Intelligent risk identification and analysis method based on multi-modal heterogeneous data fusion

By combining deep neural networks with symbolic reasoning, a multimodal heterogeneous data fusion model is constructed, which solves the problems of semantic consistency and inaccurate identification of risk propagation paths in heterogeneous data fusion, realizes efficient, accurate and quantitative evaluation of risk identification and analysis, and improves the scientific nature and timeliness of risk management.

CN120765028APending Publication Date: 2025-10-10HANGZHOU ZHUIXING VIDEO TECH CO LTD

Patent Information

Application Number
CN202511014340.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-23
Publication Date
2025-10-10

AI Technical Summary

Technical Problem

Existing risk identification and analysis methods have difficulty ensuring semantic consistency during the process of heterogeneous data fusion, cannot effectively capture the complementary relationship between multimodal data, are not accurate enough in identifying risk transmission paths, and lack a quantitative assessment mechanism, which reduces the effectiveness and pertinence of risk management.

Method used

Deep neural networks and symbolic reasoning are used to jointly process multi-source data, a symbolic semantic coding mapping mechanism is constructed, features are extracted through orthogonal matrix decomposition and bidirectional long short-term memory networks, a graph embedding algorithm is applied to construct a risk propagation topological network, the node entropy value time series sequence is calculated, and the degree of risk impact is quantified.

Benefits of technology

It achieves efficient fusion of multimodal heterogeneous data, accurately extracts spatial dimensions and temporal features, improves the accuracy of risk identification and the precision of transmission paths, provides a scientific basis for decision-making, and significantly improves the timeliness and accuracy of risk warnings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120765028A_ABST
    Figure CN120765028A_ABST
Patent Text Reader

Abstract

The invention provides an intelligent risk identification and analysis method based on multi-modal heterogeneous data fusion, which relates to the technical field of risk management, and comprises the following steps of: acquiring multi-source data, constructing symbol-nerve double-space processing, generating semantic vectors, extracting features through orthogonal matrix decomposition and a bidirectional long-short-term memory network, and obtaining an intelligent risk identification and analysis result; according to the method, multi-modal information is fused by using an adaptive weight mechanism, and a risk propagation topology network is constructed by applying a graph embedding algorithm, so that accurate identification and early warning of risks are realized, and the risk prevention and control capability and prediction accuracy are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk management, and in particular to a risk intelligent identification and analysis method based on multi-modal heterogeneous data fusion. BACKGROUND

[0002] With the rapid development of information technology, multi-source heterogeneous data presents explosive growth, providing massive information resources for risk identification and early warning. Modern risk management needs to integrate structured data (such as numerical indicators, financial statements), unstructured data (such as text, images) and semi-structured data (such as XML files, JSON format data) and other multi-modal information for comprehensive analysis. Traditional risk identification methods mainly rely on a single data source or simple feature engineering, which is difficult to fully mine the deep associations between multi-dimensional data. In recent years, methods combining deep learning and symbolic reasoning have been gradually applied to risk identification in complex scenarios, building risk propagation networks through semantic modeling and feature fusion to provide decision support for risk prediction and prevention.

[0003] There are technical difficulties in the current field of risk intelligent identification and analysis. Semantic consistency is difficult to guarantee in the process of heterogeneous data fusion. Existing methods often independently process each modal data and then simply splice them, resulting in semantic information misplacement and inability to effectively capture the complementary relationship between multi-modal data. Risk propagation path identification is not accurate enough. Traditional models fail to fully model the dynamic associations between risk nodes, have limited understanding of the formation mechanism and evolution law of risk chains, and cannot track the transmission and diffusion process of risks in complex scenarios. Existing technologies lack a quantitative evaluation mechanism for the degree of risk influence, have deficiencies in multi-dimensional risk index construction, and cannot provide intuitive and operable risk prevention and control recommendations for decision makers, reducing the effectiveness and pertinence of risk management. SUMMARY

[0004] The embodiments of the present application provide a risk intelligent identification and analysis method based on multi-modal heterogeneous data fusion, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a risk intelligent identification and analysis method based on multi-modal heterogeneous data fusion, comprising: Collecting structured data, unstructured data and semi-structured data to construct a multi-source data set; Performing joint processing of deep neural networks and symbolic reasoning on the multi-source data set, constructing a symbolic semantic encoding mapping mechanism, projecting data features into a symbolic-neural dual space, performing feature extraction based on symbolic guidance constraints, and generating semantic vectors; Obtaining spatial dimension features by orthogonal matrix decomposition of the semantic vectors, extracting time sequence features through a bidirectional long short-term memory network, fusing multi-modal information based on an adaptive weight and mutual information alignment mechanism, and generating a fusion vector; Applying a graph embedding algorithm to the fusion vector to construct a risk propagation topology network, calculating the node entropy value time series, extracting topological dynamic features through mutation detection and synchronization analysis, and forming risk indicators; Calculate risk identification results based on the risk indicators, map risk transmission paths, quantify risk impact, and output early warning recommendations.

[0006] In an optional embodiment, the multi-source dataset is processed jointly by a deep neural network and symbolic reasoning, a symbolic semantic encoding mapping mechanism is constructed, data features are projected into a symbolic-neural dual space, feature extraction is performed based on symbolic guided constraints, and the generation of semantic vectors includes: Use the preset neural network to extract features from multi-source data and obtain the initial feature vector; Build a domain ontology knowledge base based on semantic attributes and semantic rules to obtain the structured organization form of domain concepts and the semantic association rules between concepts; Performing a structured transformation on the initial feature vector according to the semantic attributes and the semantic rules to generate a symbolic space representation with an explicit semantic structure; Decomposing and mapping the semantic information in the symbolic space representation into the neural space to obtain semantic decoupling features, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain semantic recombinant features; Calculating the semantic consistency loss between the semantic decoupling feature and the semantic reorganization feature based on the semantic association rule, and performing joint optimization based on the semantic consistency loss to obtain an optimization parameter; The symbolic space representation is processed using the optimization parameters to output a semantic vector.

[0007] In an optional embodiment, decomposing and mapping the semantic information in the symbolic space representation into a neural space to obtain a semantic decoupling feature, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain a semantic reconstructed feature includes: receiving the symbolic space representation and constructing a semantic association matrix based on the semantic attributes and the semantic rules, wherein each element in the semantic association matrix corresponds to the semantic association strength between the features; Based on the structural relationship of the semantic association matrix, the symbolic space representation is decomposed into a plurality of semantic substructures, each of the semantic substructures corresponding to local structural information of a semantic attribute; Performing feature decomposition on each of the semantic substructures, extracting feature vectors and corresponding eigenvalues, sorting the eigenvalues ​​from large to small, calculating the proportion of each eigenvalue to the sum of all eigenvalues, selecting eigenvectors whose sum of eigenvalue proportions is greater than a preset feature threshold, and constructing a semantic basis vector group; Performing orthogonal projection decomposition on the semantic basis vector group to obtain multiple groups of semantic components, each group of semantic components corresponding to an independent semantic attribute, thereby obtaining a semantic decoupling feature; An initial feature vector is received, a projection matrix is ​​constructed based on the semantic basis vector group, the initial feature vector is multiplied by the projection matrix to obtain projection coordinates in a symbolic space, and a semantic reorganization feature is obtained based on a linear combination of the projection coordinates and the semantic basis vector group.

[0008] In an optional embodiment, the semantic vector is subjected to orthogonal matrix decomposition to obtain spatial dimension features, and temporal features are extracted through a bidirectional long short-term memory network. Multimodal information is fused based on an adaptive weight and mutual information alignment mechanism to generate a fused vector, which includes: Performing orthogonal matrix decomposition on the semantic vector to obtain a basis vector matrix, projecting the semantic vector into the feature space using the basis vector matrix to obtain a spatial dimension projection matrix, setting segmentation intervals for the spatial dimension projection matrix according to the feature dimension, and combining features within each segmentation interval to form a spatial slice of the semantic vector; Dynamically construct an adaptive partitioning matrix based on the semantic similarity of features in the spatial slice and perform a nonlinear transformation to obtain a local attention weight matrix; cluster the local features in the spatial slice, calculate the feature center of each cluster, perform feature mapping based on the feature center, and construct a global relevance weight matrix; concatenate the local attention weight matrix with the global relevance weight matrix to calculate the adaptive weight coefficient; Inputting the semantic vector into a bidirectional long short-term memory network to obtain temporal hidden layer features, performing wavelet decomposition on the temporal hidden layer features to obtain multi-scale temporal features; Calculating the mutual information between the spatial features weighted by the adaptive weight coefficients and the multi-scale temporal features, and constructing an alignment loss function including a mutual information maximization term and a semantic transformation matrix constraint term; Based on the alignment loss function, optimizing the semantic transformation matrix corresponding to the spatial feature and the multi-scale temporal feature to obtain a semantic alignment feature; The semantic alignment features are passed through the gating unit to calculate the fusion weight and obtain the fusion vector.

[0009] In an optional embodiment, an adaptive partitioning matrix is ​​dynamically constructed based on the semantic similarity of features in the spatial slices, and a nonlinear transformation is performed to obtain a local attention weight matrix including: Mapping the spatial slice to a high-dimensional feature space to construct a feature association graph, wherein the nodes of the feature association graph represent feature vectors in the spatial slice, and the edge weights represent the semantic similarity between the feature vectors; performing spectral clustering on the feature correlation graph, obtaining eigenvectors of a graph Laplacian matrix based on eigenvalue decomposition, and sorting the eigenvectors according to corresponding eigenvalues; According to the sorted eigenvectors, the degree of semantic jump between adjacent features in the spatial slice is calculated, and the position with a semantic jump degree greater than a preset jump degree threshold is determined as a partition boundary; Based on the partition boundary, the spatial slice is divided into multiple sub-regions, the variance of the eigenvectors in each sub-region is calculated, and the partition range of the sub-region is adjusted according to the variance; Performing nonlinear transformation in each sub-region, constructing an adaptive partition matrix, the elements of the adaptive partition matrix are determined by the cosine similarity of the eigenvectors in the sub-region and the spatial distance of the eigenvectors; Multiply the adaptive partition matrix by the sub-region eigenvectors to obtain a local attention weight matrix.

[0010] In an alternative embodiment, the fusion vector is applied to a graph embedding algorithm, a risk propagation topology network is constructed, a node entropy time sequence is calculated, and a topology dynamic feature is extracted through mutation detection and synchronicity analysis to form a risk indicator, including: The fusion vector is mapped to a low-dimensional embedding space through a knowledge graph relationship mapping algorithm, the knowledge graph relationship mapping algorithm includes constructing a distance metric function of entities and relationships and minimizing the representation error, and a risk propagation topology network is constructed based on the semantic distance between entities in the low-dimensional embedding space; A time series observation window is constructed for each node in the risk propagation topology network, and the node degree centrality in the time series observation window is calculated; Determine the initial binning boundary based on the probability density function of the node degree centrality, obtain the probability distribution sequence by adaptively merging the binning interval, and calculate the node entropy sequence based on the probability distribution sequence; Apply variational modal decomposition to the node entropy sequence to obtain an adaptive intrinsic modal function, calculate the instantaneous frequency spectrum through Hilbert-Huang transform, and identify the mutation points and trend turning points in the instantaneous frequency spectrum through singular spectrum analysis to obtain a mutation feature sequence of the risk propagation process; Based on the node entropy sequence, a time series correlation matrix is constructed and spectral decomposition is performed to obtain a network synchronization mode, a synchronization strength matrix between node pairs is calculated, and a synchronization node subgroup is obtained by grouping the synchronization strength matrix through community division; Calculate the intermediate centrality and clustering coefficient of the nodes in the risk propagation topology network, and combine the intermediate centrality, the clustering coefficient, the node entropy sequence, the mutation feature, and the synchronization strength matrix of the synchronization node subgroup to construct a multi-dimensional risk indicator.

[0011] In an optional embodiment, determining the initial binning boundaries based on the probability density function of the node degree centrality, obtaining a probability distribution sequence by adaptively merging binning intervals, and calculating the node entropy value sequence based on the probability distribution sequence includes: The kernel density estimation method is used to calculate the probability density function of the node degree centrality, and the probability density function is calculated by the Gaussian kernel function and the bandwidth parameter; Determine the derivative of the probability density function, identify the local maximum point and the local minimum point of the probability density function, use the local maximum point and the local minimum point as initial binning boundaries, and determine the binning interval; Calculate the information gain ratio before and after the adjacent bin intervals are merged, where the information gain ratio is calculated by the ratio of information gain to split information. When the information gain ratio is less than a preset gain ratio threshold, merge the adjacent bin intervals to obtain an adaptive bin boundary set. The node degree centrality is mapped into a discretized sequence based on the adaptive bin boundary set, the probability distribution of samples in each bin interval is calculated, and a node entropy value sequence is obtained based on the probability distribution.

[0012] A second aspect of an embodiment of the present invention provides a risk intelligent identification and analysis system based on multimodal heterogeneous data fusion, comprising: The first unit is used to collect structured data, unstructured data and semi-structured data to build multi-source data sets; The second unit is used to perform deep neural network and symbolic reasoning joint processing on the multi-source data set, build a symbolic semantic encoding mapping mechanism, project data features into the symbolic-neural dual space, perform feature extraction based on symbolic guidance constraints, and generate semantic vectors; The third unit is used to obtain spatial dimension features of the semantic vector through orthogonal matrix decomposition, extract temporal features through a bidirectional long short-term memory network, fuse multimodal information based on adaptive weights and mutual information alignment mechanism, and generate a fused vector; The fourth unit is used to apply a graph embedding algorithm to the fusion vector, construct a risk propagation topology network, calculate the time series of node entropy values, extract topological dynamic features through mutation detection and synchronization analysis, and form risk indicators; The fifth unit is used to calculate the risk identification results based on the risk indicators, map the risk transmission path, quantify the risk impact, and output early warning suggestions.

[0013] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] In an embodiment of the present invention, a risk intelligent identification and analysis method based on multimodal heterogeneous data fusion is used to realize the comprehensive processing of structured, unstructured and semi-structured data, and a symbolic semantic coding mapping mechanism is constructed to effectively project the characteristics of multi-source heterogeneous data into the symbol-neural dual space, thereby improving data utilization efficiency and analysis accuracy. A technical route combining orthogonal matrix decomposition and bidirectional long short-term memory network is adopted to accurately extract spatial dimension features and temporal features, and through adaptive weight and mutual information alignment mechanism, deep fusion of multimodal information is realized, thereby enhancing the model's perception and adaptability to complex risk scenarios. A risk propagation topological network is constructed and mutation detection and synchronization analysis are applied to extract dynamic features, thereby realizing accurate identification of risks, mapping of propagation paths and quantification of impact levels, providing a scientific basis for management decisions, significantly improving the timeliness and accuracy of risk warnings, and reducing enterprise operating risks and decision-making costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the process of intelligent risk identification and analysis method based on multimodal heterogeneous data fusion according to an embodiment of the present invention; Figure 2 This is a performance comparison chart of semantic decoupling technology on different datasets; Figure 3 This is a sensitivity analysis diagram of the bandwidth parameter of the adaptive binning method. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0018] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1FIG. 1 is a flow chart of a method for intelligent risk identification and analysis based on multimodal heterogeneous data fusion according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collect structured data, unstructured data, and semi-structured data to build multi-source data sets; Performing a joint process of deep neural network and symbolic reasoning on the multi-source dataset, constructing a symbolic semantic encoding mapping mechanism, projecting data features into a symbolic-neural dual space, performing feature extraction based on symbolic guided constraints, and generating semantic vectors; The semantic vector is decomposed into spatial dimension features through orthogonal matrix decomposition, and the temporal features are extracted through a bidirectional long short-term memory network. The multimodal information is fused based on the adaptive weight and mutual information alignment mechanism to generate a fused vector; Applying a graph embedding algorithm to the fusion vector to construct a risk propagation topology network, calculating the node entropy value time series, extracting topological dynamic features through mutation detection and synchronization analysis, and forming risk indicators; Calculate risk identification results based on the risk indicators, map risk transmission paths, quantify risk impact, and output early warning recommendations.

[0020] In one specific embodiment, structured, unstructured, and semi-structured data are collected to construct a multi-source dataset. Specifically, the collection of structured data involves extracting data with well-defined fields and formats from relational databases, such as risk event records, transaction flows, balance sheets, and other financial indicator data. The collection of unstructured data involves crawling text content such as news reports, social media comments, and research reports, while also collecting relevant images, audio, and video materials. The collection of semi-structured data primarily targets network data in JSON and XML formats, log files, and partially organized email content. After collection is complete, all data is cleaned to remove duplicates, missing values, and outliers, and timestamps are aligned and the format is standardized to form a multi-source dataset suitable for subsequent analysis.

[0021] Deep neural networks and symbolic reasoning are combined to process multi-source datasets. A symbolic semantic encoding mapping mechanism is constructed. Pre-trained deep neural network models are used to perform preliminary feature extraction on different data types. Context-sensitive word vector representations are generated for text data using the BERT model, visual features are extracted from image data using convolutional networks such as ResNet, and temporal dependencies are captured using recurrent neural networks for time series data. Symbolic reasoning is introduced to represent domain knowledge as formal rules, such as the causal relationship of risk propagation and the hierarchical structure of risk categories. A symbolic-neural interface layer is constructed to map low-level features extracted by the neural network to high-level concepts in the symbols. An attention mechanism is used to highlight feature regions related to specific symbolic concepts. During feature extraction, symbolic rules are used to constrain and guide the neural network's computational process. For example, the weights of relevant features are strengthened based on the domain knowledge of risk assessment, while the influence of irrelevant features is suppressed. This symbol-guided feature extraction process projects the raw data into the symbolic-neural dual space, forming a vector representation rich in semantic information.

[0022] Semantic vectors are decomposed into spatial features using orthogonal matrix decomposition. Singular value decomposition (SVD) is then used to decompose the matrix into orthogonal basis vectors and corresponding singular values. Basis vectors with large singular values ​​are selected as the primary spatial features. These features reflect the distribution and variation of data across different dimensions. A bidirectional long short-term memory (LSTM) network is used to extract temporal features. Semantic vectors are fed into the network in chronological order. Through forward and backward information flow, the network captures long-term dependencies and temporal evolution patterns within the data sequence. The forget gate, input gate, and output gate within the network work together to selectively retain or update historical information, forming a complete representation of temporal changes. Multimodal information is fused using an adaptive weighting mechanism, dynamically adjusting weight coefficients based on the quality and relevance of different modal data to perform a weighted combination of spatial and temporal features. A mutual information alignment mechanism is introduced to calculate mutual information between features from different modalities, identify areas of information redundancy and complementarity, and optimize the feature fusion process. Through this series of processes, multimodal and multidimensional information is integrated into a unified fusion vector, preserving the core information of each data type while eliminating redundancy and noise.

[0023] A graph embedding algorithm is applied to the fused vectors to construct a risk propagation topological network. The fused vectors are mapped into a low-dimensional space, preserving the similarity between the vectors to form node representations. Based on the strength of the associations between nodes, a weighted directed graph structure is constructed, in which nodes represent risk entities, edges represent risk propagation paths, and edge weights reflect the propagation probability. To capture the dynamic evolution of the network, a time series of entropy values ​​is calculated for each node. The node entropy value reflects the uncertainty in the distribution of edge weights connected to that node. Higher entropy values ​​indicate diverse and uncertain risk propagation paths, while lower entropy values ​​indicate relatively stable risk propagation paths. A mutation detection algorithm is applied to the time series of entropy values ​​to identify significant changes in entropy values, which typically correspond to transitions in risk status. Synchronicity analysis is performed to calculate the correlation and phase differences between entropy value sequences at different nodes, revealing the risk propagation pattern and degree of synchronization within the network. These topological dynamic features comprehensively reflect the distribution and evolutionary trends of risks within the network, forming a risk indicator system.

[0024] Risk identification results are calculated based on risk indicators and compared with historical risk event data to identify current abnormal patterns and potential risk points. Graph analysis algorithms are used to map risk transmission paths, identify key transmission nodes and vulnerable links, and predict the possible direction and speed of risk diffusion. Sensitivity analysis methods are used to quantify the impact of risks, assess the weight of the impact of different risk factors on overall stability, and distinguish between high-impact factors and secondary factors. Based on the comprehensive analysis results, hierarchical early warning recommendations are generated, including risk warning information, risk level assessment, intervention measures recommendations, and monitoring priorities. Early warning recommendations are prioritized according to the risk's urgency, scope of impact, and development trend, providing decision makers with timely and targeted risk management support.

[0025] In an optional embodiment, the multi-source dataset is processed jointly by a deep neural network and symbolic reasoning, a symbolic semantic encoding mapping mechanism is constructed, data features are projected into a symbolic-neural dual space, feature extraction is performed based on symbolic guided constraints, and the generation of semantic vectors includes: Use the preset neural network to extract features from multi-source data and obtain the initial feature vector; Build a domain ontology knowledge base based on semantic attributes and semantic rules to obtain the structured organization form of domain concepts and the semantic association rules between concepts; Performing a structured transformation on the initial feature vector according to the semantic attributes and the semantic rules to generate a symbolic space representation with an explicit semantic structure; Decomposing and mapping the semantic information in the symbolic space representation into the neural space to obtain semantic decoupling features, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain semantic recombinant features; The semantic consistency loss between the semantic decoupling features and the semantic recombination features is calculated based on a semantic association rule, and the optimization parameters are obtained by jointly optimizing the semantic consistency loss; The symbol space representation is processed using the optimization parameters, and a semantic vector is output.

[0026] In a specific embodiment, a pre-set deep neural network is used to extract features from multi-source data to obtain an initial feature vector. Specifically, for input image, text and video multi-source data, convolutional neural network, recurrent neural network and three-dimensional convolutional neural network are used for processing respectively. For example, for monitoring image data, ResNet50 network is used to extract features, for police report text data, BERT model is used to extract semantic features, and for monitoring video data, C3D network is used to extract spatio-temporal features. The features extracted by these networks are normalized into a unified dimensional vector, such as a 512-dimensional feature vector, as the initial feature representation for subsequent processing.

[0027] A domain ontology knowledge base is constructed based on domain semantic attributes and semantic rules. In practical applications, taking public security monitoring analysis as an example, the domain ontology includes a "scene-behavior-risk" three-layer semantic structure. The scene includes attributes such as "traffic intersection, shopping mall, school", the behavior includes attributes such as "gather, run, fight", and the risk includes attributes such as "low risk, medium risk, high risk, emergency". The semantic rules define the association relationship between these concepts, for example, "school + gather + low risk" constitutes an effective semantic combination, while "traffic intersection + fight + low risk" is an invalid combination. In this way, the domain knowledge is explicitly organized into a structured form, which is convenient for subsequent symbolic operation and reasoning.

[0028] The initial feature vector is executed structured conversion according to semantic attributes and semantic rules to generate a symbol space representation with explicit semantic structure. In a specific implementation, a feature projection network is designed, which consists of three fully connected layers, which maps the initial feature vector to the symbol space. The output dimension of the first fully connected layer is 256, using ReLU activation function; the output dimension of the second fully connected layer is 128, using ReLU activation function; the output dimension of the third fully connected layer is the total number of attributes defined in the domain ontology, using Sigmoid activation function. Through this network, the initial feature vector is converted into a probability distribution vector, where each dimension corresponds to a semantic attribute in the ontology. For example, in the public security monitoring analysis application, a traffic intersection monitoring video after processing may obtain the following symbol space representation: [0.92, 0.04, 0.03, 0.12, 0.78, 0.05, 0.08, 0.21, 0.67, 0.04], corresponding to the probability values of "traffic intersection, shopping mall, school, gather, run, fight, low risk, medium risk, high risk, emergency".

[0029] The semantic information in the symbolic space representation is decomposed and mapped to the neural space to obtain semantic decoupling features. In this embodiment, a semantic decoder network is designed, which includes two fully connected layers. The output dimension of the first layer is 128, and the ReLU activation function is used; the output dimension of the second layer is the same as the initial feature vector, which is 512. Through this decoder, the representation of the symbolic space is mapped back to the neural space to obtain semantic decoupling features. At the same time, the initial feature vector is mapped to the symbolic space for semantic reconstruction to obtain semantic reconstructed features. Specifically, an encoder network is designed, which has a structure similar to the aforementioned feature projection network, but adds a reconstruction layer to remap the symbolic space representation to a 512-dimensional vector as a semantic reconstructed feature.

[0030] The semantic consistency loss between semantically decoupled features and semantically reassembled features is calculated based on semantic association rules. In practical applications, semantic consistency loss consists of two parts: vector similarity loss and semantic rule constraint loss. Vector similarity loss uses cosine similarity to calculate the difference between two feature vectors. A smaller value indicates closer proximity in the semantic space. Semantic rule constraint loss, based on the semantic rules defined in the domain ontology, penalizes combinations in the symbolic space representation that do not conform to the rules. For example, if "traffic intersection + fight + low risk" is an invalid combination, a corresponding penalty term is added when all three attributes in the feature vector are detected to have high probability values. These two losses are combined to form the overall semantic consistency loss.

[0031] Joint optimization is performed based on the semantic consistency loss to obtain optimized parameters. During training, the Adam optimizer is used with a learning rate of 0.001, a batch size of 64, and 50 training iterations. The optimization goal is to minimize the semantic consistency loss. The parameters of the feature projection network, semantic decoder, and encoder are updated via backpropagation. After training converges, these optimized network parameters are saved for subsequent feature processing.

[0032] The symbolic space representation is processed using optimized parameters to output a semantic vector. In practical applications, when new multi-source data is input, an initial feature vector is first extracted using a preset neural network. This is then mapped to the symbolic space using an optimized feature projection network to obtain a symbolic space representation. Finally, this representation is processed through a semantic integration network consisting of two fully connected layers. The first layer has an output dimension of 256 and uses the ReLU activation function; the second layer outputs the desired semantic vector dimension, such as 384. The resulting semantic vector combines the underlying features extracted by the neural network with the high-level semantic structure introduced by symbolic reasoning, and can be used for subsequent tasks such as risk assessment and emergency response.

[0033] In this embodiment, by extracting features from multi-source data through a preset neural network, the complementary information of data of different modalities or sources can be fully mined, thereby improving the overall representation capability; by constructing a domain ontology based on semantic attributes and semantic rules, the experience of security experts and prior knowledge can be structuredly integrated into the model, which helps the model understand concepts in the public safety field and their interrelationships; the initial feature vector is structured and converted into a symbolic space representation with an explicit semantic structure, making the internal representation of the model more interpretable and traceable; by mapping semantic information to the neural space to obtain decoupled features and the symbolic space to obtain recombined features, "semantic separation" and "semantic reconstruction" in representation learning are achieved, which helps to reduce interference between different semantic dimensions and more accurately identify and warn of public safety hazards.

[0034] In an optional embodiment, decomposing and mapping the semantic information in the symbolic space representation into a neural space to obtain a semantic decoupling feature, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain a semantic reconstructed feature includes: receiving the symbolic space representation and constructing a semantic association matrix based on the semantic attributes and the semantic rules, wherein each element in the semantic association matrix corresponds to the semantic association strength between the features; Based on the structural relationship of the semantic association matrix, the symbolic space representation is decomposed into a plurality of semantic substructures, each of the semantic substructures corresponding to local structural information of a semantic attribute; Performing feature decomposition on each of the semantic substructures, extracting feature vectors and corresponding eigenvalues, sorting the eigenvalues ​​from large to small, calculating the proportion of each eigenvalue to the sum of all eigenvalues, selecting eigenvectors whose sum of eigenvalue proportions is greater than a preset feature threshold, and constructing a semantic basis vector group; Performing orthogonal projection decomposition on the semantic basis vector group to obtain multiple groups of semantic components, each group of semantic components corresponding to an independent semantic attribute, thereby obtaining a semantic decoupling feature; An initial feature vector is received, a projection matrix is ​​constructed based on the semantic basis vector group, the initial feature vector is multiplied by the projection matrix to obtain projection coordinates in a symbolic space, and a semantic reorganization feature is obtained based on a linear combination of the projection coordinates and the semantic basis vector group.

[0035] In one specific embodiment, during semantic decoupling, symbolic space representation data is received. This data can be a symbolic space representation vector of multimodal information such as text, images, or audio. Based on a predefined set of semantic attributes (such as an object's shape, color, and material) and semantic rules (such as the rule that shape and color are independent of each other), a semantic association matrix is ​​constructed. This matrix is ​​an n×n square matrix, where n is the symbolic space dimension. Each element a_ij in the matrix represents the strength of the semantic association between the i-th feature and the j-th feature. For example, in a 10-dimensional symbolic space, if the first and third dimension features both belong to the semantic attribute describing color, then a_13 might be 0.85, indicating a strong semantic association between them.

[0036] After constructing the semantic association matrix, the structural relationships within the matrix are analyzed, and the symbolic space representation is decomposed into multiple semantic substructures. Specifically, spectral clustering is used to group features with similar semantic association strengths into the same substructure. For example, in a symbolic space containing three semantic attributes: color, shape, and size, the features might be divided into three substructures: S1 = {feature 1, feature 4, feature 7} (color-related), S2 = {feature 2, feature 5, feature 8} (shape-related), and S3 = {feature 3, feature 6, feature 9} (size-related).

[0037] For each semantic substructure, eigendecomposition is performed to extract the key semantic information. Taking substructure S1 as an example, eigenvectors v1, v2, and v3 and their corresponding eigenvalues ​​λ1 = 0.8, λ2 = 0.15, and λ3 = 0.05 are extracted. The eigenvalues ​​are sorted from largest to smallest, and the proportion of each eigenvalue to the total is calculated: p1 = 0.8 / 1.0 = 0.8, p2 = 0.15 / 1.0 = 0.15, and p3 = 0.05 / 1.0 = 0.05. If the preset feature threshold is 0.9, eigenvectors v1 and v2 whose sum of eigenvalue proportions is greater than 0.9 are selected to construct the semantic basis vector set B1 = {v1, v2} for this substructure. Similarly, corresponding semantic basis vector sets B2 and B3 are constructed for the other substructures. Ultimately, the complete semantic basis vector set B = {B1, B2, B3} is obtained.

[0038] After obtaining the semantic basis vector group, the original symbol space representation is decomposed by orthogonal projection on these basis vectors. Assuming the original representation is vector x, for each basis vector v in the semantic basis vector group B1, calculate the projection component c of x on v =<x,v> ,in<x,v> represents the inner product of vectors x and v. Thus, the original representation x is decomposed into multiple sets of semantic components: x = c1·v1+c2·v2+…+ck·vk, where k is the total number of basis vectors. Each set of semantic components corresponds to an independent semantic attribute. For example, c1·v1+c2·v2 might represent the semantic information of a color attribute. In this way, the symbolic space representation is successfully decomposed into semantically decoupled features.

[0039] When performing semantic reconstruction, an initial feature vector y is received, which may come from another data domain or representation space. Based on the previously constructed semantic basis vector group B = {B1, B2, B3}, a projection matrix P is constructed, and the column vectors of the matrix P are all the basis vectors in B. The initial feature vector y is multiplied by the projection matrix P to obtain the projection coordinate α = P of y in the symbolic space. T y, where P^T represents the transposed matrix of P. For example, for a 3D initial vector y = [0.2, 0.5, 0.3] and a projection matrix P = [[0.7, 0.1], [0.2, 0.8], [0.1, 0.1]] containing 2 basis vectors, the calculated projection coordinates are α = [0.29, 0.45].

[0040] Finally, a semantically reconstructed feature z = P·α is constructed based on a linear combination of the projected coordinate α and the semantic basis vector set B. This reconstructed feature z retains information consistent with the semantic structure of the sign space in the original feature vector y while filtering out irrelevant or noisy information. For example, if P = [[0.7, 0.1], [0.2, 0.8], [0.1, 0.1]] and α = [0.29, 0.45], then the reconstructed feature z = [0.7 × 0.29 + 0.1 × 0.45, 0.2 × 0.29 + 0.8 × 0.45, 0.1 × 0.29 + 0.1 × 0.45] = [0.248, 0.418, 0.074].

[0041] Through the above steps, the semantic information in the symbolic space representation is decomposed into decoupled features, and any initial feature vector can be mapped to the symbolic space for semantic reconstruction. This method has broad application value in tasks such as multimodal information fusion, cross-domain knowledge transfer, and semantic understanding and generation. Experiments show that compared with traditional methods, this method significantly improves the interpretability and generalization of feature representation while preserving key semantic information.

[0042] like Figure 2Figure 2 shows a comparison of the semantic decoupling accuracy of our proposed solution against four mainstream semantic decoupling methods (VAE decoupling, ICA decomposition, InfoGAN, and β-VAE) on five different datasets. The data clearly demonstrates that our proposed solution achieves significant performance advantages across all tested datasets. Specifically, on the CelebA dataset, our proposed solution achieved a decoupling accuracy of 94.3%, 13.1 percentage points higher than the second-place InfoGAN method; on the ImageNet dataset, our proposed solution achieved an accuracy of 91.7%, at least 14.9 percentage points higher than the other methods; and on the COCO, VQA-v2, and ShapeNet datasets, our proposed solution achieved accuracies of 88.5%, 86.2%, and 92.8%, respectively, significantly outperforming the compared methods. These results demonstrate that our proposed solution, through constructing a semantic association matrix and performing feature decomposition, can more effectively capture the semantic structure and correlations in data, particularly on complex and diverse datasets. The consistency advantage of the semantic decoupling capability of this technical solution on various visual datasets verifies its versatility and robustness.

[0043] In an optional embodiment, the semantic vector is subjected to orthogonal matrix decomposition to obtain spatial dimension features, and temporal features are extracted through a bidirectional long short-term memory network. Multimodal information is fused based on an adaptive weight and mutual information alignment mechanism to generate a fused vector, which includes: Performing orthogonal matrix decomposition on the semantic vector to obtain a basis vector matrix, projecting the semantic vector into the feature space using the basis vector matrix to obtain a spatial dimension projection matrix, setting segmentation intervals for the spatial dimension projection matrix according to the feature dimension, and combining features within each segmentation interval to form a spatial slice of the semantic vector; Dynamically construct an adaptive partitioning matrix based on the semantic similarity of features in the spatial slice and perform a nonlinear transformation to obtain a local attention weight matrix; cluster the local features in the spatial slice, calculate the feature center of each cluster, perform feature mapping based on the feature center, and construct a global relevance weight matrix; concatenate the local attention weight matrix with the global relevance weight matrix to calculate the adaptive weight coefficient; Inputting the semantic vector into a bidirectional long short-term memory network to obtain temporal hidden layer features, performing wavelet decomposition on the temporal hidden layer features to obtain multi-scale temporal features; Calculating the mutual information between the spatial features weighted by the adaptive weight coefficients and the multi-scale temporal features, and constructing an alignment loss function including a mutual information maximization term and a semantic transformation matrix constraint term; Based on the alignment loss function, optimizing the semantic transformation matrix corresponding to the spatial feature and the multi-scale temporal feature to obtain a semantic alignment feature; The semantic alignment features are passed through the gating unit to calculate the fusion weight and obtain the fusion vector.

[0044] In a specific embodiment, after obtaining the semantic vector, an orthogonal matrix decomposition operation is performed on the semantic vector. Specifically, the singular value decomposition technique is used to decompose the original semantic vector into the product of three matrices: a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. The right singular vector matrix is ​​the required basis vector matrix with a dimension of d×d, where d is the dimension of the original semantic vector. For example, for a 256-dimensional semantic vector, the basis vector matrix has a dimension of 256×256. Using the basis vector matrix, the original semantic vector is projected into the feature space to obtain a spatial dimension projection matrix P with a dimension of n×d, where n is the number of samples. In practical applications, the semantic vectors of 1000 samples can be projected into the feature space to obtain a spatial dimension projection matrix with a dimension of 1000×256.

[0045] For the spatial dimension projection matrix, the segmentation interval is set according to the feature dimension. Specifically, the d-dimensional feature is divided into k intervals, each containing features of d / k continuous dimensions. In this embodiment, the 256-dimensional feature is divided into 8 intervals, each containing 32 continuous features. The features within each segmentation interval are combined to form a spatial slice of the semantic vector, resulting in a total of k spatial slices, each with a dimension of n×(d / k).

[0046] After obtaining a spatial slice, an adaptive partition matrix is ​​dynamically constructed based on the semantic similarity of features within the spatial slice. Specifically, cosine similarity is calculated for features within each spatial slice to construct a similarity matrix S with dimensions (d / k) × (d / k). Based on the values ​​in the similarity matrix, a threshold t (for example, t = 0.7) is set. Features with similarity greater than t are grouped into the same partition, forming an adaptive partition matrix D. Nonlinear transformations such as the ReLU activation function are applied to this adaptive partition matrix to obtain the local attention weight matrix L, which has the same dimensions as the adaptive partition matrix.

[0047] We also perform cluster analysis on local features within spatial slices. Using the K-means algorithm, we cluster features within each spatial slice into c clusters (e.g., c = 5), and calculate the feature center for each cluster. The feature center is the average of all feature vectors within that cluster, resulting in c feature center vectors, each with a dimension of d / k. Based on these feature centers, we perform feature mapping, calculate the correlation between each feature and each feature center, and construct a global correlation weight matrix G with a dimension of (d / k) × c.

[0048] The local attention weight matrix L is concatenated with the global association weight matrix G. The concatenated matrix has a dimension of (d / k) × (d / k + c). A fully connected layer is applied to the concatenated matrix to compress it into a vector of dimension (d / k) × 1. This vector is then normalized by Softmax to obtain the adaptive weight coefficient α, which is used for subsequent feature weighting.

[0049] For temporal feature extraction, the semantic vector is input into a bidirectional long short-term memory (BiLSTM) network. The BiLSTM network consists of a forward LSTM and a backward LSTM, with the hidden layer dimension set to h (for example, h = 128). After BiLSTM processing, the temporal hidden layer features H are obtained, with dimensions n × 2h, where 2h represents the concatenation of the hidden states of the forward and backward LSTMs. Wavelet decomposition is performed on these temporal hidden layer features using Daubechies wavelet basis functions. A three-layer decomposition is performed to obtain a multi-scale temporal feature set {F1, F2, F3, F4}, corresponding to low-frequency approximation coefficients and high-frequency detail coefficients at three different scales, respectively. Each feature has a dimension of n × 2h.

[0050] During the feature alignment phase, the mutual information between spatial features weighted by adaptive weight coefficients and multi-scale temporal features is calculated. Specifically, for the spatial features A (weighted by adaptive weight coefficient α, with dimension n×d) and the temporal features Fi at each scale, their joint probability distribution and marginal probability distribution are estimated, and the mutual information value I(A, Fi) is calculated using KL divergence. Based on these mutual information values, an alignment loss function is constructed, which consists of a mutual information maximization term Lmi=-(I(A, F1)+I(A, F2)+I(A, F3)+I(A, F4)) and a semantic transformation matrix constraint term Lreg. The overall alignment loss function is L=Lmi+λLreg, where λ is a trade-off coefficient set to 0.01.

[0051] The alignment loss function is optimized using a gradient descent algorithm. The semantic transformation matrices W1 and W2 corresponding to the spatial and multi-scale temporal features are updated. The optimization is iterated for 500 rounds with a learning rate of 0.001, decreasing the learning rate every 100 rounds. After optimization, the spatial and temporal features are projected into the same semantic space using the semantic transformation matrix to obtain the semantically aligned features As and At, both with dimensions n × d.

[0052] The semantically aligned features are passed through a gating unit to calculate the fusion weight. The gating unit consists of two fully connected layers and a sigmoid activation function. Its input is the concatenation of As and At, and its output is the fusion weight β (a value between 0 and 1). The final fusion vector V is calculated using the formula V = β·As + (1-β)·At, with a dimension of n×d. This fusion vector fully preserves the information of both spatial and temporal features, and can more comprehensively express the semantic content of multimodal data.

[0053] In an actual application case, for a video understanding task, this method was used to process the features of 1,000 video clips, each with an original dimension of 256. The resulting fusion vector was able to effectively combine the visual content and temporal changes of the video, and the accuracy in downstream classification tasks was 12.3 percentage points higher than that of single-modal features.

[0054] In an optional embodiment, an adaptive partitioning matrix is ​​dynamically constructed based on the semantic similarity of features in the spatial slices, and a nonlinear transformation is performed to obtain a local attention weight matrix including: Mapping the spatial slice to a high-dimensional feature space to construct a feature association graph, wherein the nodes of the feature association graph represent feature vectors in the spatial slice, and the edge weights represent the semantic similarity between the feature vectors; Performing spectral clustering on the feature association graph, obtaining eigenvectors of the graph Laplacian matrix based on eigenvalue decomposition, and sorting the eigenvectors according to the corresponding eigenvalues; According to the sorted feature vectors, the semantic jump degree between adjacent features in the spatial slice is calculated, and the position where the semantic jump degree is greater than the preset jump degree threshold is determined as the partition boundary; Dividing the spatial slice into multiple sub-regions based on the partition boundary, calculating the variance of the feature vector in each sub-region, and adjusting the partition range of the sub-region according to the variance; Perform a nonlinear transformation in each sub-region to construct an adaptive partition matrix, wherein the elements of the adaptive partition matrix are determined by the cosine similarity of the eigenvectors in the sub-region and the spatial distance of the eigenvectors; The adaptive partition matrix is ​​multiplied by the sub-region feature vector to obtain a local attention weight matrix.

[0055] In a specific embodiment, when processing the acquired spatial slice, it is necessary to map the spatial slice to a high-dimensional feature space and construct a feature association graph. The nodes of the feature association graph represent the feature vectors in the spatial slice, and the edge weights represent the semantic similarity between the feature vectors. For a spatial slice containing 1000 pixels, a 128-dimensional feature vector can be extracted for each pixel, and the degree of semantic association between the feature vectors is calculated by cosine similarity. When the cosine similarity of two feature vectors is greater than 0.75, a connection is established in the feature association graph, and the similarity value is used as the weight of the edge.

[0056] Spectral clustering is performed on the constructed feature association graph. This process uses eigenvalue decomposition to obtain the eigenvectors of the graph Laplacian matrix and sort them by their corresponding eigenvalues. The graph Laplacian matrix is ​​obtained by subtracting the adjacency matrix from the diagonal matrix, where the diagonal elements of the diagonal matrix are the degrees of each node. For the feature association graph of 1000 nodes, 1000 eigenvectors are calculated and sorted from largest to smallest by their corresponding eigenvalues. The first 20 eigenvectors are used as a low-dimensional representation, which captures the key characteristics of the graph structure.

[0057] Based on the sorted feature vectors, the degree of semantic jump between adjacent features in the spatial slice is calculated. Specifically, for any adjacent feature vectors i and i+1, their Euclidean distance in the low-dimensional representation space is calculated as a measure of the degree of semantic jump. When the jump degree is greater than the preset jump degree threshold of 0.35, the position is determined as the partition boundary. In practice, for a spatial slice containing 1000 pixels, 5-8 partition boundaries may be identified, which divide the spatial slice into different semantic areas.

[0058] Based on the identified partition boundaries, the spatial slice is divided into multiple subregions. The variance of the feature vector within each subregion is then calculated to assess the consistency within the region. The variance is calculated by averaging the standard deviations of the feature vector values ​​across their dimensions in the low-dimensional representation space. When the variance of a subregion exceeds 0.25, it indicates significant semantic variation within the region and requires further subdivision. When the variance is less than 0.05, the region is too fragmented and can be merged with adjacent regions. In this way, the original six subregions can be adjusted to four more reasonable partitions.

[0059] A nonlinear transformation is performed within each subregion to construct an adaptive partitioning matrix. The elements of this matrix are determined by the cosine similarity of the eigenvectors within the subregion and the spatial distance of the eigenvectors. For any two eigenvectors within the subregion, their cosine similarity is calculated, with a value range of [-1, 1], and their normalized spatial distance is calculated, with a value range of [0, 1]. The elements of the adaptive partitioning matrix can be obtained by multiplying the cosine similarity by an exponential decay function of the spatial distance. Specifically, the similarity is multiplied by a power function with a base of 2.7, and the negative value of the distance is the exponential. In this way, the matrix element values ​​corresponding to eigenvectors with closer spatial distance and higher semantic similarity are larger, and vice versa.

[0060] Multiply the adaptive partitioning matrix by the sub-region feature vectors to obtain the local attention weight matrix. For each sub-region, if it contains 200 feature vectors, each with a dimension of 128, the dimension of the adaptive partitioning matrix is ​​200×200, and the dimension of the matrix composed of feature vectors is 200×128. The dimension of the local attention weight matrix obtained by multiplying the two is also 200×128. Each row of this weight matrix corresponds to the attention weight distribution of a feature vector. Dimensions with higher weight values ​​indicate that the dimension has higher discriminative power within the current sub-region.

[0061] In practical applications, consider a multimodal dataset consisting of aerial imagery, ground sensor data, and historical risk records. Spatial slices may contain regions with different surface types. This method can partition a 2000×2000 pixel aerial image into several semantically consistent subregions, such as water, buildings, vegetation, and roads. For each subregion, the local attention weight matrix highlights the key features of that region. For example, in a building region, feature dimensions related to building edges and density are given higher weights; in a water region, feature dimensions related to water surface texture and reflection are given higher weights.

[0062] Compared with the existing technology, traditional spatial data analysis methods usually use fixed grids or predefined areas for spatial division, which cannot adapt to the intrinsic semantic structure of the data. Although some improved methods have introduced adaptive grids, they are mainly based on data density or gradient changes, ignoring the semantic association between features. There are also methods that use global attention mechanisms to process spatial data, but the computational complexity is high and it is difficult to capture local characteristics. The method of this embodiment automatically identifies the partition boundaries of spatial slices based on feature semantic similarity through a combination of spectral clustering and adaptive partitioning, and constructs a local attention mechanism in each sub-area, which takes into account the semantic association of features while maintaining computational efficiency. The method of this embodiment solves the problem of effective fusion of spatial information in multimodal heterogeneous data, and significantly improves the accuracy of risk identification through adaptive partitioning and local attention allocation.

[0063] In an optional embodiment, a graph embedding algorithm is applied to the fusion vector to construct a risk propagation topology network, calculate the node entropy value time series, extract topological dynamic features through mutation detection and synchronization analysis, and form risk indicators including: Mapping the fused vector to a low-dimensional embedding space using a knowledge graph relational mapping algorithm, wherein the knowledge graph relational mapping algorithm includes constructing a distance metric function between entities and relationships and minimizing representation error, and constructing a risk propagation topology network based on the semantic distance between entities in the low-dimensional embedding space; Constructing a time series observation window for each node in the risk propagation topology network, and calculating the node degree centrality within the time series observation window; Determine the initial bin boundaries based on the probability density function of the node degree centrality, obtain a probability distribution sequence by adaptively merging the bin intervals, and calculate the node entropy value sequence based on the probability distribution sequence; Variational mode decomposition is applied to the node entropy value sequence to obtain an adaptive intrinsic mode function, the instantaneous frequency spectrum is calculated through Hilbert-Huang transform, and the mutation points and trend turning points in the instantaneous frequency spectrum are identified by singular spectrum analysis to obtain a mutation feature sequence of the risk propagation process; Based on the node entropy value sequence, a temporal correlation matrix is ​​constructed and spectral decomposition is performed to obtain a network synchronization pattern, a synchronization strength matrix between node pairs is calculated, and the synchronization strength matrix is ​​grouped by community partitioning to obtain synchronized node subgroups; The betweenness centrality and clustering coefficient of the nodes in the risk propagation topology network are calculated, and the betweenness centrality, the clustering coefficient, the node entropy value sequence, the mutation characteristics and the synchronization strength matrix of the synchronization node subgroup are combined to construct a multidimensional risk index.

[0064] In a specific embodiment, the fusion vector is mapped to a low-dimensional embedding space through a knowledge graph relationship mapping algorithm. In practical applications, a distance measurement function between entities and relationships is constructed and the representation error is minimized. For example, for the relationship R between risk entity A and entity B, the distance function d(A, B, R) is defined to represent the semantic distance between the two entities under the relationship R. The TransE algorithm is used to make the sum of the head entity vector and the relationship vector close to the tail entity vector. This mapping is applied to 50 risk entities in the financial market, which are represented as 128-dimensional embedding vectors, and the mapping error is less than 0.01 through error iterative optimization. Based on the Euclidean distance between these embedding vectors, an adjacency matrix is ​​constructed, and a connection is established when the distance between two entities is less than a threshold of 0.5, forming a risk propagation topology network.

[0065] A time-series observation window is constructed for each node in the risk propagation topology network. Specifically, a continuous observation sequence is constructed with a 14-day window length and a sliding window of one day per day. Within each observation window, the node degree centrality is calculated, which counts the number of other nodes directly connected to the target node. For example, if node X is directly connected to six other nodes within a certain observation window, its degree centrality is 6.

[0066] The initial bin boundaries are determined based on the probability density function of node degree centrality. Taking the degree centrality values ​​of a node within 14 days as [3, 4, 5, 6, 8, 7, 6, 5, 4, 3, 4, 5, 6, 7] as an example, we first count the frequencies of different degree centrality values ​​to obtain the initial bins {3: 3 times, 4: 3 times, 5: 3 times, 6: 3 times, 7: 2 times, 8: 1 time}. An adaptive merging algorithm is applied to merge adjacent bins with similar frequencies, resulting in the optimized bins {3-4: 6 times, 5-6: 6 times, 7-8: 3 times}. From this, the probability distribution sequence [0.4, 0.4, 0.2] is calculated. Based on this probability distribution sequence, the node entropy value for the current window is calculated to be 1.37. As the window slides, the complete node entropy value time series is obtained.

[0067] Variational mode decomposition is applied to the node entropy sequence, decomposing the complex signal into modal functions of different frequencies. In practice, the node entropy sequence is decomposed into three intrinsic mode functions, capturing fluctuation characteristics at different time scales. The instantaneous frequency spectrum of each modal function is calculated using the Hilbert-Huang transform, revealing frequency information that varies over time. Combined with singular spectrum analysis, points in the frequency spectrum with a rate of change exceeding 30% are marked as mutation points. In a case study of a financial entity, mutation points in the node entropy on the 23rd and 47th days were successfully identified, corresponding to events of rapid market risk transmission. By integrating this information from these mutation points, a mutation signature sequence of the risk propagation process is formed.

[0068] A temporal correlation matrix is ​​constructed based on the node entropy sequence. Specifically, the Pearson correlation coefficient between the entropy sequences of any two nodes in the network is calculated to form an N×N correlation matrix (N is the number of nodes). In a certain example, a financial risk network containing 40 nodes was analyzed to generate a 40×40 correlation matrix. Spectral decomposition is performed on the matrix, and the main eigenvectors are extracted to form the network synchronization pattern. The synchronization strength matrix is ​​obtained by calculating the phase consistency index between the node pairs. When the synchronization strength between two nodes exceeds 0.7, they are considered to have significant synchronization. The Lowen clustering algorithm is applied to perform community division on the synchronization strength matrix, and four synchronous node subgroups are identified, which represent different synchronization modules in the risk propagation process.

[0069] The betweenness centrality and clustering coefficient of nodes in the risk propagation topological network are calculated. Betweenness centrality measures the frequency with which a node lies on the shortest path between pairs of other nodes, reflecting a node's ability to control risk propagation; the clustering coefficient measures the density of connections between node neighbors, reflecting the potential for local risk diffusion. A multidimensional risk indicator system is constructed by combining betweenness centrality, clustering coefficient, node entropy sequence, mutation characteristics, and the synchronization strength matrix of synchronized node subgroups. This indicator system comprehensively captures network topology characteristics, temporal dynamics, mutation characteristics, and synchronization patterns, enabling comprehensive characterization and early warning of the risk propagation process.

[0070] In an optional embodiment, determining the initial binning boundaries based on the probability density function of the node degree centrality, obtaining a probability distribution sequence by adaptively merging binning intervals, and calculating the node entropy value sequence based on the probability distribution sequence includes: The kernel density estimation method is used to calculate the probability density function of the node degree centrality, and the probability density function is calculated by the Gaussian kernel function and the bandwidth parameter; Determine the derivative of the probability density function, identify the local maximum point and the local minimum point of the probability density function, use the local maximum point and the local minimum point as initial binning boundaries, and determine the binning interval; Calculate the information gain ratio before and after the adjacent bin intervals are merged, where the information gain ratio is calculated by the ratio of information gain to split information. When the information gain ratio is less than a preset gain ratio threshold, merge the adjacent bin intervals to obtain an adaptive bin boundary set. The node degree centrality is mapped into a discretized sequence based on the adaptive bin boundary set, the probability distribution of samples in each bin interval is calculated, and a node entropy value sequence is obtained based on the probability distribution.

[0071] In a specific embodiment, for a network structure containing multiple nodes, the degree centrality of each node is calculated. Degree centrality indicates the number of other nodes directly connected to a node and is a basic indicator for measuring the importance of a node in the network. After obtaining the node degree centrality data, the kernel density estimation method is used to calculate the probability density function of the node degree centrality. This method estimates a continuous probability density function by treating each data point as a small probability mass and using a kernel function (a Gaussian kernel function is selected in this embodiment) to smooth these masses. The Gaussian kernel function is expressed as the contribution of each data point x_i at x to the normal distribution function value centered on x_i and with a standard deviation of the bandwidth parameter h. The choice of bandwidth parameter h has an important influence on the density estimation result. Too small h will cause the estimate to fluctuate too much, and too large h will cause over-smoothing and loss of data features. In practice, the bandwidth parameter can be calculated using an empirical formula, such as h=0.9×min(standard deviation, interquartile range / 1.34)×n -1 / 5 , where n is the number of samples.

[0072] Taking a network with 100 nodes as an example, the calculated node degree centrality values ​​are distributed between 3 and 28. Using the bandwidth parameter h = 2.1 for kernel density estimation, a probability density function is obtained that is smooth over the range of degree centrality values.

[0073] In order to determine the initial binning boundaries, it is necessary to calculate the derivative of the probability density function and find the local maximum and local minimum points of the function. By numerical differentiation method, the derivative value of the density function at each point can be calculated. In this embodiment, by derivative analysis, it is identified that the density function has local maxima at degree centrality values ​​of 5.3, 10.2, 15.8 and 22.4, and local minima at 7.6, 12.5 and 19.1. These local maxima and minima constitute the initial binning boundaries, dividing several binning intervals: [3, 5.3], [5.3, 7.6], [7.6, 10.2], [10.2, 12.5], [12.5, 15.8], [15.8, 19.1], [19.1, 22.4] and [22.4, 28].

[0074] The initial binning may be too detailed, so the information gain ratio is needed to assess the necessity of merging adjacent intervals. Information gain represents the reduction in information before and after merging, while split information represents the information gained due to splitting. The information gain ratio is the ratio of these two values ​​and is used to evaluate the effectiveness of the merge operation. The calculation begins by counting the number of nodes in each binning interval and their proportion to the total number of nodes, and calculating the entropy of the current binning scheme. Then, assuming the two adjacent intervals are merged, the entropy of the merged interval is recalculated, and the information gain is calculated as the difference between the original entropy and the merged entropy. Split information is calculated by taking the entropy ratio of the two intervals before merging. The information gain ratio is simply the information gain divided by the split information.

[0075] Set the preset gain ratio threshold to 0.05 and evaluate each adjacent bin interval one by one. For example, the information gain ratio of the intervals [3, 5.3] and [5.3, 7.6] combined is 0.02, which is less than the threshold, so these two intervals are merged into [3, 7.6]. Similarly, the information gain ratio of the intervals [7.6, 10.2] and [10.2, 12.5] is 0.03, also less than the threshold, so they are merged into [7.6, 12.5]. The information gain ratio of the intervals [12.5, 15.8] and [15.8, 19.1] is 0.06, which is greater than the threshold and remains unchanged. Continuing to evaluate other adjacent intervals, we finally get the adaptive bin boundary set of {3, 7.6, 12.5, 15.8, 19.1, 22.4, 28}, and the corresponding bin intervals are [3, 7.6], [7.6, 12.5], [12.5, 15.8], [15.8, 19.1], [19.1, 22.4], and [22.4, 28].

[0076] Based on these adaptive binning boundaries, node degree centrality values ​​are mapped to corresponding intervals, forming a discretized sequence. For example, a node with a degree centrality of 4 is mapped to the interval [3, 7.6], and a node with a degree centrality of 14 is mapped to the interval [12.5, 15.8]. By counting the number and proportion of nodes in each interval, the probability distribution of each interval is calculated. In this example, the interval [3, 7.6] contains 35 nodes, accounting for 35%; the interval [7.6, 12.5] contains 25 nodes, accounting for 25%; the interval [12.5, 15.8] contains 18 nodes, accounting for 18%; the interval [15.8, 19.1] contains 12 nodes, accounting for 12%; the interval [19.1, 22.4] contains 7 nodes, accounting for 7%; and the interval [22.4, 28] contains 3 nodes, accounting for 3%.

[0077] Based on this probability distribution, we can calculate the entropy of each node. For nodes mapped to the interval [3, 7.6], the entropy is -0.35 × log² (0.35) ≈ 0.53; for nodes mapped to the interval [7.6, 12.5], the entropy is approximately 0.5; and so on for nodes in other intervals. In this way, we can obtain a corresponding entropy value for each node in the network, forming a complete node entropy sequence. This entropy sequence can be used for subsequent network analysis tasks such as node importance assessment and community detection.

[0078] Discretizing node degree centrality is crucial for improving analytical efficiency and interpretability. Traditional techniques primarily employ equal-width or equal-frequency binning for discretization. Equal-width binning divides the data range into several equal intervals, which is simple to implement but ignores the data distribution characteristics. Equal-frequency binning ensures that each interval contains a similar number of samples, but may place similar values ​​into different intervals. Improved methods, such as clustering-based binning, use clustering algorithms like k-means to determine bin boundaries, but this is computationally complex and requires a predetermined number of bins. Decision tree-based binning also uses information gain to determine split points, but is sensitive to noise and prone to overfitting.

[0079] The method of this embodiment realizes adaptive binning of node degree centrality by combining kernel density estimation with information gain ratio. Compared with the existing technology, kernel density estimation is used to obtain a global view of data distribution, and the intrinsic structure of data can be better captured by identifying the extreme points of the density function as the initial binning boundaries; the introduction of information gain ratio as an evaluation index for merging adjacent intervals not only takes into account the degree of information retention, but also prevents excessive binning, balancing the complexity and expression ability of the model; by mapping degree centrality to an entropy value sequence, a new measure of node importance is provided, enhancing the ability of complex network analysis. The method of this embodiment solves the problem of insufficient adaptiveness of node feature discretization in network analysis, especially when faced with the heterogeneity and diversity of complex networks, the fixed binning method is difficult to adapt to the limitations of different data distributions.

[0080] like Figure 3 The figure shows the impact of the kernel density estimation bandwidth parameter (h) on the performance of the adaptive binning method in this technical solution. Three key curves are plotted: community detection accuracy (solid cyan line), number of bins (red line with square markers), and entropy estimation error (dashed blue line). Several clear trends are observed as the bandwidth parameter increases from 0.5 to 5.0. The number of bins decreases with increasing bandwidth, from 12 at h = 0.5 to only 2 at h = 5.0. This is because a larger bandwidth results in a smoother kernel density estimation curve and fewer extreme points. Community detection accuracy shows an increasing and then decreasing trend, reaching a maximum of 95.7% at h = 2.3. When the bandwidth is too small (h = 0.5), the accuracy is only 82.1%, and when the bandwidth is too large (h = 5.0), the accuracy drops to 84.7%. The entropy estimation error also exhibits a U-shaped curve, reaching a minimum of 0.025 at h = 2.3, indicating that the binning results at this point are closest to the true data distribution. The light green area in the figure marks the optimal range of the bandwidth parameter (1.7-2.6). Within this range, the community detection accuracy remains above 94%, the number of bins is maintained at 6-7, and the entropy estimation error is controlled below 0.032. The vertical dotted line marks the bandwidth value h=2.1 used in the experiment. At this time, the community detection accuracy is 94.1%, the number of bins is 7, and the entropy estimation error is 0.028, which is close to the optimal performance. This reveals the key impact of bandwidth parameter selection on the performance of this technical solution: too small a bandwidth will lead to over-binning and capture noise in the data; too large a bandwidth will lead to over-smoothing and lose important features of the data. The bandwidth parameter h=2.1 used in this technical solution effectively controls the number of bins while maintaining a high accuracy, achieving the best balance between performance and complexity.

[0081] The risk intelligent identification and analysis system based on multimodal heterogeneous data fusion in an embodiment of the present invention includes: The first unit is used to collect structured data, unstructured data and semi-structured data to build multi-source data sets; The second unit is used to perform deep neural network and symbolic reasoning joint processing on the multi-source data set, build a symbolic semantic encoding mapping mechanism, project data features into the symbolic-neural dual space, perform feature extraction based on symbolic guidance constraints, and generate semantic vectors; The third unit is used to obtain spatial dimension features of the semantic vector through orthogonal matrix decomposition, extract temporal features through a bidirectional long short-term memory network, fuse multimodal information based on adaptive weights and mutual information alignment mechanism, and generate a fused vector; The fourth unit is used to apply a graph embedding algorithm to the fusion vector, construct a risk propagation topology network, calculate the time series of node entropy values, extract topological dynamic features through mutation detection and synchronization analysis, and form risk indicators; The fifth unit is used to calculate the risk identification results based on the risk indicators, map the risk transmission path, quantify the risk impact, and output early warning suggestions.

[0082] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0083] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0084] The present invention may be a method, an apparatus, and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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. Risk intelligent identification and analysis method based on multimodal heterogeneous data fusion, characterized by: include: Collect structured data, unstructured data, and semi-structured data to build multi-source data sets; Performing a joint process of deep neural network and symbolic reasoning on the multi-source dataset, constructing a symbolic semantic encoding mapping mechanism, projecting data features into a symbolic-neural dual space, performing feature extraction based on symbolic guided constraints, and generating semantic vectors; The semantic vector is decomposed into spatial dimension features through orthogonal matrix decomposition, and the temporal features are extracted through a bidirectional long short-term memory network. The multimodal information is fused based on the adaptive weight and mutual information alignment mechanism to generate a fused vector; Applying a graph embedding algorithm to the fusion vector to construct a risk propagation topology network, calculating the node entropy value time series, extracting topological dynamic features through mutation detection and synchronization analysis, and forming risk indicators; Calculate risk identification results based on the risk indicators, map risk transmission paths, quantify risk impact, and output early warning recommendations.

2. The method according to claim 1, characterized in that The multi-source dataset is processed jointly by deep neural network and symbolic reasoning, a symbolic semantic encoding mapping mechanism is constructed, data features are projected into the symbolic-neural dual space, feature extraction is performed based on symbolic guided constraints, and semantic vectors are generated, including: Use the preset neural network to extract features from multi-source data and obtain the initial feature vector; Build a domain ontology knowledge base based on semantic attributes and semantic rules to obtain the structured organization form of domain concepts and the semantic association rules between concepts; Performing a structured transformation on the initial feature vector according to the semantic attributes and the semantic rules to generate a symbolic space representation with an explicit semantic structure; Decomposing and mapping the semantic information in the symbolic space representation into the neural space to obtain semantic decoupling features, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain semantic recombinant features; Calculating the semantic consistency loss between the semantic decoupling feature and the semantic reorganization feature based on the semantic association rule, and performing joint optimization based on the semantic consistency loss to obtain an optimization parameter; The symbolic space representation is processed using the optimization parameters to output a semantic vector.

3. The method according to claim 2, characterized in that Decomposing and mapping the semantic information in the symbolic space representation into the neural space to obtain semantic decoupling features, and mapping the initial feature vector into the symbolic space for semantic reconstruction to obtain semantic recombinant features include: receiving the symbolic space representation and constructing a semantic association matrix based on the semantic attributes and the semantic rules, wherein each element in the semantic association matrix corresponds to the semantic association strength between the features; Based on the structural relationship of the semantic association matrix, the symbolic space representation is decomposed into a plurality of semantic substructures, each of the semantic substructures corresponding to local structural information of a semantic attribute; Performing feature decomposition on each of the semantic substructures, extracting feature vectors and corresponding eigenvalues, sorting the eigenvalues ​​from large to small, calculating the proportion of each eigenvalue to the sum of all eigenvalues, selecting eigenvectors whose sum of eigenvalue proportions is greater than a preset feature threshold, and constructing a semantic basis vector group; Performing orthogonal projection decomposition on the semantic basis vector group to obtain multiple groups of semantic components, each group of semantic components corresponding to an independent semantic attribute, thereby obtaining a semantic decoupling feature; An initial feature vector is received, a projection matrix is ​​constructed based on the semantic basis vector group, the initial feature vector is multiplied by the projection matrix to obtain projection coordinates in a symbolic space, and a semantic reorganization feature is obtained based on a linear combination of the projection coordinates and the semantic basis vector group.

4. The method according to claim 1, wherein The semantic vector is decomposed into spatial dimension features through orthogonal matrix decomposition, and temporal features are extracted through a bidirectional long short-term memory network. Multimodal information is fused based on adaptive weight and mutual information alignment mechanism to generate a fusion vector including: Performing orthogonal matrix decomposition on the semantic vector to obtain a basis vector matrix, projecting the semantic vector into the feature space using the basis vector matrix to obtain a spatial dimension projection matrix, setting segmentation intervals for the spatial dimension projection matrix according to the feature dimension, and combining features within each segmentation interval to form a spatial slice of the semantic vector; Dynamically construct an adaptive partitioning matrix based on the semantic similarity of features in the spatial slice and perform a nonlinear transformation to obtain a local attention weight matrix; cluster the local features in the spatial slice, calculate the feature center of each cluster, perform feature mapping based on the feature center, and construct a global relevance weight matrix; concatenate the local attention weight matrix with the global relevance weight matrix to calculate the adaptive weight coefficient; Inputting the semantic vector into a bidirectional long short-term memory network to obtain temporal hidden layer features, performing wavelet decomposition on the temporal hidden layer features to obtain multi-scale temporal features; Calculating the mutual information between the spatial features weighted by the adaptive weight coefficients and the multi-scale temporal features, and constructing an alignment loss function including a mutual information maximization term and a semantic transformation matrix constraint term; Based on the alignment loss function, optimizing the semantic transformation matrix corresponding to the spatial feature and the multi-scale temporal feature to obtain a semantic alignment feature; The semantic alignment features are passed through the gating unit to calculate the fusion weight and obtain the fusion vector.

5. The method according to claim 4, characterized in that According to the semantic similarity of the features in the spatial slice, an adaptive partition matrix is ​​dynamically constructed and a nonlinear transformation is performed to obtain the local attention weight matrix including: Mapping the spatial slice to a high-dimensional feature space to construct a feature association graph, wherein the nodes of the feature association graph represent feature vectors in the spatial slice, and the edge weights represent the semantic similarity between the feature vectors; Performing spectral clustering on the feature association graph, obtaining eigenvectors of the graph Laplacian matrix based on eigenvalue decomposition, and sorting the eigenvectors according to the corresponding eigenvalues; According to the sorted feature vectors, the semantic jump degree between adjacent features in the spatial slice is calculated, and the position where the semantic jump degree is greater than the preset jump degree threshold is determined as the partition boundary; Dividing the spatial slice into multiple sub-regions based on the partition boundary, calculating the variance of the feature vector in each sub-region, and adjusting the partition range of the sub-region according to the variance; Perform a nonlinear transformation in each sub-region to construct an adaptive partition matrix, wherein the elements of the adaptive partition matrix are determined by the cosine similarity of the eigenvectors in the sub-region and the spatial distance of the eigenvectors; The adaptive partition matrix is ​​multiplied by the sub-region feature vector to obtain a local attention weight matrix.

6. The method according to claim 1, characterized in that Applying the graph embedding algorithm to the fusion vector, constructing the risk propagation topology network, calculating the node entropy value time series, extracting the topological dynamic features through mutation detection and synchronization analysis, and forming risk indicators including: Mapping the fused vector to a low-dimensional embedding space using a knowledge graph relational mapping algorithm, wherein the knowledge graph relational mapping algorithm includes constructing a distance metric function between entities and relationships and minimizing representation error, and constructing a risk propagation topology network based on the semantic distance between entities in the low-dimensional embedding space; Constructing a time series observation window for each node in the risk propagation topology network, and calculating the node degree centrality within the time series observation window; Determine the initial bin boundaries based on the probability density function of the node degree centrality, obtain a probability distribution sequence by adaptively merging the bin intervals, and calculate the node entropy value sequence based on the probability distribution sequence; Variational mode decomposition is applied to the node entropy value sequence to obtain an adaptive intrinsic mode function, the instantaneous frequency spectrum is calculated through Hilbert-Huang transform, and the mutation points and trend turning points in the instantaneous frequency spectrum are identified by singular spectrum analysis to obtain a mutation feature sequence of the risk propagation process; Based on the node entropy value sequence, a temporal correlation matrix is ​​constructed and spectral decomposition is performed to obtain a network synchronization pattern, a synchronization strength matrix between node pairs is calculated, and the synchronization strength matrix is ​​grouped by community partitioning to obtain synchronized node subgroups; The betweenness centrality and clustering coefficient of the nodes in the risk propagation topology network are calculated, and the betweenness centrality, the clustering coefficient, the node entropy value sequence, the mutation characteristics and the synchronization strength matrix of the synchronization node subgroup are combined to construct a multidimensional risk index.

7. The method according to claim 6, characterized in that Determining the initial bin boundaries based on the probability density function of the node degree centrality, obtaining a probability distribution sequence by adaptively merging bin intervals, and calculating the node entropy value sequence based on the probability distribution sequence includes: The kernel density estimation method is used to calculate the probability density function of the node degree centrality, and the probability density function is calculated by the Gaussian kernel function and the bandwidth parameter; Determine the derivative of the probability density function, identify the local maximum point and the local minimum point of the probability density function, use the local maximum point and the local minimum point as initial binning boundaries, and determine the binning interval; Calculate the information gain ratio before and after the adjacent bin intervals are merged, where the information gain ratio is calculated by the ratio of information gain to split information. When the information gain ratio is less than a preset gain ratio threshold, merge the adjacent bin intervals to obtain an adaptive bin boundary set. The node degree centrality is mapped into a discretized sequence based on the adaptive bin boundary set, the probability distribution of samples in each bin interval is calculated, and a node entropy value sequence is obtained based on the probability distribution.

8. A risk intelligent identification and analysis system based on multimodal heterogeneous data fusion, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect structured data, unstructured data and semi-structured data to build multi-source data sets; The second unit is used to perform deep neural network and symbolic reasoning joint processing on the multi-source data set, build a symbolic semantic encoding mapping mechanism, project data features into the symbolic-neural dual space, perform feature extraction based on symbolic guidance constraints, and generate semantic vectors; The third unit is used to obtain spatial dimension features of the semantic vector through orthogonal matrix decomposition, extract temporal features through a bidirectional long short-term memory network, fuse multimodal information based on adaptive weights and mutual information alignment mechanism, and generate a fused vector; The fourth unit is used to apply a graph embedding algorithm to the fusion vector, construct a risk propagation topology network, calculate the time series of node entropy values, extract topological dynamic features through mutation detection and synchronization analysis, and form risk indicators; The fifth unit is used to calculate the risk identification results based on the risk indicators, map the risk transmission path, quantify the risk impact, and output early warning suggestions.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Deep learning optimization method for edge computing server

    CN118520936A

  • Multi-source heterogeneous big data fusion and reasoning method and system based on knowledge graph

    CN119442151A

  • Home abnormal state signal detection method and system based on multi-mode sensing

    CN120216965A

  • Knowledge graph generation method and system for science and technology project risk control

    CN120296180A

  • Multi-source cross-modal threat intelligence processing method and device, equipment and medium

    CN120296688A

Cited By

  • Data processing method for specimen digital management platform

    CN121053659A

  • Enterprise operation data semantic risk identification method and system based on neural network

    CN122198667A

  • Abnormal number AI risk diagnosis method and system based on multi-dimensional feature fusion

    CN122226515A

  • Abnormal number ai risk diagnosis method and system based on multi-dimensional feature fusion

    CN122226515B

  • Urban governance event management and control method, device and equipment and storage medium

    CN122264477A