Intelligent risk identification method and system based on multi-modal material identification and strategy inference

Abstract

The invention discloses an intelligent risk identification method and system based on multi-modal material identification and strategy inference, and relates to the technical field of artificial intelligence and risk control. The method comprises the following steps: receiving a multi-source heterogeneous material; performing feature extraction and type identification; constructing a material semantic graph representing a semantic relationship between materials and encoding the material semantic graph into a graph vector; extracting scene features; inputting the features into a strategy inference engine, and carrying out comprehensive inference through evidence chain analysis, multi-channel verification and strategy node probability calculation; and outputting the risk tag and the interpretable report. The system correspondingly comprises an input layer, a core processing layer and an output layer. Through collaborative modeling and inference of the material semantic graph and the strategy node graph, structural understanding and intelligent risk decision making of the multi-modal material are achieved, the problems that in the prior art, rules are rigid, and correlation analysis and modal information splitting are lacked are solved, and the automation, accuracy and interpretability of risk identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and risk control technology, in particular to an intelligent risk identification method and system based on multi-modal material identification and strategy inference. BACKGROUND

[0002] Currently, in the scene of insurance claims, financial audit and other scenarios that need to verify the authenticity of multi-source materials, the common automatic solution is mainly based on optical character recognition (OCR) and pre-set rule system. The typical process is: extracting the text information in the image through OCR, and then using hard-coded business rules (such as "if-then" conditional judgment) to check the format or logic of the text field.

[0003] Such existing technologies mainly have the following limitations:

[0004] 1. The existing technology relies too much on OCR and can only extract text, but cannot perform high-precision visual recognition and classification on the type of materials (such as invoices, contracts, and certificates), resulting in interruption of the automatic process and often requiring manual intervention.

[0005] 2. The risk control logic in the existing technology is fixed in code, and any adjustment or expansion of business strategies must modify the program, making it difficult to quickly adapt to new scenarios, new materials or new risk patterns, with high expansion and maintenance costs.

[0006] 3. The system in the existing technology checks each material as an isolated individual, and cannot understand the internal dependencies, complementarity, timing and other logical relationships between materials, so it cannot judge whether the "evidence chain" formed by a group of materials is complete and reasonable as a whole.

[0007] 4. The existing technology processes multi-modal information such as images, text and structured data in isolation, lacks effective means to align and consistently cross-verify cross-modal information, and is difficult to find deep contradictions.

[0008] The root cause of these defects is that the existing technology architecture lacks the ability to structurally model material sets and lacks an intelligent decision-making mechanism that can be inferred and configured, thereby restricting the evolution of risk identification towards automation and intelligence. SUMMARY

[0009] The present application designs an intelligent risk identification method and system based on multi-modal material identification and strategy inference, which solves the technical problems of the existing technology, such as limited recognition ability, rigid rule system, lack of correlation analysis and fragmented modal information.

[0010] In order to solve the above technical problems, the present application adopts the following scheme:

[0011] An intelligent risk identification method based on multi-modal material identification and strategy inference, characterized in that it comprises the following steps:

[0012] Receiving multi-modal input materials, including image materials, text materials and structured data;

[0013] Feature extraction and identification are performed on the multi-modal input materials to obtain the type label and corresponding feature vector of each material;

[0014] Based on the identified materials and the feature vectors, a material semantic graph is constructed, which takes materials as nodes and establishes edges between nodes based on predefined semantic association relationships, including at least one of type hierarchy relationship, dependency relationship, complementary relationship, substitution relationship, exclusion relationship, sequence relationship and aggregation relationship;

[0015] The material semantic graph is graph structure coded to generate a graph structure feature vector representing the overall topology and semantics of the material set;

[0016] Extract and fuse scene elements to generate a scene feature vector; input the graph structure feature vector and the scene feature vector into a strategy inference engine for risk inference, the strategy inference process includes: the evidence chain integrity analysis module performs structural analysis on the material semantic graph to output an integrity score, the multi-channel consistency verification module performs cross-modal information alignment verification to output a consistency conflict score, and the strategy node probability calculation module calculates the trigger probability of each node in the strategy node graph based on the graph structure feature vector, the scene feature vector, the integrity score and the consistency conflict score; and

[0017] Based on the trigger probability, the integrity score and the consistency conflict score, a risk label and an explainability report are generated and output.

[0018] Preferably, the feature extraction and identification of the multi-modal input materials specifically includes:

[0019] The image material is identified by a computer vision (CV) classification model to output the material type label and visual feature vector;

[0020] The natural language processing (NLP) model is used for semantic understanding of the text material to output the text semantic feature vector; and the structured data is vectorized and encoded to output the structured feature vector.

[0021] Preferably, the graph structure coding of the material semantic graph is implemented by any one or combination of graph neural network (GNN), graph attention network (GAT), topology coding or path coding.

[0022] Preferably, the algorithm model adopted by the strategy node probability calculation module in calculating the trigger probability comprises any one or a combination of a parameterized scoring function, a graph structure message passing model, a Bayesian inference model or a rule-based scoring normalization model.

[0023] A computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the intelligent risk identification method based on multi-modal material identification and strategy inference as described above.

[0024] An electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the intelligent risk identification method based on multi-modal material identification and strategy inference as described above.

[0025] An intelligent risk identification system based on multi-modal material identification and strategy inference, comprising:

[0026] An input interface layer for receiving multi-modal input materials;

[0027] A core processing layer connected to the input interface layer, comprising:

[0028] A material identification module for feature extraction and identification of the multi-modal input materials to obtain the type label and corresponding feature vector of each material;

[0029] A material semantic modeling module for constructing a material semantic graph based on the identified materials and the feature vectors, and performing graph structure coding on the material semantic graph to generate a graph structure feature vector, wherein the material semantic graph establishes edges between nodes according to semantic association relationships including at least one of type hierarchical relationship, dependency relationship, complementary relationship, substitution relationship, exclusion relationship, sequence relationship and aggregation relationship;

[0030] A scene element extraction module for extracting and fusing scene elements to generate a scene feature vector; and

[0031] A strategy inference engine for risk inference based on the graph structure feature vector and the scene feature vector, the strategy inference engine internally comprising: an evidence chain integrity analysis module for performing structural analysis on the material semantic graph and outputting an integrity score; a multi-channel consistency verification module for cross-modal information alignment verification and outputting a consistency conflict score; and a strategy node probability calculation module for maintaining a strategy node graph and calculating the trigger probability of each strategy node based on the graph structure feature vector, the scene feature vector, the integrity score and the consistency conflict score; and

[0032] An output interface layer, connected with the core processing layer, is configured to generate and output a risk label and an explainability report based on the trigger probability, the integrity score, and the consistency conflict score.

[0033] Preferably, the material recognition module is integrated with a computer vision (CV) classification model, a natural language processing (NLP) model, and a structured data encoder.

[0034] Preferably, the material semantic modeling module is integrated with a graph structure encoder, which is implemented by using any one or a combination of a graph neural network (GNN), a graph attention network (GAT), topological encoding, or path encoding.

[0035] Preferably, the algorithm model used by the strategy node probability calculation module to calculate the trigger probability includes any one or a combination of a parameterized scoring function, a graph structure message passing model, a Bayesian inference model, or a rule-based scoring normalization model.

[0036] The intelligent risk identification method and system based on multi-modal material recognition and strategy inference have the following beneficial effects:

[0037] (1) The present application realizes high-precision and fine-grained automatic identification of the type of image materials (such as invoices, contracts, and certificates) by introducing a deep computer vision (CV) classification module, supports dozens or even hundreds of material types, and overcomes the limitations of traditional solutions that rely on OCR and can only extract text but cannot understand the semantics of materials. This reduces the need for human intervention from the source and makes end-to-end automation of the review process possible.

[0038] (2) The present application builds a configurable "strategy node graph" and changes the risk judgment from hard-coded "if-then" rules to probabilistic calculation based on multi-dimensional feature vectors, which enables the risk engine to have "thinking" capabilities. Business strategy adjustments can be realized by updating the strategy node graph or model parameters without modifying the core code, thereby quickly adapting to new business scenarios, risk patterns, and material types, greatly reducing system maintenance costs and improving the life cycle.

[0039] (3) The present application provides a material semantic graph modeling method that can formally define and automatically build various semantic relationships (such as dependency, complementarity, timing, and contradiction) between materials. Based on the analysis of this graph structure, the system can perform evidence chain integrity inference, not only discovering the absence of a single material, but also identifying the contradictions and abnormalities of material combinations in logic and timing, realizing a technological leap from "single-point verification" to "global evidence chain rationality assessment", and discovering more hidden and complex risks.

[0040] (4) The present application realizes the alignment and comparison of image visual features, text semantic features and structured field information in the same semantic space through the multi-channel consistency verification module. This cross-modal cross-verification mechanism can effectively find the inconsistency between different information sources (such as person and certificate inconsistency, and contradiction between ticket information and filled content), which cannot be realized by single mode or simple rule matching, thereby significantly improving the accuracy and reliability of risk identification.

[0041] (5) The present application outputs standardized risk labels and detailed "explainability reports" through the system, and clearly indicates the basis for triggering risks (such as missing which material, where information is contradictory). This not only enhances the transparency and reliability of the system, but also provides clear decision support for manual review, facilitating quick problem positioning and taking corresponding measures, and realizing efficient operation of man-machine cooperation.

[0042] (6) The core of the present application is a general multi-modal understanding and graph inference framework, rather than being bound to specific business rules. By configuring different material type libraries, semantic relationship definitions and strategy node graphs, this scheme can be widely applied to insurance claims, financial credit, government approval, compliance audit and other fields that require automatic verification of multi-source materials, and has high promotion value and reusability. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 : The flowchart of the intelligent risk identification method based on multi-modal material recognition and strategy inference of the present application;

[0044] Figure 2 : The material semantic graph construction schematic diagram of the present application;

[0045] Figure 3 : The working schematic diagram of the strategy inference engine of the present application;

[0046] Figure 4 : The system composition block diagram of the present application;

[0047] Figure 5 : The internal structure block diagram of the strategy inference engine of the present application. DETAILED DESCRIPTION

[0048] The present application will be further described below: Figures 1 to 5

[0049] ​The present application is based on a multi-modal material identification and strategy inference intelligent risk identification method and system, which aims to solve the problems of single material identification dimension, rigid rule system, lack of correlation analysis and multi-modal information fragmentation in existing risk identification technology. The core innovation lies in the introduction of a dual graph structure modeling and inference mechanism of material semantic graph and strategy node graph. By constructing a material semantic graph to formally represent the complex semantic relationship between multiple materials, and using a configurable and inferable strategy node graph for intelligent and probabilistic risk decision-making, the method realizes the deep fusion of multi-modal information and automatic risk identification. The method and system are particularly suitable for scenarios such as insurance claims, financial review, government approval, etc. that require automatic verification of multi-source evidence materials.

[0050] Embodiment 1: An intelligent risk identification method based on multi-modal material identification and strategy inference.

[0051] Please refer to Figures 1 to 3 which shows the overall flowchart of the method of the present application. The method comprises the following steps:

[0052] Step S100: receiving multi-modal input materials.

[0053] The system receives the original materials related to the to-be-reviewed task through a data interface, including image materials (such as electronic files of certificates, bills, photos), text materials (such as case descriptions, explanatory documents), and structured data in forms (such as user-filled date, amount, number, etc. fields).

[0054] Step S200: multi-modal feature extraction and material identification.

[0055] For the received image materials, a pre-trained computer vision (CV) deep learning model is used for identification, such as a CNN or Transformer architecture. This step outputs two results: one is the type label of the material, such as "front side of second-generation ID card" and "special invoice for value-added tax"; the other is the deep visual feature vector extracted from the image. At the same time, for text materials and structured data, natural language processing (NLP) models (such as BERT, ERNIE, etc.) and feature encoders are used for semantic understanding and vectorization, respectively, to form text feature vectors and structured feature vectors.

[0056] Step S300: constructing a material semantic graph.

[0057] This step is one of the core innovations of the present application. As Figure 1As shown, all the material instances identified in step S200 are abstracted as nodes in a graph structure. Each node carries its type code and feature vector attributes. Then, according to a predefined business knowledge base or rules, various types of semantic association edges are established among these nodes. Edge types include but are not limited to: type hierarchy relationship, dependency relationship, complementary relationship, substitution relationship, exclusion relationship, sequence relationship, and aggregation relationship.

[0058] Type hierarchy relationship is to represent the classification or parent-child semantic relationship between material types. For example, there is a general-to-specific relationship between "vehicle insurance policy" and "compulsory insurance policy".

[0059] Dependency relationship is to represent that material A is a prerequisite for the validity or existence of material B. For example, the submission of "claim application" must depend on the existence of "accident liability determination".

[0060] Complementary relationship is to represent that multiple materials mutually confirm each other in information and jointly constitute a complete evidence unit. For example, "loan contract main text" and "contract annex: list of collateral" jointly constitute a complete credit contract material.

[0061] Substitution relationship is to represent that under a certain review target, a certain material can be replaced by another material. For example, in proving personal income, "bank statement" and "income tax payment certificate issued by the tax department" can replace each other.

[0062] Exclusion relationship is to represent that two materials have logical conflicts in key claims or information and cannot exist at the same time. For example, two "disability identification reports" with opposite disability levels issued by different hospitals.

[0063] Sequence relationship is to represent that there is a chronological or procedural order between materials. For example, the date of "registration card" should be earlier than the recording date of "outpatient medical record", which in turn should be earlier than the date of "drug prescription".

[0064] Aggregation relationship is to represent that multiple base materials are combined to form a higher-level semantic structure or composite evidence. For example, a complete "annual audit report of an enterprise" can be aggregated from multiple sub-materials such as "balance sheet", "profit and loss statement", "cash flow statement", and "audit opinion".

[0065] The material semantic graph (a heterogeneous graph containing multiple node and edge types) constructed through the above steps is processed by a graph encoder. The graph encoder can employ any graph structure encoding technique, such as Graph Neural Network (GNN), Graph Attention Network (GAT), topological coding, path coding, or a combination thereof. This encoder aggregates information from all nodes and edges in the graph, ultimately outputting a global graph structure feature vector. This vector comprehensively represents the overall topological structure and relational semantics of the current material set.

[0066] Step S400: Extract and merge scene elements.

[0067] This step aims to create a task-level contextual representation. It integrates the textual feature vector and structured feature vector generated in step S200, and combines them with the specific business scenario context (such as "auto insurance property damage claims" or "credit loan application review") to generate a unified scenario feature vector. This vector describes the overall background and business attributes of this review task.

[0068] Scene elements include, but are not limited to, the review task type code (such as car insurance claims, credit loans) obtained from the work order system, and the applicant's basic attribute category. The scene element extraction module encodes these elements and merges them with the global text and structured features extracted from the input materials to generate a scene feature vector.

[0069] Step S500: Risk inference based on dual graphs.

[0070] like Figure 1 As shown in the flowchart, the graph structure feature vector generated in step S300 and the scene feature vector generated in step S400 are input together into the policy inference engine for comprehensive risk analysis. This step executes three sub-processes in parallel:

[0071] The S510 evidence chain integrity analysis module analyzes the structure of the material semantic graph, determines whether the evidence chain is complete, whether there are any missing necessary nodes, whether the critical path is broken, and whether there are any contradictions in the temporal or logical relationships between nodes, and finally outputs an integrity score.

[0072] The S520 multi-channel consistency verification module performs cross-modal alignment verification. For example, it compares the name recognized by CV from an ID card image, the text extracted by OCR from the same image, and the name entered in a user form, calculating a consistency score. This process quantifies conflicts between different information channels. This operation is performed on all cross-validable fields, generating a consistency conflict score.

[0073] S530 policy node probability calculation module performs: this is another core innovation. An editable policy node graph is maintained in the engine. Each policy node (such as "identity impersonation risk", "ticket tampering suspicion", "loss rationality doubt") in the graph is an independent micro-computation unit. Each node receives the above-mentioned graph structure feature vector, scene feature vector, integrity score and consistency conflict score as comprehensive input. A probability calculation unit inside the node is used to evaluate the trigger probability. The unit can use one or a combination of parameterized scoring functions (such as logistic regression, shallow neural network), graph structure message passing model, Bayesian inference model or rule-based scoring normalization model to calculate a trigger probability between 0 and 1. This way realizes the transformation from binary decision to probabilistic risk assessment, greatly improving the interpretability and decision flexibility of the system.

[0074] As a specific description of the implementation of the probability calculation unit, the following two optional implementation examples are listed:

[0075] 1. Based on parameterized scoring function: taking a logistic regression model as a specific implementation of a parameterized scoring function as an example. For any node in the policy node graph (for example, the "single document risk" node), its input features (including the normalized graph structure feature vector, scene feature vector, integrity score and consistency conflict score) are concatenated into a comprehensive feature vector X. The trigger probability P of the node is calculated by a pre-trained logistic regression model, and the calculation formula can be represented as:

[0076] P = σ (W T X + b)

[0077] Where W is the weight vector, b is the bias term, W T represents the transpose of the weight vector W, and σ(·) is the Sigmoid function. The model parameters W and b can be obtained by supervised learning from historical audit task data (annotated with whether the risk node is determined to be triggered by experts). This way realizes the smooth, differentiable mapping from multi-dimensional evidence to a single risk probability.

[0078] 2. Based on Bayesian inference model:

[0079] Take the Bayesian network as an example of a specific implementation of the Bayesian inference model. The policy node (such as "loss exaggeration risk") is regarded as a Boolean variable to be inferred, and its parent nodes can be set to several evidence variables derived from input features (such as "quote amount abnormally high", "low matching degree between photo and text description", "repair factory history credibility score"). Based on business knowledge or historical data, a conditional probability table is constructed to define the conditional probability of the risk node being true given the state of the parent nodes. When performing inference, the observed state of the evidence variables in the current task (obtained by comparing the feature vector with the threshold) is input into the Bayesian network, and the posterior probability of the risk node being triggered is calculated using the Bayesian formula or exact / approximate inference algorithms (such as variable elimination, Gibbs sampling). This method can explicitly model the dependencies between evidence and incorporate prior knowledge.

[0080] Step S600: Generate risk decision and output.

[0081] The policy inference engine aggregates the trigger probabilities of all policy nodes, and combines the integrity score and conflict score to generate the final comprehensive risk level (such as "high risk", "medium risk", "low risk") and specific risk label through a pre-set decision fusion algorithm (such as weighted average, highest risk principle). At the same time, the system automatically generates an explainability report, clearly listing one or more key reasons for the high risk determination (for example: "triggering'single certificate' risk: missing traffic police accident identification paper"; "triggering 'information contradiction' risk: invoice amount does not match contract amount"). The result is output through API or interface, driving subsequent manual review, automatic rejection or supplementary material process.

[0082] Embodiment 2: An intelligent risk identification system based on multi-modal material identification and policy inference.

[0083] Please refer to Figure 4 which shows the composition block diagram of the system of the present application. The system includes an input interface layer, a core processing layer and an output interface layer.

[0084] The input interface layer is used to receive multi-modal raw materials such as step S100 and perform necessary preprocessing. It is responsible for the access, buffering and preliminary cleaning of multi-source heterogeneous data, providing a unified format of data stream for downstream processing.

[0085] The core processing layer includes a material identification module, a material semantic modeling module, a scene element extraction module and a policy inference engine.

[0086] The material identification module is used to perform step S200, integrates CV, NLP and other models, and is responsible for material classification and feature vectorization.

[0087] The material semantic modeling module is configured to perform step S300, has a graph database or in-memory graph structure management capability, constructs a material semantic graph, and generates a graph structure feature vector.

[0088] The scene element extraction module is configured to perform step S400 and generate a scene feature vector.

[0089] The strategy inference engine is configured to perform step S500 and is an intelligent decision center of the system.

[0090] As shown in Figure 5 The strategy inference engine further includes an evidence chain integrity analysis module, a multi-channel verification module, and a strategy node graph composed of multiple strategy nodes. The models work together to realize probabilistic risk inference.

[0091] The output interface layer is configured to perform step S600 and output a risk label and an explainability report to a downstream system or a user interface.

[0092] Embodiment 3: Application Example.

[0093] Taking the car insurance claim fraud scene as an example:

[0094] The user submits materials: a driver's license picture, a close-up picture of a damaged vehicle, and a repair shop quote picture.

[0095] The material recognition module of the system identifies three types of materials and extracts features.

[0096] The material semantic modeling module of the system constructs a semantic graph, establishes edges between "driver's license" and "quote" (dependence relationship between vehicle owner and repair), and between "damage photo" and "quote" (corroboration relationship between loss and quote). At the same time, the evidence chain integrity analysis module detects the missing "on-site panoramic photo" and "police accident determination" nodes that can prove the authenticity of the accident, and gives a low integrity score.

[0097] The scene element extraction module of the system extracts repair items and high-value quote information from the text.

[0098] The strategy inference engine works: the multi-channel verification module finds that the texture of the damaged part in the close-up photo is slightly different from the common parts of the vehicle model; in the strategy node graph, the "incomplete document risk" node is triggered due to the low integrity score, and the "loss exaggeration risk" node obtains a high trigger probability due to the implicit contradiction between the photo and the quote and the high-value text features.

[0099] The output interface layer of the system outputs a "high risk" label, and the report lists: "main doubts: 1. Key accident proof materials are missing; 2. The loss part details do not match the vehicle model standard, manual review is recommended."

[0100] The application is described above with reference to the drawings, and it is obvious that the implementation of the application is not limited by the above manner, as long as various improvements are made by using the method concept and technical scheme of the application, or the concept and technical scheme of the application is directly applied to other occasions without improvement, which are all within the protection scope of the application.

Claims

1. An intelligent risk identification method based on multi-modal material identification and strategy inference, characterized in that, The method comprises the following steps: receiving multi-modal input materials, the materials comprising image materials, text materials and structured data; performing feature extraction and recognition on the multi-modal input materials to obtain type labels and corresponding feature vectors of the materials; constructing a material semantic graph based on the recognized materials and the feature vectors, the material semantic graph taking the materials as nodes and establishing edges between the nodes based on predefined semantic association relationships, the semantic association relationships comprising at least one of type hierarchical relationship, dependency relationship, complementary relationship, substitution relationship, exclusion relationship, sequence relationship and aggregation relationship; performing graph structure coding on the material semantic graph to generate a graph structure feature vector representing the overall topology and semantics of the material set; extracting and fusing scene elements to generate a scene feature vector; inputting the graph structure feature vector and the scene feature vector into a strategy inference engine for risk inference, the strategy inference process comprising: performing structural analysis on the material semantic graph by an evidence chain integrity analysis module to output an integrity score, performing cross-modal information alignment verification by a multi-channel consistency verification module to output a consistency conflict score, and calculating a trigger probability of each node in a strategy node graph based on the graph structure feature vector, the scene feature vector, the integrity score and the consistency conflict score by a strategy node probability calculation module; and generating and outputting a risk label and an explainability report based on the trigger probability, the integrity score and the consistency conflict score.

2. The intelligent risk identification method based on multi-modal material identification and strategy inference according to claim 1, characterized in that: The feature extraction and recognition on the multi-modal input materials specifically comprise: recognizing the image materials by a computer vision (CV) classification model to output material type labels and visual feature vectors; performing semantic understanding on the text materials by a natural language processing (NLP) model to output text semantic feature vectors, and vectorizing and encoding the structured data to output structured feature vectors. 3.The intelligent risk identification method based on multi-modal material identification and strategy inference according to claim 1, characterized in that: The graph structure coding on the material semantic graph is specifically implemented by any one or combination of graph neural network (GNN), graph attention network (GAT), topology coding or path coding. 4.The intelligent risk identification method based on multi-modal material identification and strategy inference according to claim 1, characterized in that: The algorithm model adopted by the strategy node probability calculation module for calculating the trigger probability comprises any one or combination of parameterized scoring function, graph structure message passing model, Bayesian inference model or rule-based scoring normalization model.

5. A computer readable storage medium storing a computer program, characterized in that: The computer program, when executed by a processor, implements the intelligent risk identification method based on multi-modal material recognition and strategy inference as claimed in any one of claims 1 to 4.

6. An electronic device, comprising: comprise: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the intelligent risk identification method based on multi-modal material recognition and strategy inference as claimed in any one of claims 1 to 4.

7. An intelligent risk identification system based on multi-modal material identification and strategy inference, characterized in that, comprise: an input interface layer for receiving multi-modal input materials; a core processing layer connected with the input interface layer, comprising: a material recognition module for performing feature extraction and recognition on the multi-modal input materials to obtain type labels and corresponding feature vectors of the materials; The material semantic modeling module is configured to construct a material semantic graph based on the identified materials and the feature vectors, and to perform graph structure coding on the material semantic graph to generate a graph structure feature vector, wherein the material semantic graph establishes edges between nodes according to semantic association relationships including at least one of type hierarchical relationship, dependency relationship, complementary relationship, substitution relationship, exclusion relationship, sequence relationship and aggregation relationship; The scene element extraction module is configured to extract and fuse scene elements to generate a scene feature vector; and The policy inference engine is configured to perform risk inference based on the graph structure feature vector and the scene feature vector, and includes the following modules inside: an evidence chain integrity analysis module configured to perform structural analysis on the material semantic graph and output an integrity score; a multi-channel consistency verification module configured to perform cross-modal information alignment verification and output a consistency conflict score; and a policy node probability calculation module configured to maintain a policy node graph, and calculate a trigger probability of each policy node based on the graph structure feature vector, the scene feature vector, the integrity score and the consistency conflict score; and An output interface layer connected to the core processing layer is configured to generate and output a risk label and an explainability report based on the trigger probability, the integrity score and the consistency conflict score. 8.The intelligent risk identification system based on multi-modal material identification and strategy inference of claim 7, wherein: The material recognition module integrates a computer vision (CV) classification model, a natural language processing (NLP) model and a structured data encoder.

9. The intelligent risk identification system based on multi-modal material identification and policy inference according to claim 8, characterized in that: The graph structure encoder integrated in the material semantic modeling module is implemented by using any one or a combination of graph neural network (GNN), graph attention network (GAT), topology coding or path coding.

10. The intelligent risk identification system based on multi-modal material identification and policy inference according to claim 9, characterized in that: The algorithm model used by the policy node probability calculation module to calculate the trigger probability includes any one or a combination of parameterized scoring function, graph structure message passing model, Bayesian inference model or rule-based scoring normalization model.

Images