Security level penetration data management system and method of multi-model collaborative engine

Abstract

The invention discloses a security hierarchy penetration data management system and method of a multi-model collaborative engine, relates to the field of data management, and is used for solving the problems that provincial data is difficult to efficiently directly reach a base layer and base layer data cannot be fully reflowed and added due to a hierarchy barrier existing in a traditional data management mode. The system comprises a large model layer, a basic service layer, a service governance layer and an application layer. According to the method, a hierarchical barrier is broken, provincial-level data can directly reach a base layer, safe and ordered flow of the data is ensured through authority control and access queue control of the MCP gateway, and meanwhile, a data asset pool shared by the whole province is formed through backflow and increment of the base layer data. The large model is enabled, through multi-model pipelined cooperation and cross-system data association, the advantages of the large model in the aspects of data cleaning, semantic understanding, early warning and prediction and the like are fully played, and the intelligent level of government affair data processing is improved.

Description

Technical Field

[0001] The present invention relates to the field of data governance, and specifically, to a security-level penetration data governance system and method for a multi-model collaborative engine. Background Art

[0002] During the digital transformation of government, traditional data governance models face hierarchical barriers, making it difficult for provincial-level data to reach grassroots organizations efficiently and effectively, and for grassroots data to be fully circulated and added value. Furthermore, when leveraging big models to power data processing, challenges arise such as difficulty in multi-model collaboration and insufficient cross-system data linkage. This hinders the full implementation of big models' scenario-based value, hindering the progress of digital transformation in government.

[0003] Patent document CN202410485715.4 discloses a data classification prediction method and system using multi-model fusion, including collecting and preprocessing labeled text data and classification lists. A large language model is used to enhance classification descriptions and save information through vectorization technology. Subsequently, a text summary is generated and data enhancement is performed through the model. For classifications with insufficient data, the model is used to synthesize data to meet the minimum data requirements. The fine-tuned BERT model and semantic similarity query are combined to verify each other to ensure correct classification of the text. The text data input by the user is cleaned and desensitized, and text summaries are extracted through the model. These summaries are semantically vectorized, and then the most matching classification is retrieved through the vector database. The present invention combines large language model technology with traditional BERT text classification technology, utilizes the accuracy of BERT text classification for fixed standard classification, and uses large language model technology to improve the robustness of the system for unknown classifications.

[0004] Patent document CN202111478942.7 discloses a method and device for constructing a network sentiment knowledge base based on multi-model data modeling, which relates to the field of Internet technology, including: dividing real-time network traffic data into real-time network sentiment data according to the content of real-time network traffic data; performing data modeling on the real-time network sentiment data based on multiple models to obtain a real-time network sentiment database and store it in a multi-model database; obtaining the network metadata and the application layer protocol parsing data from the multi-model database for association to obtain accumulated network sentiment data; performing data modeling on the accumulated network sentiment data based on multiple models to obtain an accumulated network sentiment database and store it in a multi-model database; associating the real-time network sentiment database and the accumulated network sentiment database stored in the multi-model database to obtain a network sentiment knowledge base. This method not only realizes the comprehensive storage of network sentiment information, but also realizes multi-dimensional knowledge creation.

[0005] None of the above technical solutions solve the problem of hierarchical barriers in traditional data governance methods, which makes it difficult for provincial data to reach the grassroots level efficiently, and grassroots data cannot be fully refluxed for added value. Summary of the Invention

[0006] The purpose of this invention is to provide a method to break the hierarchical barriers of data governance, realize the efficient circulation of provincial data and grassroots data, and enhance the intelligent level of data processing through big model empowerment, so as to provide a reusable implementation paradigm for the digital transformation of government affairs.

[0007] In order to achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0008] A multi-model collaborative engine security layer-penetrating data governance system, including a large model layer, a basic service layer, a service governance layer, and an application layer;

[0009] The large model layer includes various AI large models such as DeepSeek, Tongyi Qianwen, and ChatGL, which are used to provide basic capabilities for natural language processing and image recognition;

[0010] Basic service layer, including MCP service management, RAG, and data center;

[0011] The MCP service management is used to manage model context protocol related services, realize the atomic encapsulation of provincial data interfaces, and support cities to establish intelligent body modules to customize localized data services;

[0012] The RAG is used to combine retrieval technology and generative models, and work with data centers to eliminate the illusion of large models;

[0013] The data center stores various types of data and provides data support for the upper layer;

[0014] The service governance layer includes permission control, monitoring logs, service orchestration, interface management, flow control, and security auditing;

[0015] The authority control is used to build a unified data pass system for the entire province, manage user or role access rights, and achieve direct access to emergency instructions and batch distribution of non-sensitive data during off-peak hours through flow control and interface management;

[0016] The monitoring log is used to monitor the system operation status in real time and record operation logs;

[0017] The service orchestration is used to combine and schedule different services according to business needs;

[0018] The interface management is used to manage the internal and external interfaces of the system;

[0019] The flow control is used to adjust and distribute the request flow according to the system load;

[0020] The security audit is used to audit system security-related operations;

[0021] Application layer, including urban governance components, government office components;

[0022] The urban governance component covers video intelligent tagging and pan-event discovery business application fields.

[0023] The government office component is a specific functional module constructed based on a bottom-layer large model and basic services, and realizes man-machine dialogue, graphic processing, intelligent analysis, document processing, and intelligent customer service for government office.

[0024] The multi-model support dynamic loading component of the large model layer realizes multi-model collaborative processing of data.

[0025] The dynamic loading component includes a task allocation module, a model selection module, a data transmission module, and a result integration module.

[0026] The task allocation module is the starting point of multi-model collaboration. When the system receives a data processing task, the task allocation module analyzes the task and determines the required data processing capabilities, such as natural language processing, image recognition, or data analysis.

[0027] Taking a certain 12345 hotline voice data processing task as an example, the task allocation module identifies the need for voice-to-text conversion, classification, and generation of disposal suggestions.

[0028] The model selection module selects the most suitable model from a large number of models based on the capability requirements parsed by the task allocation module.

[0029] This module has a clear understanding of the advantages and applicable scenarios of each model. For example, Tongyi Qianwen excels in voice-to-text conversion, DeepSeek is good at data classification, and ChatGLM is suitable for generating text content. For the 12345 hotline voice data task, the model selection module will select Tongyi Qianwen for voice translation, DeepSeek for classification, and ChatGLM for generating disposal suggestions.

[0030] The data transmission module is responsible for accurately transmitting data between different models. During transmission, it ensures that the data format meets the requirements of the target model.

[0031] For example, after Tongyi Qianwen converts voice into text, the data transmission module will organize and transmit the text data in the format required by DeepSeek to ensure smooth data flow between models.

[0032] The result integration module integrates the output results of different models to form a complete and business-demand-compliant final result when each model completes its respective data processing task.

[0033] For the 12345 hotline voice data processing, the result integration module will organize the classification results and disposal suggestions and provide them for use in subsequent business processes.

[0034] The present invention also provides a security layer penetration data governance method for a multi-model collaborative engine, which adopts the security layer penetration data governance system for a multi-model collaborative engine as described in claim 6, and is characterized by comprising the following steps:

[0035] Step 1: Data downlink channel: This encapsulates provincial and ministerial data resources through interface / data modules, supporting direct calls from prefecture-level and municipal systems through the user-side large-scale model API. The MCP gateway dynamically allocates data downlink traffic to ensure real-time transmission of emergency and livelihood data or control data permissions and priority queues. Data is double-checked by the public security firewall and Internet behavior management, with sensitive fields automatically desensitized during transmission. Data is then transmitted from the provincial and ministerial data center to prefecture-level and municipal users via interface services, MCP local gateways, and dedicated networks.

[0036] Among them, the data sinking channel, first of all, adopts interface standardization, encapsulates provincial and ministerial data resources through interface / data modules, and supports direct calls from local and municipal systems through the large model API on the user side.

[0037] Second, by adopting intelligent routing, the MCP gateway dynamically allocates data decentralization traffic, giving priority to ensuring the real-time transmission of emergency and livelihood data, or controlling data permissions and priority queues.

[0038] Third, secure penetration: data flows through the public security firewall + Internet behavior management double review, and sensitive fields are automatically desensitized during the transmission process; data is transmitted from the provincial and ministerial data centers through interface services, MCP local gateways, and dedicated lines to users in prefecture-level cities;

[0039] Step 2: Data empowerment closed loop: grassroots data is uploaded to the integrated machine cluster via a dedicated network and stored in a local database or provincial and ministerial data center. Multi-model collaborative processing of grassroots data is used, and the processing results are transmitted back to the provincial and ministerial data center, forming a data-intelligence-decision-making enhanced closed loop.

[0040] It's important to note that the first step is to form a closed loop, and the second is to enhance it. Specifically, grassroots data is uploaded to the integrated appliance cluster via a dedicated network for storage in a local database. Alternatively, it can be uploaded to the provincial and ministerial data center for storage and processing. Multiple models are used to collaboratively process grassroots data, leveraging their respective strengths to form a closed loop. Furthermore, the processing results are transmitted back to the provincial and ministerial data center, forming an enhanced closed loop of data, intelligence, and decision-making.

[0041] Step 3: External data acquisition: The all-in-one cluster captures external public opinion data from the Internet through the public security firewall, and uses a large model to analyze it after cleaning it through Internet behavior management; the large model is used to process and output secure desensitized statistical data to support public services.

[0042] External data acquisition is divided into two channels: internal and external. Specifically, internally, external public opinion data is captured from the internet via an integrated appliance cluster through the public security firewall. After being cleansed through online behavior management, it is analyzed using large models. Externally, through large-scale processing, desensitized statistical data is securely output to support public services.

[0043] Compared with the existing technology, the present invention, first, breaks down hierarchical barriers and enables provincial data to reach the grassroots level directly. It also ensures the safe and orderly flow of data through permission control and access queue control of the MCP gateway. At the same time, the grassroots data flows back and increases in value to form a data asset pool shared by the entire province.

[0044] Second, empowering big models, through multi-model pipeline collaboration and cross-system data association, fully leveraging the advantages of big models in data cleaning, semantic understanding, early warning and prediction, and improving the intelligent level of government data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention can be further illustrated by the non-limiting examples given in the accompanying drawings;

[0046] Figure 1 It is a schematic diagram of the process of the present invention;

[0047] Figure 2 Schematic diagram of the present invention. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0049] like Figure 1 , Figure 2 As shown, a multi-model collaborative engine security layer penetration data governance system includes a large model layer, a basic service layer, a service governance layer, and an application layer;

[0050] The large model layer includes various AI large models such as DeepSeek, Tongyi Qianwen, and ChatGL, which are used to provide basic capabilities for natural language processing and image recognition;

[0051] Basic service layer, including MCP service management, RAG, and data center;

[0052] The MCP service management is used to manage model context protocol related services, realize the atomic encapsulation of provincial data interfaces, and support cities to establish intelligent body modules to customize localized data services;

[0053] The RAG is used to combine retrieval technology and generative models, and work with data centers to eliminate the illusion of large models;

[0054] The data center stores various types of data and provides data support for the upper layer.

[0055] The service governance layer includes permission control, monitoring log, service orchestration, interface management, traffic control, and security audit.

[0056] The permission control is used to build a unified data pass system in the province, manage user or role access permissions, and realize emergency command direct access and non-sensitive data batch release during idle time through traffic control and interface management.

[0057] The monitoring log is used to monitor the system running state in real time and record operation logs.

[0058] The service orchestration is used to combine and schedule different services according to business needs.

[0059] The interface management is used to manage internal and external interfaces of the system.

[0060] The traffic control is used to adjust and distribute request traffic according to system load.

[0061] The security audit is used to audit system security-related operations.

[0062] The application layer includes city governance components and government office components.

[0063] The city governance component covers video intelligent labeling and pan-event discovery business application fields.

[0064] The government office component is a specific function module built based on the bottom layer large model and basic service, which realizes man-machine dialogue, image processing, intelligent analysis, document processing, and intelligent customer service for government office.

[0065] Preferably, the multi-model of the large model layer supports dynamic loading components to realize multi-model collaborative processing of data.

[0066] The dynamic loading component includes a task allocation module, a model selection module, a data transmission module, and a result integration module.

[0067] The task allocation module is the starting point of multi-model collaboration. When the system receives a data processing task, the task allocation module analyzes the task and determines the required data processing capabilities, such as natural language processing, image recognition, or data analysis.

[0068] The model selection module selects the most suitable model from a large number of models based on the capability requirements analyzed by the task allocation module.

[0069] The data transfer module is responsible for accurately transferring data between different models. During the transfer process, it ensures that the data format meets the requirements of the target model.

[0070] The result integration module, after each model completes its own data processing task, summarizes and integrates the output results of different models to form a complete final result that meets business needs.

[0071] The collaborative processing steps of dynamically loaded components are as follows:

[0072] S1, model warehouse calling step, dynamic loading component has a model warehouse, which stores various large models. The models exist in the form of independent components, waiting to be called; such as DeepSeek, Tongyi Qianwen, ChatGLM, etc.

[0073] S2, the resource management monitoring step, is responsible for monitoring system resource usage, including memory and CPU. When a new model needs to be loaded, the resource management determines whether the current system resources are sufficient. If so, it allocates the appropriate resources for the new model. If not, it adjusts resource allocation based on task priority and resource usage to ensure the new model can be loaded smoothly.

[0074] S3, the loading control step, receives instructions from the model selection module. The dynamic loading component loads the required model from the model repository. It collaborates with resource management to initialize and configure the model, enabling it to quickly integrate into the system and begin processing tasks. When a task is completed or system resources are limited, the module promptly unloads unused models, freeing up system resources and preparing for subsequent model loading. This technical solution achieves dynamic loading and efficient collaboration of multiple models through the coordinated operation of these dynamic loading components.

[0075] Preferably, the APIs for atomically encapsulating provincial data interfaces managed by MCP service include population query and corporate credit investigation.

[0076] Preferably, RAG combines retrieval technology to retrieve information from a database to assist the large model in generating more accurate content, wherein the data comes from a data center.

[0077] Preferably, a data pass system built with permission control can apply for data access permissions based on demand at the municipal level. For example, a county can only access population data within its jurisdiction.

[0078] Preferably, monitoring logs can be used to troubleshoot, analyze system performance and user behavior.

[0079] A security layer penetration data governance method for a multi-model collaborative engine includes the following steps:

[0080] Step 1, data sinking channel, provincial data resources are encapsulated by interface / data module, supporting city systems to directly call through user-side large model API; MCP gateway dynamically allocates data downlink traffic, ensuring real-time transmission of emergency and livelihood data or controlling data rights and priority queue; data is double-checked by public security firewall and Internet behavior management, sensitive fields are automatically desensitized during transmission, and data is transmitted from provincial data center to city users through interface service, MCP local gateway, and dedicated line network.

[0081] Step 2, data empowerment closed loop, grassroots data is uploaded to all-in-one machine cluster and stored in local database or provincial data center; multi-model collaborative processing is used for grassroots data; the processing results are fed back to the provincial data center to form a data-intelligent-decision enhancement closed loop.

[0082] Step 3, external data acquisition, all-in-one machine cluster grabs external public opinion data from Internet through public security firewall, and analyzes the data through large model after cleaning by Internet behavior management; safe output desensitization statistical data is supported for public service through large model processing.

[0083] Preferably, in the data empowerment closed loop step, different models take their strengths when multi-model collaborative processing is used for grassroots data, DeepSeek serves as the core of government data cleaning, and TongYiQianWan serves as the semantic understanding bridge.

[0084] The hierarchical barriers of traditional data governance limit the efficient circulation and value-added of data between the provincial and grassroots levels. The technical solution effectively solves this problem by the following technical solutions.

[0085] Through the data sinking channel, provincial data efficiently reaches the grassroots level.

[0086] Interface standardization: the provincial data center encapsulates key resources such as provincial permanent population database and crime data through interface / data module, and converts them into standard API. This enables city systems to directly call through user-side large model API, breaking the technical barriers of data calling and realizing the opening of provincial data to grassroots in a standard and convenient way.

[0087] Intelligent routing: MCP gateway plays a core role in dynamically allocating data downlink traffic. For emergency and livelihood data, real-time transmission is prioritized to ensure that grassroots can obtain important information in a timely manner. At the same time, different types of data can be reasonably scheduled according to data rights and priority queue, such as batch downlink of regular data during night time, which improves network resource utilization and ensures the orderliness of data transmission.

[0088] Secure penetration: During data transmission, data undergoes dual scrutiny by public security firewalls and online behavior management to ensure data security and compliance. Furthermore, sensitive fields, such as the last four digits of ID numbers, are automatically desensitized, ensuring that provincial-level data reaches the grassroots level efficiently and securely while maintaining data security.

[0089] Through data empowerment closed loop, we can achieve added value by reflux of grassroots data.

[0090] Data upload: Surveillance video and other data generated by grassroots units is uploaded to the integrated machine cluster via a dedicated network. It can be stored in a local database or, as needed, uploaded to a provincial or ministerial data center. This provides a foundation for centralized processing and analysis of grassroots data.

[0091] Multi-model collaboration: Leveraging the strengths of different large-scale models, we deeply process data uploaded by grassroots organizations. For example, for traffic monitoring data, DeepSeek is responsible for data cleaning, Tongyi Qianwen performs semantic understanding and transformation, and other models complete subsequent analysis to generate valuable information such as province-wide congestion heat maps.

[0092] Results are transmitted and shared: Processed results are transmitted back to the provincial and ministerial data center, forming a closed loop of data-intelligence-decision-making enhancement. The provincial and ministerial departments make decisions based on these results and share the value-added data assets among various municipalities. For example, a municipal enterprise credit analysis model can be reused throughout the province, achieving the return and value-added of grassroots data.

[0093] The above is a detailed introduction to the security level penetration data governance system and method of a multi-model collaborative engine provided by the present invention. The description of the specific embodiment is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A security level penetration data governance system for a multi-model collaborative engine, characterized by: Including large model layer, basic service layer, service governance layer, and application layer; The large model layer includes various AI large models such as DeepSeek, Tongyi Qianwen, and ChatGL, which are used to provide basic capabilities for natural language processing and image recognition; Basic service layer, including MCP service management, RAG, and data center; The MCP service management is used to manage model context protocol related services, realize the atomic encapsulation of provincial data interfaces, and support cities to establish intelligent body modules to customize localized data services; The RAG is used to combine retrieval technology and generative models, and work with data centers to eliminate the illusion of large models; The data center stores various types of data and provides data support for the upper layer; The service governance layer includes permission control, monitoring logs, service orchestration, interface management, flow control, and security auditing; The authority control is used to build a unified data pass system for the entire province, manage user or role access rights, and achieve direct access to emergency instructions and batch distribution of non-sensitive data during off-peak hours through flow control and interface management; The monitoring log is used to monitor the system operation status in real time and record operation logs; The service orchestration is used to combine and schedule different services according to business needs; The interface management is used to manage the internal and external interfaces of the system; The flow control is used to adjust and distribute the request flow according to the system load; The security audit is used to audit system security-related operations; The application layer includes urban governance components and government office components; The city governance components cover the application areas of video intelligent tagging and pan-event discovery services; The government office components are specific functional modules built based on the underlying big model and basic services, which realize human-computer dialogue, image and text processing, intelligent analysis, document processing, and intelligent customer service in government office.

2. The security level penetration data governance system of the multi-model collaborative engine according to claim 1 is characterized in that: The multiple models of the large model layer support dynamic loading of components to achieve multi-model collaborative data processing.

3. The security level penetration data governance system of the multi-model collaborative engine according to claim 2 is characterized in that: The dynamic loading component includes a task allocation module, a model selection module, a data transmission module, and a result integration module; The task assignment module is the starting point for multi-model collaboration. When the system receives a data processing task, the task assignment module analyzes the task and identifies the data processing capabilities required for the task, such as natural language processing, image recognition, or data analysis. The model selection module selects the most suitable model from a plurality of models according to the capability requirements analyzed by the task allocation module; The data transfer module is responsible for accurately transferring data between different models and ensuring that the data format meets the requirements of the target model during the transfer process; The result integration module, after each model completes its own data processing task, summarizes and integrates the output results of different models to form a complete final result that meets business needs.

4. The security level penetration data governance system of the multi-model collaborative engine according to claim 3 is characterized in that: The collaborative processing steps of the dynamic loading component are as follows: S1, the model warehouse calling step, the dynamic loading component has a model warehouse, which stores various large models. The models exist in the form of independent components and are waiting to be called; S2, the resource management monitoring step, is responsible for monitoring system resource usage, including memory and CPU. When a new model needs to be loaded, the resource management determines whether the current system resources are sufficient. If so, it allocates the corresponding resources to the new model. If not, it adjusts resource allocation based on task priority and resource usage to ensure that the new model can be loaded smoothly. S3, the loading control step, after receiving the instruction from the model selection module, the dynamic loading component loads the required model from the model repository; it will collaborate with the resource management to complete the initialization and configuration of the model, so that it can be quickly integrated into the system and start processing tasks; when the task is completed or the system resources are tight, the module can also promptly unload the unused model, freeing up system resources and preparing for subsequent possible model loading.

5. The security level penetration data governance system of the multi-model collaborative engine according to claim 4 is characterized in that: The MCP service manages the APIs for atomically encapsulating provincial data interfaces, including population query and corporate credit investigation.

6. The security level penetration data governance system of the multi-model collaborative engine according to claim 5 is characterized in that: The RAG is combined with retrieval technology to retrieve information from the database to assist the large model in generating more accurate content, and the data comes from the data center.

7. The security level penetration data governance system of the multi-model collaborative engine according to claim 6 is characterized in that: The data pass system constructed by the permission control can apply for data call permissions on demand according to cities and counties.

8. The security level penetration data governance system of the multi-model collaborative engine according to claim 7 is characterized in that: The monitoring logs can be used to troubleshoot, analyze system performance and user behavior.

9. A security level penetration data management method for a multi-model collaborative engine, using the security level penetration data management system for a multi-model collaborative engine as claimed in claim 8, characterized in that: The following steps are involved: Step 1: Data sinking channel, encapsulating provincial and ministerial data resources through interface / data modules, supporting direct calls from prefecture-level and municipal systems through the user-side big model API; The MCP gateway dynamically allocates data downlink traffic to ensure real-time transmission of emergency and livelihood data or control data permissions and priority queues. Data is double-checked by the public security firewall and Internet behavior management, and sensitive fields are automatically desensitized during transmission. Data is transmitted from the provincial and ministerial data centers via interface services, MCP local gateways, and dedicated lines to users in prefecture-level cities. Step 2: Data empowerment closed loop: grassroots data is uploaded to the integrated machine cluster via a dedicated network and stored in a local database or provincial and ministerial data center. Multi-model collaborative processing of grassroots data is used, and the processing results are transmitted back to the provincial and ministerial data center, forming a data-intelligence-decision-making enhanced closed loop. Step 3: External data acquisition: The all-in-one cluster captures external public opinion data from the Internet through the public security firewall, cleans it through online behavior management, and then uses a large model for analysis; Support public services through large-scale model processing to securely output desensitized statistical data.

10. The security level penetration data governance method of the multi-model collaborative engine according to claim 9 is characterized in that: In the data empowerment closed-loop step, when multiple models collaborate to process grassroots data, different models each take advantage of their strengths. DeepSeek serves as the government data cleaning hub, and Tongyi Qianwen serves as a semantic understanding bridge.

Images