A large model-based cross-domain data and multi-dimensional joint analysis system and method

Abstract

This invention discloses a cross-domain data and multi-dimensional joint analysis system and method based on a large model. The system includes: a model building module, used to intelligently model at least one corresponding functional model using natural language-driven methods and based on a preset large language model; a processing module, used to: transmit the data to be queried to the functional model and perform multi-source data retrieval using deployed multi-source databases to obtain retrieval results; then perform feature fusion and intelligent analysis calculations based on the retrieval results, and generate an assessment result based on the calculation results; and an information verification module, used to perform multi-modal intelligent analysis technology and combine multiple databases to conduct multi-dimensional deep information correlation analysis on the assessment result, automatically generating a structured investigation report containing risk scores, suspicious behavior feature annotations, and handling suggestions. This invention can realize multi-dimensional intelligent correlation analysis of cross-domain data, effectively improving the intelligent level of port investigation and control.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a system and method for cross-domain data and multi-dimensional joint analysis based on a large model. Background Technology

[0002] Currently, the port inspection and control systems in use mainly employ rule engines and database query technologies. Furthermore, the various functional modules of the system operate independently, data cannot be shared or interconnected, and intelligent analysis capabilities are lacking, resulting in the following problems:

[0003] First, model building relies on manually preset rules, which require frequent modifications and have a slow response time; second, information verification only supports single-condition queries and cannot perform multi-dimensional correlation analysis; and third, risk assessment uses static scoring standards that cannot be dynamically adjusted. Summary of the Invention

[0004] In view of the technical deficiencies mentioned in the background art, the purpose of this invention is to provide a cross-domain data and multi-dimensional joint analysis system and method based on a large model, aiming to at least partially solve one of the technical problems in the related art.

[0005] To achieve the above objectives, in a first aspect, embodiments of the present invention provide a cross-domain data and multi-dimensional joint analysis system based on a large model, the system comprising:

[0006] The model building module is used to perform intelligent modeling in a natural language-driven manner and based on a preset large language model to build at least one corresponding functional model. During model building, model conditions are set, and each model condition is treated as a feature parameter and associated with a corresponding weight.

[0007] Processing module, used for:

[0008] The data to be queried is transmitted to the functional model and combined with the deployed multi-source database to perform multi-source data retrieval to obtain the retrieval results; wherein, there are multiple functional models;

[0009] Then, feature fusion and intelligent analysis calculations are performed based on the search results, and judgment results are generated based on the calculation results;

[0010] The information verification module is used for:

[0011] By employing multimodal intelligent analysis technology and combining multiple databases to conduct multi-dimensional in-depth correlation analysis of the assessment results, a structured investigation report containing risk scores, suspicious behavior feature annotations, and handling suggestions is automatically generated.

[0012] In a preferred implementation of this application, the system further includes a data analysis module, which is used for:

[0013] It integrates analysis entry points for multiple risk types and analyzes each risk type using pre-built assessment models, automatically generating corresponding personnel profiles.

[0014] In a preferred implementation of this application, the system further includes a risk assessment module, which is used for:

[0015] It supports batch uploading of personnel list files and evaluates them according to different functional models selected. Each model automatically performs multi-dimensional feature analysis on each person in the list, and finally generates an assessment report with risk level and outputs a list of key personnel, realizing fully automated processing from data input to risk output.

[0016] As a preferred implementation of this application, the system further includes a data query module, which provides an intelligent model hit risk record tracing function, supporting users to view recent model judgment records and detailed analysis data in real time;

[0017] This module uses a time-series database to store the complete trajectory of each model calculation, and users can filter and query through multi-dimensional conditions; at the same time, clicking on a specific record can view the complete analysis process.

[0018] As one specific implementation of this application, the step of clicking on a specific record to view the complete analysis process specifically includes:

[0019] Feature analysis of input data;

[0020] Key parameters and weight distribution for model computation;

[0021] Derivation of the calculation logic for risk scoring;

[0022] Final judgment criteria and confidence level analysis.

[0023] As a specific implementation of this application, the data query module also supports exporting a single judgment record into a technical report containing visual charts to facilitate subsequent model optimization and result verification; and all query operations are logged with complete audit logs to ensure the traceability of the judgment process.

[0024] Secondly, embodiments of the present invention also provide a method for cross-domain data and multi-dimensional joint analysis based on a large model, applied to the cross-domain data and multi-dimensional joint analysis system based on a large model described in the first aspect. The method includes the following steps:

[0025] Using a natural language-driven approach and based on a pre-defined large language model, intelligent modeling is performed to construct at least one corresponding functional model. During model construction, model conditions are set, and each model condition is treated as a feature parameter and associated with a corresponding weight.

[0026] The data to be queried is transmitted to the functional model and combined with the deployed multi-source database to perform multi-source data retrieval to obtain the retrieval results; wherein, there are multiple functional models;

[0027] Then, feature fusion and intelligent analysis calculations are performed based on the search results, and judgment results are generated based on the calculation results;

[0028] By employing multimodal intelligent analysis technology and combining multiple databases to conduct multi-dimensional in-depth correlation analysis of the assessment results, a structured investigation report containing risk scores, suspicious behavior feature annotations, and handling suggestions is automatically generated.

[0029] The technical solution provided by this invention uses a natural language-driven approach and performs intelligent modeling based on a preset large language model. The data to be queried is then transmitted to the model and combined with a deployed multi-source database for multi-source data retrieval, enabling intelligent cross-domain data correlation analysis. Furthermore, during risk assessment, evaluation is performed based on different selected functional models. Each model automatically performs multi-dimensional feature analysis on each person in the list. Based on the differences between the models, the static defects of current risk assessments are overcome. This solution will effectively improve the intelligence level of port inspection and control, realizing a shift from passive defense to proactive early warning. Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0031] Figure 1 This is a principle block diagram of a cross-domain data and multi-dimensional joint analysis system based on a large model provided in an embodiment of the present invention;

[0032] Figure 2 This is a flowchart of a cross-domain data and multi-dimensional joint analysis method based on a large model provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0035] Please refer to Figure 1 This invention provides a cross-domain data and multi-dimensional joint analysis system based on a large model, the system comprising:

[0036] The model building module is used to perform intelligent modeling in a natural language-driven manner and based on a preset large language model to build at least one corresponding functional model. During model building, model conditions are set, and each model condition is treated as a feature parameter and associated with a corresponding weight.

[0037] Processing module, used for:

[0038] The data to be queried is transmitted to the functional model and combined with the deployed multi-source database to perform multi-source data retrieval to obtain the retrieval results; wherein, there are multiple functional models;

[0039] Then, feature fusion and intelligent analysis calculations are performed based on the search results, and judgment results are generated based on the calculation results;

[0040] The information verification module is used for:

[0041] By employing multimodal intelligent analysis technology and combining multiple databases to conduct multi-dimensional in-depth correlation analysis of the assessment results, a structured investigation report containing risk scores, suspicious behavior feature annotations, and handling suggestions is automatically generated.

[0042] It should be noted that the model conditions involved in this invention are based on the scenario of port inspection and are specifically designed with the following parameters: selected large model, age, ethnic code, regional code, violations and irregularities within the past year, and carrying prohibited items. No restrictions are imposed here. The data to be queried includes real-time data or data received through an interface.

[0043] Because this model building module can construct corresponding functional models based on actual drug-related, fraud-related, and historical illegal and irregular scenarios, it also has model management and model square functions. Model management is used to control the use, editing, and prohibition of functional models. Model square is used to push functional models with good performance to the entire site for sharing.

[0044] During implementation, natural language-driven intelligent modeling technology is adopted. For example, after a police officer inputs an instruction (such as "screen males in xx county, aged 20-60, who have visited a certain area more than 3 times a year"), the Large Language Model (LLM) parses the instruction in real time and automatically generates structured rules (household registration / gender / travel, etc.). The configuration of SQL rules, which would normally take 4 hours, is completed within 5 minutes. Then, the rule chain is dynamically generated according to the interface (household registration matching → gender filtering → travel frequency filtering) and the parameters can be adjusted by dragging and dropping (such as changing the age to 25-55 years old).

[0045] Meanwhile, during the data retrieval process, encrypted feature vector technology is used to link the public security network (household registration / criminal record) and the Meisha system (travel frequency) to achieve zero transmission of raw data.

[0046] And by optimizing the model based on user feedback, a closed-loop model self-optimization mechanism is constructed.

[0047] In this embodiment, to facilitate the provision of targeted strategy recommendations to border inspection authorities, the system further includes a data analysis module, which is used for:

[0048] It integrates analysis entry points for multiple risk types and analyzes each risk type using pre-built assessment models, automatically generating corresponding personnel profiles.

[0049] Specifically, the multiple risk types include those with 21 unreturned permits, illegal activities, and fraud; that is, each risk type is analyzed in a specific way and connected to the bar chart data source ("the situation of people involved in telecommunications fraud at various ports"); among them, 21 permits refer to the Mainland Residents' Travel Permit to Hong Kong and Macao. Some people go to Hong Kong and Macao under the guise of tourism, and then go to other countries to work illegally. These people have not returned after their Hong Kong and Macao visas expired, so they are called people with 21 unreturned permits.

[0050] For the specific analysis of "persons with 21 certificates who have not returned," the system constructed a multi-angle assessment model as one of the assessment models. This model involves data analysis from multiple perspectives, including the household registration distribution, age, and gender of these individuals with 21 certificates. Specifically, it includes:

[0051] A dynamic early warning threshold is established based on the trend of personnel number changes. The distribution characteristics at the provincial and municipal levels are presented through GIS heat maps. The high-risk group concentration range (18-25 years old accounts for 67%) is analyzed in combination with age structure pyramid analysis. Abnormal customs clearance records at high-frequency exit ports such as Shenzhen Bay and Pudong are highlighted. The gender difference coefficient (males account for 82.3%) is introduced to assist in behavior prediction. The associated risk index of customs clearance types such as tourist visas and business visas is further subdivided. Finally, a three-dimensional personnel profile with spatiotemporal markers is generated to provide border inspection authorities with precise deployment strategies.

[0052] This module, which focuses on the analysis of individuals involved in fraud, integrates basic data such as their place of residence, age distribution, and the spatiotemporal characteristics of their document processing. Combined with dynamic information such as communication behavior, it uses machine learning algorithms to establish a multi-dimensional feature analysis model. This model automatically generates visualized profiles of individuals and gang-related maps, which can accurately identify the characteristic patterns of high-risk groups and provide intelligent investigation and apprehension strategy suggestions for public security organs.

[0053] In this embodiment, the information verification module employs multimodal intelligent analysis technology to construct a "three-in-one" verification system:

[0054] 1) Real-time connection to the immigration management database via API interface, supporting the query of personnel's historical entry and exit records in seconds by inputting key information such as name and document number, and automatically marking abnormal entry and exit frequencies (such as multiple trips to sensitive countries within 30 days).

[0055] 2) Based on the large model, conduct in-depth correlation analysis on personnel information across all dimensions, and automatically generate a structured investigation report that includes risk scores (0-100 points), suspicious behavior characteristics annotations, and handling suggestions;

[0056] 3) The biometric comparison system uses a third-generation deep learning algorithm, which can achieve millisecond-level retrieval in a database of hundreds of millions of people by uploading facial photos. It also supports cross-age recognition in scenarios where faces are covered by masks, glasses, etc.

[0057] By leveraging at least three dimensions and implementing data collaboration through a distributed message queue, a closed-loop verification chain is formed, from electronic records to biometrics. This is equivalent to re-verifying and confirming the judgment results of the aforementioned processing modules, thereby reducing false alarms.

[0058] Furthermore, the system also includes a risk assessment module, which is used for:

[0059] It supports batch uploading of personnel list files and evaluates them according to different functional models selected. Each model automatically performs multi-dimensional feature analysis on each person in the list, and finally generates an assessment report with risk level and outputs a list of key personnel, realizing the fully automated processing from data input to risk output. This can be conveniently applied to pre-assessment of personnel who are about to pass through customs, improving the efficiency of related processing.

[0060] When applied, it can also provide risk prediction models based on machine learning (such as XGBoost, random forest, etc.), and users can choose different models for assessment according to their needs; each model automatically performs multi-dimensional feature analysis on each person in the list, including historical behavior, correlation network, abnormal indicators, etc., and finally generates an assessment report with risk level (high / medium / low) and outputs a list of key personnel to be concerned.

[0061] Specifically, a multimodal embedding submodule is added before the XGBoost input layer, and the following processing is performed:

[0062] Text modality (application notes): The text is converted into a 128-dimensional vector using a lightweight BERT model (Mini-BERT), and ambiguous expressions are filtered through an attention layer; such as terms like "few" or "unknown".

[0063] Image modality: Use MobileNet to extract features of “abnormal contours” (such as irregular metal shapes) and output a 64-dimensional vector;

[0064] Feature fusion: Structured features (such as round trip frequency) are fused with multimodal embedding vectors using a gating attention mechanism. Contradictory features are captured through multimodal fusion, and the weights are dynamically adjusted.

[0065] If the application notes state "household goods" but the screenshot shows "regular rectangular metal parts", the improved model captures this contradictory feature through multimodal fusion and raises the risk score from medium risk to high risk.

[0066] It also introduces a real-time feature feedback loop, specifically including:

[0067] Daily statistics on the matching degree between seizure results and characteristics (e.g., of the 10 cases seized on a given day, 8 cases contained the characteristics of "cross-border e-commerce declaration + actual goods exceeding the quantity").

[0068] The feature importance weights are dynamically updated using the exponential moving average (EMA).

[0069] This highlights the time-sensitive nature of port risk characteristics, and the original structure failed to promptly enhance the importance of these characteristics, leading to biased assessments.

[0070] Furthermore, the system also includes a data query module, which provides an intelligent model hit risk record tracing function, supporting users to view recent model judgment records and detailed analysis data in real time;

[0071] This module uses a time-series database to store the complete trajectory of each model calculation, and users can filter and query through multiple dimensions. At the same time, clicking on a specific record will allow you to view the complete analysis process. Users can also filter and query through multiple dimensions such as time range and risk level.

[0072] Clicking on a specific record allows you to view the complete analysis process, which includes:

[0073] Feature analysis of input data;

[0074] Key parameters and weight distribution for model computation;

[0075] Derivation of the calculation logic for risk scoring;

[0076] Final judgment criteria and confidence level analysis.

[0077] Meanwhile, the data query module also supports exporting single analysis records into technical reports containing visual charts to facilitate subsequent model optimization and result verification; and all query operations are logged with complete audit logs to ensure the traceability of the analysis process.

[0078] By applying artificial intelligence and big data technologies, pioneering breakthroughs have been achieved in four aspects: model building, information verification, data analysis, and risk assessment. This provides an integrated intelligent solution for the entire process of early warning, deployment, and handling of port inspection and control. It promotes precise port inspection and control by automatically comparing and analyzing the precise profiles of cross-border criminal groups such as those involved in illegal activities, cross-border gambling, and telecommunications fraud, based on factors such as place of residence, age group, time of obtaining Chinese passport, overseas stay trajectory, and history of illegal activities, and accurately pushing out warnings to suspicious individuals.

[0079] The above solution uses a natural language-driven approach and intelligent modeling based on a pre-set large language model. It then transmits the data to be queried to the model and performs multi-source data retrieval using a deployed multi-source database. This enables intelligent cross-domain data correlation analysis. Furthermore, during risk assessment, different functional models are selected for evaluation. Each model automatically performs multi-dimensional feature analysis on each person in the list. Based on the differences between models, this overcomes the static limitations of current risk assessment methods. This solution will effectively improve the intelligence level of port inspection and control, achieving a shift from passive defense to proactive early warning.

[0080] Based on the same inventive concept, embodiments of the present invention also provide a method for cross-domain data and multi-dimensional joint analysis based on a large model, applied to the cross-domain data and multi-dimensional joint analysis system based on a large model described in the first aspect, the method comprising the following steps:

[0081] S101 uses a natural language-driven approach and performs intelligent modeling based on a pre-set large language model to construct at least one corresponding functional model. During model construction, model conditions are set, and each model condition is treated as a feature parameter and associated with a corresponding weight.

[0082] S102, the data to be queried is transmitted to the functional model and combined with the deployed multi-source database to perform multi-source data retrieval to obtain retrieval results; wherein, the number of functional models is multiple;

[0083] S103, then feature fusion and intelligent analysis calculations are performed based on the search results, and judgment results are generated based on the calculation results;

[0084] S104 employs multimodal intelligent analysis technology and combines multiple databases to conduct multi-dimensional in-depth information correlation analysis on the judgment results, automatically generating a structured investigation report containing risk scores, suspicious behavior feature annotations, and handling suggestions.

[0085] Furthermore, the method also includes:

[0086] It integrates analysis entry points for multiple risk types and analyzes each risk type using pre-built assessment models, automatically generating corresponding personnel profiles;

[0087] It supports batch uploading of personnel list files and evaluates them according to different functional models selected. Each model automatically performs multi-dimensional feature analysis on each person in the list, and finally generates an assessment report with risk level and outputs a list of key personnel to be concerned, realizing the fully automated processing from data input to risk output.

[0088] It provides an intelligent model hit risk record tracking function, allowing users to view recent model assessment records and detailed analysis data in real time;

[0089] A time-series database is used to store the complete trajectory of each model calculation, and users can filter and query through multi-dimensional conditions; at the same time, clicking on a specific record can view the complete analysis process.

[0090] It should be noted that for a more detailed description of the workflow of the method embodiments, please refer to the aforementioned system embodiments section, which will not be repeated here.

[0091] The entire solution employs a natural language-driven approach, employing intelligent modeling based on a pre-set large language model. The data to be queried is then transmitted to the model and combined with a deployed multi-source database for multi-source data retrieval, enabling intelligent cross-domain data correlation analysis. Furthermore, during risk assessment, evaluations are conducted based on the selected functional models. Each model automatically performs multi-dimensional feature analysis on each individual in the list, overcoming the static limitations of current risk assessments. This solution will effectively enhance the intelligence level of port inspection and control, achieving a shift from passive defense to proactive early warning.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A large model-based cross-domain data and multi-dimensional joint analysis system, characterized in that, The system comprises: a model construction module for intelligently modeling in a natural language driven manner and based on a preset large language model to construct at least one corresponding functional model; wherein, during model construction, model conditions are provided, and each model condition is taken as a feature parameter and is associated with a corresponding weight; a processing module for: transmitting the data to be queried to the functional model and combining the deployed multi-source database to perform multi-source data retrieval to obtain retrieval results; wherein, the number of functional models is multiple; based on the retrieval results, performing feature fusion and intelligent analysis and calculation, and generating a research and judgment result according to the calculation result; an information verification module for: adopting a multi-modal intelligent analysis technology and combining multiple databases to perform multi-dimensional information deep correlation analysis on the research and judgment result to automatically generate a structured search report containing a risk score, a suspicious behavior feature label and a disposal suggestion.

2. The system of claim 1, wherein, It also includes a data research and judgment module, which is used for: integrating analysis portals of multiple risk types, and respectively using pre-constructed research and judgment models to analyze each risk type to automatically generate a corresponding personnel portrait.

3. The system of claim 1 or 2, wherein, It also includes a risk assessment module, which is used for: supporting batch uploading of personnel list files, and performing assessment according to different selected functional models, automatically performing multi-dimensional feature analysis on each personnel in the list by each model, finally generating an assessment report with a risk level, and outputting a list of key personnel, realizing full-process automatic processing from data input to risk output.

4. The system of claim 3, wherein, It also includes a data query module, which provides an intelligent model hit risk record tracing function, and supports users to view recent model research and judgment records and detailed analysis data in real time; This module uses a time series database to store the complete trajectory of each model operation, and users can query through multi-dimensional conditions; at the same time, by clicking a specific record, the complete research and judgment process can be viewed.

5. The system of claim 4, wherein, The clicking of the specific record can view the complete research and judgment process, specifically including: feature analysis of input data; key parameters and weight distribution of model operation; calculation logic deduction of risk score; final determination basis and confidence analysis.

6. The system of claim 5, wherein, The data query module also supports exporting a single research and judgment record as a technical report containing a visualization chart, so as to facilitate subsequent model optimization and result review; and all query operations leave complete audit logs to ensure the traceability of the research and judgment process.

7. A large model-based cross-domain data and multi-dimensional joint analysis method, characterized in that, The method is applied to the cross-domain data and multi-dimensional joint analysis system based on a large model of claim 1, and the method comprises the following steps: in a natural language driven manner, intelligently modeling based on a preset large language model to construct at least one corresponding functional model; wherein, during model construction, model conditions are provided, and each model condition is taken as a feature parameter and is associated with a corresponding weight; transmitting the data to be queried to the functional model and combining the deployed multi-source database to perform multi-source data retrieval to obtain retrieval results; wherein, the number of functional models is multiple; based on the retrieval results, performing feature fusion and intelligent analysis and calculation, and generating a research and judgment result according to the calculation result; Adopting multi-modal intelligent analysis technology and combining with multi-database, the method performs multi-dimensional information deep correlation analysis on the research and judgment result, and automatically generates a structured investigation report containing risk score, suspicious behavior characteristic label and disposal suggestion.

8. The method of claim 7, wherein, The method further comprises: An analysis entrance of multiple risk types is integrated, and a pre-constructed research and judgment model is used for analysis of each risk type, thereby automatically generating a corresponding personnel portrait.

9. The method of claim 7, wherein, The method further comprises: Batch uploading of personnel list files is supported, and evaluation is performed according to different selected function models; each model automatically performs multi-dimensional feature analysis on each personnel in the list, finally generates an evaluation report with a risk level, and outputs a list of key personnel, thereby realizing full-process automatic processing from data input to risk output.

10. The method of any one of claims 7 to 9, wherein, The method further comprises: An intelligent model hit risk record tracing function is provided, which supports real-time viewing of recent model research and judgment records and detailed analysis data by the user; A time series database is used to store the complete trajectory of each model operation, and the user can perform multi-dimensional condition filtering and query; meanwhile, by clicking a specific record, the complete research and judgment process can be viewed.

Images