Sea transportation supply chain dynamic risk identification method based on generative artificial intelligence
By using generative artificial intelligence technology, a time-series dynamic knowledge graph and a multi-level influence chain deduction framework are constructed, which solves the problems of automation and real-time performance in risk identification in traditional methods, and enables efficient and accurate identification and analysis of maritime supply chain risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-10
AI Technical Summary
Traditional maritime supply chain risk management methods lack a unified understanding and correlation capabilities, are unable to perform logical reasoning and impact chain simulations, rely on human experience for decision-making, and cannot reflect real-time changes in the supply chain situation.
By employing generative artificial intelligence technology, and through a multi-layer entity recognition architecture and a large language model in the maritime field, multi-source heterogeneous information is collected, a time-series dynamic knowledge graph is constructed, event confidence is calculated and multi-level influence chain inference is performed, and a structured natural language report is generated.
It enables automated and large-scale identification and understanding of maritime supply chain risks, efficiently and accurately identifies risk events, reflects the time evolution of risk situations, and generates logically rigorous professional analysis reports.
Smart Images
Figure CN121836385A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and supply chain management, and in particular to a method for dynamic risk identification in maritime supply chains based on generative artificial intelligence. Background Technology
[0002] As the complexity and uncertainty of the global maritime supply chain continue to increase, businesses face increasingly diverse risks, including port congestion, strikes, weather delays, customs detentions, and other operational risks. Traditional maritime supply chain risk management methods rely primarily on manual monitoring and experience-based judgment, which have the following technical shortcomings: Risk information is scattered across multiple heterogeneous data sources, including news reports, port announcements, AIS data, charter contracts, and bills of lading terms. Existing systems lack a unified understanding and correlation capability. Traditional rule engines or machine learning models can only match known risk patterns and cannot understand the deep semantics of risk events, let alone perform logical reasoning and impact chain deduction. For example, the system knows that "Port A is congested" and "Route B passes through Port A," but it cannot automatically deduce a complete impact chain such as "the Evergreen 123 vessel carrying our company's key components on Route B will be delayed for 3 days, affecting production line C." Existing monitoring systems can only provide basic alerts of "what happened" and cannot answer the key questions of "what does this mean" and "what should I do?" Decision-making still heavily relies on human experience. Traditional knowledge graphs are mostly static structures and cannot reflect real-time changes in the supply chain situation, making it difficult to capture the propagation and evolution of risks. Summary of the Invention
[0003] This invention provides a dynamic risk identification method for the maritime supply chain based on generative artificial intelligence, which overcomes the problems of lack of unified understanding and correlation capabilities, inability to perform logical reasoning and influence chain deduction, decision-making still relying heavily on human experience, and inability to reflect real-time changes in the supply chain situation.
[0004] To achieve the above objectives, the technical solution of the present invention is as follows: A method for dynamic risk identification in the maritime supply chain based on generative artificial intelligence includes: S1. Collect multi-source heterogeneous information, identify the multi-source heterogeneous information through the designed multi-layer entity recognition architecture and the large language model in the maritime field to obtain complete events; standardize the complete events to obtain standardized complete events; calculate the event confidence of the standardized complete events to obtain the event confidence; obtain structured risk event data based on the standardized complete events and the event confidence. The multi-source heterogeneous information includes news, announcements, AIS data, charter contracts, and bill of lading terms; The complete event includes entities, event types, and event relationships; S2. Based on the structured risk event data, standardize it using a pre-defined unified event abstraction model to obtain standardized structured risk event data; based on the standardized structured risk event data, construct a dynamic knowledge graph, and insert time using an effective time window calculation engine to obtain a time-series dynamic knowledge graph; S3. Based on the temporal dynamic knowledge graph, the event subgraph of the temporal dynamic knowledge graph is obtained through the constructed structured prompt word engine, and prompt words are generated according to the event subgraph of the temporal dynamic knowledge graph. Based on the prompt words and the designed multi-level influence chain inference framework, chain reasoning is performed through the large language model in the maritime field to obtain a structured natural language report. The structured natural language report includes event type, severity, execution summary, overall confidence level, and recommended action; S4. Identify risk events in the maritime supply chain based on the event type and severity of the structured natural language reports.
[0005] Furthermore, the large language model in the maritime domain is a large language model with maritime professional cognitive capabilities obtained through deep domain adaptation technology; the execution steps of the deep domain adaptation technology include: S11. Collect a maritime professional corpus training set; the corpus types of the maritime professional corpus training set include: charter contracts, bill of lading terms, shipping news and port regulations; S12. Supervised training of the large language model is performed based on the training set. By adjusting the parameters of the large language model, a large language model for maritime transport is obtained. S13. Based on a professional dictionary for the maritime industry, design domain-specific prompt word templates, and call the large language model of maritime fundamentals through the domain-specific prompt word templates; The maritime transport-related professional dictionary includes keywords related to strikes, congestion, weather delays, port closures, and customs detentions. S14. Connect and combine the domain-specific prompt word templates and the large language model of maritime fundamentals to obtain the large language model of maritime domain.
[0006] Furthermore, the event confidence level is calculated using the following formula:
[0007] In the formula, The final confidence level for the risk event; Baseline confidence level; This is a confidence adjustment based on the information source type; This is the confidence adjustment based on the keywords in the text that indicate certainty or uncertainty.
[0008] Furthermore, the steps for constructing a time-series dynamic knowledge graph include: S211. Based on the structured risk event data, standardize it using a pre-defined unified event abstraction model to obtain standardized structured risk event data. S212. Using entities in standardized structured risk event data as nodes and event types and relationships in standardized structured risk event data as edges, construct a dynamic knowledge graph based on the nodes and edges. S213. Calculate the effective time window of nodes and edges using the effective time window calculation engine, thereby obtaining the effective time window of nodes and the effective time window of edges, and then obtaining the time-series dynamic knowledge graph.
[0009] Furthermore, the time-series dynamic knowledge graph is updated using an event-driven architecture, with specific steps including: S221. Calculate the similarity between entities and nodes in standardized structured risk event data to obtain location similarity; S222. Calculate the similarity between event types and edges in standardized structured risk event data to obtain event type similarity; S223. Based on the event types of standardized structured risk event data, obtain the effective time window through the effective time window calculation engine; calculate the similarity between the effective time window and the effective time of the node to obtain the time similarity. S224. Weighted fusion of location similarity, event type similarity, and time similarity yields a comprehensive similarity. S225. If the overall similarity is greater than or equal to a preset threshold, then an update operation is performed on the node; the update operation includes weighted fusion of confidence scores, merging the effective time window with the effective time of the node, and updating the node data; S226. If the overall similarity is less than a preset threshold, a new node is created in the dynamic knowledge graph.
[0010] Furthermore, the steps for generating structured natural language reports include: S31. Traverse the temporal dynamic knowledge graph through the context building module of the structured prompt word engine to obtain the event subgraph; S32. The event subgraph is structurally transformed using the task definition module and output specification module of the structured prompt word engine to obtain the final prompt word; S33. Based on the final prompt and the designed multi-level influence chain deduction framework, chain reasoning is performed through a large language model in the maritime field to obtain the deduction influence chain; S34. Based on the confidence level of the event occurrence and the confidence level of the context richness, perform a weighted comprehensive evaluation of the event confidence level to obtain the overall confidence level; based on the overall confidence level, calculate the confidence level of the inferred influence chain to obtain the inferred confidence level; S35. Based on the projection confidence level and the projection impact chain reaching the level value n in the multi-level impact chain projection framework, evaluate the projection impact chain; if the level value of the projection impact chain is greater than or equal to n and the projection confidence level is greater than the first threshold, it is high risk; if the level value of the projection impact chain is less than n but greater than or equal to n-1 and the projection confidence level is less than the first threshold but greater than the second threshold, it is medium risk; otherwise, it is low risk. S36. Combine the inferred impact chain, inferred confidence level, and risk level to generate a structured natural language report.
[0011] Furthermore, the expression for calculating the confidence level of the inferred influence chain is as follows:
[0012] In the formula, To extrapolate confidence levels; Assess the overall confidence level of risk events; The confidence score for context richness is expressed by the following formula:
[0013] In the formula, Number of affected vessels; The quantity of goods affected.
[0014] Beneficial effects: This invention provides a dynamic risk identification method for the maritime supply chain based on generative artificial intelligence. Through a large language model in the maritime field and a multi-layer entity recognition architecture, it can efficiently and accurately extract structured risk events from massive amounts of multi-source heterogeneous information (such as news, reports, shipping data, and meteorological information). It overcomes the limitations of traditional methods that rely on manual labor and are difficult to process unstructured text, and realizes the automated and large-scale collection and understanding of risk information. By constructing a time-series dynamic knowledge graph, we not only depict the complex relationships between risk entities but also reflect the evolution of the risk situation over time. Combined with a multi-level influence chain deduction framework, we can simulate risk transmission and cascading effects. By leveraging a structured prompt word engine to guide domain-specific large language models in targeted reasoning, we generate natural language reports with fixed structures and rigorous logic, transforming complex graph data into professional analyses that humans can intuitively understand. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of the dynamic risk identification method of the present invention; Figure 2 This is a visual diagram illustrating the distribution of risk severity in an embodiment of the present invention; Figure 3 This is a visual diagram of the event timeline in an embodiment of the present invention; Figure 4 This is a visual diagram illustrating the deduced influence chain in an embodiment of the present invention; Figure 5 This is a visualization diagram of confidence analysis in an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] This embodiment provides a method for dynamic risk identification in the maritime supply chain based on generative artificial intelligence, such as... Figure 1 As shown, it includes: S1. Collect multi-source heterogeneous information, identify the multi-source heterogeneous information through the designed multi-layer entity recognition architecture and the large language model in the maritime field to obtain complete events; standardize the complete events to obtain standardized complete events; calculate the event confidence of the standardized complete events to obtain the event confidence; obtain structured risk event data based on the standardized complete events and the event confidence. The multi-source heterogeneous information includes news, announcements, AIS data, charter contracts, and bill of lading terms; The complete event includes entities, event types, and event relationships; S2. Based on the structured risk event data, standardize it using a pre-defined unified event abstraction model to obtain standardized structured risk event data; based on the standardized structured risk event data, construct a dynamic knowledge graph, and insert time using an effective time window calculation engine to obtain a time-series dynamic knowledge graph; S3. Based on the temporal dynamic knowledge graph, the event subgraph of the temporal dynamic knowledge graph is obtained through the constructed structured prompt word engine, and prompt words are generated according to the event subgraph of the temporal dynamic knowledge graph. Based on the prompt words and the designed multi-level influence chain inference framework, chain reasoning is performed through the large language model in the maritime field to obtain a structured natural language report. The structured natural language report includes event type, severity, execution summary, overall confidence level, and recommended action; S4. Identify risk events in the maritime supply chain based on the event type and severity of the structured natural language reports.
[0019] Preferably, the large language model in the maritime domain is a large language model with maritime professional cognitive capabilities obtained through deep domain adaptation technology; the execution steps of the deep domain adaptation technology include: S11. Collect a maritime professional corpus training set; the corpus types of the maritime professional corpus training set include: charter contracts, bill of lading terms, shipping news and port regulations; S12. Supervised training of the large language model is performed based on the training set. By adjusting the parameters of the large language model, a large language model for maritime transport is obtained. The supervised training refers to fine-tuning the parameters of a basic large language model using a maritime professional corpus training set. S13. Based on a professional dictionary for the maritime industry, design domain-specific prompt word templates, and call the large language model of maritime fundamentals through the domain-specific prompt word templates; The maritime transport-related professional dictionary includes keywords related to strikes, congestion, weather delays, port closures, and customs detentions. S14. Connect and combine the domain-specific prompt word templates and the large language model of maritime fundamentals to obtain the large language model of maritime domain.
[0020] Specifically, the strike-related keywords include a complete terminology system encompassing strike, industrial action, walkout, and labor dispute; The congestion-related keywords include descriptive terms describing different degrees of severity, such as congestion, backlog, bottleneck, and queueing. The weather delay keywords include meteorological terms such as typhoon, storm, fog, bad weather, and gale. The port closure keywords include status changes such as closure, shutdown, suspended, and halted. The keywords related to customs detention include regulatory terms such as customs, inspection, detained, and hold.
[0021] Specifically, the multi-layer entity recognition architecture enables large language models in the maritime domain to accurately identify three types of core entities in multi-source heterogeneous information; the three types of core entities include basic entities, business entities, and relational entities. The basic entities include physical objects such as ports, ships, and cargo; the business entities include professional concepts such as Estimated Time of Arrival (ETA), demurrage, and force majeure; and the relational entities include semantic relationships such as is_berthing_at, is_carrying, and is_destined_for. A hybrid strategy is employed to identify three types of core entities. For basic entities, regular expression matching is used to ensure accuracy; for business entities and relational entities, a large language model from the maritime domain is used for deep understanding to achieve accurate extraction; and an entity normalization mechanism is used to unify the same entity with different expressions into a standard form to ensure the accuracy of subsequent processing.
[0022] Preferably, the event confidence level is calculated using the following formula:
[0023] In the formula, The final confidence level for the risk event; Baseline confidence level; This is a confidence adjustment based on the information source type; This is the confidence adjustment based on the keywords in the text that indicate certainty or uncertainty.
[0024] Specifically, a unified event abstraction model is adopted to fuse the signals extracted from the large language model in the maritime field into a standardized risk event representation. The fusion process adopts a deduplication algorithm based on similarity calculation: when the similarity of two events in the three dimensions of type, location, and time window exceeds the threshold, the fusion operation is triggered, the event information is merged and the confidence is recalculated. The standardized output adopts JSON-LD format to ensure the machine readability and semantic interoperability of the data.
[0025] In a specific embodiment, the entities, event types, event relationships, and event confidence in a standardized complete event are converted into JSON format to generate a standardized JSON-LD file.
[0026] Preferably, the steps for constructing a time-series dynamic knowledge graph include: S211. Based on the structured risk event data, standardize it using a pre-defined unified event abstraction model to obtain standardized structured risk event data. S212. Using entities in standardized structured risk event data as nodes and event types and relationships in standardized structured risk event data as edges, construct a dynamic knowledge graph based on the nodes and edges. S213. Calculate the effective time window of nodes and edges using the effective time window calculation engine, thereby obtaining the effective time window of nodes and the effective time window of edges, and then obtaining the time-series dynamic knowledge graph.
[0027] Preferably, the time-series dynamic knowledge graph is updated using an event-driven architecture, and the specific steps include: S221. Calculate the similarity between entities and nodes in standardized structured risk event data to obtain location similarity; S222. Calculate the similarity between event types and edges in standardized structured risk event data to obtain event type similarity; S223. Based on the event types of standardized structured risk event data, obtain the effective time window through the effective time window calculation engine; calculate the similarity between the effective time window and the effective time of the node to obtain the time similarity. S224. Weighted fusion of location similarity, event type similarity, and time similarity yields a comprehensive similarity. S225. If the overall similarity is greater than or equal to a preset threshold, then an update operation is performed on the node; the update operation includes weighted fusion of confidence scores, merging the effective time window with the effective time of the node, and updating the node data; S226. If the overall similarity is less than the preset threshold, a new node is created in the dynamic knowledge graph. The process of creating new nodes in the dynamic knowledge graph is described in steps S211 to S312.
[0028] Specifically, the node's validity time records the time dimension of entity state changes, including state start time, state end time, and record creation time; the edge timestamp records the relationship's validity time, including relationship establishment time and relationship end time. The effective time window calculation engine uses a combination of rule engine and statistical analysis to calculate the time window. For explicit duration statements, the numerical value and unit are extracted using regular expressions and converted into a standard time increment. For implicit duration inference, default values are assigned based on historical data of event types: strike events are assumed to last 48 hours by default, weather delays are assumed to last 24 hours by default, and port congestion is assumed to last 48 hours by default. The system analyzes keywords in the text and combines them with historical statistical patterns to scale or offset the default duration proportionally, ensuring the accuracy and practicality of the effective time window.
[0029] Specifically, the event type similarity uses a predefined classification system to calculate semantic distance; the location similarity is based on precise matching of port codes; and the time similarity considers time overlap, setting events occurring within 7 days as potentially similar events.
[0030] Preferably, the steps for generating a structured natural language report include: S31. Traverse the temporal dynamic knowledge graph through the context building module of the structured prompt word engine to obtain the event subgraph; S32. The event subgraph is structurally transformed using the task definition module and output specification module of the structured prompt word engine to obtain the final prompt word; S33. Based on the final prompt and the designed multi-level influence chain deduction framework, chain reasoning is performed through a large language model in the maritime field to obtain the deduction influence chain; S34. Based on the confidence level of the event occurrence and the confidence level of the context richness, perform a weighted comprehensive evaluation of the event confidence level to obtain the overall confidence level; based on the overall confidence level, calculate the confidence level of the inferred influence chain to obtain the inferred confidence level; S35. Based on the projection confidence level and the projection impact chain reaching the level value n in the multi-level impact chain projection framework, evaluate the projection impact chain; if the level value of the projection impact chain is greater than or equal to n and the projection confidence level is greater than the first threshold, it is high risk; if the level value of the projection impact chain is less than n but greater than or equal to n-1 and the projection confidence level is less than the first threshold but greater than the second threshold, it is medium risk; otherwise, it is low risk. In this embodiment, n takes the value 4; S36. Combine the inferred impact chain, inferred confidence level, and risk level to generate a structured natural language report.
[0031] Specifically, the structured prompt word engine includes a context building module, a task definition module, and an output specification module; the context building module extracts event subgraphs from the knowledge graph, including related entities and relationships centered on the target event, forming the factual basis for reasoning; the task definition module clarifies the inference objectives and constraints, and sets the analysis depth and consideration parameters; the output specification module strictly defines the output structure and format requirements; The typical prompts in this embodiment are as follows: "Starting from risk event X, traverse the knowledge graph and deduce its potential impact path on [supply chain entity Y]. Output format: 1. Direct impact; 2. Indirect impact; 3. Final impact; 4. Confidence analysis." Specifically, the multi-level impact chain deduction framework adopts a four-level progressive analysis structure. The first level, port operation impact analysis, focuses on the infrastructure level, assessing the impact on terminal operations and berth utilization indicators. The second level, vessel scheduling impact analysis, focuses on the transportation level, assessing the impact on shipping schedules and route operation indicators. The third level, cargo delivery impact analysis, focuses on the logistics level, assessing the impact on delivery time, inventory levels, and supply chain indicators. The fourth level, business impact analysis, focuses on the economic level, assessing the impact on contract performance and financial losses. Each level of deduction is built upon the previous level, forming a complete causal chain to ensure comprehensive coverage of risk propagation paths.
[0032] Preferably, the expression for calculating the confidence level of the inferred influence chain is:
[0033] In the formula, To extrapolate confidence levels; Assess the overall confidence level of risk events; The confidence score for context richness is expressed by the following formula:
[0034] In the formula, Number of affected vessels; The quantity of goods affected.
[0035] Specifically, a matrix assessment method is used to comprehensively determine the risk level based on the highest level reached by the influence chain and the average confidence level; the highest level reached by the influence chain is the fourth level of business impact; the average confidence level is the arithmetic mean of the confidence levels of all events involved in the influence chain deduction process; The risk level is determined as follows: if the highest level is greater than or equal to 4 and the average confidence level is greater than 0.7, it is considered high risk; if the highest level is greater than or equal to 3 and the average confidence level is greater than 0.5, it is considered medium risk; otherwise, it is considered low risk. An upgrade mechanism is also set up so that the risk level is automatically upgraded when the scope of impact exceeds the threshold, ensuring the accuracy and adaptability of the risk classification.
[0036] In a specific embodiment, the fixed-structure natural language report not only includes the conclusion, but also details the basis and reasoning path for arriving at the conclusion; an example of the fixed-structure natural language report is as follows: 1. Direct impact: Suspension of passage through the Suez Canal. Basis: Official announcement from the Suez Canal Authority; 2. Indirect impact: The EMI European route will need to detour around the Cape of Good Hope. Basis: This route traditionally relies on the Suez Canal. 3. Final Impact: Container 'CCLU1234567' is expected to arrive at the Port of Hamburg 18 days later, which will trigger the late delivery penalty clause in the sales contract; Basis: Detouring around the Cape of Good Hope will increase the sailing distance and time, and the contract stipulates that late delivery will incur penalties; 4. Confidence level: 95%; Analysis: Based on the official announcement of the Canal Authority, the confidence level of the event is high; based on the complete ship-cargo-contract relationship, the confidence level of the impact chain is high.
[0037] Specifically, the risk visualization charts include a risk severity distribution chart, an event timeline chart, an impact chain chart, and a confidence analysis chart; The risk severity distribution chart uses a bar chart to display the distribution of high, medium, and low risk events, employing color coding (red for high risk, orange for medium risk, and green for low risk). It includes detailed label descriptions and statistical data, such as... Figure 2 As shown; The event timeline chart, presented in Gantt chart format, displays the occurrence and development of risk events chronologically. Each event bar is labeled with the event type, location, confidence level, and expected duration, and includes a current time indicator line, such as... Figure 3 As shown; The influence chain diagram uses directed graph visualization technology to display the risk propagation path. Nodes represent risk events or influencing entities, edges represent influence relationships, node size reflects the scope of influence, color reflects the risk level, and edge thickness reflects the intensity of influence. Figure 4 As shown; The confidence analysis chart uses grouped bar charts to compare the confidence assessment results of different risk reports, displaying the confidence level of event occurrence, impact assessment, and overall confidence level, including an average confidence level reference line, such as... Figure 5 As shown; The decision-making recommendations are generated based on the simulation results, providing specific countermeasure suggestions. The system has a built-in knowledge base of countermeasure strategies for different risk types and levels, which can provide contextualized, targeted and actionable suggestions.
[0038] The present invention has the following beneficial effects: This invention presents a dynamic risk identification method for the maritime supply chain based on generative artificial intelligence. By using a large language model in the maritime field and a multi-layer entity recognition architecture, it can efficiently and accurately extract structured risk events from massive amounts of multi-source heterogeneous information (such as news, reports, shipping data, and meteorological information). This overcomes the limitations of traditional methods that rely on manual labor and are difficult to process unstructured text, and realizes the automated and large-scale collection and understanding of risk information. By constructing a time-series dynamic knowledge graph, we not only depict the complex relationships between risk entities but also reflect the evolution of the risk situation over time. Combined with a multi-level influence chain deduction framework, we can simulate risk transmission and cascading effects. By leveraging a structured prompt word engine to guide domain-specific large language models in targeted reasoning, we generate natural language reports with fixed structures and rigorous logic, transforming complex graph data into professional analyses that humans can intuitively understand.
[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for dynamic risk identification of a sea transportation supply chain based on generative artificial intelligence, characterized in that, The application relates to a risk event prediction method based on a maritime supply chain, and belongs to the field of maritime supply chain risk event prediction. The method comprises the following steps: S1, collecting multi-source heterogeneous information, identifying the multi-source heterogeneous information through a designed multi-layer entity recognition architecture and a large language model in the maritime field to obtain complete events; standardizing the complete events to obtain standardized complete events; calculating the event confidence of the standardized complete events to obtain event confidence; according to the standardized complete events and the event confidence, structured risk event data is obtained; the multi-source heterogeneous information comprises news, announcements, AIS data, charter contracts and bill of lading clauses; the complete events comprise entities, event types and event relationships; S2, according to the structured risk event data, the standardized processing is performed through a preset unified event abstraction model to obtain standardized structured risk event data; a dynamic knowledge graph is constructed according to the standardized structured risk event data, and time insertion is performed through an effective time window calculation engine to obtain a time sequence dynamic knowledge graph; S3, taking the time sequence dynamic knowledge graph as a fact basis, the event subgraph of the time sequence dynamic knowledge graph is obtained through a constructed structured prompt word engine, and the prompt word is generated according to the event subgraph of the time sequence dynamic knowledge graph; according to the prompt word and a designed multi-level influence chain deduction framework, chain reasoning is performed through the large language model in the maritime field to obtain a structured natural language report; the structured natural language report comprises an event type, a severity, an executive summary, an overall confidence and recommended actions; 2. The method of claim 1, wherein the method is based on generative artificial intelligence for dynamic risk identification of a maritime supply chain. S4, according to the event type and the severity of the structured natural language report, it is determined that a risk event exists in the maritime supply chain. the large language model in the maritime field is a large language model with maritime professional cognitive ability obtained through deep field self-adaption technology; the execution steps of the deep field self-adaption technology comprise the following steps: S11, collecting a maritime professional corpus training set; the corpus types of the maritime professional corpus training set comprise a charter contract, a bill of lading clause, shipping news and port regulations; S12, performing supervised training on the large language model according to the training set, and adjusting the parameters of the large language model to obtain a maritime basic large language model; S13, designing a field-specific prompt word template according to a maritime field professional dictionary, and calling the maritime basic large language model through the field-specific prompt word template; the maritime field professional dictionary comprises strike keywords, congestion keywords, weather delay keywords, port closure keywords and customs detention keywords; 3. The method of claim 1, wherein the method is based on generative artificial intelligence for dynamic risk identification of a maritime supply chain. S14, the field-specific prompt word template and the maritime basic large language model are combined to obtain the large language model in the maritime field. wherein is a final confidence for the risk event; is a base confidence; is a confidence adjustment based on information source type; is a confidence adjustment based on certainty or uncertainty keywords in the text.
4. The method of claim 1, wherein, the event confidence is calculated by the following formula: the steps of constructing the time sequence dynamic knowledge graph comprise the following steps: S211, according to the structured risk event data, the standardized processing is performed through a preset unified event abstraction model to obtain standardized structured risk event data; S212, taking the entities in the standardized structured risk event data as nodes, taking the event types and event relationships in the standardized structured risk event data as edges, and constructing a dynamic knowledge graph according to the nodes and edges; S213, calculating the effective time window of the node and the edge through the effective time window calculation engine, obtaining the effective time window of the node and the effective time window of the edge, and further obtaining the time sequence dynamic knowledge graph.
5. The method of claim 4, wherein the method is based on generative artificial intelligence for dynamic risk identification of a maritime supply chain. The time sequence dynamic knowledge graph is updated through an event-driven architecture, and the specific steps include: S221, calculating the similarity between the entity in the standardized structured risk event data and the node to obtain the position similarity; S222, calculating the similarity between the event type of the standardized structured risk event data and the edge to obtain the event type similarity; S223, obtaining the effective time window through the effective time window calculation engine according to the event type of the standardized structured risk event data; and calculating the time similarity according to the effective time window and the effective time of the node; S224, weighting and fusing the position similarity, the event type similarity and the time similarity to obtain the comprehensive similarity; S225, if the comprehensive similarity is greater than or equal to the preset threshold, the node is updated; the update operation includes weighting and fusing the confidence, merging the effective time window and the effective time of the node, and updating the node data; S226, if the comprehensive similarity is less than the preset threshold, a new node is created in the dynamic knowledge graph.
6. The method of claim 1, wherein the method is based on generative artificial intelligence for dynamic risk identification of a maritime supply chain. The steps of generating a structured natural language report include: S31, traversing the time sequence dynamic knowledge graph through the context construction module of the structured prompt word engine to obtain an event subgraph; S32, converting the structure of the event subgraph through the task definition module and the output specification module of the structured prompt word engine to obtain a final prompt word; S33, according to the final prompt word and the designed multi-level influence chain deduction framework, performing chain reasoning through a large language model in the maritime field to obtain a deduction influence chain; S34, weighting and comprehensively evaluating the event confidence according to the event occurrence confidence and the context richness confidence to obtain an overall confidence; and calculating the confidence of the deduction influence chain according to the overall confidence to obtain a deduction confidence; S35, according to the deduction confidence and the deduction influence chain reaching the level value n in the multi-level influence chain deduction framework, evaluating the deduction influence chain; if the level value of the deduction influence chain is greater than or equal to n and the deduction confidence is greater than a first threshold, it is high risk; if the level value of the deduction influence chain is less than n and greater than or equal to n-1 and the deduction confidence is less than the first threshold and greater than a second threshold, it is medium risk; otherwise, it is low risk; S36, combining the deduction influence chain, the deduction confidence and the risk level to generate a structured natural language report.
7. The method of claim 6, wherein the method is based on generative artificial intelligence for dynamic risk identification of a maritime supply chain. The expression for calculating the confidence of the deduction influence chain is: wherein is the confidence of the risk event; is the overall risk event confidence; is the contextual richness confidence, expressed by the following formula: wherein is the number of affected ships; is the number of affected cargo.