score, and timeliness guarantee score, with weights allocated as follows: reliability 0.5, cost-efficiency 0.3, and timeliness guarantee 0.2. Performance monitoring during the reconstruction process includes indicators such as reconstruction time, resource consumption, and success rate. The reconstruction time target is controlled within 3 seconds, peak memory usage does not exceed 6GB, and the reconstruction success rate is required to reach over 85%.

[0087] In a specific implementation case, the input perturbation-embedded representation is a 512-dimensional floating-point vector with a perturbation strength of 0.15, corresponding to a 768-dimensional structured demand vector. Semantic alignment analysis maps the two vectors to a 256-dimensional space, calculating a global deviation of 0.31, exceeding the preset threshold of 0.25, thus marking the solution as failed. Local deviation analysis shows a deviation of 0.28 for the supplier selection module, 0.34 for the logistics configuration module, and 0.19 for the quality assurance module, with the logistics configuration module exhibiting the highest deviation. The large language model failure mode recognition output shows a logistics delay probability of 0.72 and a supplier capacity shortage probability of 0.18, identified as the primary and secondary causes of failure, respectively. The reverse tracing algorithm locates the failure node as the logistics service provider LOG_005, with the failure relationship edge representing the timeliness guarantee relationship between this node and the downstream distribution center. The alternative path search re-executed inference from the neighboring nodes LOG_003 and LOG_007 of LOG_005. After a search depth of 4 hops, an alternative path bypassing the failed node was found. The path length was 5 nodes, including 1 redundant transit node. The semantic distance metric of the reconstructed path was 0.22, which is less than the preset threshold, and the overall score was 94% of the original path, confirming successful reconstruction. The entire failure analysis and reconstruction process took 2.4 seconds, with peak memory usage of 4.8GB, achieving the expected performance targets.

[0088] In one optional implementation, based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the neighboring nodes of the failed node to search for alternative paths that can bypass the failed node, including:

[0089] In the domain knowledge graph, locate all incoming and outgoing nodes of the failed node, analyze the semantic association strength between the incoming nodes and the structured demand vector using a large language model, and determine the incoming node with the highest semantic association strength as the starting node for alternative path search.

[0090] The process involves expanding the traversal along the relation edges in the domain knowledge graph from the starting node, performing functional completeness verification on each candidate path encountered, and marking the verified candidate paths as potential alternative paths.

[0091] For the potential alternative path, calculate the semantic matching degree between its endpoint node and the outgoing edge node. When the semantic matching degree meets the connection condition, establish a connection edge from the endpoint node to the outgoing edge node to form a complete alternative path that bypasses the failed node.

[0092] The adjacency relationship localization of failed nodes in the domain knowledge graph employs a graph traversal algorithm. This algorithm identifies all incoming and outgoing nodes of the failed node by parsing the adjacency matrix or adjacency list data structure of the graph. The graph is stored in a sparse matrix format, with node indices as 64-bit integers, edge weights as 32-bit floating-point numbers, and edge attributes stored as key-value pairs. Incoming node identification is achieved through reverse traversal, where the algorithm scans all directed edges pointing to the failed node in the graph and extracts the source node identifiers to form the incoming node set. Outgoing node identification is completed through forward traversal, scanning all directed edges originating from the failed node and collecting the target node identifiers to form the outgoing node set. The time complexity of adjacency relationship lookup is O(degree), where the degree is the sum of the in-degree and out-degree of the node, and the space complexity is O(number of nodes). Query results are cached in an in-memory hash table with a cache capacity limited to 100,000 node relationships, using an LRU eviction policy, with a target cache hit rate of over 85%.

[0093] The large language model analyzes the semantic association strength between incoming edge nodes and structured demand vectors using a multi-layer transformer architecture. The model has 13 billion parameters, and the input layer receives concatenated features from a 768-dimensional demand vector and a 256-dimensional node embedding vector. Semantic association strength is calculated using an attention mechanism with 16 attention heads, each 64-dimensional. The multi-head attention output is processed through a linear projection layer to obtain the association strength score. The score ranges from 0 to 1, with precision maintained to four decimal places. The calculation uses 16-bit mixed-precision floating-point arithmetic. Model inference employs a batch processing mode with a batch size of 32 incoming edge nodes, and the inference time per batch is controlled within 200 milliseconds. The semantic association strength score is weighted by the functional weight, quality weight, and timeliness weight in the demand vector, with a weight allocation ratio of 0.4 for function, 0.35 for quality, and 0.25 for timeliness. When the association strength scores of multiple incoming edge nodes are similar (less than 0.05), the historical reliability of the node is used as an auxiliary criterion, with nodes having higher reliability scores being prioritized.

[0094] After the starting node is determined, the expansion traversal adopts a breadth-first search strategy with a search depth limit of 6 hops to avoid excessively long paths impacting performance. During the traversal process, an open queue and an access flag set are maintained. The open queue stores the nodes to be expanded and their path information, while the access flags prevent duplicate visits from causing loops. The initial queue capacity is set to 1000 elements, with a dynamic expansion step size of 500 elements. Queue elements include fields such as node identifier, path sequence, cumulative cost, and depth count. Relationship edge expansion is filtered by edge type, retaining only edge types consistent with supply chain business logic, including predefined types such as supply and demand relationships, quality relationships, logistics relationships, and cooperation relationships. The edge weight threshold is set to 0.3; edges below this threshold are filtered to ensure the quality of the expanded paths. During the expansion process, each newly visited node is evaluated in real time, with evaluation dimensions including node type matching degree, attribute similarity, and business relevance. Nodes with a comprehensive score below 0.6 are pruned.

[0095] The functional completeness verification of candidate paths employs a combination of rule engines and semantic analysis. The rule engine contains 150 predefined business rules, formatted as predicate logic expressions, supporting conditional judgments, logical operations, and numerical comparisons. Functional completeness checks encompass three levels: supply chain integrity, business process continuity, and data consistency. Integrity verification ensures the path includes necessary steps such as raw material procurement, manufacturing, quality inspection, and logistics, with each step corresponding to a specific node type identifier. Business process continuity checks whether the relationships between adjacent nodes in the path conform to business logic, prohibiting logically contradictory connections. Data consistency verification checks whether node attribute values ​​meet constraints, including time window constraints, capacity constraints, and cost constraints. The verification process uses parallel processing, with each candidate path allocated an independent verification thread. The thread pool size is 1.5 times the number of CPU cores, and the verification time for a single path is targeted to be controlled within 100 milliseconds.

[0096] The semantic matching degree calculation of the endpoint and outgoing edge nodes of potential alternative paths employs a deep learning model based on a dual-tower structure, encoding features for both endpoint and outgoing edge nodes separately. Node features include multi-dimensional information such as node type, attribute vectors, historical behavior, and network position, totaling 384 dimensions. The dual-tower encoder uses the same network structure, containing three transformer encoding layers, each with 512 hidden dimensions and eight attention heads. Matching degree calculation is achieved using cosine similarity, with similarity values ​​ranging from -1 to +1, mapped to the 0-1 interval as the matching degree score. A matching degree threshold is set to 0.75; node pairs above this threshold are considered to satisfy the connection condition. When multiple outgoing edge nodes satisfy the condition, the node with the highest matching degree is selected to establish a connection. If there are ties for the highest matching degree, and the difference is less than 0.02, the node with the lower degree is preferred to reduce graph complexity.

[0097] The establishment of connecting edges adopts a dynamic graph update mechanism. Newly created connecting edges inherit some attributes of existing edges. The edge weight is determined by the matching degree between the endpoint node and the outgoing edge node, and the weight value is equal to the matching degree score. The edge type is automatically inferred based on the combination of node types. The inference rules are based on the business ontology library, which contains 200 combinations of node types and their corresponding edge type mappings. The attribute fields of connecting edges include metadata such as creation timestamp, confidence level, verification status, and lifetime. The confidence level is equal to the semantic matching degree score, the initial verification status is "pending verification," and the lifetime is set to 7 days. Newly created connecting edges need to pass a consistency check, which includes three aspects: the rationality of the graph topology, the correctness of the business logic, and the performance impact assessment. The topology check prevents the formation of unreasonable loops or broken chains, the business logic check ensures that the new edge conforms to the supply chain constraints, and the performance impact assessment predicts the impact of the new edge on the graph query performance.

[0098] The integrity verification of alternative paths is achieved through a path quality assessment algorithm, which comprehensively considers four dimensions: path length, node quality, edge weight, and business rationality. The path length score uses an exponential decay function, with a baseline length of 4 hops, and the score decreases by 10% for each additional hop. The node quality score is calculated based on historical performance data and includes three sub-items: reliability, response speed, and service quality, with weights allocated as follows: reliability 0.5, response speed 0.3, and service quality 0.2. The edge weight score is the weighted average of all edge weights on the path, with the weight varying according to the edge's position within the path; edges at critical positions have doubled weights. The business rationality score is calculated using an expert rule system, with a rule base containing constraints such as supply chain best practices, industry standards, and regulatory requirements. The overall path score is obtained through a weighted summation, with each dimension having a weight of 0.2 for path length, 0.4 for node quality, 0.25 for edge weight, and 0.15 for business rationality.

[0099] In a specific implementation case, assume the failed node is identified as NODE_F001, and its type is a logistics service provider. This node has 5 incoming edge nodes and 3 outgoing edge nodes. The incoming edge nodes include manufacturers PROD_A002, PROD_B005, and PROD_C003, warehousing center WARE_D001, and distribution center DIST_E004, with corresponding semantic association strength scores of 0.82, 0.76, 0.71, 0.68, and 0.63, respectively. Based on the score ranking, manufacturer PROD_A002 is determined as the starting node, with the highest association strength. A breadth-first search is performed from the starting node, and after a search depth of 4 hops, 7 candidate paths are found. Functional completeness verification passes 5 of these paths, which are marked as potential alternative paths. The semantic matching degree is calculated between the endpoint nodes of the 5 potential alternative paths and the 3 outgoing edge nodes. The matching results show that the matching degrees between the endpoint nodes of 2 paths and the outgoing edge node DIST_F002 are 0.78 and 0.76, respectively, both exceeding the 0.75 threshold. The path with the highest matching degree was selected to establish a connection edge, with the edge weight set to 0.78 and the edge type being logistics and delivery relationship. The resulting complete alternative path contains 4 nodes, with a comprehensive path score of 0.84, meeting the quality requirements for alternative paths. The entire alternative path search process took 2.1 seconds, with a peak memory usage of 3.6GB, successfully bypassing the failed node and maintaining the integrity of the supply chain functionality.

[0100] In one optional implementation, a risk propagation map is established based on the failure modes. Risk trigger points are identified by tracing the diffusion path of risk in the risk propagation map. The temporal evolution characteristics of the risk trigger points are extracted into a risk feature vector, including:

[0101] The failure node in the failure mode is set as the initial risk source of the risk propagation graph. The upstream and downstream related nodes of the failure node are extracted from the domain knowledge graph. A directed risk propagation edge is established according to the business dependency relationship between the nodes. The direction of the directed risk propagation edge points to the downstream node affected by the failure node, and a risk propagation graph is constructed.

[0102] The risk propagation diagram simulates the risk diffusion process, propagating risk signals from the initial risk source to downstream nodes along the directed risk propagation edge, calculating the cumulative risk intensity of each node, and identifying nodes whose cumulative risk intensity exceeds their carrying capacity as secondary failure nodes.

[0103] Using a large language model, trace the node sequence and propagation edge sequence from the initial risk source to the secondary failure node, determine the complete path composed of the node sequence and the propagation edge sequence as the risk diffusion path, and locate the risk trigger point in the risk diffusion path;

[0104] Based on the temporal dimensionality reduction algorithm, the state change data of the risk trigger point within multiple historical time windows are reduced in dimensionality. The dimensionality-reduced state change data is organized into temporal evolution features in chronological order. The temporal evolution features are vectorized and encoded using the large language model to generate a risk feature vector that captures the dynamic evolution law of the risk trigger point.

[0105] The risk propagation graph is constructed by setting the failure nodes in the failure modes as the initial risk sources. The attribute information of the failure nodes includes basic data such as node identifier, node type, business role, and historical reliability score. The initial risk intensity of the risk source node is set to 1.0, representing a complete failure state. The propagation coefficient is determined according to the criticality of the node in the supply chain: 0.9 for core suppliers, 0.7 for general suppliers, 0.6 for logistics nodes, and 0.4 for service nodes. The extraction of upstream and downstream related nodes is implemented through a graph traversal algorithm. The algorithm starts from the failure node and performs a breadth-first search along the edges of the domain knowledge graph, with the search depth limited to 4 hops to avoid the risk propagation graph becoming too large. During the search, edge types are filtered, retaining business-related edges such as supply and demand dependence, quality dependence, timeliness dependence, and resource sharing, while filtering weakly related edges with a weight lower than 0.2. The strength of business dependencies is comprehensively evaluated through three dimensions: historical transaction frequency, degree of cooperation, and resource dependence, with evaluation weights of 0.4, 0.3, and 0.3, respectively. The establishment of directed risk propagation edges follows the business logic of risk transmission. The propagation direction is from the risk-affecting party to the affected party, and the weight of the propagation edge is equal to the product of the original dependency strength and the propagation coefficient. The risk propagation graph is stored using an adjacency list data structure, which supports efficient graph traversal operations. The maximum node capacity is set to 500, and the maximum number of edges is set to 2000.

[0106] The risk diffusion process simulation employs a discrete event-driven time-stepping method with a time step size of 30 minutes and a total simulation duration of 120 hours, covering the complete cycle from risk outbreak to stabilization. Within each time step, the risk signal propagates from the currently failed node along the directed risk propagation edge to downstream nodes. The propagation intensity is calculated by multiplying the risk intensity of the source node by the weight of the propagation edge. The cumulative risk intensity of downstream nodes is calculated by weighted summation of multi-source risk signals, with the weights representing the credibility score of the corresponding propagation edge, which is based on historical propagation data statistics. Node carrying capacity parameters are set according to node type and size: 2.8 for large suppliers, 2.3 for medium-sized suppliers, 1.8 for small suppliers, 2.0 for logistics centers, 1.5 for delivery nodes, and 1.2 for service providers. The cumulative risk intensity calculation includes a time decay factor, decaying by 3% each time step to simulate the natural dissipation effect of risk impact. When the cumulative risk intensity of a node first exceeds its carrying capacity, the node is marked as a secondary failure node, triggering a time recording to minute-level precision. The node state transitions to a failed state and begins propagating risk signals outwards. The simulation process employs event queue management, with event types including risk propagation events, node failure events, and state update events. Events are executed in order of timestamps to ensure the determinism and reproducibility of the simulation results.

[0107] The large language model for tracing risk propagation paths employs a path search algorithm based on graph neural networks. The model has 13 billion parameters and uses a hybrid design combining transformer architecture and graph convolutional networks. Model input includes a structured representation of the risk propagation graph, identification information of initial risk sources and secondary failure nodes, and multimodal features such as time-series risk intensity data. Graph structure encoding is implemented through four graph convolutional layers, each containing 256 hidden units. The activation function uses LeakyReLU, with a dropout rate of 0.15. Path search employs an improved A* algorithm, combining the objectives of minimizing path length and maximizing risk propagation intensity in the heuristic function, with weights allocated as follows: path length 0.4, propagation intensity 0.6. The search process maintains open and closed lists. The open list is sorted by heuristic function value, prioritizing the expansion of nodes with the highest scores. Node sequences and propagation edge sequences are recorded using a path backtracking mechanism. Each search node stores its parent node pointer and corresponding propagation edge identifier; path construction is achieved through pointer backtracking. Complete path verification includes three levels: connectivity checks, temporal consistency checks, and business logic checks, ensuring the validity and rationality of the identified paths. The time complexity of path search is O(number of nodes × number of edges × log number of nodes), and the time taken for a single search is controlled within 800 milliseconds.

[0108] The risk trigger point location employs a network criticality analysis algorithm, which comprehensively assesses the impact of each node in the risk propagation path on risk transmission. The criticality score comprises three dimensions: structural importance, propagation contribution, and temporal criticality. Structural importance is calculated using the node's degree centrality and betweenness centrality; propagation contribution measures the node's contribution to downstream risk accumulation; and temporal criticality assesses the node's critical position in the risk propagation sequence. Degree centrality calculates the sum of a node's in-degree and out-degree, while betweenness centrality calculates the proportion of shortest paths passing through that node to the total number of shortest paths. Propagation contribution is calculated as the ratio of a node's total risk output to the total risk propagation along the path, and temporal criticality is determined based on the node's positional weight within the path and the urgency of its failure time. All three dimensions are normalized to a range of 0 to 1, and the overall score is calculated using a weighted average, with weights allocated as follows: structural importance 0.35, propagation contribution 0.4, and temporal criticality 0.25. The threshold for determining risk trigger points is set to 0.65. Nodes exceeding the threshold are identified as risk trigger points. When there are multiple candidate trigger points, the node with the highest score is selected. When the scores are equal, the node that is earlier in the path is selected first.

[0109] The time-series dimensionality reduction algorithm extracts features and reduces the dimensionality of state change data at risk trigger points. The state change data includes four types of time-series indicators: risk intensity, propagation rate, impact range, and recovery indicators. The historical time window is set to 45 days, with a sliding step of 6 hours. Each window contains 180 data points, corresponding to state observations every 6 hours within the 45 days. The original state data has 720 dimensions, containing observations of the four indicators at 180 time points. The dimensionality reduction algorithm uses kernel principal component analysis (KPCA), with the radial basis function chosen as the kernel function and the bandwidth parameter set to 0.5 times the data standard deviation. The dimensionality reduction process retains principal components with a cumulative variance contribution rate of 92%, typically the top 18 to 25 principal components, with the specific number dynamically determined based on data complexity. Data preprocessing includes three steps: missing value interpolation, outlier detection, and smoothing filtering. Missing values ​​are filled using cubic spline interpolation, outliers are detected using the isolated forest algorithm and replaced with local means, and smoothing filtering uses a Kalman filter to eliminate measurement noise. The reduced feature vector has a dimension of 22, retaining more than 95% of the information content of the original data.

[0110] The temporal evolution feature organization employs a sliding time window strategy, arranging the dimensionality-reduced features of 60 consecutive time windows in chronological order to form a temporal evolution feature matrix. The feature matrix has a dimension of 60×22, with rows corresponding to time windows and columns corresponding to the dimensionality-reduced feature dimensions. Feature standardization uses the Z-score method, normalizing based on the mean and standard deviation of the training data to ensure that each dimension of the feature has the same numerical scale. Trend extraction of temporal features is achieved through a local weighted regression smoothing method, with a smoothing parameter set to 0.3, eliminating short-term random fluctuations while preserving trend information. The quality assessment of the feature matrix is ​​completed through feature correlation analysis and information entropy calculation, with a correlation threshold set to 0.85. Highly correlated features are merged, and features with an information entropy below 0.1 are removed.

[0111] The large language model employs a sequence modeling architecture for vectorized encoding of temporal evolution features. The encoder includes components such as a positional encoding layer, a multi-layer bidirectional LSTM, a multi-head self-attention mechanism, and a feedforward neural network. The positional encoding layer adds positional information to the temporal features, with an encoding dimension of 22, the same as the feature dimension. The bidirectional LSTM consists of three layers, each with a hidden state dimension of 512, a dropout rate of 0.25, and a positive activation function. The multi-head self-attention mechanism contains 12 attention heads, each with a dimension of 64, for a total dimension of 768. The feedforward neural network contains two linear layers: an intermediate layer with a dimension of 2048 and an output layer with a dimension of 256. The encoding process compresses the 60×22 temporal evolution feature matrix into a 256-dimensional risk feature vector, with vector elements ranging from -2 to +2 continuous real numbers. Model training uses a contrastive learning framework, where positive samples represent features from different time periods of the same risk trigger point, and negative samples represent features from different risk trigger points. The loss function used is InfoNCE loss. The training dataset contains 150,000 risk trigger point samples, with a training batch size of 64 and an initial learning rate of 0.0003. A cosine annealing scheduling strategy is used. Model inference supports batch processing, with a maximum batch size of 32 samples and an average inference time of 120 milliseconds.

[0112] In the specific implementation case, the failure node is the core supplier SUPP_A01, whose business role is a raw material provider, with an initial risk intensity of 1.0 and a propagation coefficient of 0.9. Domain knowledge graph extraction yielded 15 associated nodes, including 6 upstream raw material suppliers, 5 downstream manufacturers, and 4 logistics service providers, constructing a risk propagation graph with 16 nodes and 28 propagation edges. The risk propagation simulation ran for 120 hours. At hour 12, manufacturer MANU_B03 reached its carrying capacity of 2.3, triggering a secondary failure; at hour 28, logistics node LOGI_C05 reached its carrying capacity of 2.0, triggering failure; and at hour 51, service provider SERV_D02 reached its carrying capacity of 1.2, triggering failure. Risk propagation path tracing identified the critical path from SUPP_A01 to MANU_B03, including the node sequence SUPP_A01→PROC_X12→MANU_B03 and the propagation edge sequence edge_01→edge_15, with a path length of 3 hops. Risk trigger point location analysis showed that the PROC_X12 node had a comprehensive criticality score of 0.72, identifying it as a major risk trigger point. 45 days of historical data from this trigger point were processed using time-series dimensionality reduction to obtain a 22-dimensional feature vector, with 60 time windows forming a 60×22 time-series evolution feature matrix. A large language model was used to generate a 256-dimensional risk feature vector with an L2 norm of 8.34 and a feature distribution mean of 0.08 and a standard deviation of 0.41, successfully capturing the periodic fluctuations and long-term evolution trends of the risk trigger point. The entire risk propagation modeling and feature extraction process took 6.8 seconds, with peak memory usage of 9.2GB and GPU utilization reaching 82%, meeting the performance requirements for real-time risk monitoring.

[0113] In one optional implementation, the state change data of the risk trigger point within multiple historical time windows is reduced in dimensionality using a time-series dimensionality reduction algorithm. The dimensionality-reduced state change data is then organized into time-series evolution features in chronological order, including:

[0114] Obtain the state change data of the risk trigger point within multiple historical time windows, construct a high-dimensional state vector from the multi-dimensional values ​​corresponding to each historical time window, and form a high-dimensional state vector sequence indexed by historical time windows.

[0115] The high-dimensional state vector sequence is subjected to time-series correlation analysis to calculate the variance of each monitoring dimension in the time series, and to identify the variance dimensions with a change magnitude lower than a preset magnitude threshold and the redundant dimensions with linear correlation in the multiple historical time windows.

[0116] The retention weights of each monitoring dimension are obtained by nonlinear mapping of the variance based on the time series dimensionality reduction algorithm. Basic retention weights are assigned to the variance dimension and the redundant dimension, and enhanced retention weights are assigned to the dominant dimension that contributes significantly to the time series change pattern. The high-dimensional state vector sequence is then projected using the basic retention weights and the enhanced retention weights to generate a low-dimensional state vector with a reduced number of dimensions.

[0117] After associating the low-dimensional state vector with the corresponding historical time window identifier, time series mapping is performed on the low-dimensional state vector to form time series evolution features.

[0118] The acquisition of risk trigger point status change data is based on a continuous monitoring mechanism. The monitoring system adopts a distributed data acquisition architecture, with monitoring agents deployed at risk trigger points collecting status indicator data every 15 minutes. The status indicators include 12 core dimensions: risk propagation intensity, node load level, connection stability, response latency, resource utilization, failure frequency, recovery time, and impact scope. The numerical range of each dimension is standardized and mapped to an integer range of 0 to 10. The historical time window is set to a length of 24 hours, with a sliding step of 1 hour, continuously collecting 90 days of historical data to form 2160 time windows. Each time window contains 96 sampling points, and through statistical aggregation, six statistical features are calculated: mean, standard deviation, maximum value, minimum value, skewness, and kurtosis. These features are combined with the original 12 dimensions to form an 18-dimensional high-dimensional state vector. The construction process of the state vector includes a data cleaning step, using the isolated forest algorithm to detect outliers, controlling the outlier rate to within 5%, and using time series interpolation methods to correct detected outliers. The high-dimensional state vector sequence is organized by time window index and uses a B+ tree index structure to support efficient range queries, with single query latency controlled within 20 milliseconds. Data storage adopts a columnar storage format, achieving a compression ratio of 4:1 and optimizing storage space usage by 60%.

[0119] Temporal correlation analysis performed cross-time statistical analysis on high-dimensional state vector sequences across 2160 time windows. The calculation process employed a sliding window method to handle the non-stationary nature of the time series data. A weighted variance method was used to calculate the variance, with the weight of recent time windows set to 1.0. The weight decayed exponentially with time distance, with a decay coefficient of 0.995, ensuring that recent data dominated the variance calculation. The variance calculation accuracy for each monitoring dimension across the 2160 time windows was maintained to eight decimal places, with a computational complexity of O(number of dimensions multiplied by number of time windows). A preset amplitude threshold was determined using historical statistical data. The quantile distribution of the variance of all 18 dimensions was calculated, and the variance value corresponding to the 30th percentile (0.08) was selected as the threshold benchmark. The magnitude of change was assessed through variance normalization, dividing the variance of each dimension by the maximum variance of that dimension to obtain the normalized magnitude, with values ​​ranging from 0 to 1. Dimensions below the preset amplitude threshold were marked as variance dimensions; these dimensions exhibited low variability in historical time series and contributed limitedly to time series pattern recognition. Linear correlation analysis employed a sliding correlation coefficient, with a window length of 720 time points and a sliding step size of 72 time points, to calculate the dynamic correlation between each dimension pair. The correlation coefficient matrix was an 18×18 symmetric matrix, with element values ​​ranging from -1 to +1, and calculation precision maintained to six decimal places. Redundant dimensions were identified through a correlation coefficient threshold filter, with a threshold set to 0.82. Dimension pairs exceeding this threshold were considered to have linear correlation and constituted redundant dimension combinations.

[0120] The temporal dimensionality reduction algorithm employs a nonlinear mapping method based on a variational autoencoder. The encoder network consists of three fully connected layers with 64, 32, and 16 hidden neurons, respectively. The ELU activation function is used, and the dropout rate is set to 0.3 to prevent overfitting. The nonlinear mapping process projects the 18-dimensional state vector onto a 16-dimensional latent space. The mapping function is trained by minimizing a combined objective function of reconstruction loss and KL divergence loss. The retention weights are calculated using information theory methods, determining dimensional importance by calculating the contribution of each dimension to the reconstruction error. The information entropy of each monitored dimension is calculated using kernel density estimation, with the bandwidth parameter automatically selected using the Silverman rule, and the entropy value ranging from 0 to log18. The basic retention weight is set to 0.15 for variance dimensions and 0.1 for redundant dimensions, reflecting the retention priority of dimensions during dimensionality reduction. The calculation of enhanced retention weights is based on principal component analysis results, extracting the loading matrix of the first eight principal components corresponding to the original dimensions. Dimensions with an absolute value of the loading coefficient greater than 0.6 are identified as dominant dimensions. The enhancement weight of the dominant dimension is calculated by adding 0.4 times the absolute value of the loading coefficient to 0.5, ensuring that the dominant dimension receives an enhancement weight value between 0.74 and 0.9. Weight normalization ensures that the sum of the weights of all dimensions equals the total number of dimensions, maintaining the numerical stability of the weight distribution.

[0121] The projection transformation process employs a weighted principal component analysis (PPA) algorithm. The weight matrix consists of a weighted average of the basic retained weights and the enhanced retained weights, with the average weight being 0.3 for the basic weights and 0.7 for the enhanced weights. Weighting factors are introduced during the calculation of the covariance matrix, giving greater influence to dimensions with higher weights in the covariance calculation. Eigenvalue decomposition is implemented using the Jacobi method, with an iteration precision of 1e-10 and a maximum of 5000 iterations to ensure the numerical stability of the decomposition results. Principal components are selected based on the cumulative variance contribution rate criterion, retaining the top k principal components with a cumulative contribution rate of 85%, typically with k varying between 10 and 14. The dimension of the low-dimensional state vector is equal to the number of retained principal components. The projection transformation is achieved through matrix multiplication, with a computational complexity of O(dimensionality squared multiplied by the number of samples). The transformation process preserves the temporal structure of the data; the transformed low-dimensional state vector sequence remains arranged according to the original time window order. The dimensionality reduction quality is evaluated using two metrics: variance retention rate and reconstruction accuracy. The target value for variance retention rate is set to 0.85, and the reconstruction accuracy is measured by the root mean square error, with a target value of less than 0.12. The projection transformation is calculated using a batch processing mode, with a batch size of 256 time windows, and the processing time for a single batch is controlled within 150 milliseconds.

[0122] The association between low-dimensional state vectors and historical time window identifiers employs a hash mapping data structure. Time window identifiers are represented by 64-bit integer timestamps with minute-level precision. The hash function uses the FNV-1a algorithm, and the hash table size is set to 1.3 times the number of time windows. The load factor is controlled below 0.77 to ensure query performance. Low-dimensional state vectors are stored in a compact array format, with array elements being 32-bit floating-point numbers. The memory layout uses a structure array approach to improve cache hit rate. The association establishment process includes data consistency checks to ensure that each time window identifier corresponds to a unique low-dimensional state vector. Duplicate identifiers are handled using a last-overwrite strategy. Concurrency access control uses a read-write lock mechanism to support multi-threaded concurrent reads, while write operations use exclusive locks to ensure data consistency. Memory management employs memory pool technology, with pre-allocated memory blocks of 1MB to reduce the overhead of frequent memory allocations.

[0123] Time series mapping reorganizes the correlated low-dimensional state vectors in chronological order, forming a continuous temporal evolution feature data structure. The mapping process employs a fixed-length sliding window method, with a window length of 168 time windows corresponding to a 7-day time span and a sliding step size of 24 time windows corresponding to a 1-day interval. The temporal evolution feature data format is a three-dimensional tensor: the first dimension is the sliding window number, the second dimension is the time point number within the window, and the third dimension is the dimension index of the low-dimensional state vector. Tensor storage uses a row-major order format, supporting efficient continuous memory access. Smoothing of the temporal features employs a Savitzky-Golay filter with a multinomial order of 3 and a window length of 15 time points, effectively removing high-frequency noise while preserving the temporal trend features. Feature standardization uses the Z-score method, calculating the mean and standard deviation of each dimension based on global statistics. The standardized feature values ​​have a mean of 0 and a standard deviation of 1. Data integrity verification includes three steps: missing value detection, time continuity check, and numerical reasonableness verification. The proportion of missing values ​​is controlled within 2%, the time interval error is controlled within 5%, and the numerical range check ensures that all feature values ​​are within a reasonable range.

[0124] In a specific implementation case, the risk trigger point NODE_RISK_201 generated 2160 time windows of state data during a 90-day monitoring period. The original monitoring dimensions consisted of 12 dimensions plus 6 statistical features, forming an 18-dimensional high-dimensional state vector sequence. Time-series correlation analysis calculated the variance of each dimension: the variance for risk propagation intensity was 0.156, for node load level it was 0.043, for connection stability it was 0.089, and for response latency it was 0.201. The variation in node load level (0.043) was below the preset threshold of 0.08 and was therefore marked as a variance dimension. Linear correlation analysis revealed a correlation coefficient of 0.87 between risk propagation intensity and impact range, exceeding the 0.82 threshold and constituting a redundant dimension pair. The risk propagation intensity dimension, with its larger variance, was retained. The nonlinear mapping calculation of the temporal dimensionality reduction algorithm yields the retention weights for each dimension. The variance dimension's node load level receives a base weight of 0.15, the redundancy dimension's influence range receives a base weight of 0.1, and the dominant dimensions' response latency and fault frequency receive enhanced weights of 0.86 and 0.78, respectively. The projection transformation retains the top 12 principal components corresponding to 85% of the cumulative variance contribution rate, generating a 12-dimensional low-dimensional state vector sequence. A hash mapping is established between the 2160 low-dimensional state vectors and the time window identifiers. The time series mapping uses a window length of 168 to generate 81 sliding window temporal evolution features, resulting in a final feature tensor dimension of 81×168×12. The entire dimensionality reduction process takes 4.7 seconds, with a peak memory usage of 3.2GB. The data compression rate after dimensionality reduction is 67%, the variance retention rate is 0.87, and the reconstruction error is 0.094, meeting the accuracy requirements for temporal evolution feature extraction.

[0125] In one optional implementation, the method of generating a hierarchical risk control strategy for the reconstructed supply chain solution based on the risk feature vector using the large language model, and transforming the hierarchical risk control strategy into newly added constraint edges to enhance the risk resilience of the solution, includes:

[0126] The numerical and temporal components in the risk feature vector are analyzed. Based on the difference in the amplitude of the numerical components and the rate of change of the temporal components, the risk is divided into multiple risk levels. For each risk level, the corresponding response time requirements and prevention and control priorities are determined.

[0127] The risk sensitivity of each node in the reconstructed supply chain scheme is assessed by using the prevention and control priority assessment, and important nodes with failure probabilities exceeding preset probability thresholds and weak edges with insufficient risk propagation blocking capabilities are identified under different risk levels. The important nodes and the weak edges are then taken as the key targets for prevention and control.

[0128] Construct a hierarchical prevention and control strategy for the key prevention and control targets, and refine the hierarchical prevention and control strategy into enhanced constraint rules for the important nodes and backup path constraint rules for the weak edges according to the response time requirements;

[0129] The strengthening constraint rules are transformed into new constraint relationship edges connecting the important nodes and their limited collaborative objects. The backup path constraint rules are transformed into new constraint relationship edges connecting the starting point of the weak edge and the ending point of the redundant path. New constraint relationship edges are added to the knowledge graph of the reconstructed supply chain solution to enhance the risk resistance capability of the solution.

[0130] The risk feature vector is analyzed using a vector decomposition algorithm, dividing the 256-dimensional risk feature vector into two subspaces based on semantic grouping: numerical components and temporal components. The numerical components occupy the first 128 dimensions, containing a quantitative representation of static risk attributes such as risk intensity, propagation range, impact depth, and duration. Vector elements range from -2 to +2, representing continuous real numbers. The temporal components occupy the last 128 dimensions, encoding dynamic features such as risk evolution trends, rate of change, periodic patterns, and abnormal fluctuations. Vector elements also range from -2 to +2, representing floating-point values. The amplitude differences of the numerical components are assessed by calculating the absolute value distribution of vector elements in each dimension. A quantile method is used to divide the amplitude values ​​into five intervals: 0-20 percentile for low amplitude, 20-40 percentile for low-to-medium amplitude, 40-60 percentile for medium amplitude, 60-80 percentile for medium-to-high amplitude, and 80-100 percentile for high amplitude. The rate of change of the time-series components is assessed by calculating the difference between vector elements of adjacent time steps. The absolute value of the difference reflects the severity of the change, and the rate of change is also classified using a quintile method. Risk levels are determined using a two-dimensional grid method, dividing the two-dimensional space formed by the amplitude of the numerical components and the rate of change of the time-series components into 25 grid cells, each corresponding to a risk level identifier. Risk levels are ranked by severity, from L1 (low risk) to L25 (extremely high risk), with higher level numbers indicating more severe risk.

[0131] The determination of response timeliness requirements and prevention and control priorities is based on the mapping relationship between risk levels and business impact. The mapping rules are stored in a configuration matrix with a dimension of 25×4. Rows correspond to risk levels, and columns correspond to four parameters: response timeliness, prevention and control priority, resource investment, and monitoring frequency. For L1 to L5 low-risk levels, the response timeliness requirement is set to 72 hours, and the prevention and control priority is set to 5. For L6 to L10 low-to-medium risk levels, the response timeliness is 48 hours, and the prevention and control priority is 4. For L11 to L15 medium-risk levels, the response timeliness is 24 hours, and the prevention and control priority is 3. For L16 to L20 high-to-medium risk levels, the response timeliness is 12 hours, and the prevention and control priority is 2. For L21 to L25 extremely high-risk levels, the response timeliness is 6 hours, and the prevention and control priority is 1. A lower prevention and control priority value indicates a higher priority, requiring more resources and attention for prevention and control. The setting of response timeliness requirements considers the urgency of supply chain operations and the speed of risk spread, ensuring that prevention and control measures can be implemented promptly before significant losses occur. The configuration matrix supports dynamic adjustment, and parameter settings can be modified according to historical prevention and control effects and business change requirements. Parameter adjustments are managed using a version control mechanism, and configuration rollback and A / B testing verification are supported.

[0132] Risk sensitivity assessment employs Monte Carlo simulation to calculate the failure probability of each node in the restructured supply chain solution under different risk levels. The simulation process constructs a node state transition model, with node states including normal operation, performance degradation, partial failure, and complete failure. The state transition probability matrix is ​​dynamically adjusted based on historical failure data and risk levels. The Monte Carlo simulation runs 10,000 independent trials, each simulating a 72-hour operating cycle, calculating the failure probability by counting the number of failures at each node under different risk levels. Preset probability thresholds are determined based on business continuity requirements: a threshold of 0.05 for core business nodes, 0.1 for important business nodes, and 0.2 for general business nodes. Nodes with failure probabilities exceeding the corresponding thresholds are marked as important nodes, requiring focused attention and control. The assessment of risk propagation blocking capability is implemented using graph theory algorithms, calculating the importance of cut edges and bottleneck coefficients for each edge in the supply chain solution graph. Cut edge importance reflects the change in graph connectivity after an edge is removed, while the bottleneck coefficient measures the critical position of an edge in the risk propagation path. The criteria for determining insufficient blocking ability are that the edge importance is greater than 0.3 and the bottleneck coefficient is greater than 0.4. Edges that meet the conditions are marked as weak edges.

[0133] The hierarchical prevention and control strategy is constructed using a hierarchical decision tree model. The root node of the decision tree determines the risk level, the intermediate nodes determine the type of the object to be controlled, and the leaf nodes represent specific prevention and control measures. The decision tree contains 5 risk level layers and 2 object type layers, forming a total of 10 decision paths. Prevention and control strategies for critical nodes include four types of measures: redundancy backup, load balancing, health monitoring, and rapid recovery. Prevention and control strategies for weak edges include four types of measures: path diversification, transmission encryption, bandwidth expansion, and fault switching. The generation of reinforcement constraint rules uses a template matching method, with 8 predefined constraint rule templates, including capacity constraints, latency constraints, reliability constraints, and security constraints. Template parameters are automatically filled using relevant dimension values ​​from the risk feature vector, and the parameter filling rule uses a combination of linear mapping and piecewise functions. The generation of backup path constraint rules is based on a graph search algorithm, searching for alternative paths functionally equivalent to the weak edge. The evaluation indicators for alternative paths include four dimensions: path length, transmission latency, reliability score, and cost. The priority ranking of constraint rules adopts the analytic hierarchy process (AHP) to construct a judgment matrix to evaluate the relative importance of different constraint rules. The consistency ratio is controlled within 0.1 to ensure the rationality of the ranking results.

[0134] The transformation of reinforced constraint rules into newly added constraint edges is implemented using a rule parsing engine, which comprises three components: a lexical analyzer, a syntax analyzer, and a semantic analyzer. The lexical analyzer decomposes the constraint rule text into token sequences, recognizing syntactic elements such as node identifiers, operators, and numerical constants. The syntax analyzer constructs an abstract syntax tree based on predefined constraint language syntax rules, which are represented in Backus normal form and support the parsing of nested constraints and complex conditions. The semantic analyzer traverses the abstract syntax tree to extract semantic information about the constraint relationships, including structured data such as constraint type, constraint object, constraint condition, and constraint parameters. The attribute fields of the newly added constraint edges include metadata such as edge type, source node, target node, constraint strength, effective condition, and expiration time. Edge types are represented using enumerated values, supporting 12 types including capacity constraint edges, delay constraint edges, and reliability constraint edges. Constraint strength values ​​range from 0 to 1, with larger values ​​indicating stricter constraints. Effective conditions use conditional expression formats, supporting logical operations and numerical comparisons.

[0135] The transformation of backup path constraint rules employs a path instantiation method, mapping abstract path constraint rules to concrete graph edge connections. Path search utilizes the k-shortest path algorithm, searching for multiple candidate paths from the weak edge's origin to the redundant path's endpoint. The number of paths, k, is set to 3 to 5 to ensure sufficient alternatives. Path evaluation employs a multi-objective optimization method, with the objective function including four sub-objectives: minimizing path length, minimizing transmission delay, maximizing reliability, and minimizing cost. The weights are allocated as follows: path length 0.3, transmission delay 0.3, reliability 0.25, and cost 0.15. Path selection uses the Pareto optimality method, choosing the path with the highest overall score from the non-dominated solution set as the backup path. The creation process of adding new constraint edges includes a conflict detection step, checking whether the new edge logically conflicts with existing constraint edges or competes for resources. Conflict resolution employs a priority scheduling method, where higher-priority constraints override lower-priority constraints.

[0136] The addition of new constraint edges in the supply chain solution knowledge graph is implemented using graph update operations, which include four steps: node verification, edge insertion, index update, and consistency check. Node verification ensures that the source and target nodes of the constraint edge exist in the graph; non-existent nodes are automatically created and set with default attributes. The edge insertion operation adds the new constraint edge to the graph's adjacency list, with insertion positions sorted by edge weight to ensure high-weight edges are accessed first. The index update synchronously modifies the graph's auxiliary index structure, including node degree indexes, edge type indexes, and constraint strength indexes to accelerate queries. The consistency check verifies the rationality of the updated graph topology, including loop detection, connectivity verification, and constraint conflict analysis. The graph update operation employs a transaction processing mechanism to ensure the atomicity of multiple constraint edge addition operations, automatically rolling back to the pre-update state in case of failure. The enhanced risk resilience is verified through comparative analysis, comparing the risk propagation simulation results of the supply chain solution before and after the update. Evaluation indicators include risk diffusion speed, impact range, recovery time, and overall loss.

[0137] In a specific implementation case, the input 256-dimensional risk feature vector was parsed to obtain 128-dimensional numerical components and 128-dimensional temporal components. The amplitude distribution of the numerical components showed that 60% of the dimensions had medium amplitude and 25% had high amplitude. The rate of change distribution of the temporal components showed that 40% had a medium-low rate of change and 35% had a medium rate of change. The risk level was classified as L18, corresponding to a medium-high risk level, with a response time requirement of 12 hours and a prevention and control priority of 2. Risk sensitivity assessment used Monte Carlo simulation on 47 nodes of the restructured supply chain, identifying 8 important nodes whose failure probability exceeded the preset threshold, including 3 supplier nodes, 2 manufacturer nodes, and 3 logistics nodes. Risk propagation blocking capability assessment identified 12 weak edges, mainly concentrated in critical logistics paths and core supplier connection edges. The hierarchical prevention and control strategy generated 14 reinforced constraint rules and 8 backup path constraint rules. The rule transformation process created 22 new constraint relationship edges, including 9 capacity constraint edges, 6 latency constraint edges, 4 reliability constraint edges, and 3 backup path edges. The update operation took 850 milliseconds and used 320MB of memory incrementally. The updated supply chain solution risk propagation simulation showed that the average risk diffusion speed was reduced by 35%, the scope of impact was reduced by 28%, and the overall risk resistance capability was significantly improved.

[0138] A second aspect of this invention provides a medical device supply chain recommendation system based on a large language model, comprising:

[0139] The first unit is used to obtain the medical device requirement description and convert the medical device requirement description into a structured requirement vector using a large language model.

[0140] The second unit is used to construct a domain knowledge graph of the medical device supply chain. It uses a large language model to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links, and the supply chain collaboration paths are converted into embedded representations of candidate supply chain solutions through graph representation learning.

[0141] The third unit is used for the embedded representation of the candidate supply chain solutions. It generates adversarial disturbances by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applies them to the embedded representation. It uses a large language model to analyze the degree of semantic deviation between the embedded representation after the disturbance and the structured demand vector to identify failure modes and generate failure cause descriptions. Based on the failure modes, it re-searches alternative paths on the domain knowledge graph or introduces redundant nodes for robust reconstruction.

[0142] The fourth unit is used to establish a risk propagation graph based on the failure modes, identify risk trigger points by tracing the diffusion path of risks in the risk propagation graph, extract the temporal evolution characteristics of the risk trigger points as risk feature vectors, use the large language model to generate hierarchical prevention and control strategies for the reconstructed supply chain solution based on the risk feature vectors, and transform the hierarchical prevention and control strategies into newly added constraint relationship edges to enhance the risk resistance capability of the solution.

[0143] A third aspect of the present invention provides an electronic device, comprising:

[0144] processor;

[0145] Memory used to store processor-executable instructions;

[0146] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0147] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0148] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical instrument supply chain recommendation method based on a large language model assistance, characterized by, include: Obtain the medical device requirement description, and use a large language model to convert the medical device requirement description into a structured requirement vector; A domain knowledge graph of the medical device supply chain is constructed. A large language model is used to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links. The supply chain collaboration paths are transformed into embedded representations of candidate supply chain solutions through graph representation learning. For the embedded representation of the candidate supply chain solution, adversarial perturbations are generated by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applied to the embedded representation. The semantic deviation between the perturbed embedded representation and the structured demand vector is analyzed using a large language model to identify failure modes and generate failure cause descriptions. Based on the failure modes, alternative paths are re-searched on the domain knowledge graph or redundant nodes are introduced for robust reconstruction. A risk propagation graph is established based on the failure modes. Risk trigger points are identified by tracing the diffusion path of risks in the risk propagation graph. The temporal evolution characteristics of the risk trigger points are extracted as risk feature vectors. The large language model is used to generate a hierarchical prevention and control strategy for the reconstructed supply chain solution based on the risk feature vector, and the hierarchical prevention and control strategy is transformed into new constraint relationship edges to enhance the risk resistance capability of the solution.

2. The method of claim 1, wherein, Constructing a domain knowledge graph for the medical device supply chain, and using a large language model to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths, includes: The structured demand vector is used to extract the implicit functional dependencies and scenario constraints in the demand through a hierarchical semantic decomposition algorithm. Based on the functional dependencies, the instrument classification node that matches the semantics of the demand is located in the domain knowledge graph as the starting point for inference. The scenario constraints are then converted into path filtering rules for graph traversal. The large language model is used to expand and traverse outward from the inference starting point along the supply and demand matching relationship edge. During the traversal, the supplier capability nodes connected by each relationship edge are evaluated in real time according to the path filtering rules to see if they meet the scenario constraints. Supplier capability nodes that meet the conditions are dynamically aggregated into the current hop reachable node set. For each supplier capability node in the set of reachable nodes, the large language model is used to extend downstream logistics service nodes along the quality traceability relationship edge and the timeliness guarantee relationship edge. During the extension process, the timeliness requirements and quality standards in the structured demand vector are combined to perform semantic similarity calculation on the reachability of each downstream node, and the logistics service nodes that can form a complete supply chain closed loop are selected. The complete path from the inference starting point through the supplier capability node to the logistics service node is identified as the supply chain collaborative path.

3. The method of claim 1, wherein, The method involves using a large language model to analyze the semantic deviation between the perturbed embedded representation and the structured requirement vector to identify failure modes and generate failure cause descriptions. Based on these failure modes, a robust reconstruction is then performed by re-searching for alternative paths or introducing redundant nodes on the domain knowledge graph. The embedded representation of the supply chain solution after applying adversarial perturbation is semantically aligned with the structured demand vector. The degree of semantic deviation is quantified by calculating the semantic distance metric between the embedded representation after perturbation and the structured demand vector. Embedded representations with a semantic deviation exceeding a preset deviation threshold are marked as failed solutions. The failure mode of the failed scheme is identified using a large language model. The perturbed embedded representation is mapped back to the corresponding path in the domain knowledge graph. By analyzing the magnitude of the representation changes of each node and relation edge in the path before and after the perturbation, the failed nodes and failed relation edges that cause semantic deviation are located. Based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the adjacent nodes of the failed node to search for alternative paths that can bypass the failed node. When there are no alternative paths that meet the conditions, intermediate nodes with redundancy guarantee capabilities are inserted between the upstream and downstream nodes of the failed relationship edge to form a redundant path. The semantic distance metric is calculated between the redundant path and the structured requirement vector. When the semantic distance metric is less than the preset deviation threshold, the reconstruction is confirmed to be successful.

4. The method of claim 3, wherein, Based on the position of the failed node in the domain knowledge graph, multi-hop reasoning is re-executed starting from the neighboring nodes of the failed node to search for alternative paths that can bypass the failed node, including: In the domain knowledge graph, locate all incoming and outgoing nodes of the failed node, analyze the semantic association strength between the incoming nodes and the structured demand vector using a large language model, and determine the incoming node with the highest semantic association strength as the starting node for alternative path search. The process involves expanding the traversal along the relation edges in the domain knowledge graph from the starting node, performing functional completeness verification on each candidate path encountered, and marking the verified candidate paths as potential alternative paths. For the potential alternative path, calculate the semantic matching degree between its endpoint node and the outgoing edge node. When the semantic matching degree meets the connection condition, establish a connection edge from the endpoint node to the outgoing edge node to form a complete alternative path that bypasses the failed node.

5. The method according to claim 1, characterized in that, A risk propagation graph is established based on the aforementioned failure modes. By tracing the diffusion path of risk in the risk propagation graph, risk trigger points are identified. The temporal evolution characteristics of the risk trigger points are extracted into a risk feature vector, including: The failure node in the failure mode is set as the initial risk source of the risk propagation graph. The upstream and downstream related nodes of the failure node are extracted from the domain knowledge graph. A directed risk propagation edge is established according to the business dependency relationship between the nodes. The direction of the directed risk propagation edge points to the downstream node affected by the failure node, and a risk propagation graph is constructed. The risk propagation diagram simulates the risk diffusion process, propagating risk signals from the initial risk source to downstream nodes along the directed risk propagation edge, calculating the cumulative risk intensity of each node, and identifying nodes whose cumulative risk intensity exceeds their carrying capacity as secondary failure nodes. Using a large language model, trace the node sequence and propagation edge sequence from the initial risk source to the secondary failure node, determine the complete path composed of the node sequence and the propagation edge sequence as the risk diffusion path, and locate the risk trigger point in the risk diffusion path; Based on the temporal dimensionality reduction algorithm, the state change data of the risk trigger point within multiple historical time windows are reduced in dimensionality. The dimensionality-reduced state change data is organized into temporal evolution features in chronological order. The temporal evolution features are vectorized and encoded using the large language model to generate a risk feature vector that captures the dynamic evolution law of the risk trigger point.

6. The method according to claim 5, characterized in that, Based on a time-series dimensionality reduction algorithm, the state change data of the risk trigger point within multiple historical time windows are reduced in dimensionality. The dimensionality-reduced state change data is then organized into time-series evolution features in chronological order, including: Obtain the state change data of the risk trigger point within multiple historical time windows, construct a high-dimensional state vector from the multi-dimensional values ​​corresponding to each historical time window, and form a high-dimensional state vector sequence indexed by historical time windows. The high-dimensional state vector sequence is subjected to time-series correlation analysis to calculate the variance of each monitoring dimension in the time series, and to identify the variance dimensions with a change magnitude lower than a preset magnitude threshold and the redundant dimensions with linear correlation in the multiple historical time windows. The retention weights of each monitoring dimension are obtained by nonlinear mapping of the variance based on the time series dimensionality reduction algorithm. Basic retention weights are assigned to the variance dimension and the redundant dimension, and enhanced retention weights are assigned to the dominant dimension that contributes significantly to the time series change pattern. The high-dimensional state vector sequence is then projected using the basic retention weights and the enhanced retention weights to generate a low-dimensional state vector with a reduced number of dimensions. After associating the low-dimensional state vector with the corresponding historical time window identifier, time series mapping is performed on the low-dimensional state vector to form time series evolution features.

7. The method according to claim 1, characterized in that, The method of generating a hierarchical risk prevention and control strategy for the reconstructed supply chain solution using the large language model based on the risk feature vector, and transforming the hierarchical risk prevention and control strategy into newly added constraint relationship edges to enhance the risk resistance capability of the solution includes: The numerical and temporal components in the risk feature vector are analyzed. Based on the difference in the amplitude of the numerical components and the rate of change of the temporal components, the risk is divided into multiple risk levels. For each risk level, the corresponding response time requirements and prevention and control priorities are determined. The risk sensitivity of each node in the reconstructed supply chain scheme is assessed by using the prevention and control priority assessment, and important nodes with failure probabilities exceeding preset probability thresholds and weak edges with insufficient risk propagation blocking capabilities are identified under different risk levels. The important nodes and the weak edges are then taken as the key targets for prevention and control. Construct a hierarchical prevention and control strategy for the key prevention and control targets, and refine the hierarchical prevention and control strategy into enhanced constraint rules for the important nodes and backup path constraint rules for the weak edges according to the response time requirements; The strengthening constraint rules are transformed into new constraint relationship edges connecting the important nodes and their limited collaborative objects. The backup path constraint rules are transformed into new constraint relationship edges connecting the starting point of the weak edge and the ending point of the redundant path. New constraint relationship edges are added to the knowledge graph of the reconstructed supply chain solution to enhance the risk resistance capability of the solution.

8. A medical device supply chain recommendation system based on a large language model, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to obtain the medical device requirement description and convert the medical device requirement description into a structured requirement vector using a large language model. The second unit is used to construct a domain knowledge graph of the medical device supply chain. It uses a large language model to perform multi-hop reasoning on the domain knowledge graph based on the structured demand vector to identify implicit supply chain collaboration paths. Each collaboration path is assigned a semantic weight that reflects the causal strength of the links, and the supply chain collaboration paths are converted into embedded representations of candidate supply chain solutions through graph representation learning. The third unit is used for the embedded representation of the candidate supply chain solutions. It generates adversarial disturbances by simulating supply interruption scenarios, quality anomaly scenarios, and demand mutation scenarios and applies them to the embedded representation. It uses a large language model to analyze the degree of semantic deviation between the embedded representation after the disturbance and the structured demand vector to identify failure modes and generate failure cause descriptions. Based on the failure modes, it re-searches alternative paths on the domain knowledge graph or introduces redundant nodes for robust reconstruction. The fourth unit is used to establish a risk propagation map based on the failure mode, identify risk trigger points by tracing the diffusion path of risk in the risk propagation map, and extract the temporal evolution characteristics of the risk trigger points as risk feature vectors. The large language model is used to generate a hierarchical prevention and control strategy for the reconstructed supply chain solution based on the risk feature vector, and the hierarchical prevention and control strategy is transformed into new constraint relationship edges to enhance the risk resistance capability of the solution.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images