Scene perception modeling and verification method and system based on parallel intelligence
By combining multimodal data acquisition and dynamic feature processing with adaptive weight allocation and domain knowledge graph verification, the problems of insufficient data fusion and low classification accuracy in existing scene perception modeling are solved, achieving comprehensive business scene perception and flexible classification optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-13
AI Technical Summary
Existing scene perception modeling methods suffer from limitations such as single data acquisition dimensions, insufficient multimodal data fusion, static feature processing with poor dynamic adaptability, lack of adaptive feature weight allocation, and lack of semantic constraints on classification results, leading to low perception accuracy and limited classification accuracy.
The method employs multimodal data acquisition, dynamic feature decoupling processing, spatiotemporal feature fusion, adaptive weight allocation, hierarchical clustering, and domain knowledge graph verification, combined with an incremental learning framework for dynamic optimization, to generate final business classification labels and trigger business processing flows.
It achieves comprehensive business scenario awareness, improves the accuracy and flexibility of classification, can adapt to dynamic changes in business scenarios, and ensures that the classification results are consistent with the actual business situation.
Smart Images
Figure CN121659779A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information modeling technology, specifically to a method and system for scene perception modeling and verification based on parallel intelligence. Background Technology
[0002] Against the backdrop of accelerated digital transformation, accurate perception and intelligent classification of business scenarios have become core requirements for enterprises to achieve refined operations, dynamic decision-making, and personalized services. Scenario-aware modeling technology analyzes multi-source data in the business environment to uncover scenario features and establish a classification system, providing crucial support for subsequent business rule triggering and resource scheduling. However, existing scenario-aware modeling methods still face the following technical bottlenecks in practical applications: First, data collection is limited to a single dimension, and multimodal data fusion is insufficient. Traditional methods primarily focus on the collection and analysis of structured data (such as transaction records and sensor values), while the utilization of semi-structured data (such as log files and configuration documents) and unstructured data (such as text, images, and audio) is relatively weak. Since the scene information carried by different modalities is complementary (such as text describing user needs and images reflecting environmental conditions), the one-sided collection of data from a single modality can easily lead to incomplete expression of scene features, thereby affecting perception accuracy.
[0003] Secondly, feature processing is static and lacks dynamic adaptability. Existing feature decoupling techniques mostly extract features from data based on fixed rules or historical experience, making it difficult to cope with the dynamic changes in data distribution over time and space in business scenarios (such as periodic fluctuations in user behavior patterns and the impact of sudden events on the scenario). At the same time, the feature fusion process often ignores spatiotemporal correlations (such as the weight differences of the same feature at different times and in different regions), resulting in the fused feature matrix failing to accurately reflect the real-time state of the scenario.
[0004] Third, the feature weight allocation lacks adaptability, limiting classification accuracy. Traditional methods often use preset weights or statistical frequencies to determine feature importance, without considering the dynamic value of features in different business scenarios (e.g., "user's historical spending amount" has a higher weight in promotional scenarios, while "abnormal operation frequency" has a higher weight in risk warning scenarios). Fixed weight allocation mechanisms can easily lead to the underestimation of key features, while non-key features interfere with classification results, reducing the accuracy of scenario classification.
[0005] Fourth, the classification results lack semantic constraints and are disconnected from actual business operations. Existing classification methods mostly rely on unsupervised clustering (such as K-means, DBSCAN) or supervised learning models (such as SVM, neural networks), classifying scenarios only from the perspective of data distribution, without verifying the classification results by incorporating domain knowledge (such as industry rules, business logic). For example, clustering based solely on user behavior data may misclassify "high-spending users" and "abnormal transaction users" into the same category, while domain knowledge can clearly distinguish the semantic differences between the two, avoiding classification bias. Summary of the Invention
[0006] To address the aforementioned technical problems, this paper provides a method and system for scene perception modeling and verification based on parallel intelligence. This technical solution solves the problems mentioned in the background section.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, a scene perception modeling and verification method based on parallel intelligence is provided, comprising: The raw data stream of the business scenario is obtained through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. Dynamic feature decoupling processing is performed on the original data stream to generate a multi-dimensional feature vector set; Based on the context association model of business scenarios, spatiotemporal feature fusion is performed on multi-dimensional feature vector sets to obtain a fused feature matrix; An adaptive weight allocation mechanism is used to evaluate the feature importance of the fused feature matrix and generate a feature weight map. Based on the feature weight map, a hierarchical clustering algorithm is used to initially classify the business scenarios and obtain a candidate classification set. The domain knowledge graph is used to perform semantic constraint verification on the candidate classification set to correct classification bias. The revised classification results are dynamically optimized using an incremental learning framework to generate the final business classification labels. The final business category tags are mapped to the preset business rule engine, triggering the corresponding business processing flow.
[0008] Preferably, the step of performing dynamic feature decoupling processing on the original data stream to generate a multi-dimensional feature vector set specifically includes: Semantic parsing is performed on the unstructured data in the original data stream to extract implicit semantic features. The semantic parsing model adopts the BERT or GPT architecture and outputs a vector with dimension n. Statistical features are extracted from structured and semi-structured data to generate explicit statistical features, including mean, variance, maximum and minimum values. Cross-modal association analysis is performed on the implicit semantic features and explicit statistical features using a feature cross-network to generate cross-modal association features. The feature cross-network employs a multilayer perceptron, and the formula is as follows: ; in, For cross-modal association features, For implicit semantic features, These are explicit statistical characteristics. , This is the weight matrix. , For bias terms, It is the ReLU activation function. Use the Sigmoid activation function; The implicit semantic features, explicit statistical features, and cross-modal association features are concatenated to form a multi-dimensional feature vector set.
[0009] Preferably, the spatiotemporal feature fusion includes: Construct a time decay function based on the time series characteristics of the business scenario. The formula is: ; in, The time decay coefficient, The interval between the current time and historical time; Construct a spatial proximity matrix based on the spatial distribution characteristics of the business scenario. Matrix elements Indicates sample With sample spatial distance The reciprocal of, that is: ; The time decay function Spatial proximity matrix Perform weighted fusion to generate a spatiotemporal weight matrix. The formula is: ; in, This represents the time-weighted ratio. Using the spatiotemporal weight matrix For the multi-dimensional feature vector set The fusion feature matrix is obtained by performing dynamic weighting. The formula is: ; in, This indicates element-wise multiplication.
[0010] Preferably, the adaptive weight allocation mechanism includes: The fusion feature matrix is calculated using an attention mechanism. The attention score ai for each feature dimension is given by the formula: ; in, This is the weight matrix. For attention vectors, For bias terms, The number of feature dimensions; Based on the attention score The weight coefficients for each feature dimension are generated using the softmax function. The formula is: ; According to the weighting coefficients For the fused feature matrix Perform weighted summation to generate the feature weight map. The formula is: .
[0011] Preferably, the hierarchical clustering algorithm includes: Initialize cluster centers The number of cluster centers The complexity is dynamically determined based on the business scenario, using the following formula: ; Where N is the number of samples; Calculate each sample With cluster center similarity Cosine similarity is used, and the formula is: ; Sample Assign to the cluster with the highest similarity ; Update cluster center The formula is: ; in, For clustering The number of samples in the sample; Repeat the above steps until the cluster centers converge, generating the candidate classification set. .
[0012] Preferably, the semantic constraint verification includes: From domain knowledge graph Extract entity relationship chains related to business scenarios ,in For a collection of entities, For a set of relations; The candidate classification set Category tags in Mapped to the entity relationship chain Generate mapping labels ; Detect the classification label With mapping labels Semantic consistency is determined using Jaccard similarity, with the formula: ; like Then, based on the entity relationship chain Correct the category labels .
[0013] Preferably, the incremental learning framework includes: Build an initial classification model The initial classification model is based on historical business data. The training process yielded a loss function using cross-entropy loss, with the following formula: ; in, For real labels, For predicting labels, Number of categories; Real-time collection of new business data Extract feature vectors from new business data ; The feature vector of the new business data Input the initial classification model Generate predicted classification results ; Calculate the predicted classification result Compared with the true classification results error The formula is: ; like Then update the initial classification model. The parameter θ is determined using the gradient descent method, and the formula is as follows: ; in, The learning rate; The optimized classification model The above steps are repeated as a new initial classification model.
[0014] Preferably, the business rule engine includes: Predefined business rule library The business rule base contains a mapping relationship between category tags l and business processing flow P. ; According to the final business classification label From the business rule base Match the corresponding business processing flow in the middle; If no matching business process is found, the exception handling mechanism is triggered, and exception handling suggestions are generated.
[0015] In a second aspect of the invention, a scene perception modeling and verification system based on parallel intelligence is also provided, comprising: The acquisition module is used to acquire the raw data stream of the business scenario through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. A decoupling module is used to perform dynamic feature decoupling processing on the original data stream to generate a multi-dimensional feature vector set. The fusion module is used to perform spatiotemporal feature fusion on a multi-dimensional feature vector set based on a context association model of the business scenario to obtain a fused feature matrix. The generation module is used to evaluate the feature importance of the fused feature matrix using an adaptive weight allocation mechanism and generate a feature weight map. The first classification module is used to perform preliminary classification of business scenarios based on feature weight maps and hierarchical clustering algorithms to obtain a candidate classification set. The correction module is used to perform semantic constraint verification on the candidate classification set using a domain knowledge graph to correct classification bias. The second classification module is used to dynamically optimize the corrected classification results through an incremental learning framework to generate the final business classification labels. The triggering module is used to map the final business category label to a preset business rule engine and trigger the corresponding business processing flow.
[0016] Compared with existing technologies, this invention provides a method and system for intelligent perception and accurate classification of business across all scenarios, which has the following beneficial effects: This invention acquires structured, semi-structured, and unstructured data through a multimodal data acquisition unit, breaking away from the traditional method's reliance on a single data type. It can collect information from a wider range of sources, fully utilizing the value inherent in various data types to provide rich material for subsequent analysis, resulting in a more comprehensive and in-depth understanding of business scenarios. An adaptive weight allocation mechanism is employed to evaluate the importance of features in the fused feature matrix, generating a feature weight map. This mechanism automatically assigns reasonable weights to different features based on data characteristics and business needs, highlighting the impact of key features on classification and reducing interference from irrelevant or secondary features, thereby improving classification accuracy. The domain knowledge graph is used to perform semantic constraint verification on the candidate classification set, correcting classification biases. The domain knowledge graph contains rich business domain knowledge and semantic information, which can judge the rationality of the classification results from the semantic level, promptly identify and correct classification errors caused by data noise or algorithm limitations, and ensure that the classification results conform to the actual business. Furthermore, the incremental learning framework is used to dynamically optimize the corrected classification results, enabling the method to automatically adjust the classification model and parameters as business data accumulates and business scenarios change, adapting to new business needs and data characteristics without having to rebuild the entire classification system, thus improving the flexibility and adaptability of the method. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the method flow of S101-S108 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S204 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S403 in this invention. Detailed Implementation
[0018] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0019] Example 1 Please refer to Figure 1 As shown, in a first aspect of the present invention, a scene perception modeling and verification method based on parallel intelligence is provided, comprising: S101. Obtain the raw data stream of the business scenario through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. S102. Perform dynamic feature decoupling processing on the original data stream to generate a multi-dimensional feature vector set; S103. Based on the context association model of business scenarios, spatiotemporal feature fusion is performed on the multi-dimensional feature vector set to obtain the fused feature matrix; S104. An adaptive weight allocation mechanism is used to evaluate the feature importance of the fused feature matrix and generate a feature weight map. S105. Based on the feature weight map, perform preliminary classification of business scenarios using a hierarchical clustering algorithm to obtain a candidate classification set; S106. Use domain knowledge graphs to perform semantic constraint verification on candidate classification sets and correct classification bias. S107. Dynamically optimize the corrected classification results using an incremental learning framework to generate the final business classification labels; S108. Map the final business category label to the preset business rule engine and trigger the corresponding business processing flow.
[0020] Those skilled in the art will understand that this invention acquires structured, semi-structured, and unstructured data through a multimodal data acquisition unit, breaking away from the traditional method's reliance on a single data type. It can collect information from a wider range of sources, fully utilizing the value inherent in various types of data to provide rich material for subsequent analysis, resulting in a more comprehensive and in-depth understanding of business scenarios. Furthermore, it employs an adaptive weight allocation mechanism to evaluate the importance of features in the fused feature matrix, generating a feature weight map. This mechanism automatically assigns reasonable weights to different features based on data characteristics and business needs, highlighting the impact of key features on classification and reducing interference from irrelevant or secondary features, thereby improving classification accuracy. The accuracy of the classification is improved by using a domain knowledge graph to perform semantic constraint verification on the candidate classification set, correcting classification biases. The domain knowledge graph contains rich business domain knowledge and semantic information, which can judge the rationality of the classification results from a semantic level, promptly identify and correct classification errors caused by data noise or algorithm limitations, and ensure that the classification results conform to the actual business. Furthermore, the method dynamically optimizes the corrected classification results through an incremental learning framework, enabling the method to automatically adjust the classification model and parameters as business data accumulates and business scenarios change, adapting to new business needs and data characteristics without having to rebuild the entire classification system, thus improving the flexibility and adaptability of the method.
[0021] Please refer to Figure 2 As shown, dynamic feature decoupling processing is performed on the original data stream to generate a multi-dimensional feature vector set, specifically including: S201. Perform semantic parsing on the unstructured data in the original data stream to extract implicit semantic features. The semantic parsing model adopts the BERT or GPT architecture and outputs a vector with dimension n. S202. Extract statistical features from structured and semi-structured data to generate explicit statistical features, including mean, variance, maximum value, and minimum value. S203. Cross-modal association analysis is performed on implicit semantic features and explicit statistical features through a feature cross-network to generate cross-modal association features. The feature cross-network uses a multilayer perceptron, and the formula is as follows: ; in, For cross-modal association features, For implicit semantic features, These are explicit statistical characteristics. , This is the weight matrix. , For bias terms, It is the ReLU activation function. Use the Sigmoid activation function; S204. Concatenate implicit semantic features, explicit statistical features, and cross-modal association features to form a multi-dimensional feature vector set.
[0022] Spatiotemporal feature fusion includes: Construct a time decay function based on the time series characteristics of the business scenario. The formula is: ; in, The time decay coefficient, The interval between the current time and historical time; Construct a spatial proximity matrix based on the spatial distribution characteristics of the business scenario. Matrix elements Indicates sample With sample spatial distance The reciprocal of, that is: ; Time decay function Spatial proximity matrix Perform weighted fusion to generate a spatiotemporal weight matrix. The formula is: ; in, This represents the time-weighted ratio. Using the spatiotemporal weight matrix For multi-dimensional feature vector sets Dynamic weighting is performed to obtain the fused feature matrix. The formula is: ; in, This indicates element-wise multiplication.
[0023] Adaptive weight allocation mechanisms include: The fusion feature matrix is calculated using an attention mechanism. The attention score ai for each feature dimension is given by the formula: ; in, This is the weight matrix. For attention vectors, For bias terms, The number of feature dimensions; Based on attention score The weight coefficients for each feature dimension are generated using the softmax function. The formula is: ; According to the weighting coefficients For the fusion feature matrix Perform weighted summation to generate a feature weight map. The formula is: .
[0024] Hierarchical clustering algorithms include: Initialize cluster centers Number of cluster centers The complexity is dynamically determined based on the business scenario, using the following formula: ; Where N is the number of samples; Calculate each sample With cluster center similarity Cosine similarity is used, and the formula is: ; Sample Assign to the cluster with the highest similarity ; Update cluster center The formula is: ; in, For clustering The number of samples in the sample; Repeat the above steps until the cluster centers converge, generating a candidate classification set. .
[0025] Please refer to Figure 3 As shown, semantic constraint verification includes: S301, From Domain Knowledge Graph Extract entity relationship chains related to business scenarios ,in For a collection of entities, For a set of relations; S302, Set the candidate classification set Category tags in Mapping to entity relationship chain Generate mapping labels ; S303, Inspection and Classification Labels With mapping labels Semantic consistency is determined using Jaccard similarity, with the formula: ; S304, if Then according to the entity relationship chain Correct category tags .
[0026] Incremental learning frameworks include: Build an initial classification model The initial classification model is based on historical business data. The training process yielded a loss function using cross-entropy loss, with the following formula: ; in, For real labels, For predicting labels, Number of categories; Real-time collection of new business data Extract feature vectors from new business data ; Feature vectors of new business data Input initial classification model Generate predicted classification results ; Calculate the predicted classification results Compared with the true classification results error The formula is: ; like Then update the initial classification model. The parameter θ is determined using the gradient descent method, and the formula is as follows: ; in, The learning rate; Optimize the classification model The above steps are repeated as a new initial classification model.
[0027] Please refer to Figure 4 As shown, the business rules engine includes: S401, Predefined Business Rule Base The business rule base contains the mapping relationship between category tags l and business processing flow P. ; S402, Based on the final business classification label From the business rules library Match the corresponding business processing flow in the middle; S403. If no corresponding business process is matched, the exception handling mechanism is triggered to generate exception handling suggestions.
[0028] In a second aspect of the invention, a scene perception modeling and verification system based on parallel intelligence is also provided, comprising: The acquisition module is used to acquire the raw data stream of the business scenario through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. The decoupling module is used to perform dynamic feature decoupling processing on the original data stream and generate a multi-dimensional feature vector set. The fusion module is used to perform spatiotemporal feature fusion on a multi-dimensional feature vector set based on a context association model of business scenarios, and obtain a fused feature matrix. The generation module is used to evaluate the feature importance of the fused feature matrix using an adaptive weight allocation mechanism and generate a feature weight map. The first classification module is used to perform preliminary classification of business scenarios based on the feature weight map and a hierarchical clustering algorithm to obtain a candidate classification set. The correction module is used to perform semantic constraint verification on the candidate classification set using the domain knowledge graph and correct classification bias. The second classification module is used to dynamically optimize the corrected classification results through an incremental learning framework to generate the final business classification labels. The trigger module maps the final business category tags to a preset business rule engine, triggering the corresponding business processing flow.
[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A scene perception modeling and verification method based on parallel intelligence, characterized in that, include: The raw data stream of the business scenario is obtained through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. Dynamic feature decoupling processing is performed on the original data stream to generate a multi-dimensional feature vector set; Based on the context association model of business scenarios, spatiotemporal feature fusion is performed on multi-dimensional feature vector sets to obtain a fused feature matrix; An adaptive weight allocation mechanism is used to evaluate the feature importance of the fused feature matrix and generate a feature weight map. Based on the feature weight map, a hierarchical clustering algorithm is used to initially classify the business scenarios and obtain a candidate classification set. The domain knowledge graph is used to perform semantic constraint verification on the candidate classification set to correct classification bias. The revised classification results are dynamically optimized using an incremental learning framework to generate the final business classification labels. The final business category tags are mapped to the preset business rule engine, triggering the corresponding business processing flow.
2. The scene perception modeling and verification method based on parallel intelligence according to claim 1, characterized in that, The dynamic feature decoupling process of the original data stream to generate a multi-dimensional feature vector set specifically includes: Semantic parsing is performed on the unstructured data in the original data stream to extract implicit semantic features. The semantic parsing model adopts the BERT or GPT architecture and outputs a vector with dimension n. Statistical features are extracted from structured and semi-structured data to generate explicit statistical features, including mean, variance, maximum and minimum values. Cross-modal association analysis is performed on the implicit semantic features and explicit statistical features using a feature cross-network to generate cross-modal association features. The feature cross-network employs a multilayer perceptron, and the formula is as follows: ； in, For cross-modal association features, For implicit semantic features, These are explicit statistical characteristics. , This is the weight matrix. , For bias terms, It is the ReLU activation function. Use the Sigmoid activation function; The implicit semantic features, explicit statistical features, and cross-modal association features are concatenated to form a multi-dimensional feature vector set.
3. The scene perception modeling and verification method based on parallel intelligence according to claim 2, characterized in that, The spatiotemporal feature fusion includes: Construct a time decay function based on the time series characteristics of the business scenario. The formula is: ； in, The time decay coefficient, The interval between the current time and historical time; Construct a spatial proximity matrix based on the spatial distribution characteristics of the business scenario. Matrix elements Indicates sample With sample spatial distance The reciprocal of, that is: ； The time decay function The spatial proximity matrix S is weighted and fused to generate a spatiotemporal weight matrix. The formula is: ； in, This represents the time-weighted ratio. Using the spatiotemporal weight matrix For the multi-dimensional feature vector set The fusion feature matrix is obtained by performing dynamic weighting. The formula is: ； in, This indicates element-wise multiplication.
4. The method for scene perception modeling and verification based on parallel intelligence for intelligent perception and accurate classification of full-scene business as described in claim 3, characterized in that, The adaptive weight allocation mechanism includes: The fusion feature matrix is calculated using an attention mechanism. The attention score ai for each feature dimension is given by the formula: ； in, This is the weight matrix. For attention vectors, For bias terms, The number of feature dimensions; Based on the attention score The weight coefficients for each feature dimension are generated using the softmax function. The formula is: ； According to the weighting coefficients For the fused feature matrix Perform weighted summation to generate the feature weight map. The formula is: 。 5. The scene perception modeling and verification method based on parallel intelligence according to claim 4, characterized in that, The hierarchical clustering algorithm includes: Initialize cluster centers The number of cluster centers The complexity is dynamically determined based on the business scenario, using the following formula: ； Where N is the number of samples; Calculate each sample With cluster center similarity Cosine similarity is used, and the formula is: ； Sample Assign to the cluster with the highest similarity ; Update cluster center The formula is: ； in, For clustering The number of samples in the sample; Repeat the above steps until the cluster centers converge, generating the candidate classification set. .
6. The scene perception modeling and verification method based on parallel intelligence according to claim 5, characterized in that, The semantic constraint verification includes: From domain knowledge graph Extract entity relationship chains related to business scenarios ,in For a collection of entities, For a set of relations; The candidate classification set Category tags in Mapped to the entity relationship chain Generate mapping labels ; Detect the classification label With mapping labels Semantic consistency is determined using Jaccard similarity, with the formula: ； like Then, based on the entity relationship chain Correct the category labels .
7. The scene perception modeling and verification method based on parallel intelligence according to claim 6, characterized in that, The incremental learning framework includes: Build an initial classification model The initial classification model is based on historical business data. The training process yielded a loss function using cross-entropy loss, with the following formula: ； in, For real labels, For predicting labels, Number of categories; Real-time collection of new business data Extract feature vectors from new business data ; The feature vector of the new business data Input the initial classification model Generate predicted classification results ; Calculate the predicted classification result Compared with the true classification results error The formula is: ； like Then update the initial classification model. The parameter θ is determined using the gradient descent method, and the formula is as follows: ； in, The learning rate; The optimized classification model The above steps are repeated as a new initial classification model.
8. The scene perception modeling and verification method based on parallel intelligence according to claim 7, characterized in that, The business rule engine includes: Predefined business rule library The business rule base contains a mapping relationship between category tags l and business processing flow P. ; According to the final business classification label From the business rule base Match the corresponding business processing flow in the middle; If no matching business process is found, the exception handling mechanism is triggered, and exception handling suggestions are generated.
9. A scene perception modeling and verification system based on parallel intelligence, used to implement the scene perception modeling and verification method based on parallel intelligence as described in any one of claims 1-8, characterized in that, include: The acquisition module is used to acquire the raw data stream of the business scenario through the multimodal data acquisition unit. The raw data stream includes structured data, semi-structured data and unstructured data. A decoupling module is used to perform dynamic feature decoupling processing on the original data stream to generate a multi-dimensional feature vector set. The fusion module is used to perform spatiotemporal feature fusion on a multi-dimensional feature vector set based on a context association model of the business scenario to obtain a fused feature matrix. A generation module is used to evaluate the feature importance of the fused feature matrix using an adaptive weight allocation mechanism, and generate a feature weight map. The first classification module is used to perform preliminary classification of business scenarios based on feature weight maps and hierarchical clustering algorithms to obtain a candidate classification set. The correction module is used to perform semantic constraint verification on the candidate classification set using a domain knowledge graph to correct classification bias. The second classification module is used to dynamically optimize the corrected classification results through an incremental learning framework to generate the final business classification labels. The triggering module is used to map the final business category label to a preset business rule engine and trigger the corresponding business processing flow.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-8.