Supply chain toughness risk assessment method and system based on AI assistance

Abstract

The invention discloses a supply chain toughness risk assessment method and system based on AI assistance. The method comprises the following steps: firstly, constructing a risk conduction topology network; secondly, obtaining entity feature topology of each entity and risk attribute topology of each risk event; and finally, based on the three types of topology training toughness evaluation model, outputting an anti-risk toughness coefficient of the supply chain entity for the risk event. By means of the design, through multi-topology fusion and AI model training, the complex interaction relation between the entity and the risk is accurately quantified, dynamic evaluation of the supply chain toughness is achieved, and a scientific basis is provided for an enterprise to make a risk coping strategy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted supply chain resilience risk assessment method and system. Background Art

[0002] Supply chains are at the core of the modern economy, and their resilience directly impacts the stability of businesses and their entire industrial chains. Traditional supply chain risk assessment methods often rely on manual experience or simple statistical models, which present significant limitations. Firstly, the complex transmission relationship between entities and risk events is often simplified into a single causal chain, making it difficult to accurately characterize the risk diffusion path. Secondly, the multidimensional attributes of an entity's multi-dimensional capacity risk are not systematically integrated, resulting in significant deviations between assessment results and actual resilience levels. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-assisted supply chain resilience risk assessment method and system.

[0004] In a first aspect, an embodiment of the present invention provides an AI-assisted supply chain resilience risk assessment method, comprising:

[0005] Obtain a risk transmission topology network; the risk transmission topology network includes multiple supply chain entity units and multiple risk event units, each supply chain entity unit represents a supply chain entity, and each risk event unit represents a risk event; if any supply chain entity among the multiple supply chain entities has a transmission relationship with any risk event among the multiple risk events, then the supply chain entity unit of any supply chain entity and the risk event unit of any risk event have a risk transmission edge in the risk transmission topology network;

[0006] Obtaining an entity feature topology corresponding to each supply chain entity in the plurality of supply chain entities; any entity feature topology includes a plurality of supply chain entity feature units, and any supply chain entity feature unit is used to characterize the supply chain entity feature of the corresponding supply chain entity on an evaluation indicator;

[0007] Obtaining a risk attribute topology corresponding to each of the multiple risk events; each risk attribute topology includes multiple risk event feature units, and each risk event feature unit is used to characterize the risk event feature of the corresponding risk event on an evaluation indicator;

[0008] A resilience assessment model is trained based on the risk transmission topology network, the entity feature topology of each supply chain entity and the risk attribute topology of each risk event to obtain a trained resilience assessment model; the trained resilience assessment model is used to evaluate the risk resilience coefficient of the supply chain entity against risk events.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0010] Compared to existing technologies, the present invention offers the following benefits: It employs the AI-assisted supply chain resilience risk assessment method and system disclosed in the present invention to construct a risk transmission topology network; secondly, it obtains the entity feature topology of each entity and the risk attribute topology of each risk event; and finally, it trains a resilience assessment model based on three types of topologies to output the risk resilience coefficient of the supply chain entity against risk events. This design, through multi-topology fusion and AI model training, accurately quantifies the complex interactive relationship between entities and risks, achieves a dynamic assessment of supply chain resilience, and provides a scientific basis for enterprises to formulate risk response strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0012] Figure 1 A schematic diagram of the steps of the AI-assisted supply chain resilience risk assessment method provided in an embodiment of the present invention;

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0015] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0016] In order to solve the technical problems in the above background technology, Figure 1 A flow chart of the AI-assisted supply chain resilience risk assessment method provided in an embodiment of the present disclosure is provided below. The AI-assisted supply chain resilience risk assessment method is introduced in detail.

[0017] Step S201: Obtain a risk transmission topology network; the risk transmission topology network includes multiple supply chain entity units and multiple risk event units, each supply chain entity unit represents a supply chain entity, and each risk event unit represents a risk event; if any supply chain entity among the multiple supply chain entities has a transmission relationship with any risk event among the multiple risk events, then the supply chain entity unit of any supply chain entity and the risk event unit of any risk event have a risk transmission edge in the risk transmission topology network;

[0018] Step S202: Obtain an entity feature topology corresponding to each supply chain entity in the plurality of supply chain entities; any entity feature topology includes a plurality of supply chain entity feature units, and any supply chain entity feature unit is used to represent the supply chain entity feature of the corresponding supply chain entity on an evaluation indicator;

[0019] Step S203: Obtain a risk attribute topology corresponding to each of the multiple risk events; any risk attribute topology includes multiple risk event feature units, and any risk event feature unit is used to characterize the risk event feature of the corresponding risk event on an evaluation indicator;

[0020] Step S204: training a resilience assessment model based on the risk transmission topology network, the entity feature topology of each supply chain entity, and the risk attribute topology of each risk event to obtain a trained resilience assessment model; the trained resilience assessment model is used to evaluate the risk resilience coefficient of the supply chain entity against risk events.

[0021] In an embodiment of the present invention, illustratively, first, the server needs to obtain a risk transmission topology network. The network is a graph structure containing supply chain entity units and risk event units. The nodes are connected by risk transmission edges, and the existence of the edges is determined by the actual transmission relationship between the supply chain entities and the risk events. For example, in the supply chain scenario of the consumer electronics industry, the server analyzes historical data (such as supply chain interruption records and accident reports in the past three years) and industry knowledge graphs (such as risk impact relationships annotated by experts) to identify that chip supplier A has a direct transmission relationship with the "chip production line fire" risk event F1 (the fire causes A to be unable to supply), mobile phone assembly plant C has an indirect transmission relationship with F1 (A's supply interruption causes C to stop working), logistics company D has a direct transmission relationship with the "port strike" risk event F3 (the strike will cause D's transportation delays), and retailer E has a direct transmission relationship with the "sudden drop in consumer demand" risk event F4 (the decline in demand causes E's inventory backlog). Based on this, the server constructs a risk transmission topology network, which includes 5 supply chain entity units (A, B, C, D, E) and 4 risk event units (F1-F4). Risk transmission edges are established between A-F1, C-F1, D-F3, and E-F4, and there are no edges between the other nodes without direct transmission relationships. Secondly, the server needs to obtain the entity feature topology corresponding to each supply chain entity. This topology is a subgraph containing multiple supply chain entity feature units. Each feature unit corresponds to an evaluation indicator (such as inventory level, number of backup suppliers, production elasticity, etc.). The edges between units are determined by the actual transmission relationship between features (that is, the impact of a feature change on another feature). It is specifically quantitatively verified by the following methods: Statistical correlation analysis: Calculate the Pearson correlation coefficient between features. If |r| ≥ 0.5, it is considered that there is a transmission relationship (for example, the correlation coefficient between 'inventory level' and 'production elasticity' is r = -0.62, indicating that for every 10% decrease in inventory level, the average time of production elasticity decreases). Causal inference verification: Structural equation modeling (SEM) was used to verify the causal paths between characteristics. Edges were retained if the path coefficient was ≥ 0.3 (for example, the path coefficient from 'number of backup suppliers' to 'inventory level' was 0.45, indicating that each additional backup supplier increased inventory levels by an average of 8%). Expert calibration: The subjective ratings of the impact of the characteristics (1-5) were combined with industry experts. Edges were retained if the rating was ≥ 4 (for example, the expert rating from 'financial liquidity' to 'number of backup suppliers' was 4.2, indicating that sufficient funds significantly support backup supplier cooperation). Taking mobile phone assembly plant C as an example, the server determined its evaluation indicators, including "inventory level (chip / battery inventory)", "number of backup suppliers", "production flexibility (production line switching time)", "financial liquidity (cash / debt ratio)", and "geographic location (multiple ports)" by analyzing inventory records in its ERP system, the backup supplier list in its supplier management system, and historical downtime event data.Further analysis of the transmission relationship between features reveals the following: low inventory levels can lead to decreased production elasticity (when inventory is insufficient, switching production lines requires longer waiting times for materials); a small number of backup suppliers can exacerbate the impact of low inventory levels (when there are no alternative suppliers, inventory cannot be quickly replenished after depletion); poor financial liquidity can limit the number of backup suppliers (when funds are insufficient, more backup agreements cannot be signed); and a remote geographical location can prolong the time it takes to recover production elasticity (increased transportation time leads to slower arrival of backup materials). Based on this, the server constructs an entity feature topology for C, which includes the five feature units mentioned above and establishes transmission edges between "inventory level-production elasticity," "number of backup suppliers-inventory level," "financial liquidity-number of backup suppliers," and "geographic location-production elasticity." Finally, the server needs to obtain the risk attribute topology corresponding to each risk event. This topology is a subgraph containing multiple risk event feature units. Each feature unit corresponds to an evaluation metric (such as probability of occurrence, scope of impact, duration, etc.). The edges between units are determined by the actual transmission relationship between features (i.e., the amplification or mitigation effect of a change in one feature on another). This is quantitatively verified through the following methods: Event log analysis: Extract characteristic data of historical risk events (such as duration and scope of impact) and calculate conditional probabilities (for example, if the duration is ≥ 72 hours and the probability of expanding the scope of impact is 70%, the "duration-scope of impact" edge is retained); Regression analysis: Fit a linear regression model with "severity" as the dependent variable and "scope of impact" as the independent variable. If the regression coefficient is ≥ 0.5, the transmission edge is retained (for example, if the severity increases by an average of 10% for each additional chip model in the "scope of impact," the regression coefficient is 0.6, so the "scope of impact-severity" edge is retained); Domain knowledge integration: Refer to industry risk assessment guidelines. If the guidelines clearly state the impact mechanism of one feature on another (for example, "low predictability" increases the "probability of occurrence"), the transmission edge is retained. Taking risk event F1 (a chip production line fire) as an example, the server analyzed semiconductor factory fire data recorded by the fire department and insurance company risk assessment reports to determine its evaluation indicators, including "probability of occurrence (annual frequency of similar fires in the past five years)", "scope of impact (number of affected chip models)", "duration (time from fire to resumption of production)", "severity (direct economic losses from downstream shutdowns)", and "predictability (whether an early warning system is in place)". Further analysis of the transmission relationship between these characteristics revealed that low predictability (no early warning system) increases the probability of occurrence (making it impossible to detect hidden dangers in advance); long duration increases the scope of impact (prolonged shutdowns affect more downstream customers); a larger scope of impact increases severity (supply disruptions to more chip models lead to greater economic losses); and a high probability of occurrence reduces predictability (frequent events may go unnoticed).Based on this, the server constructs a risk attribute topology for F1, which includes the above five feature units, and establishes conductive edges between "predictability-probability of occurrence", "duration-scope of impact", "scope of impact-severity", and "probability of occurrence-predictability". The criteria for determining the transmission relationship are as follows: direct transmission: there is a direct causal relationship between the entity and the risk event, and the risk event in historical data causes the entity to be shut down for ≥3 days or suffer economic losses ≥5% of the entity's monthly income (for example, chip supplier A was directly shut down for 7 days due to 'chip production line fire' F1, and the loss accounted for 8% of its monthly income, so A and F1 have a direct transmission relationship); indirect transmission: the entity is associated with the risk event through at least one intermediate entity, and the indirect impact causes the entity to be shut down for ≥2 days or suffer economic losses ≥3% of the entity's monthly income (for example, mobile phone assembly plant C was indirectly shut down for 4 days due to A's supply interruption, and the loss accounted for 4% of its monthly income, so C and F1 have an indirect transmission relationship); quantitative threshold: the probability of the entity being affected by the risk event is statistically calculated through historical data. If the impact probability is ≥20%, it is determined that a transmission relationship exists (for example, the probability of logistics provider D being affected by 'port strike' F3 is 25%, so D and F3 have a transmission relationship).

[0022] Next, the server trains the resilience assessment model based on the above three types of topologies. In specific implementation, the risk conduction topology network is first converted into a risk conduction association matrix. The rows of the matrix represent the supply chain entity units, the columns represent the risk event units, and the element values ​​are 1 (there is a conduction edge) or 0 (no conduction edge). For example, for the risk conduction topology network constructed above, the server generates a 5×4 matrix M, where M[C][F1]=1 (assembly plant C and F1 have a conduction edge), M[A][F1]=1 (supplier A and F1 have a conduction edge), and the rest such as M[B][F1]=0 (battery supplier B has no direct relationship with F1).

[0023] Subsequently, the server calls the resilience assessment model initialized as a graph neural network (GNN). Specifically, it can call the resilience assessment model initialized as a graph attention network (GAT), and achieve two-way information transmission between entities and risk events by weighting neighborhood features through the attention mechanism. The neighborhood aggregation function of GAT is defined as in, is the feature vector of node i in layer l, W (l) is the learnable weight matrix of layer l, α ij is the attention coefficient of nodes i and j (through α ij =softmax j (LeakyReLU(a T [Wh i ||Wh j])) calculation, where a is a learnable attention vector and || represents a concatenation operation). Through the interaction between the "risk information transmission matrix" learned by the model and the risk conduction association matrix M, bidirectional information transmission between entities and risk events is achieved: on the one hand, risk events transmit information to entities (for example, F1's "long duration" feature is transmitted to A and C, prompting A to quickly resume production and C to activate the backup plan); on the other hand, entities transmit information to risk events (for example, C's "large number of backup suppliers" feature is transmitted to F1, suggesting that the actual impact of the risk on C may be weakened). Through multiple rounds (for example, three rounds) of bidirectional transmission, multiple risk conduction comprehensive representation matrices are generated, and the value of each matrix element represents the combined impact strength of the entity and the risk event after that round of transmission.

[0024] To integrate information from multiple rounds of transmission, the server applies a multi-head attention mechanism to multiple risk transmission comprehensive representation matrices, assigning higher weights to key rounds (such as the second round, focusing on indirect impacts) to generate a fused risk transmission comprehensive representation matrix. This matrix is ​​then transformed through a fully connected layer to reduce its dimensionality, resulting in a target risk transmission comprehensive representation matrix. Based on this matrix, the server extracts the "entity-based risk representation vector" for each entity (e.g., the vector for C contains its basic impact strength in risk events such as F1 and F2) and the "risk-based transmission representation vector" for each risk event (e.g., the vector for F1 contains its basic impact strength on entities such as A and C).

[0025] For the entity feature topology, the server calls the model to screen key transmission edges: by calculating the importance score of the edge (such as gradient-based importance analysis), retaining high-scoring edges (such as the "number of backup suppliers-inventory level" edge with the highest score in C's entity feature topology), and generating the first core feature interaction topology; based on this topology, the "unit feature" of each feature unit is calculated through GNN neighborhood aggregation (such as the "number of backup suppliers" unit feature value is 3, and the "inventory level" unit feature value is 0.8), and the "entity multi-dimensional capability representation vector" of the entity is obtained by splicing (such as C's vector is [3, 0.8, 48 hours, 2.5, 1]).

[0026] Similarly, for the risk attribute topology, the server calls the model to filter key transmission edges (for example, the "duration-impact range" edge has the highest score in the risk attribute topology of F1) to generate the second core feature interaction topology; based on this topology, the "unit characteristics" of each risk event feature unit are calculated (for example, the "duration" unit characteristic value of F1 is 72 hours, and the "impact range" unit characteristic value is 5), and the "risk multi-dimensional attribute representation vector" of the risk event is obtained by splicing (for example, the vector of F1 is [5%, 5, 72, 5000, 0]).

[0027] Finally, the server trains the model using four vectors: entity-based risk, entity-multidimensional capabilities, risk-based transmission, and risk-multidimensional attributes. By calculating synergy bias (the mismatch between entity-based risk and multidimensional capabilities, and between risk-based transmission and multidimensional attributes), risk transmission assessment bias (the difference between the model's outputted resilience coefficient and the historical transmission relationship), and complexity constraints (L2 regularization), a multi-task loss function is constructed. Model parameters (such as the GNN weight matrix and attention head parameters) are optimized through backpropagation until the loss converges, resulting in a fully trained resilience assessment model.

[0028] When the model is applied, the server obtains the assessed supply chain entity (e.g., e-commerce platform E) and the assessed risk event (e.g., a sudden drop in consumer demand F4). The model then outputs E's risk resilience coefficient for F4. If the coefficient falls below a threshold (e.g., 0.6), the server further analyzes E's multidimensional capability vector (e.g., inventory turnover days of 30 days, pre-sale ratio of 10%) and F4's multidimensional risk attribute vector (e.g., a 30% drop in demand, lasting for 6 months), generates targeted recommendations (e.g., "Increase pre-sale ratio to 30%, shorten inventory turnover days to 20 days"), and synchronizes them to E's management system.

[0029] In summary, this embodiment achieves quantitative assessment of supply chain resilience risks and generation of targeted enhancement suggestions by constructing a multi-dimensional topological network and combining it with AI model training, effectively improving the intelligence level of supply chain resilience management.

[0030] In an embodiment of the present invention, the resilience assessment model is trained based on the risk transmission topology network, the entity feature topology of each supply chain entity and the risk attribute topology of each risk event to obtain a trained resilience assessment model, which can be implemented through the following examples.

[0031] Calling the resilience assessment model to obtain an entity-based risk representation vector of each supply chain entity and a risk-based transmission representation vector of each risk event based on the risk transmission topology network;

[0032] Calling the resilience assessment model to obtain an entity multi-dimensional capability representation vector of each supply chain entity based on the entity feature topology of each supply chain entity;

[0033] Calling the resilience assessment model to obtain a risk multi-dimensional attribute representation vector of each risk event based on the risk attribute topology of each risk event;

[0034] The resilience assessment model is trained based on the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event to obtain the trained resilience assessment model.

[0035] In an embodiment of the present invention, illustratively, after the server completes the construction of the risk conduction topology network, entity feature topology, and risk attribute topology, it enters the resilience assessment model training phase. The following describes in detail the specific implementation process of the training steps in conjunction with the consumer electronics supply chain scenario (including entities such as chip supplier A, mobile phone assembly plant C, and "chip production line fire" risk event F1, etc.): The server first converts the risk conduction topology network into a risk conduction association matrix (a 5×4 matrix, where rows represent entities A, B, C, D, and E, columns represent risk events F1-F4, and element values ​​1 or 0 indicate whether there is a conduction edge). A resilience assessment model initialized as a graph neural network (GNN) is then called, and risk information is bidirectionally transmitted based on this matrix: Risk events transmit information to entities. For example, risk event F1 (chip fire) transmits its basic risk characteristics (such as "historical occurrence frequency" and "average impact duration") to associated entities A (direct supplier) and C (downstream assembly plant) through conductive edges. The model calculates the initial risk perception values ​​of A and C through neighborhood aggregation (A's initial value is 0.9, due to being directly affected by F1; C's initial value is 0.7, due to being indirectly affected by A's supply disruption). Entities transmit information to risk events: Entities A and C transmit their basic risk resistance capabilities (such as A's "completeness of firefighting facilities" and "number of backup production lines"; C's "safety inventory days" and "backup supplier contact records") to F1 through conductive edges. The model updates F1's basic transmission strength (F1's transmission strength to A is 0.8, and its transmission strength to C is 0.6). After three rounds of information transmission, the server obtains a comprehensive risk transmission representation matrix (for example, in the third round matrix, the comprehensive value of A-F1 is 0.85, and the comprehensive value of C-F1 is 0.72). The multi-head attention mechanism fuses the matrices from multiple rounds (assigning a weight of 0.6 to the third round because it focuses more on long-term impacts) to generate a fused matrix. Then, through a fully connected layer dimensionality reduction transformation, the "entity-based risk representation vector" for each entity is extracted (for example, the vector for C is [0.72, 0.2, 0.1, 0.3], corresponding to the basic risk strengths of F1-F4, respectively) and the "risk-based transmission representation vector" for each risk event (for example, the vector for F1 is [0.85, 0.1, 0.72, 0.1], corresponding to the basic transmission strengths of A, B, C, and D, respectively). Take the entity feature topology of mobile phone assembly plant C (including five feature units: "inventory level", "number of backup suppliers", "production elasticity", "financial liquidity", and "geographic location", and there are conduction edges such as "number of backup suppliers-inventory level" and "inventory level-production elasticity") as an example: the server call model screens the key conduction edges of C's entity feature topology.The model calculates the "gradient importance score" of the edges (such as the score of the "number of backup suppliers-inventory level" edge is 0.85, and the score of the "geographic location-production elasticity" edge is 0.3), retains the edges with scores ≥ 0.5, and generates the first core feature interaction topology (which only includes "number of backup suppliers-inventory level", "inventory level-production elasticity", and "financial liquidity-number of backup suppliers"). Based on the core topology, the model calculates the "unit characteristics" of each feature unit through GNN neighborhood aggregation: "Number of backup suppliers" unit: Based on C's supplier management system data (actually, there are three backup chip suppliers), the unit characteristic value is 3; "Inventory level" unit: Based on ERP system inventory records (current chip inventory can sustain 8 days of production), the unit characteristic value is 0.8 (normalized to the maximum inventory days of 10 days); "Production flexibility" unit: Based on historical production line switching records (switching to backup chip models takes 48 hours), the unit characteristic value is 48; "Financial liquidity" unit: Based on financial statements (cash / debt ratio is 2.5), the unit characteristic value is 2.5; "Geographic location" unit: Because C is located in a multi-port city (marked as 1), the unit characteristic value is 1. Finally, these unit characteristics are spliced ​​into C's "entity multidimensional capability representation vector" [3, 0.8, 48, 2.5, 1], which comprehensively reflects C's actual capabilities in inventory, supplier management, production flexibility, and other dimensions. Taking the risk attribute topology of risk event F1 (a chip production line fire) as an example (containing five feature units: "probability of occurrence," "scope of impact," "duration," "severity," and "predictability," with conductive edges between "duration-scope of impact" and "scope of impact-severity"), the server invokes the model to screen key conductive edges in F1's risk attribute topology. The model calculates the "gradient importance score" of each edge (e.g., a score of 0.9 for the "duration-scope of impact" edge and 0.4 for the "probability of occurrence-predictability" edge), retains edges with a score ≥ 0.5, and generates a second core feature interaction topology (containing only the "duration-scope of impact" and "scope of impact-severity" edges). Based on the core topology, the model calculates the "unit characteristics" of each feature unit through GNN neighborhood aggregation: "Occurrence probability" unit: based on historical data (the annual occurrence frequency of similar fires in the past five years is 5%), the unit characteristic value is 0.05; "Impact range" unit: based on the accident report (the fire affects 5 chip models), the unit characteristic value is 5; "Duration" unit: based on the recovery record (production resumed 72 hours after the fire), the unit characteristic value is 72; "Severity" unit: based on downstream loss statistics (causing C to stop production and suffer direct losses of 50 million yuan), the unit characteristic value is 5000; "Predictability" unit: because A has no fire warning system (marked as 0), the unit characteristic value is 0.Finally, these unit features are spliced ​​into the “risk multidimensional attribute representation vector” [0.05, 5, 72, 5000, 0] of F1, which comprehensively reflects the attribute characteristics of F1 in dimensions such as probability and impact range. The server inputs the entity basic risk representation vector (such as C's [0.72, 0.2, 0.1, 0.3]), the entity multi-dimensional capability representation vector (such as C's [3, 0.8, 48, 2.5, 1]), the risk basic transmission representation vector (such as F1's [0.85, 0.1, 0.72, 0.1]), and the risk multi-dimensional attribute representation vector (such as F1's [0.05, 5, 72, 5000, 0]) into the model and trains it through the multi-task loss function: Synergy deviation: calculate the cosine similarity between the entity basic risk vector and the capability vector (such as the similarity between C's basic vector and the capability vector is 0.6, the target is 1, and the deviation is 0.4), as well as the similarity between the risk basic transmission vector and the attribute vector (such as the similarity between F1's basic vector and the attribute vector is 0.5, and the deviation is 0.5), and the weighted summation is used to obtain the synergy deviation (0.4×0.5+0.5×0.5=0.45). Transmission Assessment Bias: Based on the basis vectors, the model predicts C's resilience coefficient for F1 as 0.6 (the actual historical value is 0.7, with a bias of 0.1). Based on the capability and attribute vectors, the coefficient is predicted to be 0.65 (the actual value is 0.7, with a bias of 0.05). The weighted summation yields the transmission bias (0.1 × 0.6 + 0.05 × 0.4 = 0.08). Complexity Constraint: An L2 regularization term is added (sum of squared parameters × 0.01, with a value of 0.02). The total assessment bias is 0.45 + 0.08 + 0.02 = 0.55. The server adjusts model parameters (such as the neighborhood aggregation weights of the GNN and the parameters of the attention head) through backpropagation. After multiple rounds of iteration (e.g., 100 rounds), the bias gradually decreases to below 0.1, the model converges, and the trained resilience assessment model is obtained. This model accurately outputs the resilience coefficient of supply chain entities to risk events, providing a quantitative basis for supply chain resilience management.

[0036] In an embodiment of the present invention, the calling of the resilience assessment model to obtain the entity-based risk characterization vector of each supply chain entity and the risk-based conduction characterization vector of each risk event based on the risk conduction topology network can be implemented through the following examples.

[0037] The risk transmission topology network is characterized as a risk transmission association matrix; the risk transmission association matrix is ​​used to indicate the risk transmission edge relationship between the risk event unit and the supply chain entity unit in the risk transmission topology network;

[0038] The resilience assessment model is called to obtain a risk information transmission matrix, and based on the risk information transmission matrix and the risk conduction association matrix, two-way information transmission is performed on the supply chain entity characteristics of the multiple supply chain entities and the risk event characteristics of the multiple risk events to obtain a risk conduction comprehensive representation matrix corresponding to the multiple supply chain entities and the multiple risk events;

[0039] Based on the risk transmission comprehensive characterization matrix, the entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event are obtained.

[0040] In an embodiment of the present invention, exemplarily, in a consumer electronics supply chain scenario, the server has constructed a risk conduction topology network comprising five supply chain entities (chip supplier A, battery supplier B, mobile phone assembly plant C, logistics provider D, retailer E) and four risk events (chip production line fire F1, battery raw material price increase F2, port strike F3, and demand drop F4). The following specifically describes how the server obtains entity-based risk characterization vectors and risk-based conduction characterization vectors based on the network: the server first converts the graph structure of the topological network into a numerical risk conduction association matrix. The row index of the matrix corresponds to the supply chain entities (A, B, C, D, E), the column index corresponds to the risk events (F1, F2, F3, F4), and the matrix element value is 1 or 0, respectively indicating whether there is a risk conduction edge between the corresponding entity and the risk event (that is, whether there is an actual impact association). In practice, the server analyzes historical data (e.g., supply chain disruption records from the past three years) and industry knowledge graphs (e.g., expert-annotated risk impact relationships) to determine the transmission relationship: A is directly related to F1 (a chip fire) (the fire caused A to be unable to supply), so the value in the F1 column of row A is 1; C is indirectly related to F1 (a supply disruption caused C to shut down), so the value in the F1 column of row C is 1; D is directly related to F3 (a port strike) (the strike caused D's shipment delays), so the value in the F3 column of row D is 1; E is directly related to F4 (a sudden drop in demand) (the drop in demand caused E's inventory backlog), so the value in the F4 column of row E is 1. Other entities and risk events without direct transmission relationships (e.g., B and F1, C and F3) have corresponding values ​​of 0. This ultimately generates a 5×4 risk transmission association matrix M, where, for example, M[A][F1] = 1, M[C][F1] = 1, M[D][F3] = 1, M[E][F4] = 1, and all other elements are 0. The server invokes a resilience assessment model initialized as a graph neural network (GNN). The model includes a learnable "risk information transfer matrix" W (with dimensions matching M and initially random parameters). Through the interaction between W and M, the server enables bidirectional information transfer between entities and risk events: Risk events transfer information to entities: The model transmits the initial characteristics of each risk event (such as F1's "historical probability of occurrence 5%" and "average duration 72 hours") to associated entities via conductive edges. For example, F1 transmits its "long duration" characteristic to A (the direct supplier) and C (the downstream assembly plant) through the value of 1 in row A, column F1 of M. This updates A's initial risk perception value to 0.8 (indicating a significant impact from F1), and C's initial risk perception value to 0.6 (due to indirect influence). Entities transfer information to risk events: The model transmits the initial characteristics of each entity (such as A's "number of backup production lines 2" and "high-level fire protection facilities") to associated risk events via conductive edges.For example, A transfers its "large number of backup production lines" feature to F1 through the value of 1 in row A, column F1 in M, updating F1's initial transmission strength to 0.7 (indicating that the actual impact on A may be weakened); C transfers its "8-day safety stock" feature to F1, updating F1's transmission strength to C to 0.5 (indicating that the impact on C may be delayed). The server iterates through three rounds of transmission (updating the entity and risk event features in each round) to generate three comprehensive risk transmission representation matrices (one for each round). For example, in the first round matrix, the A-F1 value is 0.75 (A's risk perception after risk transmission), and the C-F1 value is 0.55 (C's risk perception after risk transmission); in the second round matrix, the A-F1 value is 0.7 (the transmission strength of F1 to A after entity transmission), and the C-F1 value is 0.6 (the transmission strength of F1 to C after entity transmission); in the third round matrix, the A-F1 value is 0.68 (the final impact strength combined from the first two rounds), and the C-F1 value is 0.62. The server applies a multi-head attention mechanism to the three risk transmission comprehensive representation matrices, assigning higher weights to key rounds (e.g., round 3 focuses on long-term impacts, with a weight of 0.6; round 2 focuses on indirect impacts, with a weight of 0.3; and round 1 has a weight of 0.1), generating a fused risk transmission comprehensive representation matrix. For example, the fused A-F1 value is 0.68 × 0.6 + 0.7 × 0.3 + 0.75 × 0.1 = 0.69, and the C-F1 value is 0.62 × 0.6 + 0.6 × 0.3 + 0.55 × 0.1 = 0.61. The server then performs a dimensionality reduction transformation on the fused matrix using a fully connected layer (e.g., converting a 5 × 4 matrix into 5 × 2 and 4 × 2 lower-dimensional matrices), resulting in the target risk transmission comprehensive representation matrix. Finally, the "entity-based risk representation vector" for each entity (e.g., C's vector is [0.61, 0.1, 0.2, 0.3], corresponding to the basic risk intensities of F1-F4, respectively) and the "risk-based conduction representation vector" for each risk event (e.g., F1's vector is [0.69, 0.1, 0.61, 0.1], corresponding to the basic conduction intensities of A, B, C, and D, respectively) are extracted from this matrix. At this point, the server completes the conversion from the risk conduction topology network to the entity-based risk representation vector and the risk-based conduction representation vector, providing key input for subsequent model training.

[0041] In an embodiment of the present invention, there are multiple risk conduction comprehensive characterization matrices; obtaining the entity-based risk characterization vector of each supply chain entity and the risk-based conduction characterization vector of each risk event based on the risk conduction comprehensive characterization matrix can be implemented through the following examples.

[0042] Perform multi-dimensional feature fusion on multiple risk conduction comprehensive representation matrices based on the attention mechanism to obtain a fused risk conduction comprehensive representation matrix;

[0043] Performing feature conversion on the fused risk conduction comprehensive representation matrix to obtain a target risk conduction comprehensive representation matrix;

[0044] The entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event are extracted from the target risk transmission comprehensive characterization matrix.

[0045] In an exemplary embodiment of the present invention, in a consumer electronics supply chain scenario, the server has generated three risk transmission comprehensive representation matrices (denoted as M1, M2, and M3) through three rounds of bidirectional information transmission, corresponding to the results of information transmission rounds 1 through 3. The following uses the association between chip supplier A, mobile phone assembly plant C, and risk event F1 (a chip production line fire) as an example to explain in detail how the server derives entity-based risk representation vectors and risk-based transmission representation vectors through multi-matrix fusion, transformation, and extraction. The server first assigns attention weights to the three risk transmission comprehensive representation matrices to distinguish the importance of information from different rounds. The core of the attention mechanism is to calculate the contribution of each round's matrix to the final representation, specifically achieved through the "attention score" learned by the model. M1 (Round 1): focuses on the direct impact of the entity and the risk event (e.g., if A is directly affected by the F1 fire, M1[A][F1] = 0.75; if C is indirectly affected by A's supply disruption, M1[C][F1] = 0.55). M2 (Round 2): Focuses on the spread of indirect impacts (e.g., after F1 causes A to be out of supply, C further affects retailer E due to insufficient inventory, M2[C][F1]=0.6; M2[E][F1]=0.3). M3 (Round 3): Focuses on the stable value of long-term impacts (e.g., C alleviates the impact of F1 by activating backup suppliers, M3[C][F1]=0.62; A resumes production through backup production lines, M3[A][F1]=0.68). The server calculates the weight of each round of the matrix through a multi-head attention mechanism (e.g., 2 attention heads): the first head focuses on short-term direct impacts and assigns a weight of 0.2 to M1 (because direct impacts can be easily alleviated by subsequent measures); the second head focuses on long-term stable impacts and assigns a weight of 0.6 to M3 (because long-term impacts can better reflect actual resilience); M2 is the intermediate round and is assigned a weight of 0.2. Finally, the integrated risk conduction representation matrix M fused The element value of is the weighted sum of the corresponding elements of each round matrix. For example: M fused [A][F1]=M1[A][F1]×0.2+M2[A][F1]×0.2+M3[A][F1]×0.6=0.75×0.2+0.7×0.2+0.68×0.6=0.69; M fused[C][F1]=M1[C][F1]×0.2+M2[C][F1]×0.2+M3[C][F1]×0.6=0.55×0.2+0.6×0.2+0.62×0.6=0.61. To reduce the dimension and extract more abstract features, the server performs fused (5×4 matrix, 5 entities × 4 risk events) for feature conversion. The conversion is achieved through a fully connected layer (such as input dimension 20, output dimension 10), flattening the two-dimensional matrix into a one-dimensional vector, and then generating a low-dimensional, high-abstraction target matrix M through linear transformation and activation function (such as ReLU) target . M fused Flattened into a vector of length 20 ([0.69, 0, 0, 0, 0, 0, 0, 0.61, 0, 0, 0, 0, 0.5, 0, 0, 0, 0, 0.4], only key elements are shown); the fully connected layer parameter matrix W (20×10) is initialized by model learning, and the flattened vector is linearly transformed: M flat ×W+b (b is the bias term); apply the ReLU activation function to filter negative values ​​and obtain a vector of length 10; reshape it into a 5×2 matrix (M target ), where each row corresponds to an entity and each column corresponds to the risk feature dimension after abstraction. For example, M target The row vector of A is [0.58, 0.12] (representing the risk characteristics of A in abstract dimension 1 and dimension 2 respectively), and the row vector of C is [0.49, 0.11]. target The “entity-based risk representation vector” of each supply chain entity (each row corresponds to a vector of an entity), and the “risk-based transmission representation vector” of each risk event (each column corresponds to a vector of a risk event, which needs to be extracted after transposing the matrix) are directly extracted. The basic risk representation vector of entity A is M target The row vector of A is [0.58, 0.12], which represents the comprehensive intensity of the impact of various risk events on A in the abstract dimension; the basic risk representation vector of entity C is [0.49, 0.11], which reflects the risk perception of C in the abstract dimension; the basic conduction representation vector of risk event F1 is M target The column vector corresponding to F1 after transposition, [0.58, 0, 0.49, 0, 0] (only key elements are exemplified), represents the transmission strength of F1 to each entity. Through the above steps, the server completes the conversion from the multi-risk transmission comprehensive representation matrix to entity-based risk representation vectors and risk-based transmission representation vectors. These vectors integrate the key features of multiple rounds of information transmission and provide the core input for the subsequent training of the resilience assessment model by combining entity capabilities and risk attributes.

[0046] In an embodiment of the present invention, any one of the multiple supply chain entities is characterized as a target supply chain entity, and a risk transmission edge exists between any two supply chain entity feature units in the entity feature topology of the target supply chain entity;

[0047] The calling of the resilience assessment model to obtain the entity multi-dimensional capability representation vector of each supply chain entity based on the entity feature topology of each supply chain entity can be implemented through the following example.

[0048] Calling the resilience assessment model to perform key transmission edge screening on the risk transmission edges in the entity feature topology of the target supply chain entity to obtain a first core feature interaction topology of the entity feature topology of the target supply chain entity;

[0049] Based on the first core feature interaction topology, risk information is transmitted for the supply chain entity features of the target supply chain entity on multiple evaluation indicators to obtain unit features corresponding to each supply chain entity feature unit of the target supply chain entity in the first core feature interaction topology;

[0050] An entity multi-dimensional capability representation vector of the target supply chain entity is obtained according to the unit features corresponding to each supply chain entity feature unit of the target supply chain entity.

[0051] In an embodiment of the present invention, illustratively, taking a mobile phone assembly plant C (target supply chain entity) in a consumer electronics supply chain as an example, its entity feature topology includes five supply chain entity feature units: "inventory level" (chip / battery reserve), "number of backup suppliers" (number of alternative chip suppliers), "production flexibility" (time for production line switching - backup models), "financial liquidity" (cash / liability ratio), "geographic location" (whether it is located in a multi-port city), and there are risk transmission edges between units (such as "number of backup suppliers" is connected to "inventory level", and "inventory level" is connected to "production flexibility"). The following details how the server generates a multi-dimensional capability representation vector for C's entity based on this topology: the server calls a resilience assessment model (graph neural network GNN) to evaluate the importance of the risk transmission edges in C's entity feature topology and screen key edges. This is achieved specifically by calculating the "gradient importance score" of each edge (the degree of influence of the edge on the final resilience assessment result during model training): Data input: The server inputs the entity feature topology of C into the model, including the initial value of each feature unit (such as the initial value of "number of backup suppliers" is 3, the initial value of "inventory level" is 0.8 (normalized based on the maximum inventory of 10 days)) and the connection relationship of the edges (such as "number of backup suppliers-inventory level" and "inventory level-production elasticity"). Importance calculation: The model calculates the absolute value of the gradient of each edge to the loss function through backpropagation (the larger the gradient, the more critical the impact of the edge on the result). For example, the gradient of the edge between "number of backup suppliers and inventory level" is 0.85 (indicating that adding backup suppliers can significantly improve inventory stability); the gradient of the edge between "inventory level and production elasticity" is 0.72 (indicating that sufficient inventory can shorten production line switching time); the gradient of the edge between "financial liquidity and number of backup suppliers" is 0.65 (indicating that sufficient funds can support the cooperation of more backup suppliers); and the gradient of the edge between "geographic location and production elasticity" is 0.3 (indicating that cities with multiple ports have a smaller impact on production elasticity because backup suppliers already cover transportation demand). The server retains edges with a gradient score ≥ 0.5 and generates a "first core feature interaction topology," consisting of three key edges: "number of backup suppliers and inventory level," "inventory level and production elasticity," and "financial liquidity and number of backup suppliers." The edge between "geographic location and production elasticity" (scoring 0.3 < 0.5) is removed. Based on the first core feature interaction topology, the server invokes the model to perform risk information transfer (GNN neighborhood aggregation), calculating the "unit feature" for each feature unit (reflecting the actual impact of the feature in the key transmission relationship): First round of transfer: Information is transferred from "financial liquidity" to "number of backup suppliers."The initial value of financial liquidity is 2.5 (cash / debt ratio). This information is transferred via the edge "Financial Liquidity - Number of Backup Suppliers," updating the unit feature of "Number of Backup Suppliers" to 3×(1+2.5×0.1)=3.75 (0.1 is the transfer coefficient learned by the model, indicating that for every increase in financial liquidity, the number of backup suppliers increases by 10%). The second round of transfer: Information is transferred from "Number of Backup Suppliers" to "Inventory Level." The updated "Number of Backup Suppliers" is 3.75. This information is transferred via the edge "Number of Backup Suppliers - Inventory Level," updating the unit feature of "Inventory Level" to 0.8×(1+3.75×0.05)=0.9875 (0.05 is the transfer coefficient, indicating that for every additional backup supplier, the inventory level increases by 5%). The third round of transfer: Information is transferred from "Inventory Level" to "Production Elasticity." The updated "Inventory Level" is 0.9875. This is transferred through the edge "Inventory Level - Production Elasticity," updating the unit feature of "Production Elasticity" to 48 × (1 - 0.9875 × 0.1) = 47.52 hours (0.1 is the transfer coefficient, meaning that for every 10% increase in inventory level, the production elasticity time decreases by 10%). The server concatenates the final unit feature values ​​of each feature unit in a fixed order to obtain C's "Entity Multidimensional Capability Representation Vector." The specific values ​​are derived as follows: "Number of backup suppliers": the value after information transfer is 3.75 (the actual number of backup suppliers is 3, adjusted to 3.75 after financial liquidity enhancement); "Inventory level": the value after information transfer is 0.9875 (the actual inventory can last for 8 days, adjusted to 9.875 days after backup supplier enhancement, and normalized to 0.9875); "Production flexibility": the value after information transfer is 47.52 hours (the actual switching time is 48 hours, shortened to 47.52 hours after inventory adequacy optimization); "Financial liquidity": the original value is 2.5 (taken directly from financial statements); "Geographic location": the original value is 1 (C is located in a city with multiple ports, marked as 1). Ultimately, the entity multidimensional capability representation vector of C is [3.75, 0.9875, 47.52, 2.5, 1]. This vector comprehensively reflects C's actual risk tolerance in dimensions such as supplier management, inventory, and production flexibility, providing a key basis for subsequent assessments of its resilience to risk events.

[0052] In an embodiment of the present invention, any one of the multiple risk events is characterized as a target risk event, and a risk transmission edge exists between any two risk event feature units in the risk attribute topology of the target risk event;

[0053] The calling of the resilience assessment model to obtain the risk multi-dimensional attribute representation vector of each risk event based on the risk attribute topology of each risk event can be implemented through the following example.

[0054] Calling the resilience assessment model to perform key transmission edge screening on the risk transmission edges in the risk attribute topology of the target risk event to obtain a second core feature interaction topology of the risk attribute topology of the target risk event;

[0055] Based on the second core feature interaction topology, risk information is transmitted for the risk event features of the target risk event on multiple evaluation indicators to obtain unit features corresponding to each risk event feature unit of the target risk event in the second core feature interaction topology;

[0056] A risk multi-dimensional attribute representation vector of the target risk event is obtained according to the unit features corresponding to each risk event feature unit of the target risk event.

[0057] In an embodiment of the present invention, illustratively, taking the "chip production line fire" risk event F1 (target risk event) in the consumer electronics supply chain as an example, its risk attribute topology contains 5 risk event characteristic units: "Probability of occurrence" (annual frequency of similar fires in the past five years), "Scope of impact" (number of affected chip models), "Duration" (time from the occurrence of the fire to the resumption of production), "Severity" (direct economic losses from shutdown of downstream factories), "Predictability" (whether there is a fire warning system), and there are risk conduction edges between units (such as "Duration-Scope of impact" and "Scope of impact-Severity"). The following details how the server generates the risk multidimensional attribute representation vector of F1 based on this topology: the server calls the resilience assessment model (graph neural network GNN) to evaluate the importance of the risk conduction edges in the risk attribute topology of F1 and screen the key edges. This is achieved by calculating the "gradient importance score" of each edge (the degree of influence of the edge on the final resilience assessment results during model training): Data input: The server inputs the F1 risk attribute topology into the model, including the initial values ​​of each feature unit (such as the initial value of "duration" is 72 hours, and the initial value of "impact range" is 5 chip models) and the connection relationship of the edges (such as "duration-impact range", "impact range-severity", and "probability of occurrence-predictability"). Importance calculation: The model calculates the absolute value of the gradient of each edge with respect to the loss function through backpropagation (the larger the gradient, the more critical the edge's impact on the result). For example, the gradient of the "Duration - Impact Range" edge is 0.9 (indicating that the longer the fire lasts, the more chip models are affected); the gradient of the "Impact Range - Severity" edge is 0.85 (indicating that the more chip models affected, the greater the downstream losses); the gradient of the "Probability of Occurrence - Predictability" edge is 0.3 (indicating that the correlation between fire probability and the early warning system is weak, as historical probabilities have been determined through statistical data); and the gradient of the "Predictability - Occurrence Probability" edge is 0.4 (indicating that the lack of an early warning system has a limited effect on improving fire probability). The server retains edges with a gradient score ≥ 0.5 and generates a "second core feature interaction topology," consisting of the two key edges "Duration - Impact Range" and "Impact Range - Severity." The edges "Probability of Occurrence - Predictability" and "Predictability - Occurrence Probability" (both with scores < 0.5) are discarded. Based on the second core feature interaction topology, the server calls the model to perform risk information transmission (GNN neighborhood aggregation operation) and calculate the "unit feature" of each feature unit (reflecting the actual impact value of the feature in the key transmission relationship): First round of transmission: Information is transmitted from "duration" to "influence range". The initial value of duration is 72 hours. Through the edge "duration-influence range", the unit feature of "influence range" is updated to 5×(1+72×0.01)=12.2 (0.01 is the transmission coefficient learned by the model, indicating that the influence range increases by 1% for every hour of extension).The second round of transfer: Information is transferred from "Affected Scope" to "Severity." The updated "Affected Scope" is 12.2 chip models. Through the edge "Affected Scope - Severity," the unit feature of "Severity" is updated to 5000 × (1 + 12.2 × 0.1) = 111 million yuan (0.1 is the transfer coefficient, indicating that the loss increases by 10% for each additional affected model). The server concatenates the final unit feature values ​​of each feature unit in a fixed order to obtain the "risk multidimensional attribute representation vector" of F1. The specific values ​​are derived as follows: "Probability of Occurrence": Original value: 5% (directly derived from historical fire frequency statistics); "Scope of Impact": After information transfer, the value is 12.2 (actually affecting 5 models, adjusted to 12.2 after the duration is extended); "Duration": Original value: 72 hours (directly derived from fire recovery records); "Severity": After information transfer, the value is 111 million yuan (actual losses: 50 million yuan, adjusted to 111 million yuan after the scope of impact is expanded); "Predictability": Original value: 0 (F1 has no fire warning system, marked as 0). Ultimately, the multidimensional risk attribute representation vector for F1 is [0.05, 12.2, 72, 11100, 0]. This vector comprehensively reflects the actual risk attributes of F1 in terms of probability, scope of impact, and loss severity, providing a core basis for subsequent assessments of supply chain entities' resilience to F1.

[0058] In an embodiment of the present invention, the resilience assessment model is trained based on the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based conduction representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event to obtain a trained resilience assessment model, which can be implemented through the following examples.

[0059] Obtaining an assessment deviation value of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity-multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk-multidimensional attribute representation vector of each risk event;

[0060] The model parameters of the toughness assessment model are optimized and adjusted based on the assessment deviation value to obtain the toughness assessment model that has completed training.

[0061] In an embodiment of the present invention, for example, in a consumer electronics supply chain scenario, the server has obtained the entity-based risk representation vector ([0.61, 0.1, 0.2, 0.3]) and entity-based multidimensional capability representation vector ([3.75, 0.9875, 47.52, 2.5, 1]) of mobile phone assembly plant C, as well as the risk-based transmission representation vector ([0.69, 0.1, 0.61, 0.1]) and risk-based multidimensional attribute representation vector ([0.05, 12.2, 72, 11100, 0]) of risk event F1 (chip fire). The following details how the server calculates the evaluation deviation value and optimizes the model parameters based on these vectors: The server calculates the evaluation deviation value using a multi-task loss function, which includes three parts: coordination deviation, transmission evaluation deviation, and complexity constraints. The server must ensure that the entity's "basic risk representation" and "multidimensional capability" and the risk's "basic transmission representation" and "multidimensional attributes" are consistent (i.e., highly matched) in the feature space. Specifically, it is achieved by calculating the inverse of the cosine similarity (mismatch): Entity synergy deviation: The entity basic risk representation vector of C ([0.61, 0.1, 0.2, 0.3]) and the entity multidimensional capability representation vector ([3.75, 0.9875, 47.52, 2.5, 1]) are reduced to 2 dimensions (such as [0.5, 0.3] and [0.6, 0.4]) through the fully connected layer, and the calculated cosine similarity is 0.98 (close to 1 indicates a match), and the mismatch is 1-0.98=0.02. Risk Synergy Deviation: The risk-based transmission representation vector of F1 ([0.69, 0.1, 0.61, 0.1]) and the risk multi-dimensional attribute representation vector ([0.05, 12.2, 72, 11100, 0]) were reduced to two dimensions (e.g., [0.7, 0.2] and [0.75, 0.25]). The cosine similarity was 0.95, and the mismatch was 1-0.95 = 0.05. Total Synergy Deviation: The weighted average of the two (each with a weight of 0.5) was 0.02 × 0.5 + 0.05 × 0.5 = 0.035. The server calculates the deviations based on the difference between the model's output of the resilience coefficient and its historical actual value: First transmission deviation: Based on the entity's underlying risk vector and the risk-based transmission vector, the model predicts C's resilience coefficient for F1 to be 0.6 (the historical actual value was 0.7, as C had previously quickly resumed production through a backup supplier). The deviation is |0.6 - 0.7| = 0.1. Second transmission deviation: Based on the entity's multidimensional capability vector and the risk multidimensional attribute vector, the model predicts C's resilience coefficient for F1 to be 0.65 (the actual value was 0.7). The deviation is |0.65 - 0.7| = 0.05. Total transmission assessment deviation: Take the weighted average of the two (weight 0.6 for the first deviation and 0.4 for the second deviation), that is, 0.1 × 0.6 + 0.05 × 0.4 = 0.08.To prevent the model from overfitting the training data, the server adds an L2 regularization term, calculates the sum of the squares of the model parameters (such as the neighborhood aggregation weights of the GNN and the attention head parameters), and multiplies it by the regularization coefficient 0.01. Assuming that the sum of the squares of the parameters is 2, the complexity constraint value is 2×0.01=0.02. Adding the three parts of the deviation, the total evaluation deviation value is 0.035 (synergy) + 0.08 (conduction) + 0.02 (complexity) = 0.135. The server uses the total evaluation deviation value as the objective function and adjusts the model parameters (such as the neighborhood aggregation weight matrix of the GNN and the linear transformation parameters of the attention head) through the backpropagation algorithm. The specific steps are as follows: Gradient calculation: Calculate the gradient of the total deviation with respect to each model parameter (such as the gradient of the weight matrix W of the GNN layer is. deviation / ). For example, if the weight gradient of the "number of backup suppliers - inventory level" edge is 0.05 (indicating that increasing the weight of this edge can reduce the deviation), then the weight needs to be increased. Parameter update: Use an optimizer (such as Adam) to update the parameters according to the gradient, and the learning rate is set to 0.001. For example, the original value of the weight parameter of a GNN layer is 0.8, the gradient is -0.03, and after the update it is 0.8-0.001×(-0.03)=0.80003 (gradient descent). Iterative convergence: Repeatedly calculate the deviation value and update the parameters until the deviation value stabilizes below the threshold (such as 0.05). After about 200 rounds of iterations, the server observed that the deviation value dropped to 0.045, and the model converged. Finally, the server obtained a resilience assessment model that completed training. The model can accurately output the risk resilience coefficient of the supply chain entity for risk events (for example, the prediction coefficient of C for F1 is increased from 0.6 to 0.68, close to the actual value of 0.7), providing a reliable quantitative tool for supply chain resilience management.

[0062] In an embodiment of the present invention, the assessment deviation value of the resilience assessment model is obtained based on the entity-based risk characterization vector of each supply chain entity, the entity multidimensional capability characterization vector of each supply chain entity, the risk-based transmission characterization vector of each risk event, and the risk multidimensional attribute characterization vector of each risk event. This can be implemented through the following example.

[0063] According to the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event, a characteristic synergy deviation value of the resilience assessment model is obtained; the characteristic synergy deviation value is used to indicate the representation mismatch between the entity-based risk representation vector and the entity multidimensional capability representation vector of each supply chain entity, and is used to indicate the representation mismatch between the risk-based transmission representation vector and the risk multidimensional attribute representation vector of each risk event;

[0064] Obtaining a first risk transmission assessment deviation value of the resilience assessment model according to the entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event;

[0065] Obtaining a second risk conduction assessment deviation value of the resilience assessment model based on the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event;

[0066] Obtaining a complexity constraint value of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity-multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk-multidimensional attribute representation vector of each risk event;

[0067] The evaluation deviation value is determined based on the feature synergy deviation value, the first risk conduction evaluation deviation value, the second risk conduction evaluation deviation value and the complexity constraint value.

[0068] In an embodiment of the present invention, for example, in a consumer electronics supply chain scenario, the server has obtained the entity basic risk characterization vector of mobile phone assembly plant C (reflecting the basic exposure of C to historical risks), the entity multidimensional capability characterization vector (reflecting C's actual risk resistance in inventory, supplier management, etc.), and the risk basic conduction characterization vector of risk event F1 (fire on chip production line) (reflecting the basic impact intensity of F1 on each entity), and the risk multidimensional attribute characterization vector (reflecting the actual risk attributes of F1 in terms of probability, impact range, etc.). The following details how the server calculates the assessment deviation value based on these four types of vectors to optimize the resilience assessment model: The feature synergy deviation value is used to measure the degree of matching between the entity's "basic risk" and "multidimensional capability", and the risk's "basic conduction" and "multidimensional attribute" in the feature space. The higher the matching degree, the smaller the deviation; the lower the matching degree, the larger the deviation. For mobile phone assembly plant C: The server projects C's entity-level risk representation vector (e.g., [0.61, 0.1, 0.2, 0.3], representing C's underlying exposure to risks like F1) and its entity-level multidimensional capability representation vector (e.g., [3.75, 0.9875, 47.52, 2.5, 1], representing C's capabilities, such as the number of backup suppliers and inventory levels) onto the same dimensional space (e.g., 2D) using the model's "feature alignment layer." For example, the projection of the underlying risk vector yields [0.5, 0.3], while the projection of the multidimensional capability vector yields [0.6, 0.4]. The server calculates the "directional consistency" of these two projected vectors. If the two vectors are close in direction (e.g., pointing in nearly the same direction), this indicates that C's underlying risk exposure matches its actual capabilities (e.g., high underlying risk but strong capabilities, effectively mitigating risk), indicating a low synergy bias. If the directions differ significantly (e.g., high underlying risk but weak capabilities), the bias is high. Calculations show that the directions of C's basic risk and capability projection vectors are highly consistent, with a synergy deviation of 0.01 (smaller deviations are better). For risk event F1, the server projects F1's basic risk transmission vector (e.g., [0.69, 0.1, 0.61, 0.1], representing the basic impact strength of F1 on entities A and C) and its multi-dimensional risk attribute vector (e.g., [0.05, 12.2, 72, 11100, 0], representing F1's probability of occurrence and impact range) into a two-dimensional space (e.g., [0.7, 0.2] and [0.75, 0.25]). Calculating the directional consistency of these two vectors shows a high match between F1's basic transmission and attributes (e.g., strong basic transmission and a large impact range), with a synergy deviation of 0.01. Finally, the server averages the synergy deviations of the entities and risks, resulting in a total feature synergy deviation of 0.01. The first risk transmission assessment deviation value reflects the difference between the risk resilience coefficient predicted by the model based on the entity-based risk representation and the risk-based transmission representation and the actual historical value.The server invokes the model's "Basic Conduction Prediction Module," inputting C's entity-based risk representation vector ([0.61, 0.1, 0.2, 0.3]) and F1's risk-based transmission representation vector ([0.69, 0.1, 0.61, 0.1]). It outputs C's first-assessed risk resilience coefficient for F1 (e.g., 0.6). Simultaneously, the server retrieves C's risk resilience coefficient from the historical database when F1 actually occurred (e.g., 0.7, based on C's record of quickly resuming production through a backup supplier). If the difference between the model's predicted value (0.6) and the actual value (0.7) is small, the bias is small; a large difference indicates a large bias. Calculated, the difference is 0.1 (the predicted value is 0.1 lower than the actual value), resulting in a first-assessment risk resilience coefficient bias of 0.01 (calculated using squared difference: smaller differences indicate smaller biases). The second-assessment risk resilience coefficient bias reflects the difference between the model's predicted risk resilience coefficient, based on the entity's multidimensional capability representation and the risk's multidimensional attribute representation, and the historical actual value. The server invokes the model's "Capability Attribute Prediction Module," inputting C's entity multidimensional capability representation vector ([3.75, 0.9875, 47.52, 2.5, 1]) and F1's risk multidimensional attribute representation vector ([0.05, 12.2, 72, 11100, 0]). It outputs C's second-assessed risk resilience coefficient for F1 (e.g., 0.65). Similarly, when comparing the predicted value (0.65) to the actual historical value (0.7), if the difference between the predicted value and the actual value is small (only 0.05 lower), the bias is small. The calculated difference is 0.05, resulting in a bias of 0.0025 for the second risk transmission assessment (calculated using the squared difference). Complexity constraints limit the complexity of model parameters, preventing the model from over-reliance on noise in the training data, which could reduce its predictive power for new data. The server statistically measures the size of all learnable parameters in the model (such as the weights and biases of each neural network layer). Larger parameters indicate higher model complexity and a greater tendency to overfit. For example, if the weight parameters of a layer in the model are [0.8, 0.3, -0.5], its "size" can be calculated by the sum of squares (0.8. 2 +0.3 2 +(-0.5) 2=0.64+0.09+0.25=0.98). The server adds the squares of all parameters (assuming the total is 10) and multiplies it by a small coefficient (such as 0.001) to obtain the complexity constraint value (10×0.001=0.01). The larger the value, the more complex the model is, and the parameters need to be adjusted to make it smaller. The server weights the feature synergy deviation value, the first risk conduction assessment deviation value, the second risk conduction assessment deviation value and the complexity constraint value according to certain weights (such as synergy accounts for 30%, first conduction accounts for 40%, second conduction accounts for 20%, and complexity accounts for 10%) and sums them to obtain the total assessment deviation value. Feature synergy deviation value: 0.01 (30%: 0.01 × 0.3 = 0.003); first risk transmission assessment deviation value: 0.01 (40%: 0.01 × 0.4 = 0.004); second risk transmission assessment deviation value: 0.0025 (20%: 0.0025 × 0.2 = 0.0005); complexity constraint value: 0.01 (10%: 0.01 × 0.1 = 0.001). The total assessment deviation value is 0.003 + 0.004 + 0.0005 + 0.001 = 0.0085. A smaller value indicates more accurate model predictions, more reasonable feature matching, and lower parameter complexity. Through the above steps, the server completes the calculation of the assessment deviation values ​​based on the four types of vectors. The total deviation value of 0.0085 indicates that the model's current predictions are highly consistent with the actual values, with good feature synergy and controllable parameter complexity. The subsequent server will reduce the deviation value by adjusting the model parameters (such as optimizing the weight of the "basic conduction prediction module") until the model converges, and finally obtain a completed training model that can accurately evaluate the resilience of the supply chain.

[0069] In an embodiment of the present invention, the characteristic synergy deviation value of the resilience assessment model is obtained based on the entity-based risk characterization vector of each supply chain entity, the entity multidimensional capability characterization vector of each supply chain entity, the risk-based transmission characterization vector of each risk event, and the risk multidimensional attribute characterization vector of each risk event. This can be implemented through the following examples.

[0070] Obtaining a first synergy deviation value for the supply chain entity representation vector according to the entity basic risk representation vector and the entity multidimensional capability representation vector of each supply chain entity;

[0071] Obtaining a second synergy deviation value for the embedded feature of the risk event based on the risk basic transmission representation vector and the risk multidimensional attribute representation vector of each risk event;

[0072] The characteristic synergy deviation value is obtained according to the first synergy deviation value and the second synergy deviation value.

[0073] In an exemplary embodiment of the present invention, in a consumer electronics supply chain scenario, the server has obtained the entity's basic risk representation vector (reflecting C's basic exposure to historical risks), the entity's multidimensional capability representation vector (reflecting C's actual risk tolerance), as well as the risk-based transmission representation vector (reflecting F1's basic impact on the entity), and the risk multidimensional attribute representation vector (reflecting F1's actual risk attributes) for mobile phone assembly plant C. The following details how the server calculates the feature synergy deviation value based on these vectors: The first synergy deviation value measures the degree of match between mobile phone assembly plant C's "basic risk representation" and "multidimensional capability representation." If C's basic risk is high (susceptible to risk) but its multidimensional capabilities are also strong (able to effectively resist risk), the match is high and the deviation is small; otherwise, the deviation is large. Basic Risk Representation Vector: C's basic risk vector is [0.61, 0.1, 0.2, 0.3], where 0.61 indicates that C has a high basic exposure to F1 (chip production line fire) (historically, it is prone to shutdowns due to supplier disruptions). Multidimensional Capability Representation Vector: C's multidimensional capability vector is [3.75, 0.9875, 47.52, 2.5, 1]. 3.75 indicates C has 3.75 backup chip suppliers (actually 3, but due to strong financial capabilities, expanded collaboration is possible), 0.9875 indicates sufficient inventory (enough to sustain 9.875 days of production), and 47.52 indicates a short production line switchover time (only 47.52 hours). The server uses the model's "feature alignment layer" to project these two vectors (basic risk is 4-dimensional, capability is 5-dimensional) into the same low-dimensional space (e.g., 2-dimensional) for direct comparison. For example, after projection, the base risk is [0.5, 0.3] (representing the degree of exposure in abstract dimensions 1 and 2), while after projection, the capability is [0.6, 0.4] (representing the strength of capability in abstract dimensions 1 and 2). The server calculates the "directional consistency" of these two projection vectors. If they point in the same direction (e.g., both pointing upward and to the right), it indicates that C's underlying risk and capabilities are well-matched (high exposure but high capabilities); if the directions differ significantly (e.g., underlying risk pointing to the right, capability pointing to the left), the match is poor. The calculated projection vectors for C are highly consistent (almost overlapping), resulting in a very small first-order synergy deviation (e.g., 0.01). The second synergy deviation measures the degree of match between the "basic transmission representation" and the "multidimensional attribute representation" of risk event F1. If F1's underlying transmission is strong (significant impact on the entity) and its multidimensional attributes indicate a serious risk (e.g., wide impact, significant losses), the match is high and the deviation is small; otherwise, the deviation is large. The underlying transmission representation vector for F1 is [0.69, 0.1, 0.61, 0.1], where 0.69 indicates a high underlying impact on chip supplier A (directly causing A's shutdown), and 0.61 indicates a high underlying impact on assembly plant C (indirectly causing C's supply disruption).Multidimensional attribute representation vector: The multidimensional attribute vector for F1 is [0.05, 12.2, 72, 11100, 0]. The value 12.2 indicates that F1 affects 12.2 chip models (actually 5, due to the prolonged fire), and the value 111 million yuan indicates significant downstream losses (actually 50 million yuan, due to the wide impact). The server also projects these two vectors (basic conduction is 4-dimensional, attribute is 5-dimensional) into a 2D space (e.g., [0.7, 0.2] and [0.75, 0.25]). The directional consistency of the projected vectors is calculated: if the two directions are similar (e.g., both pointing upward and to the right), it indicates that F1's basic conduction matches the attribute (strong conduction corresponds to high-risk attributes). If the directions differ significantly (e.g., strong basic conduction but low-risk attributes), the match is poor. The calculated direction of F1's projected vectors indicates high consistency, resulting in a very low second-order synergy deviation (e.g., 0.01). The server averages the first synergy deviation value (0.01) and the second synergy deviation value (0.01) to obtain the final feature synergy deviation value (0.01). The smaller the value, the more reasonable the feature match between the entity and the risk, and the more consistent the model's use of the features. Through the above steps, the server completes the calculation of the feature synergy deviation value. The basic risk and capability of mobile phone assembly plant C are highly matched (high exposure but high capability), and the basic conduction and attributes of risk event F1 are also highly matched (strong conduction corresponds to high risk). Therefore, the synergy deviation value is extremely small, and the model's representation of the features is consistent, providing a reliable basis for subsequent optimization.

[0074] In an embodiment of the present invention, the first risk conduction assessment deviation value of the resilience assessment model is obtained based on the entity-based risk characterization vector of each supply chain entity and the risk-based conduction characterization vector of each risk event, which can be implemented through the following example.

[0075] Obtaining a first assessed risk resilience coefficient for each supply chain entity for each risk event based on the entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event;

[0076] The first risk transmission assessment deviation value is obtained based on the first assessed risk resilience coefficient of each supply chain entity for each risk event and the transmission relationship of each supply chain entity for each risk event, where the transmission relationship is at least one of the historical impact records or pre-marked causal relationships between the supply chain entity and the risk event.

[0077] In an embodiment of the present invention, exemplarily, in a consumer electronics supply chain scenario, the server has obtained the entity basic risk characterization vector of mobile phone assembly plant C (reflecting C's basic exposure to historical risks, such as [0.61, 0.1, 0.2, 0.3], where 0.61 indicates that C has a higher basic exposure intensity to F1 (a fire on a chip production line)), and the risk basic conduction characterization vector of risk event F1 (reflecting the basic impact intensity of F1 on each entity, such as [0.69, 0.1, 0.61, 0.1], where 0.61 indicates that F1 has a higher basic conduction intensity on C). The following details how the server calculates the first risk conduction assessment deviation value based on these two vectors: the server calls the "basic conduction prediction module" of the resilience assessment model, which is responsible for outputting the risk resilience coefficient of the supply chain entity to risk events based on the entity basic risk characterization and the risk basic conduction characterization (the higher the value, the stronger the entity's ability to resist risks). The server inputs C's entity-based risk representation vector ([0.61, 0.1, 0.2, 0.3]) and F1's risk-based conduction representation vector ([0.69, 0.1, 0.61, 0.1]) into the "Basic Conductance Prediction Module." This module fuses the two vectors using an internal fully connected layer (e.g., a two-layer neural network) (for example, concatenating the two vectors into an 8-dimensional vector [0.61, 0.1, 0.2, 0.3, 0.69, 0.1, 0.61, 0.1]). It then extracts key features using a nonlinear activation function (e.g., ReLU) and ultimately outputs a value between 0 and 1 as the first assessed risk resilience coefficient. In specific execution, the server's calculation process is as follows: The first fully connected layer maps the 8-dimensional input to 4 dimensions (e.g., a 4×8 weight matrix and a 4-dimensional bias term). After performing a linear transformation on the input vector, negative values ​​are filtered using the ReLU function, resulting in the intermediate features [0.5, 0.3, 0.2, 0.1]. The second fully connected layer maps the 4-dimensional intermediate features to 1 dimension (e.g., a 1×4 weight matrix and a 1-dimensional bias term). After another linear transformation, the result is compressed to the range of 0-1 using the Sigmoid function. The final output is C's first estimated risk resilience coefficient for F1: 0.6 (indicating that the model predicts that C has moderate risk resilience in the event of F1). The server needs to obtain the "conduction relationship" between C and F1 as a reference, namely, C's actual risk resilience coefficient for F1 in the past, or the causal relationship strength pre-annotated by experts (e.g., a value between 0 and 1, with higher values ​​indicating greater resilience). The server retrieves historical interaction records between C and F1 from the supply chain risk database.For example, three years ago, when F1 (a chip production line fire) occurred, C was shut down for five days due to insufficient backup suppliers (only one at the time) and low inventory levels (maintaining only three days of production). Its actual resilience coefficient was 0.4 (a low value, indicating weak capabilities). However, in the past year, C has optimized its supplier management (increasing its backup suppliers to three) and increased its safety inventory (maintaining eight days of production). In the most recent F1 simulation, C was shut down for only two days, and its actual resilience coefficient increased to 0.7 (a high value, indicating strong capabilities). The server, integrating recent data, determines that the actual value of the transmission relationship between C and F1 is 0.7 (serving as an assessment benchmark). The server compares the model's output of the first-assessed resilience coefficient (0.6) with the actual value of the transmission relationship (0.7), calculating the difference between the two as the first risk transmission assessment deviation. Smaller differences indicate more accurate model predictions; larger differences indicate that the model needs adjustment. The server uses "absolute error" to calculate the deviation (i.e., the absolute difference between the predicted value and the actual value): [{First Risk Transmission Assessment Deviation} = |0.6 - 0.7| = 0.1]; if the difference between the model's predicted value (0.6) and the actual value (0.7) is 0.1, the deviation is 0.1. This value will serve as the basis for model optimization. The server will subsequently adjust the weight parameters of the "Basic Transmission Prediction Module" (such as the weight matrix of the first fully connected layer) to bring the predicted value closer to the actual value (for example, from 0.6 to 0.68), thereby reducing the deviation (for example, from 0.1 to 0.02). Through these steps, the server completes the calculation of the deviation value for the first risk transmission assessment. Mobile phone assembly plant C's first assessment of the risk resilience coefficient for F1 is 0.6, which deviates by 0.1 from the actual value of 0.7, indicating that the model's current prediction of the relationship between basic risk and transmission has some error. The server will subsequently adjust model parameters to reduce this deviation and improve the accuracy of the model's assessment of supply chain resilience.

[0078] In an embodiment of the present invention, the second risk conduction assessment deviation value of the resilience assessment model is obtained based on the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event, which can be implemented through the following example.

[0079] Obtaining a second assessed risk resilience coefficient for each supply chain entity for each risk event based on the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event;

[0080] The second risk transmission assessment deviation value is obtained based on the second assessed risk resilience coefficient of each supply chain entity for each risk event and the transmission relationship of each supply chain entity for each risk event.

[0081] In an embodiment of the present invention, for example, in a consumer electronics supply chain scenario, the server has obtained the entity multi-dimensional capability representation vector of the mobile phone assembly plant C (reflecting C's actual risk resistance, such as [3.75, 0.9875, 47.52, 2.5, 1], which respectively represent the number of backup suppliers, inventory levels, production flexibility, financial liquidity, and geographical location), and the risk multi-dimensional attribute representation vector of the risk event F1 (chip production line fire) (reflecting the actual risk attributes of F1, such as [0.05, 12.2, 72, 11100, 0], which respectively represent the probability of occurrence, scope of impact, duration, severity, and predictability). The following details how the server calculates the second risk conduction assessment deviation value based on these two vectors: the server calls the "capability attribute prediction module" of the resilience assessment model, which is responsible for outputting the risk resistance resilience coefficient of the supply chain entity to risk events based on the entity multi-dimensional capability representation and the risk multi-dimensional attribute representation (the higher the value, the stronger the entity's ability to resist risks). The server inputs C's entity multidimensional capability representation vector ([3.75, 0.9875, 47.52, 2.5, 1]) and F1's risk multidimensional attribute representation vector ([0.05, 12.2, 72, 11100, 0]) into the "Capability Attribute Prediction Module." This module fuses the two vectors using an internal fully connected layer (e.g., a two-layer neural network) (for example, concatenating the two vectors into a 10-dimensional vector [3.75, 0.9875, 47.52, 2.5, 1, 0.05, 12.2, 72, 11100, 0]). It then extracts key features using a nonlinear activation function (e.g., ReLU) and ultimately outputs a value between 0 and 1 as the second assessment of risk resilience. During execution, the server's calculation process is as follows: The first fully connected layer maps the 10-dimensional input to 5 dimensions (e.g., a 5×10 weight matrix and a 5-dimensional bias term). After performing a linear transformation on the input vector, negative values ​​are filtered using the ReLU function, resulting in the intermediate features [2.1, 1.8, 0.9, 0.5, 0.3]. The second fully connected layer maps the 5-dimensional intermediate features to 1 dimension (a 1×5 weight matrix and a 1-dimensional bias term). After another linear transformation, the result is compressed to the range of 0-1 using the Sigmoid function. The final output is C's second-assessed risk resilience coefficient for F1, which is 0.65 (indicating that the model predicts that C has strong risk resilience when F1 occurs). The server needs to obtain the "conduction relationship" between C and F1 as a reference, that is, C's actual risk resilience coefficient for F1 in the past (or the causal relationship strength pre-annotated by experts). The server retrieves the latest historical records of C and F1 from the supply chain risk database. For example, in a recent F1 simulation exercise, C only stopped production for 2 days because it had sufficient backup suppliers (3), high inventory levels (maintaining 9.8 days of production), and short production line switching time (47.5 hours). Its actual risk resilience coefficient was 0.7 (a relatively high value, indicating strong capabilities).The server uses this value as the evaluation benchmark. The server compares the model's output of the second-assessed risk resilience coefficient (0.65) with the actual value of the transmission relationship (0.7), calculating the difference between the two as the second-assessment risk transmission assessment deviation. The smaller the difference, the more accurate the model's prediction; the larger the difference, the more likely the model needs adjustment. The server uses "absolute error" to calculate the deviation (i.e., the absolute difference between the predicted value and the actual value): [{second-assessment risk transmission deviation value} = |0.65 - 0.7| = 0.05]; if the difference between the model's predicted value (0.65) and the actual value (0.7) is 0.05, the deviation value is 0.05. This value will serve as the basis for model optimization. The server will subsequently adjust the weight parameters of the "capability attribute prediction module" (such as the weight matrix of the first fully connected layer) to bring the predicted value closer to the actual value (e.g., from 0.65 to 0.68), thereby reducing the deviation value (e.g., from 0.05 to 0.02). Through the above steps, the server completes the calculation of the second-assessment risk transmission assessment deviation value. The second-assessed resilience coefficient for F1 at mobile phone assembly plant C was 0.65, which deviated by 0.05 from the actual value of 0.7. This indicates that the model's predictions of entity capabilities and risk attributes are relatively accurate, but there is still room for improvement. Subsequent server adjustments will reduce this deviation by adjusting model parameters, further improving the model's accuracy in assessing supply chain resilience.

[0082] In the embodiments of the present invention, the following implementation modes are also provided.

[0083] Acquire and evaluate supply chain entities and assess risk events;

[0084] Calling the trained resilience assessment model to assess the risk resilience coefficient of the assessed supply chain entity with respect to the assessed risk event;

[0085] If the risk resilience coefficient of the evaluated supply chain entity for the evaluated risk event is greater than or equal to the risk resilience coefficient threshold, then based on the entity multidimensional capability representation vector of the evaluated supply chain entity and the risk multidimensional attribute representation vector of the risk event, targeted resilience enhancement suggestions are generated and synchronized to the evaluated supply chain entity.

[0086] In an exemplary embodiment of the present invention, in a consumer electronics supply chain scenario, the server has completed training of a resilience assessment model. Taking the example of a mobile phone assembly plant C (the assessment supply chain entity) requesting an assessment of its resilience to a "chip production line fire" (assessment risk event F1), the server's entire process of executing the assessment and generating recommendations is described in detail: The server receives an assessment request through the enterprise management system. For example, the supply chain manager of mobile phone assembly plant C logs into the system, selects the "Risk Resilience Assessment" function, enters the assessment object (C itself) and the target risk event (F1), and uploads C's latest operational data (such as the number of backup suppliers and inventory levels) and F1's latest risk attribute data (such as the probability of a recent fire and the scope of impact). Upon receiving the request, the server extracts the identification information of the assessment supply chain entity (C) and the assessment risk event (F1), and retrieves C's entity feature topology (including feature units and conductive edges such as "number of backup suppliers" and "inventory level") and F1's risk attribute topology (including feature units and conductive edges such as "scope of impact" and "severity") from the database, completing the assessment preparation. The server calls the trained resilience assessment model, inputs the entity feature topology of C and the risk attribute topology of F1, and the model outputs the risk resilience coefficient through the following steps: Feature extraction and characterization: The model first screens the key transmission edges of the entity feature topology of C (such as retaining key edges such as "number of backup suppliers-inventory level" and "inventory level-production elasticity"), and calculates the entity multi-dimensional capability representation vector of C (such as [4,1.0,48,2.8,1], indicating 4 backup suppliers, inventory can be maintained for 10 days, production line switching is 48 hours, financial liquidity is 2.8, and multiple port geographical locations); at the same time, the model screens the key transmission edges of the risk attribute topology of F1 (such as retaining key edges such as "duration-impact range" and "impact range-severity"), and calculates the risk multi-dimensional attribute representation vector of F1 (such as [0.04,10,60,8000,1], indicating a 4% probability of occurrence, affecting 10 chip models, lasting 60 hours, a loss of 80 million yuan, and a new early warning system predictability of 1). Resilience coefficient calculation: The model inputs C's entity multi-dimensional capability vector and F1's risk multi-dimensional attribute vector into the "capability-attribute prediction module", fuses features through the fully connected layer (such as splicing into a 10-dimensional vector), and outputs the risk resilience coefficient after nonlinear activation. Finally, the model outputs C's risk resilience coefficient for F1 as 0.75 (the threshold is set to 0.7, 0.75 ≥ 0.7, which meets the standard but still has room for optimization). Since C's risk resilience coefficient (0.75) ≥ the threshold (0.7), the server further analyzes C's entity multi-dimensional capability vector and F1's risk multi-dimensional attribute vector, identifies the matching shortcomings of capabilities and risk attributes, and generates specific resilience enhancement suggestions.Capability and risk attribute matching analysis: C's entity multidimensional capability vector shows that the "number of backup suppliers" is 4 (covering conventional chip models), but F1's risk attribute vector shows that the "scope of impact" is 10 chip models (including some special models not covered by C); C's "inventory level" is 1.0 (can maintain 10 days of production), but F1's "duration" is 60 hours (about 2.5 days). If the fire causes the supplier's recovery time to be extended (such as 72 hours in reality), C's inventory can only support 10 days, and there may be gaps due to subsequent replenishment delays; F1's "predictability" is 1 (new early warning system), but C's "production flexibility" is 48 hours (it takes 2 days to switch production lines). If the early warning is 3 days in advance (72 hours), C's production line switching time can be further shortened to match the early warning time. Targeted Recommendation Generation: Based on the above analysis, the server generated three recommendations: Expand backup supplier coverage: Sign backup agreements with two new suppliers for specialized chip models (the current four only cover conventional models; these new suppliers will cover eight of the ten models affected by F1); Establish a dynamic inventory adjustment mechanism: Based on the duration of F1 (60 hours), adjust safety stock from 10 days to 12 days (a 20% increase), and agree with suppliers on a "48-hour emergency replenishment" clause to avoid inventory gaps; Optimize the production line switching process: Leverage F1's early warning system (72 hours of advance warning) to reduce switching time from 48 hours to 36 hours (by preloading backup materials and streamlining quality inspection processes), improving response speed. The server pushes the generated resilience enhancement recommendations to C's supply chain management platform via the enterprise management system and also sends an email notification to C's management. The recommendations include specific measures (e.g., "Add two new backup suppliers for specialized chips"), expected results (e.g., "Increase backup supplier coverage from 60% to 80%"), and implementation priorities (e.g., "The dynamic inventory adjustment mechanism must be completed within one month"). C's managers can directly view the recommendations in the system and initiate the implementation process. Through these steps, the server completes the entire process, from receiving the assessment request to outputting the resilience coefficient and then generating targeted recommendations. Mobile phone assembly plant C's risk resilience coefficient meets the requirements, but the server analyzes the shortcomings in matching its capabilities with its risk attributes and provides specific optimization directions, helping C further enhance its resilience to F1 and achieve dynamic management of supply chain risks.

[0087] An embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned AI-assisted supply chain resilience risk assessment method. Figure 2 As shown, Figure 2This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0088] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. The AI-assisted supply chain resilience risk assessment method is characterized by: include: Obtain a risk transmission topology network; the risk transmission topology network includes multiple supply chain entity units and multiple risk event units, each supply chain entity unit represents a supply chain entity, and each risk event unit represents a risk event; If any supply chain entity among the multiple supply chain entities has a transmission relationship with any risk event among the multiple risk events, then the supply chain entity unit of any supply chain entity and the risk event unit of any risk event have a risk transmission edge in the risk transmission topology network; Obtaining an entity feature topology corresponding to each supply chain entity in the plurality of supply chain entities; any entity feature topology includes a plurality of supply chain entity feature units, and any supply chain entity feature unit is used to characterize the supply chain entity feature of the corresponding supply chain entity on an evaluation indicator; Obtaining a risk attribute topology corresponding to each of the multiple risk events; each risk attribute topology includes multiple risk event feature units, and each risk event feature unit is used to characterize the risk event feature of the corresponding risk event on an evaluation indicator; A resilience assessment model is trained based on the risk transmission topology network, the entity feature topology of each supply chain entity and the risk attribute topology of each risk event to obtain a trained resilience assessment model; the trained resilience assessment model is used to evaluate the risk resilience coefficient of the supply chain entity against risk events.

2. The method according to claim 1, characterized in that The training of the resilience assessment model based on the risk transmission topology network, the entity feature topology of each supply chain entity, and the risk attribute topology of each risk event to obtain a trained resilience assessment model includes: The risk transmission topology network is characterized as a risk transmission association matrix; the risk transmission association matrix is ​​used to indicate the risk transmission edge relationship between the risk event unit and the supply chain entity unit in the risk transmission topology network; The resilience assessment model is called to obtain a risk information transmission matrix, and based on the risk information transmission matrix and the risk conduction association matrix, two-way information transmission is performed on the supply chain entity characteristics of the multiple supply chain entities and the risk event characteristics of the multiple risk events to obtain a risk conduction comprehensive representation matrix corresponding to the multiple supply chain entities and the multiple risk events; Obtaining an entity-based risk representation vector of each supply chain entity and a risk-based transmission representation vector of each risk event based on the risk transmission comprehensive representation matrix; Calling the resilience assessment model to obtain an entity multi-dimensional capability representation vector of each supply chain entity based on the entity feature topology of each supply chain entity; Calling the resilience assessment model to obtain a risk multi-dimensional attribute representation vector of each risk event based on the risk attribute topology of each risk event; The resilience assessment model is trained based on the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event to obtain the trained resilience assessment model.

3. The method according to claim 2, characterized in that There are multiple risk transmission comprehensive representation matrices; obtaining the entity-based risk representation vector of each supply chain entity and the risk-based transmission representation vector of each risk event based on the risk transmission comprehensive representation matrix includes: Perform multi-dimensional feature fusion on multiple risk conduction comprehensive representation matrices based on the attention mechanism to obtain a fused risk conduction comprehensive representation matrix; Performing feature conversion on the fused risk conduction comprehensive representation matrix to obtain a target risk conduction comprehensive representation matrix; The entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event are extracted from the target risk transmission comprehensive characterization matrix.

4. The method according to claim 2, characterized in that Any one of the multiple supply chain entities is characterized as a target supply chain entity, and a risk transmission edge exists between any two supply chain entity feature units in the entity feature topology of the target supply chain entity; The calling of the resilience assessment model to obtain the entity multi-dimensional capability representation vector of each supply chain entity based on the entity feature topology of each supply chain entity includes: Calling the resilience assessment model to perform key transmission edge screening on the risk transmission edges in the entity feature topology of the target supply chain entity to obtain a first core feature interaction topology of the entity feature topology of the target supply chain entity; Based on the first core feature interaction topology, risk information is transmitted for the supply chain entity features of the target supply chain entity on multiple evaluation indicators to obtain unit features corresponding to each supply chain entity feature unit of the target supply chain entity in the first core feature interaction topology; An entity multi-dimensional capability representation vector of the target supply chain entity is obtained according to the unit features corresponding to each supply chain entity feature unit of the target supply chain entity.

5. The method according to claim 2, characterized in that Any one of the multiple risk events is characterized as a target risk event, and a risk transmission edge exists between any two risk event feature units in the risk attribute topology of the target risk event; The calling of the resilience assessment model to obtain the risk multi-dimensional attribute representation vector of each risk event based on the risk attribute topology of each risk event includes: Calling the resilience assessment model to perform key transmission edge screening on the risk transmission edges in the risk attribute topology of the target risk event to obtain a second core feature interaction topology of the risk attribute topology of the target risk event; Based on the second core feature interaction topology, risk information is transmitted for the risk event features of the target risk event on multiple evaluation indicators to obtain unit features corresponding to each risk event feature unit of the target risk event in the second core feature interaction topology; A risk multi-dimensional attribute representation vector of the target risk event is obtained according to the unit features corresponding to each risk event feature unit of the target risk event.

6. The method according to claim 2, characterized in that The training of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based conduction representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event to obtain a trained resilience assessment model includes: Obtaining an assessment deviation value of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity-multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk-multidimensional attribute representation vector of each risk event; The model parameters of the toughness assessment model are optimized and adjusted based on the assessment deviation value to obtain the toughness assessment model that has completed training.

7. The method according to claim 6, characterized in that Obtaining the assessment deviation value of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk multidimensional attribute representation vector of each risk event includes: Obtaining a first synergy deviation value for the supply chain entity representation vector according to the entity basic risk representation vector and the entity multidimensional capability representation vector of each supply chain entity; Obtaining a second synergy deviation value for the embedded feature of the risk event based on the risk basic transmission representation vector and the risk multidimensional attribute representation vector of each risk event; A characteristic synergy deviation value is obtained based on the first synergy deviation value and the second synergy deviation value; the characteristic synergy deviation value is used to indicate the degree of representation mismatch between the entity-based risk representation vector and the entity-multidimensional capability representation vector of each supply chain entity, and is used to indicate the degree of representation mismatch between the risk-based transmission representation vector and the risk-multidimensional attribute representation vector of each risk event; Obtaining a first assessed risk resilience coefficient for each supply chain entity for each risk event based on the entity-based risk characterization vector of each supply chain entity and the risk-based transmission characterization vector of each risk event; Obtaining a first risk transmission assessment deviation value based on a first assessed risk resilience coefficient of each supply chain entity for each risk event and a transmission relationship of each supply chain entity for each risk event, where the transmission relationship is at least one of a historical impact record or a pre-marked causal relationship between the supply chain entity and the risk event; Obtaining a second risk conduction assessment deviation value of the resilience assessment model based on the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event; Obtaining a complexity constraint value of the resilience assessment model based on the entity-based risk representation vector of each supply chain entity, the entity-multidimensional capability representation vector of each supply chain entity, the risk-based transmission representation vector of each risk event, and the risk-multidimensional attribute representation vector of each risk event; The evaluation deviation value is determined based on the feature synergy deviation value, the first risk conduction evaluation deviation value, the second risk conduction evaluation deviation value and the complexity constraint value.

8. The method according to claim 7, characterized in that Obtaining a second risk conduction assessment deviation value of the resilience assessment model according to the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event includes: Obtaining a second assessed risk resilience coefficient for each supply chain entity for each risk event based on the entity multidimensional capability representation vector of each supply chain entity and the risk multidimensional attribute representation vector of each risk event; The second risk transmission assessment deviation value is obtained based on the second assessed risk resilience coefficient of each supply chain entity for each risk event and the transmission relationship of each supply chain entity for each risk event.

9. The method according to claim 1, characterized in that The method further comprises: Acquire and evaluate supply chain entities and assess risk events; Calling the trained resilience assessment model to assess the risk resilience coefficient of the assessed supply chain entity with respect to the assessed risk event; If the risk resilience coefficient of the evaluated supply chain entity for the evaluated risk event is greater than or equal to the risk resilience coefficient threshold, then based on the entity multidimensional capability representation vector of the evaluated supply chain entity and the risk multidimensional attribute representation vector of the risk event, targeted resilience enhancement suggestions are generated and synchronized to the evaluated supply chain entity.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images