Supply chain risk analysis method and system based on knowledge graph

Abstract

The invention relates to the technical field of fusion of multi-modal data analysis and a knowledge graph, in particular to a supply chain risk analysis method and system based on the knowledge graph. The method is used for multi-modal data acquisition, and the acquired multi-modal data is subjected to data preprocessing to obtain standardized data; performing anomaly detection and signal extraction according to the standardized data to perform multivariable anomaly aggregation; analyzing and capturing risk signals in advance through the multivariable abnormal aggregation; establishing a risk propagation model according to the risk signal, performing comprehensive risk assessment, and obtaining a supply chain risk transfer alternative scheme through a multi-objective decision optimization knowledge graph; key nodes in the knowledge graph are identified, a specific destructive scene is generated according to a historical event mode and expert input, and an anti-factual reasoning model is established; the anti-factual reasoning model is used for modeling by presetting a low-probability high-risk event and updating a knowledge graph, and a standby scheme is made in advance through prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of integrating multi-modal data analysis and knowledge graph, and specifically to a supply chain risk analysis method and system based on knowledge graph. Background Art

[0002] Traditional supply chain risk analysis mostly relies on structured data and known risk patterns, such as order records, inventory reports, and supplier bankruptcy probability models. However, it has significant limitations when facing two key challenges: the hidden cold signals in unstructured data and the unpredictability of low-probability high-risk events. Although existing supply chain risk prediction methods can process some text-based data, such as policy news, they lack sufficient depth in mining other modal data, such as geospatial, biosensing, and acoustic data. Traditional risk assessment relies on historical statistical laws, lacks the ability to deduce sudden and cross-domain chain reaction events, and at the same time, existing models are mostly based on historical training data and are difficult to simulate extreme scenarios that have never occurred but are physically possible. Summary of the Invention

[0003] (1) Technical Problems to be Solved

[0004] The purpose of the present invention is to provide a supply chain risk analysis method and system based on knowledge graph, which are used for anomaly detection and signal extraction of multi-modal data in supply chain risk analysis, optimize and update the knowledge graph through signal modeling; model and update the knowledge graph by presetting low-probability high-risk events, predict potential risks in the supply chain and make a preparatory plan.

[0005] (2) Technical Solutions

[0006] To achieve the above purpose, on the one hand, the present invention provides a supply chain risk analysis method based on knowledge graph, and the method includes the following steps:

[0007] Step S1, perform multi-modal data collection, and the multi-modal data includes geospatial data, biosensing data, and acoustic data; preprocess the collected multi-modal data to obtain standardized data.

[0008] Step S2, perform anomaly detection and signal extraction according to the standardized data, and perform multi-variable anomaly aggregation; analyze through the multi-variable anomaly aggregation to capture risk signals in advance.

[0009] Step S3, establish a risk propagation model according to the risk signals, perform comprehensive risk assessment, and optimize the knowledge graph through multi-objective decision-making to obtain a supply chain risk transfer alternative plan.

[0010] Step S4: Identify the key nodes in the knowledge graph, generate specific disruptive scenarios based on historical event patterns and expert input, and establish a counterfactual reasoning model. The counterfactual reasoning model is built by presetting low-probability and high-risk events to model and update the knowledge graph, and making backup plans in advance through prediction.

[0011] Further, the multi-modal data is represented as: where m is the modality, m ∈ M = {geospatial, biosensing, acoustic}; is the i-th data point of modality m.

[0012] Further, the method includes:

[0013] Convert different modality data through an encoder and map it into a unified embedding space, which is represented as:

[0014]

[0015] where, is the representation of the i-th data point of modality m in the unified embedding space; f m is the modality-specific encoder for mapping modality data into the unified embedding space; is the original input data of the i-th data point of modality m; θ m is the encoder parameter of modality m; R d represents that the dimension of the unified embedding space is d.

[0016] Adjust the alignment degree between different modalities by establishing an inter-modal alignment degree formula. The inter-modal alignment degree formula is:

[0017]

[0018] where, L a is the alignment loss function; M is the set of all modalities; P m,n is the set of index pairs of corresponding data points in modality m and modality n; (i, j) is the specific data point index in corresponding modality m and modality n; d is the distance function for calculating the distance between two embedding points; represents the embedding representation of the i-th data point of modality m; represents the embedding representation of the j-th data point of modality n.

[0019] Further, the anomaly detection method includes:

[0020] Identify the anomaly data points under specific context conditions by calculating the anomaly score, and identify potential risks or problems through the anomaly data points. The formula for calculating the anomaly score is:

[0021]

[0022] Among them, represents the anomaly score of the data point of modality m t under context C ; represents the data point of modality m at time t; C t represents the context information at time t; represents the data point and the distance between the expected value μ(C t ) of the given context; μ(C t ) represents the expected value under context C t ; σ(C t ) represents the standard deviation under context C t .

[0023] Extract the anomaly signal through the above calculation formula of the anomaly score, aggregate the extracted anomaly signals of different modalities to obtain the comprehensive anomaly detection result; use the comprehensive anomaly detection result as the input of the decision support system in the knowledge graph, identify the potential risks across modalities, and make a risk response plan.

[0024] Furthermore, the knowledge graph triple is represented as: T1 = {(s, r, o) | s ∈ E1, r ∈ R1, o ∈ E1 ∪ L1}; where, T1 is the set of triples in the knowledge graph, representing the basic unit of knowledge; s is the subject in the triple, representing an entity; r is the relationship in the triple, representing the relationship between the subject and the object; o is the object in the triple, which can be another entity or a literal value; E1 is the entity set, containing all relevant entities; R1 is the relationship set, containing all possible relationship types; L1 is the literal value set, containing all literal values.

[0025] According to the confidence score formula, first calculate the confidence of each generated triple (s, r, o). When the calculated confidence value exceeds the preset confidence threshold, the corresponding triple is reliable; add all triples that meet the confidence condition to the current knowledge graph to update the knowledge graph; at the same time, calculate the obsolescence degree of each triple in the current knowledge graph. When the calculated obsolescence degree exceeds the preset obsolescence threshold, remove the corresponding triple from the knowledge graph; through the iteration of the above process, the knowledge graph is updated.

[0026] Furthermore, the confidence score formula is expressed as: conf(s, r, o) = ψ(s, r, o | D, K); where conf(s, r, o) is the confidence score of the triple (s, r, o); ψ is the confidence function used to calculate the confidence of the triple; s is the subject in the triple; r is the relation in the triple; o is the object in the triple; D is the original data containing the data source for generating the triple; and K is the background knowledge containing additional knowledge or rules.

[0027] Furthermore, the method for establishing the risk propagation model is to quantify the impact of the connection strength between nodes on the risk propagation speed and scope according to the risk propagation dynamics equation, and the risk propagation dynamics equation is:

[0028]

[0029] where, R ii (t) represents the risk level of entity ii at time t; α represents the risk propagation rate; N1 ii represents the adjacency set of entity ii; A iijj is a matrix representing the connection strength between entity ii and entity jj; μ ii (t) represents the internal risk generation rate of entity ii at time t.

[0030] Furthermore, key nodes in the knowledge graph are identified through multi-dimensional evaluation and analysis; the multi-dimensional evaluation method includes: centrality analysis, irreplaceability evaluation, cascade impact simulation, and recovery difficulty estimation; a preset destructive scenario is generated based on the key nodes; the preset destructive scenario will be identified according to historical event patterns, or integrate the risk pre-judgment of industry experts, or generate reasonable innovative scenario events.

[0031] For the preset destructive scenario, a corresponding counterfactual reasoning model is constructed, the affected nodes in the knowledge graph are removed or the node functions are reduced, the propagation mode of risk in the supply network is calculated, the secondary risk effect is evaluated, and the event consequences of the destructive scenario are predicted.

[0032] According to the analysis results of the counterfactual reasoning model, the risk score is updated, the risk links in the supply network are identified, and specific countermeasures are formulated.

[0033] Based on the same inventive concept, on the other hand, the present invention also provides a supply chain risk analysis system based on a knowledge graph, and the system includes: a first data acquisition module, an abnormal signal aggregation module, a comprehensive risk assessment module, and a counterfactual reasoning module, and each module is connected in sequence.

[0034] The first data acquisition module is used to collect multi-modal data, where the multi-modal data includes geospatial data, biosensing data, human behavior data, and acoustic data; and preprocess the collected multi-modal data to obtain standardized data.

[0035] The abnormal signal aggregation module is used to perform anomaly detection and signal extraction based on the standardized data, and perform multi-variable anomaly aggregation; analyze through the multi-variable anomaly aggregation to capture risk signals in advance.

[0036] The comprehensive risk assessment module is used to establish a risk propagation model based on the risk signals, perform comprehensive risk assessment, and obtain a supply chain risk transfer alternative plan by optimizing the knowledge graph through multi-objective decision-making.

[0037] The counterfactual reasoning module is used to identify key nodes in the knowledge graph, generate specific destructive scenarios based on historical event patterns and expert inputs, and establish a counterfactual reasoning model; the counterfactual reasoning model is modeled and updates the knowledge graph by presetting low-probability and high-risk events, and makes backup plans in advance through prediction.

[0038] (3) Beneficial effects

[0039] Compared with the prior art, the beneficial effects of the present invention are: providing a supply chain risk analysis method for multi-modal data of unstructured data; modeling and updating the knowledge graph for extreme scenarios with low probability and high risk, predicting potential supply chain risks and making preparatory plans. Brief description of the drawings

[0040] Figure 1 It is a flowchart of the supply chain risk analysis method based on the knowledge graph according to Embodiment 1 of the present invention;

[0041] Figure 2 It is a schematic diagram of the module composition of the supply chain risk analysis system based on the knowledge graph according to Embodiment 2 of the present invention. Detailed implementation manners

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Embodiment 1: As Figure 1 shown, this embodiment provides a supply chain risk analysis method based on a knowledge graph, and the method includes the following steps:

[0044] Step S1, perform multi-modal data collection. The multi-modal data includes geospatial data, biosensing data, and acoustic data; perform data preprocessing on the collected multi-modal data to obtain standardized data.

[0045] The multi-modal data is represented as: where m is the modality, m ∈ M = {geospatial, biosensing, acoustic}; is the i-th data point of modality m.

[0046] Convert data of different modalities through an encoder and map them into a unified embedding space, which is represented as:

[0047]

[0048] where, is the representation of the i-th data point of modality m in the unified embedding space; f m is a modality-specific encoder for mapping modality data into the unified embedding space; is the original input data of the i-th data point of modality m; θ m is the encoder parameter of modality m; R d indicates that the dimension of the unified embedding space is d.

[0049] It should be noted that different modal data are converted through an encoder so that heterogeneous data can be compared and analyzed in the same dimensional space. Geospatial data refers to data information related to geographical locations. In supply chain risk analysis, such data usually includes: geographical distributions of suppliers and factories, transportation routes and logistics networks, natural disaster risk areas, and satellite images and remote sensing data. Biosensing data mainly involves the monitoring of product quality, environmental conditions, and biological characteristics in supply chain risk analysis, and usually includes: raw material and product quality parameters, production environment monitoring data, biosafety indicators, and monitoring of biological characteristic changes; biosensing data is used for biological and chemical changes that cannot be captured by traditional sensors, providing early warnings for product quality and safety risks. Acoustic data is information collected through sound and vibration signals, and usually includes: sound characteristics of equipment, acoustic detection of product quality, environmental acoustic monitoring, acoustic monitoring of transportation tools, and acoustic characteristics of the processing process; the unique advantage of acoustic data is that it is convenient to collect (usually only a non-contact microphone is required), can continuously monitor, and the performance of some failure modes in acoustic characteristics is earlier than other observable parameters, providing valuable early warning signals for equipment failures, product quality, and processing abnormalities. These three types of multimodal data complement each other in supply chain risk analysis and jointly provide data support for supply chain risk analysis. Geospatial data provides a macro-level perspective on regional and network risks, biosensing data focuses on the dimensions of product quality and biosafety, while acoustic data is good at capturing equipment health and processing abnormalities. By integrating data from these different sources and characteristics, supply chain risk analysis can more comprehensively and timely identify potential risks, including early signals of low-probability high-risk events that are difficult to capture by traditional methods.

[0050] By establishing an inter-modal alignment formula to adjust the alignment degree between different modalities, the inter-modal alignment formula is as follows:

[0051]

[0052] where L a is the alignment loss function; M is the set of all modalities; P m,n is the set of index pairs of corresponding data points in modality m and modality n; (i, j) is the specific data point index in corresponding modalities m and n; d is the distance function used to calculate the distance between two embedding points; represents the embedding representation of the i-th data point in modality m; represents the embedding representation of the j-th data point in modality n.

[0053] For example, for a certain supply chain network, multi-modal data is collected, including geospatial data (GPS coordinates of 217 suppliers, satellite images of 43 main transportation routes, weather data for nearly 12 months), biosensing data (8 indicators of raw material quality parameters, sampled 3 times a day), and acoustic data (equipment noise spectrum from 20 Hz to 20,000 Hz). In the data preprocessing stage, the geospatial longitude and latitude are uniformly converted to the UTM coordinate system, the biosensing data is normalized to the range of 0 - 1, the acoustic data is passed through a band-pass filter to remove environmental noise, and the key frequency band features from 78 Hz to 330 Hz are extracted. For the injection molding equipment of Supplier B, the original acoustic data shows a 13.2% increase in noise. After being converted by a dedicated encoder, it is mapped to a 128-dimensional unified embedding space, achieving alignment with other modal data, and the alignment loss is reduced from the initial value of 0.87 to 0.21.

[0054] Step S2: Based on the standardized data, perform anomaly detection and signal extraction, and conduct multivariate anomaly aggregation; analyze through the multivariate anomaly aggregation to capture risk signals in advance.

[0055] The anomaly detection method includes:

[0056] Identify abnormal data points under specific context conditions by calculating the anomaly score, and identify potential risks or problems through the abnormal data points; the formula for calculating the anomaly score is:

[0057]

[0058] where represents the anomaly score of the data point t of modality m under context C ; represents the data point of modality m at time t; C t represents the context information at time t; represents the distance between the data point and the expected value μ(C t ) of the given context; μ(C t ) represents the expected value under context C t ; σ(C t ) represents the standard deviation under context C t .

[0059] Extract anomaly signals through the formula for calculating the anomaly score, aggregate the extracted anomaly signals of different modalities to obtain a comprehensive anomaly detection result; use the comprehensive anomaly detection result as the input to the decision support system in the knowledge graph, identify cross-modal potential risks, and formulate risk response plans.

[0060] It should be noted that context-based anomaly definition allows for the identification of data patterns that do not meet expectations in specific situations, rather than simply identifying outliers outside static thresholds. Through the aggregated analysis of multimodal anomaly signals, cross-modal risk signals can be identified, providing input for the decision support system of the knowledge graph to pre-capture potential supply chain risks.

[0061] For example: After performing anomaly detection analysis on the standardized data, it is found that the bearing hardness index of Supplier C shows anomalies. Under normal conditions, the expected value of this index is 87.5, with a standard deviation of 2.3. However, the measured value on November 17 dropped to 80.2, and the calculated anomaly score was 3.17, significantly exceeding the preset threshold of 2.5. At the same time, it is found that this anomaly appears simultaneously with two other modal anomalies: satellite images of the area where C is located show continuous heavy rain for 96 hours (rainfall is 217% higher than the historical average), and the sound intensity of C's production equipment at a frequency band of 270 Hz has increased by 31.7% compared to the reference value. Multivariate anomaly aggregation analysis shows that the three anomalies have a correlation of 95.3%. Records show that within 48 hours after the heavy rain, C's actual production capacity decreased by 52.8%, verifying the accuracy of the risk warning.

[0062] Step S3, establish a risk propagation model based on the risk signals, conduct comprehensive risk assessment, and optimize the knowledge graph through multi-objective decision-making to obtain a supply chain risk transfer alternative plan.

[0063] The knowledge graph triple is represented as: T1 = {(s, r, o) | s ∈ E1, r ∈ R1, o ∈ E1 ∪ L1}; where T1 is the set of triples in the knowledge graph, representing the basic unit of knowledge; s is the subject in the triple, representing an entity; r is the relationship in the triple, representing the relationship between the subject and the object; o is the object in the triple, which can be another entity or a literal value; E1 is the entity set, containing all relevant entities; R1 is the relationship set, containing all possible relationship types; L1 is the literal value set, containing all literal values.

[0064] According to the confidence score formula, first calculate the confidence of each generated triple (s, r, o). When the calculated confidence value exceeds the preset confidence threshold, the corresponding triple is reliable; add all triples that meet the confidence conditions to the current knowledge graph to update the knowledge graph; at the same time, by calculating the obsolescence degree of each triple in the current knowledge graph, when the calculated obsolescence degree exceeds the preset obsolescence threshold, remove the corresponding triple from the knowledge graph; through the iteration of this process, the knowledge graph is updated.

[0065] The confidence score formula is expressed as: conf(s, r, o) = ψ(s, r, o | D, K); where conf(s, r, o) is the confidence score of the triple (s, r, o); ψ is the confidence function used to calculate the confidence of the triple; s is the subject in the triple; r is the relationship in the triple; o is the object in the triple; D is the original data containing the data source for generating the triple; K is the background knowledge containing additional knowledge or rules.

[0066] The method for establishing the risk propagation model is to quantify the impact of the connection strength between nodes on the risk propagation speed and range according to the risk propagation dynamics equation, and the risk propagation dynamics equation is:

[0067]

[0068] Where, R ii (t) represents the risk level of entity ii at time t; α represents the risk propagation rate; N1 ii represents the adjacency set of entity ii; A iijj is a matrix representing the connection strength between entity ii and entity jj; μ ii (t) represents the internal risk generation rate of entity ii at time t.

[0069] For example: Construct a knowledge graph containing 2,783 nodes and 9,456 edges, and the triple examples it contains are "(Supplier C, supplies, Bearing D)" and "(Bearing D, used in, Transmission E)". Calculate the confidence of the newly generated triple "(Supplier C, faces, production capacity decline risk)" to be 0.82 (based on the weights of multi-modal abnormal evidence: geospatial 0.45, biosensing 0.28, acoustics 0.27), which exceeds the preset threshold of 0.75 and is successfully added to the graph. At the same time, it is detected that the obsolescence degree of "(Supplier C, inventory level, sufficient)" is 0.91, which exceeds the threshold of 0.8 and has been removed from the graph. The risk propagation model shows that the impact of a 52.8% decline in the production capacity of C spreads to Bearing D (a 43.2% decline in supply volume) within 72 hours, to Transmission E (a 37.6% decline in production volume) within 168 hours, and causes a 26.3% decline in the assembly rate of the whole vehicle F within 12 days, and the deviation of the actual observed data is less than 8.5%.

[0070] Step S4, identify the key nodes in the knowledge graph, generate specific disruptive scenarios according to the historical event patterns and expert input, and establish a counterfactual reasoning model; the counterfactual reasoning model is to model and update the knowledge graph by presetting low-probability high-risk events, and make backup plans in advance through prediction.

[0071] Identify key nodes in the knowledge graph through multi-dimensional evaluation analysis; the multi-dimensional evaluation method includes: centrality analysis, irreplaceability assessment, cascade impact simulation and recovery difficulty estimation; generate preset destructive scenarios based on the key nodes; the preset destructive scenarios will be based on historical event pattern identification, or integrate industry experts' risk predictions, or generate reasonable innovative scenario events.

[0072] A corresponding counterfactual reasoning model is constructed for the preset destructive scenario, the affected nodes in the knowledge graph are removed or the node functions are reduced, the risk propagation mode in the supply network is calculated, the secondary risk effects are evaluated, and the consequences of the destructive scenario are predicted.

[0073] The risk score is updated based on the analysis results of the counterfactual reasoning model, and the risk links in the supply network are identified to formulate specific response plans.

[0074] It should be noted that pre-set disruptive scenarios are generated based on these key nodes. These scenarios can be based on historical event patterns, expert risk assessments, or innovative scenarios. For each disruptive scenario, the system constructs a counterfactual reasoning model. By simulating node impairment or failure, it assesses how risks propagate within the network and their secondary effects, ultimately predicting the consequences of the event and formulating targeted response plans. Centrality analysis assesses the importance of nodes in a supply chain network by calculating each node's "centrality" within the overall network. Centrality analysis can identify key suppliers, even if they are small. Non-substitutability assessment measures the difficulty of replacing a specific node in the supply chain, considering factors such as technological proprietary technology, market structure, switching costs, and contractual restrictions. Cascading impact simulation is a predictive analysis that studies how the failure of a node will have a chain reaction throughout the supply network. Recovery difficulty estimation assesses the complexity and time required to restore normal functionality once a node fails. The combined use of these four analytical methods not only identifies key nodes in the supply chain but also predicts the propagation path and impact of potential risks, providing data support and scientific basis for risk management and emergency response.

[0075] For example: Multidimensional assessment and analysis identified Bearing D as a critical node: centrality score of 0.87 (top 3%), connected to 5 critical systems; irreplaceability score of 0.92 (only 3 companies have production capacity); cascade impact simulation showed that its shortage affected 82.7% of the product lines; recovery difficulty analysis indicated that the restart production cycle was 17.5 days. Based on historical event analysis (referring to the global chip shortage data in 2021) and interviews with industry experts (Delphi method research of 12 supply chain experts), a destructive scenario was preset: "The global rare earth material supply interruption causes Bearing D to be completely unavailable for procurement for 90 days". By forcibly reducing the node function of Bearing D to 0, detailed risk propagation data was recorded: the production of Transmission E decreased by 100% in the first week, the production of Vehicle F decreased by 87.3% in the second week, the corporate revenue decreased by 62.6% in the third week, and the cumulative market share loss was 29.4% from the fourth to the twelfth week (calculated based on comparison with historical similar events). The experimental team implemented response plan development for the vulnerable link of the single-source design of Bearing D: developed an alternative design plan (the test sample reached 94.6% of the performance indicators), established a strategic reserve for 45 days' demand (the cost-to-potential loss ratio was 1:17), established backup agreements with another 5 suppliers (covering 70% of the demand), and improved the product design (the bearing dependence decreased by 31%).

[0076] Embodiment 2: Based on the same inventive concept, as Figure 2 shown, this embodiment also provides a supply chain risk analysis system based on a knowledge graph. The system includes: a first data acquisition module, an abnormal signal aggregation module, a comprehensive risk assessment module, and a counterfactual reasoning module, and the modules are connected in sequence.

[0077] The first data acquisition module is used to collect multimodal data. The multimodal data includes geospatial data, biosensing data, human behavior data, and acoustic data; the collected multimodal data is preprocessed to obtain standardized data.

[0078] The abnormal signal aggregation module is used to perform anomaly detection and signal extraction based on the standardized data, and perform multivariate anomaly aggregation; risk signals are captured in advance through the multivariate anomaly aggregation for analysis.

[0079] The comprehensive risk assessment module is used to establish a risk propagation model based on the risk signals, perform comprehensive risk assessment, and obtain a supply chain risk transfer alternative plan by optimizing the knowledge graph through multi-objective decision-making.

[0080] The counterfactual reasoning module is used to identify critical nodes in the knowledge graph, generate specific destructive scenarios based on historical event patterns and expert inputs, and establish a counterfactual reasoning model; the counterfactual reasoning model is modeled and updates the knowledge graph by presetting low-probability high-risk events, and makes backup plans in advance through prediction.

[0081] It should be noted that regarding the system in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0082] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A supply chain risk analysis method based on a knowledge graph, characterized in that, The method includes: Step S1, perform multi-modal data collection, where the multi-modal data includes geospatial data, biosensing data, and acoustic data; preprocess the collected multi-modal data to obtain standardized data; Step S2, perform anomaly detection and signal extraction based on the standardized data, and perform multi-variable anomaly aggregation; analyze through the multi-variable anomaly aggregation to capture risk signals in advance; Step S3, establish a risk propagation model based on the risk signals, perform comprehensive risk assessment, and obtain a supply chain risk transfer alternative plan by optimizing the knowledge graph through multi-objective decision-making; Step S4, identify key nodes in the knowledge graph, generate specific disruptive scenarios based on historical event patterns and expert input, and establish a counterfactual reasoning model; the counterfactual reasoning model is modeled and updates the knowledge graph by presetting low-probability high-risk events, and makes backup plans in advance through prediction.

2. The supply chain risk analysis method based on a knowledge graph according to claim 1, wherein The multimodal data is represented as: where m is the modality, and m ∈ M = {Geospatial, Biosensing, Acoustics}; is the i-th data point of modality m.

3. The supply chain risk analysis method based on a knowledge graph according to claim 2, wherein The method includes: Convert different modal data through an encoder and map it into a unified embedding space, expressed as: Among them, is the representation of the $i$-th data point of modality $m$ in the unified embedding space; $f$ m is a modality-specific encoder for mapping modality data to the unified embedding space; is the original input data of the $i$-th data point of modality $m$; $\theta$ m are the encoder parameters of modality $m$; $R$ d indicates that the dimension of the unified embedding space is $d$; Adjust the alignment degree between different modalities by establishing an inter-modal alignment degree formula, and the inter-modal alignment degree formula is: Among them, L a is the alignment loss function; M is the set of all modalities; P m,n is the set of index pairs of corresponding data points in modality m and modality n; (i, j) is the index of a specific data point in corresponding modalities m and n; d is a distance function used to calculate the distance between two embedded points; represents the embedded representation of the i-th data point in modality m; represents the embedded representation of the j-th data point in modality n.

4. The supply chain risk analysis method based on a knowledge graph according to claim 1, characterized in that The method of the anomaly detection includes: Identify anomaly data points under specific context conditions by calculating anomaly scores, and identify potential risks or problems through the anomaly data points; the formula for calculating the anomaly scores is: Among them, represents the anomaly score of the data point of modality m t under context C ; represents the data point of modality m at time t; C t represents the context information at time t; represents the distance between the data point and the expected value μ(C t ) of the given context; μ(C t ) represents the expected value under context C t ; σ(C t ) represents the standard deviation under context C t ; Extract anomaly signals through the formula for calculating the anomaly scores, aggregate the extracted anomaly signals of different modalities to obtain a comprehensive anomaly detection result; use the comprehensive anomaly detection result as the input of the decision support system in the knowledge graph, identify cross-modal potential risks, and make risk response plans.

5. The method for analyzing supply chain risks based on a knowledge graph according to claim 4, wherein The knowledge graph triple is expressed as: T1 = {(s, r, o) | s ∈ E1, r ∈ R1, o ∈ E1 ∪ L1}; where, T1 is the set of triples in the knowledge graph, representing the basic unit of knowledge; s is the subject in the triple, representing an entity; r is the relationship in the triple, representing the relationship between the subject and the object; o is the object in the triple, which can be another entity or a literal value; E1 is the entity set, containing all relevant entities; R1 is the relationship set, containing all possible relationship types; L1 is the literal value set, containing all literal values; According to the confidence score formula, first calculate the confidence of each generated triple (s, r, o). When the calculated confidence value exceeds the preset confidence threshold, the corresponding triple is reliable; add all triples that meet the confidence conditions to the current knowledge graph to update the knowledge graph; at the same time, calculate the obsolescence degree of each triple in the current knowledge graph. When the calculated obsolescence degree exceeds the preset obsolescence threshold, remove the corresponding triple from the knowledge graph; through the iteration of this process, the knowledge graph is updated.

6. The method for supply chain risk analysis based on a knowledge graph according to claim 5, wherein The confidence score formula is expressed as: conf(s, r, o) = ψ(s, r, o | D, K); where conf(s, r, o) is the confidence score of the triple (s, r, o); ψ is the confidence function used to calculate the confidence of the triple; s is the subject in the triple; r is the relation in the triple; o is the object in the triple; D is the original data, including the data source used to generate the triple; K is the background knowledge, including additional knowledge or rules.

7. The supply chain risk analysis method based on a knowledge graph according to claim 1, characterized in that The method for establishing the risk propagation model is to quantify the impact of the connection strength between nodes on the risk propagation speed and scope according to the risk propagation dynamics equation, and the risk propagation dynamics equation is: Among them, R ii (t) represents the risk level of entity ii at time t; α represents the risk propagation rate; N1 ii represents the adjacency set of entity ii; A iijj is a matrix representing the connection strength between entity ii and entity jj; μ ii (t) represents the internal risk generation rate of entity ii at time t.

8. The supply chain risk analysis method based on a knowledge graph according to claim 1, wherein, Identify key nodes in the knowledge graph through multi-dimensional evaluation and analysis; The multi-dimensional evaluation method includes: centrality analysis, irreplaceability evaluation, cascade impact simulation, and recovery difficulty estimation; generate a preset destructive scenario based on the key nodes; the preset destructive scenario will be identified according to historical event patterns, or integrate industry experts' predictions of risks, or generate reasonable innovative scenario events; Construct a corresponding counterfactual reasoning model for the preset destructive scenario, remove or reduce the functions of the affected nodes in the knowledge graph, calculate the propagation mode of risks in the supply network, evaluate the secondary risk effects, and predict the event consequences of the destructive scenario; Update the risk score according to the analysis results of the counterfactual reasoning model, identify the risk links in the supply network, and formulate specific response plans.

9. A supply chain risk analysis system based on a knowledge graph, for performing the method according to any one of claims 1-8, characterized in that, The system includes: a first data acquisition module, an abnormal signal aggregation module, a comprehensive risk assessment module, and a counterfactual reasoning module, and each module is connected in sequence; The first data acquisition module is used for multi-modal data collection, and the multi-modal data includes geospatial data, biosensing data, human behavior data, and acoustic data; perform data preprocessing on the collected multi-modal data to obtain standardized data; The abnormal signal aggregation module is used for abnormal detection and signal extraction according to the standardized data, and perform multi-variable abnormal aggregation; capture risk signals in advance through the multi-variable abnormal aggregation analysis; The comprehensive risk assessment module is used to establish a risk propagation model according to the risk signals, perform comprehensive risk assessment, and obtain a supply chain risk transfer alternative plan by optimizing the knowledge graph through multi-objective decision-making; The counterfactual reasoning module is used to identify key nodes in the knowledge graph, generate specific destructive scenarios according to historical event patterns and expert inputs, and establish a counterfactual reasoning model; the counterfactual reasoning model is modeled and updates the knowledge graph by presetting low-probability and high-risk events, and makes backup plans in advance through prediction.

Images