Distributed hierarchical digital intelligent central station system based on large language model

Abstract

A large language model-based distributed and layered digital intelligent central station system comprises an infrastructure service layer, and the infrastructure service layer carries out global computing power resource scheduling to realize operation and maintenance intelligence of the distributed and layered digital intelligent central station system and improve the resource utilization rate, and provides a stable operation environment for other layers; the data service layer provides general data and customized data service for the model service layer and the software service layer under the driving of a large model for the data acquired from each network system of the infrastructure service layer; the model service layer manages a model through a model training platform and a model reasoning platform based on the data of the data service layer, establishes mapping and alignment from a task to a small model, and supports intelligent application of the software service layer through model operation reasoning; the software service layer provides various intelligent data application services according to typical application scenes, and provides various data products directly for users. The method has the characteristics of multi-modal interaction, operation and maintenance intelligentization, collection agility, governance autonomy and product generation.

Description

Technical Field

[0001] The present invention relates to the field of data management technology, and in particular to a distributed hierarchical digital intelligence middle platform system based on a large language model. Background Art

[0002] Through data assetization and data sharing, the data middle platform can transform data silos into data services, effectively realizing the goal of "making data usable." However, with the development of intelligent technology and changes in business needs, the construction of the data middle platform has also ushered in new opportunities and challenges.

[0003] On the one hand, the cost of data life cycle management in the data middle platform is high, and new intelligent technologies need to be introduced to achieve efficient data management: the data life cycle runs through the entire process of data collection, aggregation, governance, service, operation and maintenance, and destruction. At present, the data middle platform is still highly dependent on manual management and fails to achieve efficient and automated operation of data. It is necessary to study new driving mechanisms to achieve a closed loop of the data life cycle and turn global, massive, multi-source, and heterogeneous data resources into data assets.

[0004] On the other hand, the overall intelligence level of the data middle platform is limited and it is difficult to adapt to flexible and changing business needs: the data middle platform still has problems such as a single human-computer interaction mode, limited data analysis capabilities, limited data efficiency mining, limited data intelligence applications, and insignificant intelligent decision-making effectiveness. It is unable to provide timely and accurate data insights and decision-making support. It needs to transform from a product-centric to a service-centric approach, rely on intelligent technology to achieve agile development, and realize data intelligence to ultimately empower decision-making. Summary of the Invention

[0005] The purpose of the present invention is to provide a distributed hierarchical digital intelligence middle platform system based on a large language model, which overcomes the problems of loose connection between the existing data middle platform and the business and limited intelligence level.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A distributed hierarchical digital intelligence middle platform system based on a large language model, comprising: an infrastructure service layer, a data service layer, a model service layer and a software service layer; the infrastructure service layer performs global computing power resource scheduling through containers, virtual machines and bare metal to realize intelligent operation and maintenance of the distributed hierarchical digital intelligence middle platform system and improve resource utilization, and provide a stable and reliable operating environment for the data service layer, model service layer and software service layer; the data service layer obtains data from various networks of the infrastructure service layer, and provides general data services and customized data services to the model service layer and the software service layer under the drive of a large model according to the data structure, application field and application scenario characteristics; the model service layer manages the model through the model training platform and the model inference platform based on the data of the data service layer, establishes a mapping and alignment of tasks to small models, and supports the intelligent application of the software service layer through model operation reasoning; the software service layer provides various intelligent data application services according to typical application scenarios, and directly provides a variety of data products to users.

[0008] Furthermore, the infrastructure service layer is represented as Cloud(N,C,U), where N represents communication; C represents computing power; U represents container; and the data service layer is represented as Represents customized data services, where m represents the number of customized data types according to the business type; G represents general data services; the model service layer is represented by LLM(s)·[m i,j ], where LLM(s) is a large language model group, [m i,j ] represents various general AI models and business model set matrices, i represents the business type, j represents the jth one, that is, Mi,j represents the jth small model of the i-th business, and the software service layer is represented by p~N(μ,σ 2 ), the distributed hierarchical digital intelligence middle platform system adopts an autonomous loosely coupled full-level architecture, which is expressed as f(x)~N(0,1) indicates that the application emerges in a point-like manner and gradually tends to a normal distribution.

[0009] Furthermore, the infrastructure service layer includes various heterogeneous network facilities and a data cloud platform that integrates virtualization, containerization, image repository and monitoring services for disaster recovery and backup.

[0010] Furthermore, the data service layer includes a multimodal database with data collection, data storage, data processing, and data analysis functions. The general data service classifies data according to the data structure, and divides data in PDF, doc, docx, txt, mp3, mp4, png, and jpg formats into unstructured data, semi-structured data, structured data, and streaming data to support general data services; the customized data service is classified and organized according to business type and classified based on application scenario characteristics to support customized data services.

[0011] Furthermore, the model service layer includes various large language models, general AI models and business models. Driven by various large language models, it leads various general AI models and business model set matrices, and establishes mapping and alignment from tasks to small models through semantic description of general AI models or business models.

[0012] Furthermore, the software service layer includes a big data task platform, an application portal and interactive window, an application scheduling system and a block generation system for various typical scenarios. It adopts a natural language interaction mode, conducts human-computer interaction with users through natural language, pushes data in the form of graphics and text, model visualization services and reasoning analysis results, and realizes block generation.

[0013] Furthermore, the multimodal database specifically includes analytical database, relational database, spatiotemporal database, graph library, vector library and index library; the general data service specifically includes data collection, data governance, data fusion, data storage and computing, data assets and data services.

[0014] Furthermore, the various types of large language models include domain-specific large language models, basic general large language models and multimodal large models; small models include general AI models for text recognition, video summarization, human detection, and face recognition, and reusable business models for specific business scenarios.

[0015] Furthermore, the block generation is performed with the block as the minimum granularity, and data products are reproduced according to user needs, and data application services are flexibly customized.

[0016] Furthermore, the model training platform is used for model management and optimization, including model management, model construction, model training, model optimization and model evaluation modules; the model inference platform is used to call various large language models and small model sets for inference, including model visualization, model deployment, model inference, model sharing and model encapsulation modules.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1) This invention adopts generative artificial intelligence technology represented by large language models to achieve deep integration of data and AI driven by large models. Relying on its outstanding semantic understanding, content generation, and tool calling capabilities, it can achieve a leap from traditional experience-based decision-making to intelligent decision-making.

[0019] 2) The present invention generates content through a large language model to obtain various data analysis reports, thereby improving data analysis capabilities. It can form a "picture-based" data display interactive interface based on traditional chart display statistics, and accurately present the comprehensive results of data statistics, analysis, and reasoning in a more diverse visual way, effectively improving the diversity and intuitiveness of services.

[0020] 3) The present invention can realize intelligent data annotation, agile data collection, and autonomous data governance through a large language model. It can enrich the amount of data and the form of data expression through the generation of vector data and knowledge graphs, and realize intelligent and agile data life cycle management, making the data management mechanism more intelligent.

[0021] In summary, the present invention effectively integrates the traditional "cloud platform, data middle platform, AI middle platform, and business middle platform", and has significant features such as multimodal interaction, intelligent operation and maintenance, agile collection, autonomous governance, and product generation. It can realize the four generation loops of computing power generation, data generation, model generation, and application generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a functional architecture diagram of the present invention.

[0023] Figure 2 This is the overall architecture diagram of the system of the present invention.

[0024] Figure 3 The data annotation workflow diagram of the present invention is shown. DETAILED DESCRIPTION

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, they can also obtain drawings of other embodiments based on these drawings without making any creative efforts.

[0026] like Figure 1As shown, a distributed hierarchical digital intelligence middle platform system based on a large language model, from bottom to top, can be understood that the lower layer in the functional architecture is the infrastructure of the upper layer, and the services implemented by the upper layer need to rely on the lower layer to be implemented, which are the infrastructure service layer, data service layer, model service layer and software service layer. The infrastructure service layer is to ensure the system computing power and network infrastructure, the data service layer is the data service base of the middle platform, the model service layer is the key to realize intelligence, and the software service layer is the interface of intelligent interaction. Specifically, it can be expressed as Where: LLM(s) is a large language model group; f(x)~N(0,1) represents the application of point-like emergence and is normally distributed; [m i,j ] represents various business and AI small model matrices; Represents customized data services, specifically divided according to business types; G represents generalized data services, specifically including data collection and data governance; Cloud represents network-cloud integration; N represents communication; C represents computing power; and U represents container. This invention is driven by a large model group. The infrastructure layer (cloud part) is the foundation (in the denominator). The numerator represents (driven by the large model) the data layer to complete "general + customized" agile data collection and drive small model sets, ultimately realizing the application emergence of the software service layer (with development and emergence, it will show a normal distribution).

[0027] The four service layers can be described as "clouds and islands". The infrastructure layer uses the "cloud" to connect the distributed data middle platform (such as the "island"); the characteristics of the data service layer are: data is like the "sea", scattered and diverse; the model service layer is driven by large models; the software service layer is an interactive interface.

[0028] The infrastructure service layer provides a stable and reliable operating environment for the data service layer, model service layer, and software service layer. The data service layer retrieves stored data from the infrastructure service layer and processes, integrates, and analyzes it, transforming raw data into valuable information. It then provides this data as services to the model service layer and software service layer. The data service layer further optimizes and processes the data based on feedback from the model service layer and software service layer. The model service layer uses various machine learning and deep learning algorithms and models for training and inference based on the data provided by the data service layer, enabling intelligent decision-making and forecasting. It provides model capabilities as services to the software service layer, providing intelligent support for business applications. The software service layer is an interface for users and businesses, integrating the capabilities of the data service layer and model service layer into specific business applications, providing various functions and services. By invoking interfaces from the data service layer and model service layer, the software service layer implements functions such as data display, analysis, forecasting, and business process automation to meet user business needs.

[0029] From the perspective of relationship, from bottom to top (from the user's perspective), they are the foundation, data, model, and software layers. It can be understood that the foundation layer is the hardware computing power foundation, data is the data base (providing data support for applications), and the model is the reasoning foundation (providing intelligent support for applications). The three support the software service layer to provide users with final applications, but the first three also have their own businesses.

[0030] Specifically: the software service layer is the application interface of the entire middle platform system that provides services based on specific businesses (benchmarked as the business middle platform); the model service layer mainly implements model management (benchmarked as the AI middle platform), and is the key to realizing intelligent applications of the software service layer; the data service layer undergoes a series of data processing such as data collection, data labeling, and data governance, and ultimately provides data support for the software service layer (benchmarked as the data middle platform); the basic service layer is the network and computing power resource base for the system operation (benchmarked as the cloud platform).

[0031] like Figure 2 As shown, for geographically dispersed distributed middle platforms, this model adopts an autonomous, loosely coupled, full-level architecture, driven by the generative AI technology of the large language model. It has significant features such as multimodal interaction, intelligent operation and maintenance, agile collection, autonomous governance, and product generation. It includes:

[0032] The Infrastructure as a Service (IaaS) layer, which can be represented as Cloud (N, C, U), mainly includes various heterogeneous network facilities and a data cloud platform that integrates virtualization, containerization, image repositories, and monitoring services for disaster recovery and backup. In response to the requirements of high performance and high availability under high concurrency conditions, the key to the IaaS layer is to use the dual-engine cloud base technology of virtualization + containers for distributed cloud deployment, and use container cloud and bare metal for deployment, upgrade and operation. It can effectively ensure the service quality and resource utilization of the distributed middle-office architecture through state keying and elastic scaling. Faced with geographically dispersed distributed data middle-offices, based on cloud collaboration, intelligent scheduling, cloud resource virtualization, containers and other technologies, it can provide a cloud operating system based on cloud native architecture and integrated data, models, and application "ecosystem" for the digital intelligence middle-office system. It has dual engines of virtualization and containerization; multiple perspectives from cluster perspective, data perspective, and application perspective; hyper-convergence, lightweight, and intelligence, and effectively solves problems such as high data dimensions, complex data types, large data volumes, and difficulties in real-time data collection.

[0033] Data as a Service (DaaS) can be represented as It mainly includes multimodal databases and customized and generalized data services, and has data collection, data storage, data processing, and data analysis functions. In view of the characteristics of data distribution, multi-source distribution, fragmented and diverse, and multimodal (text, video, vector, etc.), the key to the DaaS layer is to adopt an agile data organization method of "general + customized", comprehensively considering the characteristics of data structure, application field, application scenario, etc., in order to provide general data services and special data services respectively. The general data organization method is to classify data according to the data structure, and divide data in various formats such as PDF, doc, docx, txt, mp3, mp4, png, jpg, etc. into unstructured data, semi-structured data, structured data, and streaming data, which can support general data services such as data collection, data governance, data fusion, and data labeling. The customized data organization method is to classify and organize according to business type and classify based on application scenario characteristics, which can support special data services.

[0034] like Figure 3 As shown, taking data annotation as an example:

[0035] 1) Data collection and preparation: First, collect a large amount of raw data from various sources, clean, process and pre-process the data for subsequent labeling work;

[0036] 2) Build a labeling system: Build a labeling system based on the labeling system design plan and principles for the collected and labeled data, and create a label library based on the three-level labeling standards;

[0037] 3) Label vectorization: Use pre-trained language embedding models to convert text labels into vectors;

[0038] 4) Build a tag library: store the vectorized tag data into the tag library;

[0039] 5) Abstract vectorization: vectorize the abstract text;

[0040] 6) Vector similarity retrieval: Using the cosine similarity method, the vectorized summary is compared with the tag information in the tag library after the vectorization. At this time, a similarity threshold can be specified, such as the threshold ≥

[0041] If the threshold is less than 0.8, it is classified as the final label attribute. If the threshold is less than 0.8, it is eliminated. This method can filter out the final accurate and concise labels. This method can quickly find the labels most relevant to the summary;

[0042] 7) Prompt word template: The large model generates the label that best matches the summary based on the prompt word template;

[0043] 8) Label generation: After comprehensively analyzing the information in the prompt word, the large language model uses its internal knowledge and language processing capabilities to generate labels that are accurate and meaningful.

[0044] Model as a Service (MaaS) can be expressed as LLM(s)·[m i,j ], mainly including various large language models, general AI models and various small models for specific businesses. Among them, the large language models include domain-specific large language models, basic general large language models and multimodal large models. The model service layer mainly manages the models, which are divided into model training platform and model reasoning platform, with algorithm library, model management, algorithm optimization, model encapsulation, model deployment, model monitoring and other modules. According to the user's perspective classification, the model training platform mainly realizes model management and optimization, and the model reasoning platform calls various large language models and small model sets for reasoning. The model training platform includes model management, model construction, model training, model optimization and model evaluation modules; the model reasoning platform includes model visualization, model deployment, model reasoning, model sharing and model encapsulation modules.

[0045] For complex business logic, the key technology of the MaaS layer lies in the matrix algorithm of large and small model alignment, that is, various large models (LLM(s)) are used as drivers to lead various general AI models and business model set matrices ([m i,j By semantically describing the small models, we can establish a mapping and alignment between tasks and business models. After receiving user instructions, the large model first semantically understands the user's needs, decomposes the task into multiple subtasks, and then selects the small models corresponding to the subtasks based on the semantic descriptions of the small models. The large model then comprehensively integrates the execution results of each small model. The large model is generally a series of general large models and domain-specific large language models that have been fine-tuned for the domain. The small models mainly include general AI models such as text recognition, video summarization, human detection, and face recognition, as well as reusable business models for specific business scenarios.

[0046] Software as a Service (SaaS) can be represented as p ~N(μ,σ 2), mainly including big data platforms, application portals and interactive windows, application scheduling centers, and block generation systems for various typical scenarios. It can provide various data application services based on typical application scenarios, and provide a variety of data products such as large-screen visualization, BI reports, data insights, and data intelligence products, fully aligned with business goals as a window for information input, using a natural language interaction mode to replace the traditional mouse and key operation mode. With the support of a large language model, the middle platform conducts human-computer interaction with users through natural language. As an information display window, the visualization interface of the SaaS layer pushes data, model visualization services, and reasoning analysis results in the form of "pictures and texts". The specific implementation method is as follows:

[0047] (1) System input: Input requirements from the software service layer to the LLM;

[0048] (2) Type parsing: Based on semantic understanding, LLM parses the specific data analysis report type that needs to be drafted;

[0049] (3) Task planning: Based on the agent capabilities, LLM determines the small model (from the model service layer) and related data (from the data service layer) that need to be called according to the analysis report type, mapping relationship and semantics;

[0050] (4) System execution: The relevant small models input the execution results to the LLM, and the LLM further generates a data analysis report;

[0051] (5) Comprehensive integration: Based on the large model reasoning and content generation capabilities, LLM combines preset templates with relevant cases in the knowledge base as a reference to comprehensively integrate the execution results and accurate data to form the final analysis report.

[0052] In addition, the key to realizing f(x) lies in block generation. By using various data products as blocks, the SaaS layer can reproduce products according to user needs, flexibly customize applications, and have the ability to flexibly configure products and flexibly orchestrate business logic, effectively improving the customization level of the middle platform.

[0053] The present invention mainly uses engineering integration through intelligent technologies such as large language models, speech recognition, and speech synthesis. By integrating Baidu's ASR model and TTS voice package, it realizes multimodal capabilities, further enriches natural language interaction modes, and improves service perception. At the same time, combined with collaborative filtering and intelligent recommendation algorithms, it can form customized and refined data applications for users. The data services are intuitive and diverse, support intelligent decision-making, further deeply explore the value of data, improve the level of intelligence, and make the interaction mode and application services more intelligent.

[0054] The present invention can realize computing power generation, data generation, model generation, and application generation. Figure 2The corresponding generation loop on the right side: the IaaS layer generates computing power through the cycle of "analysis, scheduling, application, and release"; the DaaS layer generates data through the cycle of "collection-fusion-analysis-release"; the MaaS layer generates models through "induction-training-validation-iteration"; and the SaaS layer uses configuration blocks as the minimum granularity and generates applications through the method of "recommendation-selection-job-generation".

Claims

1. A distributed hierarchical digital intelligence middle platform system based on a large language model, characterized by: include: Infrastructure service layer, data service layer, model service layer and software service layer; the infrastructure service layer uses containers, virtual machines and bare metal to perform global computing resource scheduling to realize the intelligent operation and maintenance of the distributed and layered digital intelligence middle-office system and improve resource utilization, and provide a stable and reliable operating environment for the data service layer, model service layer and software service layer; the data service layer will obtain data from various networks of the infrastructure service layer, and provide general data services and customized data services to the model service layer and software service layer under the drive of large models according to the data structure, application field and application scenario characteristics; the model service layer manages the model through the model training platform and model inference platform based on the data of the data service layer, establishes the mapping and alignment of tasks to small models, and supports the intelligent application of the software service layer through model operation reasoning; the software service layer provides various intelligent data application services according to typical application scenarios, and provides a variety of data products directly to users.

2. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 1, characterized in that: The infrastructure service layer is represented as Cloud(N,C,U), where N represents communication, C represents computing power, and U represents container. The data service layer is represented as represents customized data services, where m represents the number of customized data types according to the business type; G represents general data services; LLM(s) represents the large language model group, and the model service layer is represented as LLM(s)·[m i,j ],[m i,j ] represents various general AI models and business model set matrices, i represents the business type, j represents the jth one, that is, Mi,j represents the jth small model of the i-th business, and the software service layer is represented by p~N(μ,σ 2 ), the distributed hierarchical digital intelligence middle platform system adopts an autonomous loosely coupled full-level architecture, which is expressed as f(x)~N(0,1) indicates that the application emerges in a point-like manner and gradually tends to a normal distribution.

3. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 2, characterized in that: The infrastructure service layer includes various heterogeneous network facilities and a data cloud platform that integrates virtualization, containerization, image warehouse and monitoring services for disaster recovery and backup.

4. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 2, characterized in that: The data service layer includes a multimodal database with data collection, data storage, data processing, and data analysis functions. The general data service classifies data according to the data structure, and divides data in PDF, doc, docx, txt, mp3, mp4, png, and jpg formats into unstructured data, semi-structured data, structured data, and streaming data to support general data services; the customized data service is classified and organized according to business type and classified based on application scenario characteristics to support customized data services.

5. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 2, characterized in that: The model service layer includes various large language models, general AI models and business models. It is driven by various large language models and leads various general AI models and business model set matrices. By semantically describing the general AI model or business model, it establishes a mapping and alignment from task to small model.

6. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 2, characterized in that: The software service layer includes a big data task platform, an application portal and interactive window, an application scheduling system, and a block generation system for various typical scenarios. It adopts a natural language interaction mode, conducts human-computer interaction with users through natural language, pushes data in the form of graphics and text, model visualization services and reasoning analysis results, and realizes block generation.

7. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 4, characterized in that: The multimodal database specifically includes analytical database, relational database, spatiotemporal database, graph library, vector library and index library; the general data service specifically includes data collection, data governance, data fusion, data storage and computing, data assets and data services.

8. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 5, characterized in that: The various types of large language models include domain-specific large language models, basic general large language models and multimodal large models; small models include general AI models for text recognition, video summarization, human detection, face recognition and reusable business models for specific business scenarios.

9. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 6, characterized in that: The block generation is based on the block as the minimum granularity, and data products are reproduced according to user needs, and data application services are flexibly customized.

10. A distributed hierarchical digital intelligence middle platform system based on a large language model according to claim 1, characterized in that: The model training platform is used for model management and optimization, including model management, model construction, model training, model optimization and model evaluation modules; the model inference platform is used to call various large language models and small model sets for inference, including model visualization, model deployment, model inference, model sharing and model encapsulation modules.

Images