Intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration

Abstract

This invention discloses an intelligent traffic data analysis and decision-making system based on dynamic multi-model collaboration. The system includes a multimodal acquisition module that collects voice, text, video, and structured data and loads it into a traffic knowledge graph; a feature processing module that extracts semantic, temporal, and spatial features and generates a fusion vector; a semantic modeling module that combines the knowledge graph with interaction records to generate an intent vector; a dynamic routing module that selects an edge or cloud model based on the intent vector to generate a query statement; a prompt memory module that fuses historical templates and alignment parameters to generate a query draft; a query validation module that performs semantic and structural validation and outputs a correction statement; a query execution module that generates a response vector; a causal interpretation module that generates an explanation vector; and a response generation module that outputs multi-granularity results based on user roles. This invention improves the collaborative processing capability and semantic parsing accuracy of complex traffic queries.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation systems and artificial intelligence technologies, and in particular to an intelligent transportation data analysis and decision-making system based on multi-model dynamic collaboration. Background Art

[0002] With the continuous development of intelligent transportation systems, the means of acquiring traffic data are becoming increasingly diverse. Multimodal data, such as voice commands, text input, video surveillance, and real-time structured data, are widely used to assist in analysis and decision-making. However, in the collaborative processing of multimodal data, the system still faces three key gaps, which have become a major bottleneck restricting the further application of existing technologies.

[0003] First, there's the issue of discontinuities in intent-based queries. Currently, many systems attempt to convert user voice or text queries into structured database query statements through large language models or natural language processing technologies. However, in real-world scenarios, due to semantic misalignment between the natural language expressions used by users and the field structure, entity labels, and logical relationships in the underlying database, the accuracy of the generated SQL statements is difficult to guarantee. This is especially true when complex spatiotemporal conditions or multi-table join queries are involved, resulting in significantly higher error rates, with some studies showing they can exceed 30%. This issue prevents the system from accurately identifying and responding to users' actual query intent, impacting the system's practicality.

[0004] Secondly, there's a modal alignment gap in the multimodal data fusion process. In traffic management scenarios, data sources such as video streams, text notifications, sensor data, and device status logs are not only heterogeneous but also vary significantly in time scale, spatial location, and semantic granularity. Existing systems lack an effective unified alignment strategy, resulting in difficulty synchronizing data time series, inconsistent spatial references, and inability to align semantic structures during the information fusion phase. Ultimately, the fused data cannot support high-quality joint analysis and query construction.

[0005] Finally, there's an execution gap at the decision-making and response level. Most current systems return results based on static query statements, lacking the ability to dynamically perceive and respond to real-time traffic conditions. For example, in the face of unexpected traffic jams, temporary traffic controls, or traffic light outages, the system struggles to dynamically adjust query intent, response content, or interaction methods to suit the new scenario. This results in delayed system feedback, a lack of flexibility and timeliness, and difficulty meeting the real-time decision-making requirements of traffic management.

[0006] Therefore, how to provide an intelligent traffic data analysis and decision-making system based on dynamic collaboration of multiple models is an urgent problem that needs to be solved by technical personnel in this field. Summary of the Invention

[0007] One purpose of the present invention is to propose an intelligent traffic data analysis and decision-making system based on dynamic collaboration of multiple models. The present invention fully integrates traffic knowledge graphs, lightweight large language models, Prompt memory sharing and multimodal semantic modeling technology, and describes in detail the unified intent parsing, query generation and causal response generation process for multi-source traffic input data such as voice, text, and video. It has the advantages of high semantic understanding accuracy, strong modal alignment capability and good dynamic response flexibility.

[0008] The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to an embodiment of the present invention includes:

[0009] The multimodal acquisition module is used to collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph containing entity types, field labels, and structural relationships;

[0010] Feature processing module, which is used to extract semantic, temporal and spatial features, perform spatiotemporal alignment and normalization, and generate fused feature vectors;

[0011] The semantic modeling module, including the semantic parsing module, graph enhancement unit, spatiotemporal context unit, and historical behavior unit, is used to generate intent vectors based on the fused feature vectors, traffic knowledge graph, and user interaction records;

[0012] A dynamic routing module is used to select whether to deploy a lightweight model on the edge node or execute a multi-model collaborative chain in the cloud to generate query statements based on the intent vector;

[0013] Prompt memory module, used to load the Prompt memory shared pool, fuse historical SQL templates and alignment parameters to generate memory-enhanced query drafts;

[0014] The query verification module is used to perform semantic consistency and structural logic verification based on the traffic knowledge graph and output the corrected query statement;

[0015] A query execution module, used to execute the modified query statement and generate a statement response vector;

[0016] The causal explanation module is used to generate a causal explanation vector by combining the sentence response vector, the intention vector and the causal structure fragment;

[0017] The response generation module is used to generate and output multi-granularity response results based on user roles and response configuration templates.

[0018] Optionally, modules can be connected using the following methods:

[0019] S1. Collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph;

[0020] S2. performing spatiotemporal alignment and normalization processing on the multimodal input dataset, extracting semantic, temporal, and spatial features, and generating a fused feature vector;

[0021] S3: Generate an intention vector with spatiotemporal constraints based on the fusion of feature vectors and traffic knowledge graph and user historical interactions;

[0022] S4. Input the intent vector into the dynamic routing module, and select a lightweight model deployed on the edge node or a multi-model collaborative chain deployed on the cloud to generate a query statement according to the complexity;

[0023] S5. Load the Prompt memory shared pool in the multi-model collaborative chain, fuse the historical SQL template and alignment parameters, and generate a memory-enhanced query draft;

[0024] S6. Performing a semantic and logical double-loop check based on the traffic knowledge graph on the memory-enhanced query draft, and outputting a revised query statement;

[0025] S7. Execute the corrected query statement to obtain the query result, and construct a causal explanation vector by combining the intent vector and the fused feature vector;

[0026] S8. Generate a multi-granularity response result based on the causal explanation vector and the user role, and output a query response.

[0027] Optionally, the Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent paradigm set, wherein:

[0028] The historical SQL template set is used to store SQL templates generated based on historical query tasks, and to establish an index relationship with the query target based on the semantic structure, spatiotemporal context and intent vector;

[0029] The alignment parameter set is used to record structure mapping parameters related to multimodal input data, wherein the structure mapping parameters include video frame time tags, voice segment identifiers, text segment numbers, and structured field alignment relationships, and are uniformly encoded into a standardized alignment parameter vector;

[0030] The intention paradigm set is used to represent the characteristic patterns of the historical intention vectors and is used to perform similarity matching operations between the current intention and the historical paradigms.

[0031] Optionally, the S2 specifically includes:

[0032] S21, extracting initial semantic features, temporal features, and spatial features from the voice commands, text queries, video streams, and structured traffic data in the multimodal input dataset;

[0033] S22. Align the features by timestamp and fuse them to generate a feature matrix:

[0034] ,in ;

[0035] in, is the fusion feature vector sequence, For the The fusion feature unit of time steps, is the semantic feature vector, is the time series characteristic scalar, is the spatial eigenvector, ;

[0036] S23. Normalize and align the channels of the fused feature vector sequence to generate a fused feature vector.

[0037] Optionally, the S3 specifically includes:

[0038] S31, inputting the fused feature vector into the semantic parsing module, extracting the semantic label, relationship structure and context window information, and generating a semantic structure vector;

[0039] S32, fusing the semantic structure vector with entity types, field labels, and structural relationships in the traffic knowledge graph to generate a graph enhancement vector;

[0040] S33, constructing a spatiotemporal context vector based on the time tag and spatial position included in the fused feature vector;

[0041] S34. Load the user's historical interaction records, extract historical keywords, time features, and target preferences, and generate a historical behavior vector;

[0042] S35. Concatenate the graph enhancement vector, spatiotemporal context vector, and historical behavior vector, input them into the multi-layer perceptron structure, and generate the intent vector:

[0043] ;

[0044] in, is the intention vector, is the graph enhancement vector, is the spatiotemporal context vector, is the historical behavior vector, Represents vector concatenation operation, Represents the multi-layer perceptron structure;

[0045] S36. Synchronize the intention vector with the timestamp and use it as input to the dynamic routing module.

[0046] Optionally, the S5 specifically includes:

[0047] S51. Loading a Prompt memory shared pool into the multi-model collaborative chain, wherein the Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent paradigm set;

[0048] S52, calculating the similarity between the current intention vector and the intention pattern set, and selecting the matching unit with the highest similarity;

[0049] S53. Extract the SQL template and alignment parameters from the matching unit, fill the current intent vector into the corresponding variable slot in the SQL template, use the alignment parameters to replace the field name, table structure identifier and screening conditions in the template, and generate a memory-enhanced query draft according to the preset assembly rules.

[0050] Optionally, the S6 specifically includes:

[0051] S61. Based on the field labels in the traffic knowledge graph, perform semantic consistency check on the field names and table names in the memory-enhanced query draft to determine whether the referenced fields match the label definitions and whether the used table structure corresponds to the entity type.

[0052] S62. For semantically inconsistent field names, perform a replacement operation to replace the fields in the draft with standard field names that match the field labels;

[0053] S63. Based on the structural relationships in the traffic knowledge graph, perform structural relationship verification to determine whether the table connection paths, field combination methods, and screening conditions in the draft violate the spatial connection relationships or control dependency rules between entities;

[0054] S64. Construct a conflict detection item for each query statement based on the structural relationship verification rules, and count the number and severity of all violated structural relationship rules as a logical conflict measurement indicator for the statement;

[0055] S65. For query drafts whose logical conflict measures exceed a preset threshold, perform field combination reorganization, conditional clause rewriting, and structural logic adjustment to generate a revised query statement that conforms to the structural relationship of the traffic knowledge graph.

[0056] Optionally, the S7 specifically includes:

[0057] S71. Execute the modified query statement to obtain a structured query result;

[0058] S72. Extract key field values, constraint expressions, and target indicators from the query results and construct a statement response vector.

[0059] S73, jointly processing the sentence response vector and the intention vector to generate an initial interpretation vector;

[0060] S74, based on the semantic features, temporal features, and spatial features contained in the fused feature vector, retrieving causal structure fragments related to the current query event;

[0061] S75. Perform semantic matching and structural alignment on the initial explanation vector and the retrieved causal structure fragment to generate a causal explanation vector.

[0062] Optionally, the S8 specifically includes:

[0063] S81. Generate an initial vector of the response content based on the semantic path, spatiotemporal pattern, and event causal chain structure in the causal explanation vector;

[0064] S82. Obtain a role tag of the current user, where the user roles include traffic management personnel, ordinary travel users, and platform system users;

[0065] S83. Match a corresponding response configuration template based on the user role, where the response configuration template defines the output type, information granularity, and interaction mode;

[0066] S84: Fusing the response content initial vector with the response configuration template to generate a multi-granularity response result, wherein the multi-granularity response result includes heat map display content, text prompt content, and voice broadcast content;

[0067] S85: Perform format generation and terminal adaptation on the multi-granularity response result to complete the output of the query response.

[0068] The beneficial effects of the present invention are:

[0069] This invention addresses the current challenges of collaborative multimodal data processing in the transportation sector, building a comprehensive solution from intent parsing and modal alignment to response generation. Through a semantic modeling module, the system integrates natural language input, such as voice and text, with entity types, field labels, and structural relationships in the transportation knowledge graph. Combined with historical user behavior data, it generates an intent vector with spatiotemporal context constraints. This mechanism effectively mitigates the semantic misalignment between user query intent and the database schema, reducing the mismatching common in traditional text-to-SQL approaches.

[0070] To process heterogeneous modal data, the system has designed a multimodal acquisition and feature processing module that supports unified spatiotemporal alignment and normalization of voice commands, video streams, structured data, and text queries. This module extracts semantic, temporal, and spatial features and generates a fused feature vector. This vector serves as the unified input for subsequent model calculations, effectively enabling the coordinated processing of heterogeneous modalities. This solves the problem of unified processing of video, text, and real-time data, and enhances the system's overall multimodal perception capabilities.

[0071] To address the real-time bottleneck in traffic decision-making, this invention introduces a dynamic routing mechanism. Based on the complexity of the intent vector, lightweight models are deployed at edge nodes, or a multi-model collaborative chain is invoked in the cloud to rapidly generate queries. Furthermore, by leveraging the Prompt memory shared pool, historical SQL templates and structure alignment parameters are incorporated into query draft generation, accelerating query construction while improving query accuracy and structural rationality.

[0072] In addition, the system combines the statement response vector and the intention vector to further construct a causal explanation vector, and generates multi-granular response content based on the user role and response configuration template, including three forms: heat map, text prompts and voice broadcast. It meets the diverse understanding and application needs of users with different roles for traffic information, thereby achieving more flexible and intelligent traffic query response output. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 This is a flowchart of the method for the intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration proposed by the present invention;

[0075] Figure 2 This is the intention vector generation structure diagram of the intelligent traffic data analysis and decision generation system based on multi-model dynamic collaboration proposed by the present invention;

[0076] Figure 3 This is a structural diagram of the Prompt memory sharing pool of the intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration proposed by the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0078] refer to Figure 1-3, an intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration, including:

[0079] The multimodal acquisition module is used to collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph containing entity types, field labels, and structural relationships;

[0080] Feature processing module, which is used to extract semantic, temporal and spatial features, perform spatiotemporal alignment and normalization, and generate fused feature vectors;

[0081] The semantic modeling module, including the semantic parsing module, graph enhancement unit, spatiotemporal context unit, and historical behavior unit, is used to generate intent vectors based on the fused feature vectors, traffic knowledge graph, and user interaction records;

[0082] A dynamic routing module is used to select whether to deploy a lightweight model on the edge node or execute a multi-model collaborative chain in the cloud to generate query statements based on the intent vector;

[0083] Prompt memory module, used to load the Prompt memory shared pool, fuse historical SQL templates and alignment parameters to generate memory-enhanced query drafts;

[0084] The query verification module is used to perform semantic consistency and structural logic verification based on the traffic knowledge graph and output the corrected query statement;

[0085] A query execution module, used to execute the modified query statement and generate a statement response vector;

[0086] The causal explanation module is used to generate a causal explanation vector by combining the sentence response vector, the intention vector and the causal structure fragment;

[0087] The response generation module is used to generate and output multi-granularity response results based on user roles and response configuration templates.

[0088] The present invention constructs a complete intelligent traffic analysis and response system by integrating a multimodal acquisition module, a feature processing module, a semantic modeling module, a dynamic routing module, a prompt memory module, a query verification module, a query execution module, a causal interpretation module and a response generation module. It can achieve efficient linkage between multi-source input, multi-model collaboration and semantic enhancement, and provide users with query response results with clear structure, consistent semantics and diverse forms.

[0089] In this embodiment, the modules are connected through the following methods:

[0090] S1. Collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph;

[0091] S2. performing spatiotemporal alignment and normalization processing on the multimodal input dataset, extracting semantic, temporal, and spatial features, and generating a fused feature vector;

[0092] S3: Generate an intention vector with spatiotemporal constraints based on the fusion of feature vectors and traffic knowledge graph and user historical interactions;

[0093] S4. Input the intent vector into the dynamic routing module, and select a lightweight model deployed on the edge node or a multi-model collaborative chain deployed on the cloud to generate a query statement according to the complexity;

[0094] S5. Load the Prompt memory shared pool in the multi-model collaborative chain, fuse the historical SQL template and alignment parameters, and generate a memory-enhanced query draft;

[0095] S6. Performing a semantic and logical double-loop check based on the traffic knowledge graph on the memory-enhanced query draft, and outputting a revised query statement;

[0096] S7. Execute the corrected query statement to obtain the query result, and construct a causal explanation vector by combining the intent vector and the fused feature vector;

[0097] S8. Generate a multi-granularity response result based on the causal explanation vector and the user role, and output a query response.

[0098] The present invention implements module cascading and collaborative work through eight steps, ensuring a clear input and output logic for the entire process from data collection to response generation, avoiding information loss or misjudgment in the processing chain, and supporting intent-driven query strategies to enhance the system's response accuracy and interpretation capabilities between user intent and actual traffic semantics.

[0099] In this embodiment, the Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent pattern set, wherein:

[0100] The historical SQL template set is used to store SQL templates generated based on historical query tasks, and to establish an index relationship with the query target based on the semantic structure, spatiotemporal context and intent vector;

[0101] The alignment parameter set is used to record structure mapping parameters related to multimodal input data, wherein the structure mapping parameters include video frame time tags, voice segment identifiers, text segment numbers, and structured field alignment relationships, and are uniformly encoded into a standardized alignment parameter vector;

[0102] The intention paradigm set is used to represent the characteristic patterns of the historical intention vectors and is used to perform similarity matching operations between the current intention and the historical paradigms.

[0103] The present invention structures the Prompt memory shared pool into three parts: a historical SQL template set, an aligned parameter set, and an intent paradigm set, and constructs a matching mechanism based on paradigm indexing and parameter alignment, which significantly improves the generation efficiency and structural accuracy of query drafts, while enhancing the adaptability of the memory module to different query contexts.

[0104] In this embodiment, S2 specifically includes:

[0105] S21, extracting initial semantic features, temporal features, and spatial features from the voice commands, text queries, video streams, and structured traffic data in the multimodal input dataset;

[0106] S22. Align the features by timestamp and fuse them to generate a feature matrix:

[0107] ,in ;

[0108] in, is the fusion feature vector sequence, For the The fusion feature unit of time steps, is the semantic feature vector, is the time series characteristic scalar, is the spatial eigenvector, ;

[0109] S23. Normalize and align the channels of the fused feature vector sequence to generate a fused feature vector.

[0110] The present invention ensures the spatiotemporal consistency and structural unity of multimodal data fusion by setting three types of feature extraction paths: semantic, temporal and spatial in the feature processing process, and adopts timestamp alignment and channel standardization fusion strategies, laying a high-quality input foundation for subsequent intent modeling and semantic reasoning.

[0111] In this embodiment, S3 specifically includes:

[0112] S31, inputting the fused feature vector into the semantic parsing module, extracting the semantic label, relationship structure and context window information, and generating a semantic structure vector;

[0113] S32, fusing the semantic structure vector with entity types, field labels, and structural relationships in the traffic knowledge graph to generate a graph enhancement vector;

[0114] S33, constructing a spatiotemporal context vector based on the time tag and spatial position included in the fused feature vector;

[0115] S34. Load the user's historical interaction records, extract historical keywords, time features, and target preferences, and generate a historical behavior vector;

[0116] S35. Concatenate the graph enhancement vector, spatiotemporal context vector, and historical behavior vector, input them into the multi-layer perceptron structure, and generate the intent vector:

[0117] ;

[0118] in, is the intention vector, is the graph enhancement vector, is the spatiotemporal context vector, is the historical behavior vector, Represents vector concatenation operation, Represents the multi-layer perceptron structure;

[0119] S36. Synchronize the intention vector with the timestamp and use it as input to the dynamic routing module.

[0120] The present invention constructs semantic structure vectors, graph enhancement vectors, spatiotemporal context vectors and historical behavior vectors, and uses a multi-layer perception structure to generate intent vectors, so that the system can fully integrate semantic hierarchy, context and behavioral preferences when understanding the user query background, thereby improving the pertinence of routing and query generation strategies.

[0121] In this embodiment, the S5 specifically includes:

[0122] S51. Loading a Prompt memory shared pool into the multi-model collaborative chain, wherein the Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent paradigm set;

[0123] S52, calculating the similarity between the current intention vector and the intention pattern set, and selecting the matching unit with the highest similarity;

[0124] S53. Extract the SQL template and alignment parameters from the matching unit, fill the current intent vector into the corresponding variable slot in the SQL template, use the alignment parameters to replace the field name, table structure identifier and screening conditions in the template, and generate a memory-enhanced query draft according to the preset assembly rules.

[0125] The present invention embeds the Prompt memory sharing pool in the multi-model collaborative chain and introduces a paradigm-based intent matching and field alignment mechanism to achieve dynamic reorganization and parameter mapping of SQL templates, effectively reducing the query generation cost under complex intents and improving the semantic fit and structural standardization of generated statements.

[0126] In this embodiment, S6 specifically includes:

[0127] S61. Based on the field labels in the traffic knowledge graph, perform semantic consistency check on the field names and table names in the memory-enhanced query draft to determine whether the referenced fields match the label definitions and whether the used table structure corresponds to the entity type.

[0128] S62. For semantically inconsistent field names, perform a replacement operation to replace the fields in the draft with standard field names that match the field labels;

[0129] S63. Based on the structural relationships in the traffic knowledge graph, perform structural relationship verification to determine whether the table connection paths, field combination methods, and screening conditions in the draft violate the spatial connection relationships or control dependency rules between entities;

[0130] S64. Construct a conflict detection item for each query statement based on the structural relationship verification rules, and count the number and severity of all violated structural relationship rules as a logical conflict measurement indicator for the statement;

[0131] S65. For query drafts whose logical conflict measures exceed a preset threshold, perform field combination reorganization, conditional clause rewriting, and structural logic adjustment to generate a revised query statement that conforms to the structural relationship of the traffic knowledge graph.

[0132] The present invention introduces a dual-loop mechanism of semantic consistency verification and structural logic verification to perform field standardization, path reorganization and condition rewriting operations on query drafts, thereby ensuring that the generated statements are semantically accurate and structurally compliant, reducing the risks of logical conflicts and execution errors, and enhancing the stability and fault tolerance of the system.

[0133] In this embodiment, the S7 specifically includes:

[0134] S71. Execute the modified query statement to obtain a structured query result;

[0135] S72. Extract key field values, constraint expressions, and target indicators from the query results and construct a statement response vector.

[0136] S73, jointly processing the sentence response vector and the intention vector to generate an initial interpretation vector;

[0137] S74, based on the semantic features, temporal features, and spatial features contained in the fused feature vector, retrieving causal structure fragments related to the current query event;

[0138] S75. Perform semantic matching and structural alignment on the initial explanation vector and the retrieved causal structure fragment to generate a causal explanation vector.

[0139] The present invention constructs a semantic mapping relationship between statement response vectors and causal explanation vectors, and performs semantic alignment in combination with causal structure fragments, so that the query results have stronger explanatory capabilities and causal guidance functions. It is particularly suitable for scenarios such as traffic event backtracking, behavior analysis and response path tracing.

[0140] In this embodiment, the S8 specifically includes:

[0141] S81. Generate an initial vector of the response content based on the semantic path, spatiotemporal pattern, and event causal chain structure in the causal explanation vector;

[0142] S82. Obtain a role tag of the current user, where the user roles include traffic management personnel, ordinary travel users, and platform system users;

[0143] S83. Match a corresponding response configuration template based on the user role, where the response configuration template defines the output type, information granularity, and interaction mode;

[0144] S84: Fusing the response content initial vector with the response configuration template to generate a multi-granularity response result, wherein the multi-granularity response result includes heat map display content, text prompt content, and voice broadcast content;

[0145] S85: Perform format generation and terminal adaptation on the multi-granularity response result to complete the output of the query response.

[0146] The present invention sets user role tags and matches response configuration templates to construct multi-granularity response content such as heat maps, text prompts and voice broadcasts, thereby achieving differentiated services for different users. It can not only meet the information overview needs of management, but also provide travel users with real-time, accurate and personalized query feedback.

[0147] Example 1:

[0148] To verify the feasibility of this invention, we applied it to an intelligent traffic data analysis and decision-making platform. This platform, designed for urban travel services, supports users submitting traffic query requests via various methods, including voice and text. It integrates traffic knowledge graphs with multimodal input datasets to complete the entire process of semantic understanding, query construction, and multi-granularity response output. Traditional systems often suffer from semantic parsing errors, difficulty aligning modal data, and overly static response content, resulting in low system response accuracy and an inability to meet users' real-time, personalized travel query needs.

[0149] In actual deployment, the multimodal acquisition module first acquires user voice commands, text queries, video streams, and structured traffic data, constructs a multimodal input dataset, and loads it into the traffic knowledge graph. The feature processing module normalizes and spatially aligns the multimodal data to generate a fused feature vector. Based on the fused feature vector, the semantic modeling module combines the traffic knowledge graph with user interaction records to generate an intent vector. The dynamic routing module schedules models based on the complexity of the intent vector, invoking lightweight models on the edge or cloud-based collaborative chains to generate query statements.

[0150] During the query construction phase, the Prompt Memory Module loads the Prompt Memory Shared Pool, calls historical SQL templates, and aligns parameters to generate a memory-enhanced query draft. The Query Verification Module verifies the draft for both semantic consistency and structural logic, outputting the revised query statement. Finally, the Query Execution Module retrieves the query results. The Causal Explanation Module constructs a Causal Explanation Vector by combining the intent vector and the statement response vector. The Response Generation Module generates multi-granular response content based on the user role and the response configuration template, and then outputs the result.

[0151] To verify the actual difference between the present invention and the traditional solution, the key indicators of the two systems in general usage scenarios are statistically compared, as shown in Table 1:

[0152] Table 1 System comparison experimental results

[0153] ;

[0154] As can be seen in Table 1, this invention achieves significant improvements in intent vector generation, query construction, response time control, and semantic consistency. In particular, the introduction of the field labels and structural relationships of the traffic knowledge graph into the query validation module significantly reduces query failures due to structural logic conflicts. Furthermore, the introduction of the Prompt memory sharing mechanism further shortens query construction time and improves SQL draft accuracy.

[0155] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration, characterized by: include: The multimodal acquisition module is used to collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph containing entity types, field labels, and structural relationships; Feature processing module, which is used to extract semantic, temporal and spatial features, perform spatiotemporal alignment and normalization, and generate fused feature vectors; The semantic modeling module, including the semantic parsing module, graph enhancement unit, spatiotemporal context unit, and historical behavior unit, is used to generate intent vectors based on the fused feature vectors, traffic knowledge graph, and user interaction records. Specifically, it includes: The fused feature vector is input into the semantic parsing module to extract the semantic label, relationship structure and context window information to generate a semantic structure vector; Fusing the semantic structure vector with entity types, field labels, and structural relationships in the traffic knowledge graph to generate a graph enhancement vector; Constructing a spatiotemporal context vector based on the time tags and spatial positions contained in the fused feature vector; Load user historical interaction records, extract historical keywords, time features, and target preferences, and generate historical behavior vectors; The graph enhancement vector, spatiotemporal context vector, and historical behavior vector are concatenated and input into the multi-layer perceptron structure to generate the intent vector: ； in, is the intention vector, is the graph enhancement vector, is the spatiotemporal context vector, is the historical behavior vector, Represents vector concatenation operation, Represents the multi-layer perceptron structure; Synchronize the intent vector with the timestamp and use it as input to the dynamic routing module; A dynamic routing module is used to select whether to deploy a lightweight model on the edge node or execute a multi-model collaborative chain in the cloud to generate query statements based on the intent vector; Prompt memory module, used to load the Prompt memory shared pool, fuse historical SQL templates and alignment parameters to generate memory-enhanced query drafts; The query verification module is used to perform semantic consistency and structural logic verification based on the traffic knowledge graph and output the corrected query statement; A query execution module, used to execute the modified query statement and generate a statement response vector; The causal explanation module is used to generate a causal explanation vector by combining the sentence response vector, the intention vector and the causal structure fragment; The response generation module is used to generate and output multi-granularity response results based on user roles and response configuration templates.

2. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 1 is characterized in that: The modules are implemented as follows: S1. Collect voice commands, text queries, video streams, and structured traffic data, build a multimodal input dataset, and load the traffic knowledge graph; S2. performing spatiotemporal alignment and normalization processing on the multimodal input dataset, extracting semantic, temporal, and spatial features, and generating a fused feature vector; S3: Generate an intention vector with spatiotemporal constraints based on the fusion of feature vectors and traffic knowledge graph and user historical interactions; S4. Input the intent vector into the dynamic routing module, and select a lightweight model deployed on the edge node or a multi-model collaborative chain deployed on the cloud to generate a query statement according to the complexity; S5. Load the Prompt memory shared pool in the multi-model collaborative chain, fuse the historical SQL template and alignment parameters, and generate a memory-enhanced query draft; S6. Performing a semantic and logical double-loop check based on the traffic knowledge graph on the memory-enhanced query draft, and outputting a revised query statement; S7. Execute the corrected query statement to obtain the query result, and construct a causal explanation vector by combining the intent vector and the fused feature vector; S8. Generate a multi-granularity response result based on the causal explanation vector and the user role, and output a query response.

3. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent pattern set, wherein: The historical SQL template set is used to store SQL templates generated based on historical query tasks, and to establish an index relationship with the query target based on the semantic structure, spatiotemporal context and intent vector; The alignment parameter set is used to record structure mapping parameters related to multimodal input data, wherein the structure mapping parameters include video frame time tags, voice segment identifiers, text segment numbers, and structured field alignment relationships, and are uniformly encoded into a standardized alignment parameter vector; The intention paradigm set is used to represent the characteristic patterns of the historical intention vectors and is used to perform similarity matching operations between the current intention and the historical paradigms.

4. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The S2 specifically includes: S21, extracting initial semantic features, temporal features, and spatial features from the voice commands, text queries, video streams, and structured traffic data in the multimodal input dataset; S22. Align the features by timestamp and fuse them to generate a feature matrix: ； in, is the fusion feature vector sequence, For the The fusion feature unit of time steps, is the semantic feature vector, is the time series characteristic scalar, is the spatial eigenvector, ; S23. Normalize and align the channels of the fused feature vector sequence to generate a fused feature vector.

5. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The S5 specifically includes: S51. Loading a Prompt memory shared pool into the multi-model collaborative chain, wherein the Prompt memory shared pool includes a historical SQL template set, an alignment parameter set, and an intent paradigm set; S52, calculating the similarity between the current intention vector and the intention pattern set, and selecting the matching unit with the highest similarity; S53. Extract the SQL template and alignment parameters from the matching unit, fill the current intent vector into the corresponding variable slot in the SQL template, use the alignment parameters to replace the field name, table structure identifier and screening conditions in the template, and generate a memory-enhanced query draft according to the preset assembly rules.

6. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The S6 specifically includes: S61. Based on the field labels in the traffic knowledge graph, perform semantic consistency check on the field names and table names in the memory-enhanced query draft to determine whether the referenced fields match the label definitions and whether the used table structure corresponds to the entity type. S62. For semantically inconsistent field names, perform a replacement operation to replace the fields in the draft with standard field names that match the field labels; S63. Based on the structural relationships in the traffic knowledge graph, perform structural relationship verification to determine whether the table connection paths, field combination methods, and screening conditions in the draft violate the spatial connection relationships or control dependency rules between entities; S64. Construct a conflict detection item for each query statement based on the structural relationship verification rules, and count the number and severity of all violated structural relationship rules as a logical conflict measurement indicator for the statement; S65. For query drafts whose logical conflict measures exceed a preset threshold, perform field combination reorganization, conditional clause rewriting, and structural logic adjustment to generate a revised query statement that conforms to the structural relationship of the traffic knowledge graph.

7. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The S7 specifically includes: S71. Execute the modified query statement to obtain a structured query result; S72. Extract key field values, constraint expressions, and target indicators from the query results and construct a statement response vector. S73, jointly processing the sentence response vector and the intention vector to generate an initial interpretation vector; S74, based on the semantic features, temporal features, and spatial features contained in the fused feature vector, retrieving causal structure fragments related to the current query event; S75. Perform semantic matching and structural alignment on the initial explanation vector and the retrieved causal structure fragment to generate a causal explanation vector.

8. The intelligent traffic data analysis and decision-making system based on multi-model dynamic collaboration according to claim 2 is characterized in that: The S8 specifically includes: S81. Generate an initial vector of the response content based on the semantic path, spatiotemporal pattern, and event causal chain structure in the causal explanation vector; S82. Obtain a role tag of the current user, where the user roles include traffic management personnel, ordinary travel users, and platform system users; S83. Match a corresponding response configuration template based on the user role, where the response configuration template defines the output type, information granularity, and interaction mode; S84: Fusing the response content initial vector with the response configuration template to generate a multi-granularity response result, wherein the multi-granularity response result includes heat map display content, text prompt content, and voice broadcast content; S85: Perform format generation and terminal adaptation on the multi-granularity response result to complete the output of the query response.

Images