Intelligent association and risk early warning system for supply chain data based on knowledge graph

Abstract

This invention relates to the field of supply chain risk early warning technology, and discloses a supply chain data intelligent association and risk early warning system based on knowledge graphs. The system includes: a supply chain knowledge graph construction module; a multi-dimensional stress state quantification module, used to map risk disturbance events into a multi-dimensional stress state vector containing inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components; a stress propagation path parsing module, used to propagate the stress vector hop-by-hop along directed edges to obtain a cumulative vector; a risk transmission backbone generation module, used to generate a backbone link by tracing the maximum value of the multi-dimensional stress gradient; and an early warning evidence chain output module, used to extract the excessive stress components to form a directed link early warning evidence chain labeled with node identifiers, component names, values, and risk levels. This system models the risk transmission process as multi-dimensional stress field propagation, generating an interpretable structured early warning evidence chain.

Description

Technical Field

[0001] This invention relates to the field of supply chain risk early warning technology, and in particular to a supply chain data intelligent association and risk early warning system based on knowledge graphs. Background Technology

[0002] As global supply chain networks become increasingly complex, the risks and disturbances they face are becoming more diversified, including supplier shutdowns, logistical disruptions, geopolitical conflicts, natural disasters, and quality recalls. Building supply chain relationship networks based on knowledge graphs and conducting intelligent analysis and early warning of risk events has become an important technological direction in the field of supply chain management. Existing technologies include solutions for supply chain risk prediction and decision optimization using knowledge graphs. These solutions calculate node risk scores by constructing supply chain knowledge graphs and combining them with graph neural networks and other models, identifying risk propagation paths and generating response strategies.

[0003] However, existing technologies generally simplify supply chain risk early warning to a problem of "risk value quantification and propagation probability calculation." When a risk disturbance event occurs, existing solutions can only output node risk scores or propagation probabilities, failing to reveal how the multi-dimensional impact of the risk event is structurally transmitted along the material flow and supply dependencies of the supply chain network. They also cannot generate interpretable early warning evidence to explain to users "which dimension of risk is transmitted from which nodes, through which key links, to the affected nodes." This fundamental deficiency leads to a lack of interpretability in early warning results, making it difficult for supply chain managers to accurately locate critical links and weak points in risk propagation, and hindering the efficient development of targeted risk mitigation measures.

[0004] Therefore, this invention proposes a knowledge graph-based intelligent association and risk warning system for supply chain data. Summary of the Invention

[0005] This invention provides a knowledge graph-based intelligent data association and risk early warning system for the supply chain. By modeling the supply chain risk transmission process as a multi-dimensional stress field propagation, it generates a structured and interpretable early warning evidence chain classified and labeled according to the stress component type. This enables supply chain managers to trace the source and key links of risk transmission from multiple dimensions such as inventory, delivery, finance, and alternative feasibility.

[0006] This invention provides a knowledge graph-based intelligent data association and risk early warning system for supply chains, comprising: The supply chain knowledge graph construction module is used to extract supply chain entities and relationships between entities from multi-source heterogeneous data, and construct a supply chain knowledge graph in which nodes represent suppliers, manufacturers, logistics hubs, and distribution nodes, and directed edges represent the direction of material flow and supply dependence between entities. The multidimensional stress state quantification module is used to map real-time perceived multi-type risk disturbance events to the affected nodes of the supply chain knowledge graph. It generates a multidimensional stress state vector for the affected nodes, which includes inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components. The multidimensional stress state vector is then superimposed on the corresponding node attributes of the supply chain knowledge graph. The stress propagation path parsing module is used to propagate the multidimensional stress state vector of the affected node to the adjacent nodes hop by hop along the directed edges of the supply chain knowledge graph according to the stress conduction attenuation rule, so as to obtain the cumulative multidimensional stress state vector of each node. The risk transmission backbone generation module is used to calculate the multidimensional stress gradient between each adjacent node and the current node's cumulative multidimensional stress state vector by hop along the directed edge direction of the supply chain knowledge graph, starting from the source point. The adjacent node corresponding to the maximum value of the multidimensional stress gradient is selected as the next hop node to generate the risk transmission backbone. The early warning evidence chain output module is used to extract stress components that exceed the early warning component threshold from the cumulative multidimensional stress state vector of each node on the risk transmission backbone link as a risk transmission early warning evidence chain. The risk transmission early warning evidence chain is output in the form of a directed link labeled with node identifier, stress component name, stress component value and risk level.

[0007] Furthermore, the multidimensional stress state quantization module is also used to perform correlation enhancement on multimodal weak signals. The correlation enhancement process includes: Continuously monitor low-confidence signals in multimodal data sources, including ambiguous sentiment statements in social media texts, delayed delivery status updates in supplier portal data streams, and irregular pauses between transportation nodes in logistics trajectory data streams. The low-confidence signals detected are grouped according to time windows and related entities to form multiple sets of weak signals to be confirmed. The entity identifiers of the related entities are shared with the node identifiers in the supply chain knowledge graph. For each set of weak signals to be confirmed, the source modality, semantic vector, and timestamp features of the signal are extracted to construct a weak signal association graph. The weak signal association graph uses each low-confidence signal as a node and the modal complementarity, semantic similarity, and temporal proximity between signals as edges. The graph diffusion algorithm is run on the weak signal association graph to calculate the association enhancement confidence score of each node. When the association enhancement confidence score exceeds the weak signal confirmation threshold, the corresponding node is upgraded from a low confidence signal to a confirmed early risk signal. The confirmed early risk signals are input into the stress component weight allocation process to generate the pre-multidimensional stress state vector of the corresponding affected node. The pre-multidimensional stress state vector is superimposed on the corresponding node before the core risk disturbance event triggers formal stress quantification.

[0008] Furthermore, in the stress propagation path analysis module, the stress propagation is propagated hop-by-hop to adjacent nodes according to the stress conduction attenuation rule. The stress conduction attenuation rule is set as follows: The initial transmission attenuation coefficient is obtained by weighting and summing the supply share ratio, the normalized inventory buffer days index, and the normalized alternative supplier availability index according to preset weights. The cooperative fluctuation pattern of the multidimensional stress state vector of the two nodes of the directed edge is monitored within a preset observation window. The cooperative fluctuation pattern is characterized by the Pearson correlation coefficient of the inventory stress component of the two nodes, the dynamic time warping distance of the delivery stress component, and the Granger causality of the financial stress component. The cooperative wave pattern features are input into a pre-trained binary classifier, which outputs the probability score of directed edges belonging to cascaded coupling links. Subtracting the probability score of the directed edge belonging to the cascaded coupling link from 1 gives the cascade suppression factor. Multiplying the cascade suppression factor by the initial conduction attenuation coefficient gives the corrected conduction attenuation coefficient.

[0009] Furthermore, the positive samples used in the training of the pre-trained binary classifier are the set of directed edges in historical risk events that have been confirmed to have cascade propagation, and the negative samples are the set of directed edges in historical risk events that have not had cascade propagation. The training features include the Pearson correlation coefficient of the historical multidimensional stress state vectors of the two endpoints, the dynamic time warping distance, the Granger causality, and the supply share of directed edges.

[0010] Furthermore, the risk transmission backbone generation module is also used to perform path confidence verification on the generated risk transmission backbone. The verification process includes: The path confidence assessment model inputs the corrected transmission attenuation coefficient of each directed edge on the risk transmission backbone, the total number of hops in the link, and the cumulative multidimensional stress state vector magnitude of the end node of the link. The path confidence assessment model outputs the overall confidence score of the backbone link. When the overall confidence score is lower than the path confidence threshold, abandon the current backbone link and backtrack to the suboptimal multidimensional stress gradient direction to re-trace until a backbone link that meets the path confidence threshold is generated, or the number of backtracking reaches the preset maximum number of backtracking, or there are no more unvisited suboptimal directions. If no backbone link that meets the path confidence threshold is generated when the backtracking terminates, the backbone link with the highest path confidence score during the backtracking process will be used as the output. The risk transmission backbone generation module is also used to: when the multidimensional stress gradients of multiple adjacent nodes of the source point all exceed the preset gradient threshold, simultaneously track these adjacent nodes to generate multiple parallel risk transmission branch links, and each branch link performs path confidence verification.

[0011] Furthermore, the early warning evidence chain output module extracts stress components exceeding the early warning component threshold from the accumulated multidimensional stress state vector as the risk transmission early warning evidence chain. It also includes problem attribution chain generation, the process of which includes: For each marked node in the risk transmission early warning evidence chain, the upstream propagation path that caused each excessive stress component to exceed the early warning component threshold is traced. The upstream propagation path consists of directed edges on the main risk transmission link from the source node of the risk disturbance event to the corresponding marked node. Extract the key propagation edge with the largest stress contribution in the upstream propagation path; Each labeled node and its corresponding critical transmission edge are connected in the structural order of the risk transmission backbone to generate a problem attribution chain. The problem attribution chain is presented in the form of alternating node identifiers and critical transmission edge identifiers, indicating the source of the maximum stress for each affected node.

[0012] Furthermore, the stress contribution ratio of each key transmission edge in the problem attribution chain is calculated using the Shapley value attribution method. On the set of directed edges from the source node of the risk disturbance event to the labeled node, each directed edge in the set is regarded as a participant, and the corresponding excess stress component value of the labeled node is taken as the total contribution target. The marginal contribution average of each directed edge to the corresponding excess stress component value is calculated as the stress contribution ratio.

[0013] Furthermore, it also includes a risk transmission comparison and analysis module across the supply chain network, used for: Obtain supply chain knowledge graphs for at least two independent supply chain networks; Using a pre-set set of standard test risk events as input, the stress propagation path analysis module and the risk transmission backbone generation module are run on each supply chain network to obtain their respective risk transmission backbones. The risk transmission backbone links of different supply chain networks are standardized and aligned under a unified coordinate system. The similarity of the link topology and the difference of the stress attenuation mode are calculated. A preset number of nodes with the largest cumulative multidimensional stress state vector magnitude in each supply chain network are extracted as the set of weak nodes. Output a comparative analysis report across supply chain networks. The report should include the similarity values ​​of the link topology of different supply chain networks, the differences in stress attenuation patterns, and the node identifiers and types included in the weak node sets of each supply chain network.

[0014] Furthermore, it also includes a stress component radar chart overlay module, which is used to present the numerical proportions of inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components of each node in real time on the topological display base map of the supply chain knowledge graph in the form of a radar chart. Each axis scale of the radar chart corresponds to the warning component threshold of each stress component, and the axis segment exceeding the standard is highlighted with a warning color.

[0015] Furthermore, it also includes a warning evidence chain narrative generation module, which is used to convert the risk transmission warning evidence chain and problem attribution chain into natural language warning narrative text. The natural language warning narrative text starts from the source node of the risk disturbance event and describes the changes in multidimensional stress state and the main causes of the transmitted stress node by node along the main risk transmission link. The early warning evidence chain narrative generation module is also used to receive input natural language query text, perform semantic parsing on the natural language query text, retrieve matching content in the natural language early warning narrative text based on the parsing results, and output the corresponding narrative text fragments.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: Existing technologies simplify supply chain risk early warning to a problem of "risk score calculation and propagation probability deduction," failing to reveal how the multi-dimensional impacts of risk events, such as inventory shocks, delivery delays, financial losses, and deterioration of alternative feasibility, are structurally transmitted along material flow and supply dependence relationships. This results in a lack of process transparency and dimensional interpretability in the early warning results. This invention reconstructs the supply chain risk transmission process as a problem of "multi-dimensional stress field propagation modeling." By quantifying risk events into a multi-dimensional stress state vector containing inventory stress components, delivery stress components, financial stress components, and alternative feasibility stress components, it defines a three-parameter stress transmission attenuation rule based on supply share, inventory buffer, and alternative suppliers. It then traces the risk transmission backbone link hop-by-hop according to the maximum value of the multi-dimensional stress gradient, finally extracting the excessive stress components and labeling them with node identifiers, component names, values, and risk levels. This forms an interpretable early warning evidence chain classified by component type, enabling supply chain managers to trace the source and path of risk transmission according to inventory, delivery, financial, and alternative feasibility dimensions, achieving precise dimensional positioning and interpretable early warning.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the core module and data flow of the knowledge graph-based supply chain data intelligent association and risk warning system in this embodiment of the invention; Figure 2 This is a flowchart of the multimodal weak signal correlation enhancement process in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] refer to Figure 1 and Figure 2 This invention provides an embodiment of a knowledge graph-based intelligent correlation and risk warning system for supply chain data, comprising: The supply chain knowledge graph construction module is used to extract supply chain entities and relationships between entities from multi-source heterogeneous data, and construct a supply chain knowledge graph in which nodes represent suppliers, manufacturers, logistics hubs, and distribution nodes, and directed edges represent the direction of material flow and supply dependence between entities. The multidimensional stress state quantification module is used to map real-time perceived multi-type risk disturbance events to the affected nodes of the supply chain knowledge graph. It generates a multidimensional stress state vector for the affected nodes, which includes inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components. The multidimensional stress state vector is then superimposed on the corresponding node attributes of the supply chain knowledge graph. The stress propagation path parsing module is used to propagate the multidimensional stress state vector of the affected node to the adjacent nodes hop by hop along the directed edges of the supply chain knowledge graph according to the stress conduction attenuation rule, so as to obtain the cumulative multidimensional stress state vector of each node. The risk transmission backbone generation module is used to calculate the multidimensional stress gradient between each adjacent node and the current node's cumulative multidimensional stress state vector by hop along the directed edge direction of the supply chain knowledge graph, starting from the source point. The adjacent node corresponding to the maximum value of the multidimensional stress gradient is selected as the next hop node to generate the risk transmission backbone. The early warning evidence chain output module is used to extract stress components that exceed the early warning component threshold from the cumulative multidimensional stress state vector of each node on the risk transmission backbone link as a risk transmission early warning evidence chain. The risk transmission early warning evidence chain is output in the form of a directed link labeled with node identifier, stress component name, stress component value and risk level.

[0022] In this embodiment, multi-source heterogeneous data refers to structured data such as inventory ledgers and purchase orders from enterprise resource planning systems, semi-structured logistic data from warehouse management systems and transportation management systems, supply plan data from supplier portals, and unstructured text data such as news and public opinion texts and meteorological disaster announcements from publicly available internet channels.

[0023] In this embodiment, supply chain entities and relationships between entities are extracted from multi-source heterogeneous data to construct a supply chain knowledge graph. The process is as follows: for structured data, entities and relationships are extracted through field mapping; for unstructured text data, entity names and supply and transportation relationships between entities are identified through named entity recognition and relationship extraction technologies; names pointing to the same entity from different data sources are aligned and semantically disambiguated and uniformly mapped to standard entity identifiers; entities are stored as nodes and labeled with entity type labels; and relationships are stored as directed edges and labeled with relationship type labels and weight attributes.

[0024] In this embodiment, a supply chain entity refers to an organization or facility node that performs an independent function in the supply chain network.

[0025] In this embodiment, the inter-entity relationship refers to the connection between supply chain entities formed through business transactions such as supply, transportation, and procurement cooperation.

[0026] In this embodiment, a supplier refers to an enterprise node that provides raw materials or components downstream of the supply chain, a manufacturer refers to a factory node that processes and assembles raw materials and components, a logistics hub refers to a node that provides warehousing and transit transportation services, and a distribution node refers to a node that sells products to end customers.

[0027] In this embodiment, the material flow direction refers to the physical flow direction of raw materials, components, or finished products from upstream to downstream among various entities in the supply chain.

[0028] In this embodiment, supply dependency refers to a one-way relationship in which the production activities of downstream nodes depend on the timely and sufficient supply of materials from upstream nodes.

[0029] In this embodiment, the real-time perceived multi-type risk disturbance events refer to events that may affect the normal operation of the supply chain, which are identified in real time by continuously monitoring multimodal data sources and using event detection technology. Event types include supplier production stoppage events, logistics channel disruption events, geopolitical events, natural disaster events, quality recall events, and financial credit events.

[0030] In this embodiment, the affected node refers to a node in the supply chain knowledge graph that directly corresponds to the location of the detected risk disturbance event or the entity involved through attribute matching.

[0031] In this embodiment, real-time perceived multi-type risk disturbance events are mapped to affected nodes in the supply chain knowledge graph, and multi-dimensional stress state vectors are generated for the affected nodes. The execution process is as follows: a preset stress component weight allocation template is matched according to the event type, and different event types correspond to different four-component weight ratios; the event impact magnitude is normalized to obtain the stress amplitude coefficient, and the impact magnitude is determined by the weighted sum of the geographical range level, the expected duration level, and the severity level; each preset weight is multiplied by the stress amplitude coefficient to obtain the values ​​of the inventory stress component, delivery stress component, financial stress component, and alternative feasible stress component.

[0032] In this embodiment, the inventory stress component characterizes the impact of a risk disturbance event on the current inventory level of the affected node, and is determined by the deviation of the current inventory level from the safety stock threshold; the delivery stress component characterizes the impact of the event on the node's on-time delivery capability, and is determined by the ratio of the number of delayed delivery orders to the total number of orders; the financial stress component characterizes the pressure of the event on the node's cash flow, and is determined by the combined ratio of overdue accounts payable and delayed accounts receivable collection; the alternative feasibility stress component characterizes the difficulty for the node to find alternative supply sources or alternative logistics channels, and is determined by the inverse of the number of available alternative suppliers and the detour cost of alternative channels.

[0033] In this embodiment, the inventory stress component can also be called the inventory shock index, the delivery stress component can also be called the delivery delay index, the financial stress component can also be called the capital pressure index, and the alternative feasibility stress component can also be called the alternative elasticity index. The above alternative names can be used interchangeably in the following description.

[0034] In this embodiment, superimposing the multidimensional stress state vector onto the corresponding node attributes of the supply chain knowledge graph means adding four stress component attribute fields to the attribute list of the affected node to store the values ​​of each stress component. Subsequent stress propagation calculations directly read the values ​​from the stress component attribute fields of each node.

[0035] In this embodiment, the stress transmission attenuation rule defines the attenuation calculation method when the multidimensional stress state vector propagates along the directed edge. The transmission attenuation coefficient used for attenuation calculation is obtained by weighting the supply share ratio of the directed edge, the normalized value of the upstream node's inventory buffer days, and the normalized value of the downstream node's available alternative suppliers according to preset weights.

[0036] In this embodiment, the multidimensional stress state vector of the source node is propagated hop-by-hop along the directed edge according to the stress propagation attenuation rule to obtain the cumulative multidimensional stress state vector of each node, with the affected node as the source node. The execution method is as follows: the initial multidimensional stress state vector of the source node is assigned to the cumulative multidimensional stress state vector of the source node; all outgoing edge adjacent nodes of the current propagation node are traversed, and the cumulative multidimensional stress state vector of the current node is multiplied by the propagation attenuation coefficient of the connecting edge to obtain the path stress component and accumulated to the cumulative multidimensional stress state vector of the adjacent node; after the propagation of the current node is completed, the adjacent nodes are added to the next round of propagation queue and traversed hop-by-hop until no new node is visited or the path stress component is lower than the preset propagation termination threshold.

[0037] In this embodiment, starting from the source node, the multidimensional stress gradient between each adjacent node and the current node is calculated hop-by-hop along the directed edge direction of the supply chain knowledge graph. The calculation method is as follows: the four gradient components of the inventory stress component difference, delivery stress component difference, financial stress component difference, and alternative feasibility stress component difference between the adjacent node and the current node are combined into a multidimensional stress gradient vector. The magnitude of this vector is the multidimensional stress gradient, which represents the degree of change of the multidimensional stress state in the direction from the current node to the adjacent node.

[0038] In another embodiment, the multidimensional stress gradient is calculated as follows: the component differences between adjacent nodes and the current node are calculated independently for the inventory stress component, delivery stress component, financial stress component, and alternative feasible stress component, generating a four-dimensional stress gradient direction vector. The component differences are compared with preset component gradient thresholds. When the difference in any dimension exceeds the gradient threshold corresponding to that dimension, that dimension is selected as the main tracking direction. After tracking to the next node along this direction, the gradient of each component is calculated independently in the same way and the tracking direction is determined, generating a risk transmission backbone link with independent tracking paths in each component dimension.

[0039] As another optional gradient calculation method, the differences in inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components can be multiplied by preset gradient weights and then summed to obtain a scalar value as a multidimensional stress gradient. Each gradient weight can be configured according to the sensitivity of the supply chain network to risks in different dimensions.

[0040] In this embodiment, the adjacent node corresponding to the maximum value of the multidimensional stress gradient is selected as the next hop node to generate the risk transmission backbone link. The execution method is as follows: taking the source node of the risk disturbance event as the tracking starting point, the magnitude of the multidimensional stress gradient pointing from the tracking starting point to each outgoing edge adjacent node is calculated. The adjacent node with the largest magnitude is selected as the next node of the tracking path and the connecting edge is recorded. The next node is updated as the tracking starting point, and the gradient calculation and maximum value selection steps are repeated until the current node has no outgoing edge adjacent nodes or the magnitude of the cumulative multidimensional stress state vector of all outgoing edge adjacent nodes is lower than the preset propagation termination threshold or the number of tracking hops reaches the preset maximum number of hops. The nodes along the way are connected in sequence to generate the risk transmission backbone link.

[0041] In this embodiment, the warning component threshold is an independent numerical threshold preset for the inventory stress component, delivery stress component, financial stress component, and alternative feasibility stress component. Each threshold is set according to the acceptable risk boundaries of the supply chain industry for inventory security, on-time delivery, financial health, and supply substitutability.

[0042] In this embodiment, stress components exceeding the warning component threshold in the cumulative multidimensional stress state vector of each node on the risk transmission backbone are extracted as the risk transmission warning evidence chain. Specifically, each node on the backbone is traversed, and the inventory stress component, delivery stress component, financial stress component, and alternative feasible stress component are compared with their respective warning component thresholds one by one. The excess stress components whose values ​​exceed the thresholds are selected, and the node identifier, component name, and component value of the excess stress component are recorded.

[0043] In this embodiment, the risk transmission early warning evidence chain is output in the form of a directed link with labeled node identifier, stress component name, stress component value and risk level. The risk level is divided into three levels: high risk, medium risk and low risk according to the exceedance range. The exceedance range is the proportion of stress component value exceeding the early warning component threshold.

[0044] Furthermore, the multidimensional stress state quantization module is also used to perform correlation enhancement on multimodal weak signals. The correlation enhancement process includes: Continuously monitor low-confidence signals in multimodal data sources, including ambiguous sentiment statements in social media texts, delayed delivery status updates in supplier portal data streams, and irregular pauses between transportation nodes in logistics trajectory data streams. The low-confidence signals detected are grouped according to time windows and related entities to form multiple sets of weak signals to be confirmed. The entity identifiers of the related entities are shared with the node identifiers in the supply chain knowledge graph. For each set of weak signals to be confirmed, the source modality, semantic vector, and timestamp features of the signal are extracted to construct a weak signal association graph. The weak signal association graph uses each low-confidence signal as a node and the modal complementarity, semantic similarity, and temporal proximity between signals as edges. The graph diffusion algorithm is run on the weak signal association graph to calculate the association enhancement confidence score of each node. When the association enhancement confidence score exceeds the weak signal confirmation threshold, the corresponding node is upgraded from a low confidence signal to a confirmed early risk signal. The confirmed early risk signals are input into the stress component weight allocation process to generate the pre-multidimensional stress state vector of the corresponding affected node. The pre-multidimensional stress state vector is superimposed on the corresponding node before the core risk disturbance event triggers formal stress quantification.

[0045] In this embodiment, a multimodal weak signal refers to a low-confidence signal captured from different modal data sources, where a single signal is insufficient to trigger the confirmation of a risk event, but the association of multiple signals may indicate an early risk.

[0046] In this embodiment, ambiguous sentiment statements in social media text refer to text content where the probability of positive or negative sentiment output by the text sentiment analysis model is close and cannot be clearly determined as positive or negative; delayed delivery status updates in the supplier portal data stream refer to the estimated delivery time field on the portal platform not being refreshed even after exceeding the preset update time window; and irregular stops between transportation nodes in the logistics trajectory data stream refer to the duration of a transport vehicle or goods staying between two transportation nodes exceeding a preset multiple of the historical average duration of that route.

[0047] In this embodiment, the low-confidence signals detected are grouped according to time windows and associated entities to form multiple sets of weak signals to be confirmed. The execution method is as follows: using a preset duration as a sliding window, low-confidence signals that appear in the same time window and are associated with the same entity or have business relationships are grouped into the same group, and each group is a set of weak signals to be confirmed.

[0048] In this embodiment, the entity identifier of the associated entity is shared with the node identifier in the supply chain knowledge graph. This means that the entity identifier field used when constructing the weak signal association graph adopts the same encoding system as the identifier field of the corresponding node in the supply chain knowledge graph, ensuring that the entity in the weak signal association graph can be directly mapped to the corresponding node in the knowledge graph.

[0049] In this embodiment, for each set of weak signals to be confirmed, the signal source modality, signal semantic vector, and signal timestamp features are extracted and a weak signal association graph is constructed. The execution method is as follows: taking each low-confidence signal in the set as a node, the data source channel of the signal is extracted as the signal source modality, the corresponding text content of the signal is encoded into a fixed-dimensional signal semantic vector through a pre-trained language model, the acquisition time of the signal is recorded as the signal timestamp feature, and the above features are stored as the attributes of the node.

[0050] In this embodiment, the modal complementarity relationship between signals refers to the degree to which two signals from different data sources corroborate or complement each other in risk description. The value is obtained by querying a preset modal complementarity relationship matrix. Semantic similarity refers to the cosine distance between the corresponding semantic vectors of two signals. Temporal proximity refers to the normalized value of the difference between the acquisition timestamps of two signals. The smaller the difference, the higher the proximity.

[0051] In this embodiment, using the modal complementarity, semantic similarity, and temporal proximity as edges means adding an undirected edge between any two nodes in the weak signal association graph. The weight of the edge is obtained by multiplying the modal complementarity value, semantic similarity value, and temporal proximity value.

[0052] In this embodiment, a graph diffusion algorithm is run on the weak signal association graph to calculate the association enhancement confidence score of each node. A random walk algorithm with restart is adopted, with the node with the highest modal confidence weight as the initial seed node. During the propagation process, the edge weight determines the walk transition probability, and the restart probability is set to a preset fixed value. The algorithm continues to iterate until the node score distribution converges to a stable state, and the final score of each node is the association enhancement confidence score.

[0053] In this embodiment, the weak signal confirmation threshold is a pre-set numerical standard. When the association enhancement confidence score of a node exceeds the threshold, the corresponding low confidence signal is upgraded to a confirmed early risk signal.

[0054] In this embodiment, the stress component weight allocation process refers to the process of generating a pre-existing multidimensional stress state vector by matching the pre-existing early risk signals with the pre-existing stress component weight allocation template according to the pre-set event types. The execution method is as follows: according to the event types corresponding to the pre-existing early risk signals, the pre-existing stress component weight allocation template is matched to obtain the ratio of inventory stress weight, delivery stress weight, financial stress weight, and alternative feasible stress weight; the influence magnitude of the pre-existing early risk signals is normalized to obtain the stress amplitude coefficient; and each weight is multiplied by the stress amplitude coefficient to obtain the component values ​​of the pre-existing multidimensional stress state vector.

[0055] In this embodiment, the confirmed early risk signal is input into the stress component weight allocation process to generate the corresponding pre-multidimensional stress state vector of the affected node. The execution method is as follows: determine the supply chain entity associated with the confirmed early risk signal, mark the corresponding node of the entity in the supply chain knowledge graph as the pre-vector receiving node, and write the generated stress component values ​​into the pre-multidimensional stress state vector.

[0056] In this embodiment, core risk disturbance events refer to risk disturbance events with a large impact or clear nature that are directly confirmed with high confidence through conventional risk event detection channels, and are distinguished from early risk signals that are indirectly confirmed through weak signal correlation enhancement.

[0057] In this embodiment, triggering formal stress quantization refers to the process of generating a formal multidimensional stress state vector and superimposing it onto the affected node after the core risk disturbance event is detected and confirmed, following the normal processing flow of the multidimensional stress state quantization module in the first embodiment.

[0058] In this embodiment, the pre-existing multidimensional stress state vector is superimposed onto the corresponding node before the core risk disturbance event triggers formal stress quantization. This means that in the early stage before the core risk disturbance event is confirmed by the conventional detection channel, the stress component values ​​of the pre-existing multidimensional stress state vector are written into the corresponding attribute fields of the affected node. When the core risk disturbance event subsequently triggers formal stress quantization, the pre-existing stress component values ​​that have been superimposed on the node are cumulatively updated. That is, the stress component values ​​generated by formal stress quantization are arithmetically added to the pre-existing stress component values ​​according to the corresponding components.

[0059] Furthermore, in the stress propagation path analysis module, the stress propagation is propagated hop-by-hop to adjacent nodes according to the stress conduction attenuation rule. The stress conduction attenuation rule is set as follows: The initial transmission attenuation coefficient is obtained by weighting and summing the supply share ratio, the normalized inventory buffer days index, and the normalized alternative supplier availability index according to preset weights. The cooperative fluctuation pattern of the multidimensional stress state vector of the two nodes of the directed edge is monitored within a preset observation window. The cooperative fluctuation pattern is characterized by the Pearson correlation coefficient of the inventory stress component of the two nodes, the dynamic time warping distance of the delivery stress component, and the Granger causality of the financial stress component. The cooperative wave pattern features are input into a pre-trained binary classifier, which outputs the probability score of directed edges belonging to cascaded coupling links. Subtracting the probability score of the directed edge belonging to the cascaded coupling link from 1 gives the cascade suppression factor. Multiplying the cascade suppression factor by the initial conduction attenuation coefficient gives the corrected conduction attenuation coefficient.

[0060] In this embodiment, the supply share ratio refers to the proportion of the quantity of materials purchased by the downstream node from the current upstream node in the total quantity of that material purchased by the downstream node, and the value is a percentage between zero and one.

[0061] In this embodiment, the normalized inventory buffer days index refers to the value obtained by dividing the number of days that the current inventory of the upstream node can sustain production by the preset maximum buffer days and mapping it to the range of zero to one.

[0062] In this embodiment, the normalized alternative supplier availability index refers to the value obtained by dividing the actual number of alternative suppliers that downstream nodes can contact by a preset sufficient number of alternative suppliers and mapping it to the range of zero to one.

[0063] In this embodiment, the preset weight refers to the weight set for the supply share ratio, the normalized inventory buffer days indicator, and the normalized alternative supplier availability indicator, respectively. The sum of the three weights is one, and the weights are preset according to the degree of influence of the three parameters on the risk transmission attenuation.

[0064] In this embodiment, the initial conduction attenuation coefficient is obtained by weighting and summing the supply share ratio, the normalized inventory buffer days index, and the normalized alternative supplier availability index according to preset weights. The execution method is to multiply the values ​​of the three parameters by their respective preset weights and then add them together. The result is an initial conduction attenuation coefficient between zero and one. The larger the coefficient, the less attenuation the stress is subjected to when it propagates along the directed edge, and the more fully the downstream node is affected by the stress of the upstream node.

[0065] In this embodiment, the preset observation window refers to a fixed time interval that is set in advance. When monitoring the cooperative fluctuation mode of the multidimensional stress state vector of the two ends of the directed edge, the historical stress component data is extracted by using this time interval as a sliding window.

[0066] In this embodiment, the Pearson correlation coefficient of the inventory stress components at both ends of the directed edge refers to the degree of linear correlation between the inventory stress components at both ends of the directed edge and their time series values ​​within a preset observation window. The dynamic time warping distance of the delivery stress components refers to the minimum distance between the time series of delivery stress components at both ends of the directed edge after time axis bending and alignment. The Granger causality of the financial stress components refers to the predictive ability of the historical value of the financial stress component at one end of the directed edge to the current value of the financial stress component at the other end of the directed edge. These three indicators together characterize the degree of synchronous linkage of stress state changes at both ends of the directed edge from three dimensions: correlation, shape similarity, and causal driving force.

[0067] In this embodiment, the cooperative fluctuation pattern features are input into a pre-trained binary classifier, which outputs a probability score indicating whether a directed edge belongs to a cascaded coupling link. The pre-trained binary classifier is specifically a random forest classifier, constructed as follows: Directed edges from historical risk events that have been manually labeled and confirmed to have stress cascade propagation effects are collected as positive samples, while directed edges confirmed not to have cascade propagation effects are collected as negative samples. The Pearson correlation coefficient, dynamic time warp distance, and Granger causality of the nodes at both ends of each sample edge are extracted as input features. Whether or not it is a cascaded coupling link is used as the classification label for training the random forest model. After training, the model outputs a probability score between zero and one for the input features. The closer the probability score is to one, the more likely the directed edge is to be a cascaded coupling link.

[0068] In this embodiment, the probability score of a directed edge belonging to a cascaded coupling link is subtracted from 1 to obtain the cascade suppression factor. This cascade suppression factor is then multiplied by the initial conduction attenuation coefficient to obtain the corrected conduction attenuation coefficient. The corrected conduction attenuation coefficient is used for subsequent stress propagation calculations. Specifically, the higher the probability score of the cascaded coupling link, the more spurious synchronization components driven by common external factors in historical risk events are present in the stress at the two ends of the directed edge, and the weaker the independent conduction capability of the directed edge itself. In this case, the cascade suppression factor is closer to zero, the corrected conduction attenuation coefficient is closer to zero, and the attenuation of stress propagating along the link is more significant. Conversely, the lower the probability score of the cascaded coupling link, the more the stress synchronization fluctuations at the two ends of the directed edge mainly originate from real stress conduction. The closer the cascade suppression factor is to one, the closer the corrected conduction attenuation coefficient is to the initial conduction attenuation coefficient, and the stress is normally conducted. Subtracting the cascaded coupling probability score from one as the cascade suppression factor quantifies and eliminates spurious conduction components caused by common external factors in the coupling link, making the corrected conduction attenuation coefficient more accurately reflect the independent stress conduction capability of the directed edge itself.

[0069] Furthermore, the positive samples used in the training of the pre-trained binary classifier are the set of directed edges in historical risk events that have been confirmed to have cascade propagation, and the negative samples are the set of directed edges in historical risk events that have not had cascade propagation. The training features include the Pearson correlation coefficient of the historical multidimensional stress state vectors of the two endpoints, the dynamic time warping distance, the Granger causality, and the supply share of directed edges.

[0070] In this embodiment, cascading propagation refers to a linkage pattern in which the components of the multidimensional stress state vectors of the two ends of a directed edge exhibit synchronous and violent fluctuations in the time series during a historical risk disturbance event. This linkage pattern is confirmed by manual annotation or post-event analysis to be the direct driving factor of the stress change of the upstream node.

[0071] In this embodiment, the set of directed edges confirmed to have cascading propagation in historical risk events refers to the sample set composed of all directed edges selected from historical risk event records that have been manually labeled or confirmed to have cascading propagation effects through post-event analysis.

[0072] In this embodiment, the supply share ratio of a directed edge refers to the proportion of the quantity of materials purchased by a downstream node from the upstream node corresponding to the directed edge in the total purchase quantity of that material by the downstream node. This is the same parameter used in the above embodiment to calculate the initial conduction attenuation coefficient.

[0073] In this embodiment, the positive samples used for training the pre-trained binary classifier are the set of directed edges in historical risk events where cascading propagation has been confirmed, and the negative samples are the set of directed edges in historical risk events where cascading propagation has not occurred. The training features include the Pearson correlation coefficient, dynamic time warping distance, Granger causality, and supply share of the directed edge in the historical multidimensional stress state vectors of the two end nodes. Specifically, the following steps are taken: extract the historical inventory stress component time series, historical delivery stress component time series, and historical financial stress component time series of the two end nodes of each sample edge during the corresponding event period from the historical risk event records, calculate the Pearson correlation coefficient, dynamic time warping distance, and Granger causality respectively, and read the supply share of the sample edge from the supply chain knowledge graph. The above four feature values ​​are combined into a training feature vector, and whether it is a cascading coupling link is used as the classification label. The vector is then input into a random forest classifier for model training.

[0074] Furthermore, the risk transmission backbone generation module is also used to perform path confidence verification on the generated risk transmission backbone. The verification process includes: The path confidence assessment model inputs the corrected transmission attenuation coefficient of each directed edge on the risk transmission backbone, the total number of hops in the link, and the cumulative multidimensional stress state vector magnitude of the end node of the link. The path confidence assessment model outputs the overall confidence score of the backbone link. When the overall confidence score is lower than the path confidence threshold, abandon the current backbone link and backtrack to the suboptimal multidimensional stress gradient direction to re-trace until a backbone link that meets the path confidence threshold is generated, or the number of backtracking reaches the preset maximum number of backtracking, or there are no more unvisited suboptimal directions. If no backbone link that meets the path confidence threshold is generated when the backtracking terminates, the backbone link with the highest path confidence score during the backtracking process will be used as the output. The risk transmission backbone generation module is also used to: when the multidimensional stress gradients of multiple adjacent nodes of the source point all exceed the preset gradient threshold, simultaneously track these adjacent nodes to generate multiple parallel risk transmission branch links, and each branch link performs path confidence verification.

[0075] In this embodiment, the total number of hops in the link refers to the total number of directed edges traversed on the risk propagation backbone from the source node of the risk disturbance event to the end node of the link.

[0076] In this embodiment, the path confidence assessment model refers to a predictive model used to evaluate the overall reliability of the risk transmission backbone. This model is constructed using a logistic regression classifier. The input features are the mean of the corrected transmission attenuation coefficient for each directed edge on the risk transmission backbone, the total number of hops in the link, and the magnitude of the cumulative multidimensional stress state vector of the link's terminal node. The model output is an overall confidence score between zero and one. During model training, backbone links from historical risk events whose transmission paths have been manually verified as accurate are used as positive samples, and backbone links with confirmed deviations in the transmission paths are used as negative samples. The above three features are used as training inputs, and accuracy is used as the classification label. The closer the overall confidence score is to one, the more reliable the backbone link.

[0077] In this embodiment, the corrected conduction attenuation coefficient of each directed edge on the risk transmission backbone, the total number of hops in the link, and the magnitude of the cumulative multidimensional stress state vector of the link's end node are input into the path confidence assessment model. The path confidence assessment model outputs the overall confidence score of the backbone. The execution method is as follows: first, the arithmetic mean of the corrected conduction attenuation coefficients of all directed edges on the backbone is calculated. The three values ​​of the corrected conduction attenuation coefficient of the arithmetic mean, the total number of hops in the link, and the magnitude of the cumulative multidimensional stress state vector of the link's end node are combined to form a feature vector, which is then input into a pre-trained logistic regression classifier to obtain the overall confidence score output by the model.

[0078] In this embodiment, the path confidence threshold is a pre-set numerical standard used to determine whether the overall confidence of the risk transmission backbone meets the requirements, and the value is between zero and one.

[0079] In this embodiment, the current backbone link is abandoned and backtracked to the suboptimal multidimensional stress gradient direction to re-track until a backbone link that meets the path confidence threshold is generated. The execution method is as follows: when the overall confidence score of the current backbone link is lower than the path confidence threshold, the current backbone link is abandoned, and the adjacent node with the second largest value of the multidimensional stress gradient magnitude is selected as the alternative tracking direction from the previous tracking fork node. The hop-by-hop gradient tracking is re-executed along this direction, and the path confidence evaluation model is run again to calculate the overall confidence score for the regenerated backbone link. The above process is repeated until a backbone link that meets the path confidence threshold is generated.

[0080] In this embodiment, the backtracking count refers to the cumulative number of times the path confidence verification process involves retracing from the current branch node to the alternative gradient direction.

[0081] In this embodiment, the preset maximum number of backtracking operations is a pre-set upper limit value for the number of backtracking operations. When the number of backtracking operations reaches this upper limit, the backtracking will stop even if a backbone link that meets the path confidence threshold is still not generated.

[0082] In this embodiment, "no unvisited suboptimal direction" means that the multidimensional stress gradient directions corresponding to all adjacent nodes on the current branch node have been visited in the current or previous tracing or backtracking operations, and there are no alternative gradient directions that have not yet been tried.

[0083] In this embodiment, if no backbone link that meets the path confidence threshold is generated when the backtracking is terminated, the backbone link with the highest path confidence score during the backtracking process will be used as the output. The execution method is as follows: when the backtracking is terminated, the backbone link corresponding to the maximum overall confidence score is selected from all backbone links generated in this round of path confidence verification as the final output risk transmission backbone link.

[0084] In this embodiment, the risk transmission backbone generation module is also used for multi-branch parallel tracking. The execution method is as follows: after calculating the multidimensional stress gradient of each outgoing edge adjacent node with the source node of the risk disturbance event as the tracking starting point, if the magnitude of the multidimensional stress gradient of multiple adjacent nodes exceeds the preset gradient threshold, then an independent hop-by-hop tracking process is started simultaneously for these adjacent nodes. Each tracking process selects the next node hop-by-hop according to the direction of its own maximum multidimensional stress gradient, generating multiple parallel risk transmission branch links. Each branch link performs path confidence verification. The branch links that pass the verification are output in parallel with the backbone link, providing multi-directional transmission path references for supply chain managers.

[0085] Furthermore, the early warning evidence chain output module extracts stress components exceeding the early warning component threshold from the accumulated multidimensional stress state vector as the risk transmission early warning evidence chain. It also includes problem attribution chain generation, the process of which includes: For each marked node in the risk transmission early warning evidence chain, the upstream propagation path that caused each excessive stress component to exceed the early warning component threshold is traced. The upstream propagation path consists of directed edges on the main risk transmission link from the source node of the risk disturbance event to the corresponding marked node. Extract the key propagation edge with the largest stress contribution in the upstream propagation path; Each labeled node and its corresponding critical transmission edge are connected in the structural order of the risk transmission backbone to generate a problem attribution chain. The problem attribution chain is presented in the form of alternating node identifiers and critical transmission edge identifiers, indicating the source of the maximum stress for each affected node.

[0086] In this embodiment, the labeled node refers to a node in the cumulative multidimensional stress state vector on the risk transmission backbone where at least one component exceeds the threshold of the corresponding early warning component. The labeled node records the name and value of the stress component that exceeds the standard.

[0087] In this embodiment, the risk disturbance event source node refers to the affected node directly mapped to the initial risk disturbance event detected in the supply chain knowledge graph, and the hop-by-hop tracing of the risk transmission backbone starts from the risk disturbance event source node.

[0088] In this embodiment, for each marked node in the risk transmission early warning evidence chain, the upstream propagation path that caused each excessive stress component to exceed the early warning component threshold is traced. The upstream propagation path is composed of directed edges on the risk transmission backbone link from the source node of the risk disturbance event to the corresponding marked node. The execution method is as follows: taking the source node of the risk disturbance event as the starting point of the path and the marked node as the ending point of the path, all directed edges between the starting point and the ending point are extracted from the risk transmission backbone link and arranged in the direction order of the directed edges to form the upstream propagation path.

[0089] In this embodiment, the critical transmission edge with the largest stress contribution ratio refers to the directed edge with the highest contribution ratio to the propagation of stress components among all directed edges included in the upstream propagation path of the labeled node. The contribution ratio is determined by calculating the proportion of the stress contribution of each directed edge to the stress component value exceeding the standard of the labeled node to the total contribution of all directed edges. The directed edge with the highest contribution ratio is the critical transmission edge.

[0090] In this embodiment, each labeled node and its corresponding key transmission edge are connected in series according to the structural order of the risk transmission backbone to generate a problem attribution chain. The problem attribution chain is presented in the form of alternating node identifiers and key transmission edge identifiers, indicating the source of the maximum stress for each affected node. The execution method is as follows: the elements in the problem attribution chain are arranged sequentially from the source node to the end node along the risk transmission backbone. For each labeled node, the node identifier is output first, followed by the key transmission edge identifier pointing to that node, and then the node identifier and key transmission edge identifier of the next labeled node are output, forming a sequence structure in which node identifiers and key transmission edge identifiers are alternating. The problem attribution chain intuitively shows which transmission edge mainly transmits the excessive stress of each affected node.

[0091] Furthermore, the stress contribution ratio of each key transmission edge in the problem attribution chain is calculated using the Shapley value attribution method. On the set of directed edges from the source node of the risk disturbance event to the labeled node, each directed edge in the set is regarded as a participant, and the corresponding excess stress component value of the labeled node is taken as the total contribution target. The marginal contribution average of each directed edge to the corresponding excess stress component value is calculated as the stress contribution ratio.

[0092] In this embodiment, the excessive stress component value refers to the specific value of the stress component in the cumulative multidimensional stress state vector of the labeled node that exceeds the corresponding warning component threshold.

[0093] In this embodiment, the stress contribution ratio of each key propagation edge in the problem attribution chain is calculated using the Shapley value attribution method. On the set of directed edges from the source node of the risk disturbance event to the labeled node, each directed edge in the set is considered a participant. The corresponding excess stress component value of the labeled node is used as the total contribution target. The average marginal contribution of each directed edge to the corresponding excess stress component value is calculated as the stress contribution ratio. Specifically, the participant set is first determined to be all directed edges included in the upstream propagation path from the source node of the risk disturbance event to the labeled node. The excess stress component value is then considered as the total contribution target. The stress component value is taken as the total output value of all participants working together. Then, all participants are arranged in any order, and the marginal contribution of each participant before and after joining the alliance is calculated. The marginal contribution is equal to the total output value of the alliance after joining the participant minus the total output value of the alliance before joining the participant. The total output value of the alliance is determined by the cumulative value of the stress components of the corresponding paths of the participants in the alliance. Finally, the arithmetic mean of the marginal contribution of each participant under all arrangement orders is calculated to obtain the mean marginal contribution value of the participant. The mean marginal contribution value is the stress contribution ratio of the directed edge to the over-stress component value.

[0094] In another implementation, when the number of directed edges in the upstream propagation path exceeds a preset scale threshold, the stress contribution ratio is calculated using an approximate Shapley value algorithm based on Monte Carlo sampling. The marginal contribution average is estimated by randomly sampling a preset number of participants in the order of their arrangement, thereby reducing computational overhead while ensuring consistency in attribution ranking.

[0095] Furthermore, it also includes a risk transmission comparison and analysis module across the supply chain network, used for: Obtain supply chain knowledge graphs for at least two independent supply chain networks; Using a pre-set set of standard test risk events as input, the stress propagation path analysis module and the risk transmission backbone generation module are run on each supply chain network to obtain their respective risk transmission backbones. The risk transmission backbone links of different supply chain networks are standardized and aligned under a unified coordinate system. The similarity of the link topology and the difference of the stress attenuation mode are calculated. A preset number of nodes with the largest cumulative multidimensional stress state vector magnitude in each supply chain network are extracted as the set of weak nodes. Output a comparative analysis report across supply chain networks. The report should include the similarity values ​​of the link topology of different supply chain networks, the differences in stress attenuation patterns, and the node identifiers and types included in the weak node sets of each supply chain network.

[0096] In this embodiment, two independent supply chain networks refer to two complete supply chain systems operated by different corporate entities or different industry alliances that do not have material supply or financial transactions with each other. For example, an automobile manufacturing supply chain network and an electronics manufacturing supply chain network, or the respective supply chain networks of two companies in the same industry that have no business relationship.

[0097] In this embodiment, the supply chain network refers to a complete supply relationship topology structure represented in the form of a supply chain knowledge graph, covering the entire supply relationship from upstream suppliers through manufacturers and logistics hubs to downstream distribution nodes.

[0098] In this embodiment, the stress propagation path analysis module and the risk transmission backbone generation module are run for each supply chain network to obtain their respective risk transmission backbones. The execution method is as follows: the same type and the same level of impact risk disturbance events are input into the supply chain knowledge graphs of the two supply chain networks, and each network independently executes the complete process from stress quantification to stress propagation to gradient tracking. Each of the two networks generates a risk transmission backbone.

[0099] In this embodiment, the preset standard test risk event set refers to a predefined set of standardized risk disturbance events containing different event types and different impact levels. The event type, impact level and occurrence location of each test event in the standard test risk event set have been pre-set. The same standard test risk event set is input into two supply chain networks respectively to ensure that the two networks generate the backbone link and compare the transmission path under the same risk disturbance conditions.

[0100] In this embodiment, the unified coordinate system refers to a standardized description framework constructed using supply chain entity types as coordinate dimensions. The coordinate dimensions are defined according to four types of entity roles: suppliers, manufacturers, logistics hubs, and distribution nodes. Each risk transmission backbone link is represented in the unified coordinate system as an alternating arrangement of node type sequences and edge type sequences. From the start point to the end point of the sequence, the entity type of each node and the material flow relationship or supply dependency relationship type of each directed edge are labeled in sequence.

[0101] In this embodiment, the risk transmission backbone links of different supply chain networks are standardized and aligned under a unified coordinate system. The similarity of the link topology and the difference of the stress attenuation mode are calculated. The execution method is as follows: the two risk transmission backbone links are converted into node type sequences under a unified coordinate system, and the edit distance between the two node type sequences is calculated as the similarity of the link topology. The smaller the edit distance, the more similar the two links are in terms of node type distribution. The difference of the corrected transmission attenuation coefficients at corresponding positions on the two links is calculated sequentially, and the sum of the absolute values ​​of all differences is taken as the difference of the stress attenuation mode. The similarity of the link topology and the difference of the stress attenuation mode together constitute a comparable quantitative index of the two links in terms of transmission path structure and attenuation characteristics.

[0102] In this embodiment, a preset number of nodes with the largest cumulative multidimensional stress state vector magnitude in each supply chain network are extracted as a set of weak nodes. The execution method is as follows: for each supply chain network, the preset number of nodes with the largest cumulative multidimensional stress state vector magnitude are selected from the main risk transmission link and all nodes affected by stress propagation. The node identifier and node type are recorded to form a set of weak nodes for the corresponding supply chain network. The set of weak nodes identifies the nodes that are under the most severe pressure under the same type of risk disturbance event.

[0103] Furthermore, it also includes a stress component radar chart overlay module, which is used to present the numerical proportions of inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components of each node in real time on the topological display base map of the supply chain knowledge graph in the form of a radar chart. Each axis scale of the radar chart corresponds to the warning component threshold of each stress component, and the axis segment exceeding the standard is highlighted with a warning color.

[0104] In this embodiment, the topology display base map of the supply chain knowledge graph refers to a visual graphical interface rendered based on the nodes and directed edges of the supply chain knowledge graph. Nodes are represented by graphic symbols and labeled with node identifiers, and directed edges are connected by arrows to indicate the direction of material flow or supply dependence. The layout positions of nodes and directed edges in the interface are automatically arranged according to the topology structure of the graph.

[0105] In this embodiment, on the topological display base map of the supply chain knowledge graph, the numerical proportions of the inventory stress component, delivery stress component, financial stress component, and alternative feasible stress component of each node are presented in real time in the form of a radar chart. The execution method is as follows: a four-axis radar chart is drawn next to the graphic symbol of each node. The four axes of the radar chart correspond to the inventory stress component axis, delivery stress component axis, financial stress component axis, and alternative feasible stress component axis, respectively. The data points on each axis are marked according to the real-time values ​​of the corresponding stress components. The quadrilateral formed by connecting the four data points intuitively shows the numerical proportion relationship between the four stress components.

[0106] In this embodiment, each axis of the radar chart corresponds to the warning threshold of each stress component. The segment exceeding the standard is highlighted in warning color. The execution method is as follows: a red threshold scale line is marked on each axis of the radar chart. The position of the red threshold scale line on the axis corresponds to the warning threshold of the stress component. When the real-time value of a stress component exceeds the corresponding warning threshold, the segment on the axis from the red threshold scale line to the data point is automatically rendered as red and highlighted. This allows monitoring personnel to intuitively identify the type and magnitude of the current stress component exceeding the standard in four stress dimensions at the same time.

[0107] Furthermore, it also includes a warning evidence chain narrative generation module, which is used to convert the risk transmission warning evidence chain and problem attribution chain into natural language warning narrative text. The natural language warning narrative text starts from the source node of the risk disturbance event and describes the changes in multidimensional stress state and the main causes of the transmitted stress node by node along the main risk transmission link. The early warning evidence chain narrative generation module is also used to receive input natural language query text, perform semantic parsing on the natural language query text, retrieve matching content in the natural language early warning narrative text based on the parsing results, and output the corresponding narrative text fragments.

[0108] In this embodiment, the risk transmission early warning evidence chain and the problem attribution chain are converted into natural language early warning narrative text. The execution method is as follows: taking the risk transmission early warning evidence chain as the main content and the problem attribution chain as the attribution supplement, a pre-trained natural language generation model is called to convert the structured directed link data into a continuous text description that conforms to the expression habits of natural language. The input of the natural language generation model is the node sequence of the risk transmission backbone link, the name and value of the excessive stress component of each node, and the key transmission edge identifier of each labeled node. The output is the natural language early warning narrative text.

[0109] In this embodiment, the natural language early warning narrative text starts from the source node of the risk disturbance event and describes the changes in multidimensional stress state and the main causes of the incoming stress node by node along the main risk transmission link. The text content includes: first, describing the type, occurrence time and magnitude of the risk disturbance event; then, describing the changes in multidimensional stress state of each node in sequence along the main link; additionally describing the name and magnitude of the exceeding stress component for the labeled node; and explaining the main source transmission edge identifier of the incoming stress according to the problem attribution chain.

[0110] In this embodiment, natural language query text refers to the query question text that the user enters in everyday natural language in response to the warning situation.

[0111] In this embodiment, the input natural language query text is received, and semantic parsing of the natural language query text is performed. The execution method is as follows: the natural language query text is input into a pre-trained semantic understanding model for intent recognition and entity extraction, and the query intent type and entity information such as node identifiers, stress component names or time ranges involved in the query are identified.

[0112] In this embodiment, based on the parsing results, matching content is retrieved in the natural language early warning narrative text, and the corresponding narrative text fragment is output. The execution method is as follows: according to the node identifier and stress component name obtained by semantic parsing, the matching paragraph is located in the natural language early warning narrative text, and the text of the matching paragraph and its context with a preset length is extracted as the narrative text fragment for output.

[0113] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A knowledge graph-based supply chain data intelligent association and risk early warning system, characterized in that, include: The supply chain knowledge graph construction module is used to extract supply chain entities and relationships between entities from multi-source heterogeneous data, and construct a supply chain knowledge graph in which nodes represent suppliers, manufacturers, logistics hubs, and distribution nodes, and directed edges represent the direction of material flow and supply dependence between entities. The multidimensional stress state quantification module is used to map real-time perceived multi-type risk disturbance events to the affected nodes of the supply chain knowledge graph. It generates a multidimensional stress state vector for the affected nodes, which includes inventory stress components, delivery stress components, financial stress components, and alternative feasible stress components. The multidimensional stress state vector is then superimposed on the corresponding node attributes of the supply chain knowledge graph. The stress propagation path parsing module is used to propagate the multidimensional stress state vector of the affected node to the adjacent nodes hop by hop along the directed edges of the supply chain knowledge graph according to the stress conduction attenuation rule, so as to obtain the cumulative multidimensional stress state vector of each node. The risk transmission backbone generation module is used to calculate the multidimensional stress gradient between each adjacent node and the current node's cumulative multidimensional stress state vector by hop along the directed edge direction of the supply chain knowledge graph, starting from the source point. The adjacent node corresponding to the maximum value of the multidimensional stress gradient is selected as the next hop node to generate the risk transmission backbone. The early warning evidence chain output module is used to extract stress components that exceed the early warning component threshold from the cumulative multidimensional stress state vector of each node on the risk transmission backbone link as a risk transmission early warning evidence chain. The risk transmission early warning evidence chain is output in the form of a directed link labeled with node identifier, stress component name, stress component value and risk level. 2.The knowledge graph-based supply chain data intelligent association and risk early warning system according to claim 1, characterized in that, The multidimensional stress state quantization module is also used to perform correlation enhancement on multimodal weak signals. The correlation enhancement process includes: Continuously monitor low-confidence signals in multimodal data sources, including ambiguous sentiment statements in social media texts, delayed delivery status updates in supplier portal data streams, and irregular pauses between transportation nodes in logistics trajectory data streams. The low-confidence signals detected are grouped according to time windows and related entities to form multiple sets of weak signals to be confirmed. The entity identifiers of the related entities are shared with the node identifiers in the supply chain knowledge graph. For each set of weak signals to be confirmed, the source modality, semantic vector, and timestamp features of the signal are extracted to construct a weak signal association graph. The weak signal association graph uses each low-confidence signal as a node and the modal complementarity, semantic similarity, and temporal proximity between signals as edges. The graph diffusion algorithm is run on the weak signal association graph to calculate the association enhancement confidence score of each node. When the association enhancement confidence score exceeds the weak signal confirmation threshold, the corresponding node is upgraded from a low confidence signal to a confirmed early risk signal. The confirmed early risk signals are input into the stress component weight allocation process to generate the pre-multidimensional stress state vector of the corresponding affected node. The pre-multidimensional stress state vector is superimposed on the corresponding node before the core risk disturbance event triggers formal stress quantification.

3. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 1, characterized in that, In the stress propagation path analysis module, stress propagation is performed hop-by-hop to adjacent nodes according to the stress conduction attenuation rule. The stress conduction attenuation rule is set as follows: The initial transmission attenuation coefficient is obtained by weighting and summing the supply share ratio, the normalized inventory buffer days index, and the normalized alternative supplier availability index according to preset weights. The cooperative fluctuation pattern of the multidimensional stress state vector of the two nodes of the directed edge is monitored within a preset observation window. The cooperative fluctuation pattern is characterized by the Pearson correlation coefficient of the inventory stress component of the two nodes, the dynamic time warping distance of the delivery stress component, and the Granger causality of the financial stress component. The cooperative wave pattern features are input into a pre-trained binary classifier, which outputs the probability score of directed edges belonging to cascaded coupling links. Subtracting the probability score of the directed edge belonging to the cascaded coupling link from 1 gives the cascade suppression factor. Multiplying the cascade suppression factor by the initial conduction attenuation coefficient gives the corrected conduction attenuation coefficient.

4. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 3, characterized in that, The positive samples used in the training of the pre-trained binary classifier are the set of directed edges in historical risk events that have been confirmed to have cascade propagation, and the negative samples are the set of directed edges in historical risk events that have not had cascade propagation. The training features include the Pearson correlation coefficient of the historical multidimensional stress state vectors of the two end nodes, the dynamic time warping distance, the Granger causality, and the supply share of directed edges.

5. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 3, characterized in that, The risk transmission backbone generation module is also used to perform path confidence verification on the generated risk transmission backbone. The verification process includes: The path confidence assessment model inputs the corrected transmission attenuation coefficient of each directed edge on the risk transmission backbone, the total number of hops in the link, and the cumulative multidimensional stress state vector magnitude of the end node of the link. The path confidence assessment model outputs the overall confidence score of the backbone link. When the overall confidence score is lower than the path confidence threshold, abandon the current backbone link and backtrack to the suboptimal multidimensional stress gradient direction to re-trace until a backbone link that meets the path confidence threshold is generated, or the number of backtracking reaches the preset maximum number of backtracking, or there are no more unvisited suboptimal directions. If no backbone link that meets the path confidence threshold is generated when the backtracking terminates, the backbone link with the highest path confidence score during the backtracking process will be used as the output. The risk transmission backbone generation module is also used to: when the multidimensional stress gradients of multiple adjacent nodes of the source point all exceed the preset gradient threshold, simultaneously track these adjacent nodes to generate multiple parallel risk transmission branch links, and each branch link performs path confidence verification.

6. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 1, characterized in that, The early warning evidence chain output module extracts stress components exceeding the early warning component threshold from the accumulated multidimensional stress state vector as the risk transmission early warning evidence chain. It also includes problem attribution chain generation, the process of which includes: For each marked node in the risk transmission early warning evidence chain, the upstream propagation path that caused each excessive stress component to exceed the early warning component threshold is traced. The upstream propagation path consists of directed edges on the main risk transmission link from the source node of the risk disturbance event to the corresponding marked node. Extract the key propagation edge with the largest stress contribution in the upstream propagation path; Each labeled node and its corresponding critical transmission edge are connected in the structural order of the risk transmission backbone to generate a problem attribution chain. The problem attribution chain is presented in the form of alternating node identifiers and critical transmission edge identifiers, indicating the source of the maximum stress for each affected node.

7. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 6, characterized in that, The stress contribution ratio of each key transmission edge in the problem attribution chain is calculated using the Shapley value attribution method. On the set of directed edges from the source node of the risk disturbance event to the labeled node, each directed edge in the set is regarded as a participant. The corresponding excess stress component value of the labeled node is used as the total contribution target. The marginal contribution average of each directed edge to the corresponding excess stress component value is calculated as the stress contribution ratio.

8. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 1, characterized in that, It also includes a risk transmission comparison and analysis module across supply chain networks, used for: Obtain supply chain knowledge graphs for at least two independent supply chain networks; Using a pre-set set of standard test risk events as input, the stress propagation path analysis module and the risk transmission backbone generation module are run on each supply chain network to obtain their respective risk transmission backbones. The risk transmission backbone links of different supply chain networks are standardized and aligned under a unified coordinate system. The similarity of the link topology and the difference of the stress attenuation mode are calculated. A preset number of nodes with the largest cumulative multidimensional stress state vector magnitude in each supply chain network are extracted as the set of weak nodes. Output a comparative analysis report across supply chain networks. The report should include the similarity values ​​of the link topology of different supply chain networks, the differences in stress attenuation patterns, and the node identifiers and types included in the weak node sets of each supply chain network.

9. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 1, characterized in that, It also includes a stress component radar chart overlay module, which is used to present the numerical proportions of inventory stress component, delivery stress component, financial stress component, and alternative feasible stress component of each node in real time on the topological display base map of the supply chain knowledge graph in the form of a radar chart. Each axis scale of the radar chart corresponds to the warning component threshold of each stress component, and the axis segment exceeding the standard is highlighted with a warning color.

10. The knowledge graph-based intelligent association and risk early warning system for supply chain data as described in claim 6, characterized in that, It also includes a warning evidence chain narrative generation module, which is used to convert the risk transmission warning evidence chain and problem attribution chain into natural language warning narrative text. The natural language warning narrative text starts from the source node of the risk disturbance event and describes the multidimensional stress state changes and the main causes of the transmitted stress node by node along the main risk transmission link. The early warning evidence chain narrative generation module is also used to receive input natural language query text, perform semantic parsing on the natural language query text, retrieve matching content in the natural language early warning narrative text based on the parsing results, and output the corresponding narrative text fragments.

Images