Data processing method and device based on building block digital human model and storage medium

Abstract

The application belongs to the technical field of large models, and discloses a data processing method and device based on a building block type digital human model and a storage medium. Target data from multiple heterogeneous data sources is acquired and preprocessed to obtain structured target data. Then, multiple universally callable general business capability interfaces are constructed. A task request input by a user is received, a user intention and required task parameters for generating a response are determined through a pre-trained digital human model, one or more general business capability interfaces are dynamically assembled according to the user intention and the task parameters determined by the digital human model to form a task processing chain, the task processing chain is subjected to permission verification and audit control based on a model context protocol to ensure that only authorized users can access the corresponding interfaces, the task processing chain is executed to generate a processing result, and the result is returned to the user through at least one information push channel based on a push rule preset by the user. The safety and intelligent level of digital human services are improved.

Description

Technical Field

[0001] This invention belongs to the field of large model technology, specifically relating to data processing methods, devices and storage media based on modular digital human models. Background Technology

[0002] Leveraging deep learning frameworks and optimized advanced algorithms, digital humans have achieved a high degree of human-likeness in behavior, language, and emotional expression. The combination of cloud computing and edge computing provides strong technical support for real-time interaction of digital humans. Digital human technology has shown broad prospects in multiple industries, particularly excelling in areas such as virtual anchors, education, healthcare, urban management, and government services. It can provide efficient and accurate work assistance, enhance practicality and adaptability, and lay the foundation for future innovation.

[0003] Despite rapid development, existing technologies have significant limitations that restrict their in-depth application. Firstly, there is a lack of efficient and dynamic security management mechanisms, making it difficult to dynamically adjust permissions based on user context and audit information; the interconnected nature of large models may lead to the leakage of sensitive information and infringement of user privacy. The system lacks fine-grained permission control based on "event, person, time, and reason." Secondly, the depth of intelligent interaction is insufficient. While digital humans can handle basic dialogues and perform simple tasks, they struggle with complex situations and tasks requiring high emotional engagement. There is still significant room for improvement in the natural fluency, emotional expression, and comprehension capabilities for dialogues involving complex semantics, metaphors, polysemous words, and strong contextual dependencies. The root cause lies in the overly simplistic scenario design, lacking flexibility and depth. Furthermore, the robustness and accuracy of core algorithms and models are insufficient, resulting in weaknesses in key areas such as complex query processing, large-scale data analysis, and advanced decision support. It is difficult to comprehensively cover the diverse needs of different user groups. While there have been preliminary achievements in personality trait simulation, there is still a gap to bridge before achieving a natural, fluent, and context-adaptive interactive experience. Finally, the system's intelligence level is low, exhibiting a passive approach. When executing tasks across different scenarios, task tags are not precise or flexible enough, hindering service efficiency and impacting user communication efficiency and decision-making quality. The lack of efficient proactive information push and real-time feedback functions weakens the user experience and reduces the timeliness and accuracy of information acquisition. Summary of the Invention

[0004] To improve the security and intelligence of digital human services, in a first aspect, this invention proposes a data processing method based on a modular digital human model, the method comprising: Target data from multiple heterogeneous data sources is acquired, and the target data is preprocessed to obtain structured target data; Based on the structured target data, multiple independent callable general business capability interfaces are constructed. The general business capability interfaces cover at least one business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. The system receives a task request input from a user and performs natural language understanding and generation processing on the task request using a pre-trained digital human model to determine the user's intent and the task parameters required to generate a response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities. Based on the user intent and task parameters determined by the digital human model, one or more of the general business capability interfaces are dynamically assembled to form a task processing chain. The task processing chain is subject to permission verification and audit control based on the model context protocol to ensure that only authorized users can access the corresponding interfaces. The task processing chain is executed to generate processing results, and the results are returned to the user through at least one information push channel based on the user's preset push rules.

[0005] In one possible implementation, acquiring target data from multiple heterogeneous data sources includes: When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

[0006] In one possible implementation, the preprocessing of the target data includes: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data.

[0007] In one possible implementation, after the step of constructing multiple independently callable general business capability interfaces based on the structured target data, the method further includes: Traffic monitoring, logging, and security auditing are performed on the aforementioned general business capability interfaces; The API gateway is used to authenticate and verify the permissions of received data access requests. After successful verification, if the traffic level of the corresponding general business capability interface is within a preset threshold, the data access request is forwarded to the corresponding general business capability interface, and the access log of the corresponding general business capability interface is updated; if the traffic level of the corresponding general business capability interface exceeds the preset threshold, the data access request is intercepted and an error response is returned.

[0008] In one possible implementation, the digital human model is trained in the following manner: Multi-source data with labeled information is input into the multimodal large model to be trained. The multimodal large model to be trained is trained based on the labeled information and the output of the multimodal large model to be trained to obtain a digital human model. The multimodal large model integrates a text generation model, an image recognition model, a speech recognition model and a deep thinking reasoning model.

[0009] In one possible implementation, different general business capability interfaces are used to handle different businesses, including: personal affairs category, vehicle service category, specific task category, material asset category, specific location category, organization category, event information category, and knowledge base category; wherein: The categories of personal affairs include personnel information management and personnel business information inquiry; The vehicle service categories include queries for basic vehicle information, transaction information, maintenance information, incident information, vehicle status and trajectory information; The specific task category includes the management of basic event information, change information, progress information, and workflow information; The material asset categories include basic item information, asset data, performance data, value, and transaction flow management; The specific location category includes basic information about the location, environmental information, information on the implementation of security measures, and visitor access management. The organization category includes information management of organization name, type, date of establishment, and place of registration; The event information categories include event time, geographical location, event participants, and event content information management; The knowledge base categories include knowledge updates and deep learning resource queries.

[0010] Secondly, this application provides a data processing device based on a modular digital human model, the device comprising: The acquisition module is used to acquire target data from multiple heterogeneous data sources and preprocess the target data to obtain structured target data; The construction module is used to build multiple independently callable general business capability interfaces based on the structured target data. The general business capability interfaces cover at least one type of business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. The input module is used to receive task requests input by the user, and to perform natural language understanding and generation processing on the task requests through a pre-trained digital human model to determine the user's intention and the task parameters required to generate the response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities. An assembly module is used to dynamically assemble one or more of the general business capability interfaces according to the user intent and task parameters determined by the digital human model, forming a task processing chain; The security module is used to perform permission verification and audit control on the task processing chain based on the model context protocol to ensure that only authorized users can access the corresponding interfaces. The output module is used to execute the task processing chain, generate processing results, and return the results to the user through at least one information push channel based on the user's preset push rules.

[0011] In one possible implementation, the acquisition module is specifically used for: When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

[0012] In one possible implementation, the acquisition module is further configured to: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data.

[0013] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described data processing methods.

[0014] The data processing method based on a modular digital human model provided in this application embodiment acquires target data from multiple heterogeneous data sources and preprocesses the target data to obtain structured target data. Based on the structured target data, multiple independently callable general business capability interfaces are constructed. These general business capability interfaces cover at least one business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. A task request input by a user is received, and the task request is processed by a pre-trained digital human model through natural language understanding and generation to determine the user's intent and the task parameters required to generate a response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition, and deep reasoning capabilities. Based on the user's intent and task parameters determined by the digital human model, one or more of the general business capability interfaces are dynamically assembled to form a task processing chain. The task processing chain is subject to permission verification and audit control based on a model context protocol to ensure that only authorized users can access the corresponding interfaces. The task processing chain is executed to generate processing results, and based on user-preset push rules, the results are returned to the user through at least one information push channel. This improves the security and intelligence level of digital human services. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the data processing method based on a modular digital human model provided in an embodiment of the present invention; Figure 2 A schematic diagram of the system structure for implementing a data processing method based on a modular digital human model, provided in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will be described in detail below through embodiments.

[0017] Leveraging deep learning frameworks and optimized advanced algorithms, digital humans have achieved a high degree of human-likeness in behavior, language, and emotional expression. The combination of cloud computing and edge computing provides strong technical support for real-time interaction of digital humans. Digital human technology has shown broad prospects in multiple industries, particularly excelling in areas such as virtual anchors, education, healthcare, urban management, and government services. It can provide efficient and accurate work assistance, enhance practicality and adaptability, and lay the foundation for future innovation.

[0018] Despite rapid development, existing technologies have significant limitations that restrict their in-depth application. Firstly, there is a lack of efficient and dynamic security management mechanisms, making it difficult to dynamically adjust permissions based on user context and audit information; the interconnected nature of large models may lead to the leakage of sensitive information and infringement of user privacy. The system lacks fine-grained permission control based on "event, person, time, and reason." Secondly, the depth of intelligent interaction is insufficient. While digital humans can handle basic dialogues and perform simple tasks, they struggle with complex situations and tasks requiring high emotional engagement. There is still significant room for improvement in the natural fluency, emotional expression, and comprehension capabilities for dialogues involving complex semantics, metaphors, polysemous words, and strong contextual dependencies. The root cause lies in the overly simplistic scenario design, lacking flexibility and depth. Furthermore, the robustness and accuracy of core algorithms and models are insufficient, resulting in weaknesses in key areas such as complex query processing, large-scale data analysis, and advanced decision support. It is difficult to comprehensively cover the diverse needs of different user groups. While there have been preliminary achievements in personality trait simulation, there is still a gap to bridge before achieving a natural, fluent, and context-adaptive interactive experience. Finally, the system's intelligence level is low, exhibiting a passive approach. When executing tasks across different scenarios, task tags are not precise or flexible enough, hindering service efficiency and impacting user communication efficiency and decision-making quality. The lack of efficient proactive information push and real-time feedback functions weakens the user experience and reduces the timeliness and accuracy of information acquisition.

[0019] To improve the security and intelligence of digital human services, firstly, see... Figure 1 This invention proposes a data processing method based on a modular digital human model, the method comprising: S101, acquire target data from multiple heterogeneous data sources, and preprocess the target data to obtain structured target data.

[0020] Data source types include files, databases, API interfaces, and Kafka message queues. Target data from multiple heterogeneous data sources can be collected in real-time or periodically from big data platforms and service platforms using big data technologies and services. Then, depending on the data source type, ETL (Extract, Transform, Load) tools or custom scripts are used to extract the data.

[0021] When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

[0022] After obtaining the target data, it is stored in an intermediate storage layer, such as HDFS (Hadoop Distributed File System) or a cloud storage service, to ensure data security and reliability. Next, the target data undergoes preprocessing, including the following steps: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data.

[0023] In one example, after data cleaning of the target data, word segmentation and part-of-speech tagging can be performed. During the data annotation stage, semantic categories and sentiment tendencies are labeled on the text data, and vector analysis models are used to improve the accuracy and efficiency of the annotation.

[0024] The purpose of data alignment is to resolve inconsistencies in data formats, transforming data into a standardized format. Finally, structuring aims to establish unified data standards and specifications, ensuring a high degree of consistency in both form and content for each type of data, thereby enhancing data usability and interoperability.

[0025] Furthermore, relevant legal, business, and document knowledge can be collected from multiple channels, categorized into the knowledge base category, and structured to ensure the comprehensiveness and authority of the knowledge base, facilitating subsequent integration and application.

[0026] Excessive data not only increases storage costs but can also lead to excessive consumption of computing resources, impacting overall processing speed. Therefore, it is necessary to reasonably compress the data size, eliminate redundant information, and retain the most core and valuable parts, while ensuring data accuracy, to accommodate the processing capabilities and efficiency of large models.

[0027] S102, Based on the structured target data, construct multiple independent callable general business capability interfaces, which cover at least one type of business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases.

[0028] Each type of general business capability interface corresponds to a type of target data. By calling these interfaces, users can access target data corresponding to at least one business theme from among people, vehicles, events, things, organizations, locations, incidents, and knowledge bases. To better manage and optimize these API interfaces, an API gateway is centrally managed and controlled through a service platform. The API gateway is responsible for API routing, security, and monitoring, ensuring that each request reaches the target service efficiently and securely. It also monitors API performance and usage in real time, promptly identifying and resolving issues. Furthermore, it supports traffic control, authentication, and logging, further enhancing system stability and reliability.

[0029] Furthermore, to facilitate rapid integration and invocation of these API interfaces, the service platform provides detailed API documentation and abundant sample code. The API documentation details the function, parameters, return values, and usage methods of each interface, ensuring that developers and users can quickly understand and master them. The sample code covers various common programming languages, such as Python, Java, and JavaScript, helping developers get started quickly and begin practical development work.

[0030] S103, receive a task request input by the user, and perform natural language understanding and generation processing on the task request through a pre-trained digital human model to determine the user's intention and the task parameters required to generate the response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities.

[0031] Large-scale models refer to artificial intelligence models with millions to hundreds of billions of parameters, also known as large-scale pre-trained models or ultra-large-scale language models. These models learn rich language and knowledge representations through unsupervised or weakly supervised pre-training on massive amounts of data, and possess strong generalization and adaptability. They typically employ deep learning architectures, using multi-layered self-attention mechanisms and feedforward neural networks to capture complex patterns and dependencies in the input data. Large-scale pre-trained models can achieve superior performance on specific tasks through techniques such as fine-tuning or cue learning, and are widely used in various fields such as natural language processing, computer vision, and speech recognition.

[0032] Based on their functions and application areas, models are mainly divided into four categories: text generation models, image recognition models, speech recognition models, and deep reasoning models. Each type of model has its unique technical characteristics and a wide range of application scenarios. Text generation models are mainly used to generate natural language text, covering everything from simple sentences to complex passages. They can generate coherent and logical text content through large-scale text data pre-training. Image recognition models are mainly used to identify and classify objects, scenes, and features in images. Based on convolutional neural network architecture, they can achieve high-precision image recognition through training with large amounts of image data. Speech recognition models are mainly used to convert speech signals into text. Deep reasoning models are used to solve complex problems and tasks involving logical reasoning, planning, and decision-making. They typically combine deep learning and symbolic reasoning techniques and can handle more advanced cognitive tasks.

[0033] The pre-trained digital human model employs a large-scale model capability encapsulation, including natural language processing (NLP), computer vision, and speech recognition capabilities. NLP, a crucial branch of artificial intelligence, aims to enable computers to understand, interpret, and generate human natural language. It combines knowledge from computer science, artificial intelligence, and linguistics to achieve intelligent processing of text or speech information.

[0034] Pre-trained digital human models encapsulate natural language processing capabilities, aiming to provide powerful language understanding and generation capabilities for multiple application scenarios, including: Intent recognition, which analyzes user input to determine the specific intent behind it. For example, determining whether a user is querying information, asking a question, requesting a service, or making a complaint. Semantic analysis, which deeply analyzes the meaning and context of text, understanding the relationships between words, phrases, and sentences to more accurately capture the deeper meaning of the text. Text classification, which automatically categorizes text into predefined categories, such as people, vehicles, events, things, organizations, places, events, and knowledge bases. Named entity recognition, which identifies and extracts specific types of entities from text, such as names of people, ID numbers, ID types, place names, and organization names. Question answering systems, which find the most relevant answers from large amounts of text data based on user questions, providing fast and accurate information retrieval services. Text generation, which automatically generates high-quality text content, such as news summaries, reports, and articles.

[0035] With the development of artificial intelligence technology, computer vision, as an important branch, has shown broad application prospects in many fields. The main goal of computer vision is to simulate human visual perception, enabling digital human models to "see" the world and make decisions or take actions based on what they see. However, traditional computer vision technologies typically focus on single tasks, such as object recognition and image classification, making it difficult to meet complex and ever-changing practical needs. Therefore, this invention encapsulates computer vision capabilities, aiming to provide flexible visual recognition capabilities for multiple application scenarios, including: object recognition, capable of identifying objects in images, such as vehicles, pedestrians, and animals; facial recognition, providing simple face recognition capabilities to distinguish personal identity information; document analysis, processing scanned documents to extract key information, such as tabular data and text content; OCR (Optical Character Recognition) recognition, recognizing and extracting text information from images; and form recognition, automatically recognizing and parsing information in forms.

[0036] Speech recognition is the technology that converts human speech into text. This invention encapsulates speech recognition capabilities to convert user speech input into text in real time, enhancing the human-computer interaction experience. By integrating speech recognition algorithms and models, this invention can efficiently and accurately convert speech data into text information that can be understood and further processed by large models, solving the challenges of human-computer interaction scenarios involving voice and large models. Its core advantages lie in convenience and the universality of dialect recognition. Specifically, in mobile office scenarios, it provides users with a scenario for interacting with digital humans through voice, similar to human-to-human communication, achieving rapid speech-to-text conversion, greatly facilitating the human-computer interaction, note-taking, and record-keeping needs of mobile office work. In addition, this capability also has powerful dialect recognition capabilities, supporting multiple dialects including but not limited to Mandarin, Cantonese, Sichuanese, Shanghainese, Hunanese, Hubeiese, Northeastern Mandarin, Fujianese, Hakka, and Minnanese, significantly improving the applicability of the invention in cross-regional and multilingual contexts.

[0037] S104. Based on the user intent and task parameters determined by the digital human model, dynamically assemble one or more of the general business capability interfaces to form a task processing chain.

[0038] Based on the specific intentions of different users, one or more of the aforementioned general business capability interfaces can be flexibly invoked to form a task processing chain. Based on deep thinking and emotion computing, users can build their own digital human personal assistants for various scenarios such as information query, data analysis, daily task assistance, and internal management according to their specific needs. Because it is assembled on demand, personalized needs can be realized, and the effects that the digital human can achieve are also highly customized, achieving a "personalized experience for every user" effect.

[0039] User intent reflects the input provided by the user during the interaction process, aiming to help the digital human accurately understand user needs and respond appropriately. Task parameters are instructions or templates used internally by the digital human, designed to guide the digital human in generating responses and ensure the accuracy and consistency of the output. Through explicit user intent, the digital human designed in this invention effectively improves the accuracy of understanding user needs, reduces erroneous responses caused by ambiguity, and reduces inaccurate information in the generated content, i.e., hallucinations.

[0040] S105, Based on the model context protocol, perform permission verification and audit control on the task processing chain to ensure that only authorized users can access the corresponding interfaces.

[0041] MCP (Model Context Protocol) is a specification for managing and transmitting machine learning models and their related contextual information, enabling the characterization of information across multiple dimensions, including people, vehicles, events, things, organizations, locations, and activities. It provides users with convenient access to various common business capability interfaces.

[0042] Traditional large-scale model authentication mainly involves access authentication, which verifies the user's identity and determines whether they have permission to use the digital human chatbot. Once approved, the user can ask the model any question, and the model will use all its knowledge to answer it, which can easily lead to the leakage of sensitive information.

[0043] This application not only verifies identity, but also determines whether the user can access the information in the current specific dialogue scenario, and what specific data they can access. For example, if a user wants to view the finance department's report data, it is necessary to check whether the user is a finance staff member, whether they currently have a task to process expense reimbursements, whether they can only view their own department's reports, and whether their access has expired today.

[0044] Permission verification is responsible for the most basic identity authentication. It confirms the user's identity, typically through methods such as username / password, SMS verification code, and single sign-on.

[0045] Audit controls maintain a library of permission rules, defined based on the principles of "task, person, reason, and time." "Task" refers to whether the user currently has a specific task, such as being authorized to handle a traffic accident. "Person" refers to the user's role—a traffic police officer, a regular citizen, or a bank employee. "Reason" refers to the purpose of the query—is it for official business or personal curiosity? "Time" refers to whether the user's permission is valid.

[0046] When a digital human needs to call an interface to query vehicle information, it will perform permission verification and audit control on the request "User A, currently executing task T, requesting to query vehicle information for license plate number 12345". The returned results may include: allow, but only able to view basic information, not the vehicle owner's private information; or deny, insufficient permissions.

[0047] In this application, permission verification and audit control are encapsulated into a unified authentication interface. This unified authentication interface can be called before performing sensitive operations.

[0048] S106, execute the task processing chain, generate processing results, and return the results to the user through at least one information push channel based on the user's preset push rules.

[0049] To achieve efficient information generation and processing, this invention encapsulates a rule engine that generates relevant information such as query results, task information, and reminder messages based on user-defined rules regarding tasks, personnel, events, time, and frequency, and formats these information into formats suitable for different channels. For message push notifications, it supports returning results to users via SMS, email, instant messaging software, etc.

[0050] For example, a user manually inputs: "Check the traffic violations for vehicle with license plate number 12345." The digital human, after calling the general business capability interface, queries and replies: "This vehicle has one unprocessed traffic violation record, resulting in 3 points and a fine of 200 yuan." The rules engine determines there is an unprocessed violation and triggers a push notification rule. The digital human then immediately adds a system prompt: "You can process this violation online through the Traffic Management 12123 APP. Do you need me to push a processing guide for you?" The data processing method based on a modular digital human model provided in this application embodiment acquires target data from multiple heterogeneous data sources and preprocesses the target data to obtain structured target data. Based on the structured target data, multiple independently callable general business capability interfaces are constructed. These general business capability interfaces cover at least one business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. A task request input by a user is received, and the task request is processed by a pre-trained digital human model through natural language understanding and generation to determine the user's intent and the task parameters required to generate a response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition, and deep reasoning capabilities. Based on the user's intent and task parameters determined by the digital human model, one or more of the general business capability interfaces are dynamically assembled to form a task processing chain. The task processing chain is subject to permission verification and audit control based on a model context protocol to ensure that only authorized users can access the corresponding interfaces. The task processing chain is executed to generate processing results, and based on user-preset push rules, the results are returned to the user through at least one information push channel. This improves the security and intelligence level of digital human services.

[0051] In one example, a traffic police officer, Zhang San, is handling a traffic accident and needs to query the detailed information of license plate number "12345". Officer Zhang San logs into the digital human system. The digital human development platform verifies his identity and returns a token. Zhang San tells the digital human to check the vehicle information and owner information for license plate B12345. The digital human understands the intent through natural language and determines that it needs to call the general business capability interfaces for vehicle query and personnel information query. Authentication is triggered, checking if user Zhang San has an ongoing task and whether that task requires querying vehicle information; whether user Zhang San's role is a traffic police officer; whether the vehicle query is for handling a traffic accident; and whether the task's permissions are valid for the current day. Based on all these conditions, the approval center returns: querying all information of the vehicle, including owner information, is permitted. Only after receiving the authentication approval response does the digital human call the general business capability interfaces for vehicle query and personnel information query, retrieve the data, and organize it into natural language to reply to Officer Zhang San. If authentication fails, the digital human will directly reply: "Sorry, you currently do not have permission to query this vehicle information" and will not perform any further operations.

[0052] This application's solution significantly enhances the security of digital human services, achieving dynamic, fine-grained, and context-aware access control, eliminating the risk of sensitive information leakage. Furthermore, it improves development efficiency; developers no longer need to concern themselves with complex access logic, only needing to call a unified interface, simplifying the development process and avoiding code duplication. All access rules are centrally managed in a security approval center. When business rules change (e.g., new regulations are introduced), only the rules need to be modified in the center, without modifying the code for each business function. All authentication requests are recorded, tracking who, when, what they attempted to access, and whether they were successful—all traceable. This application transforms security capabilities from a decentralized, reactive measure into a centralized, orchestratable, and proactive core foundational capability, ultimately supporting the reliable application of digital humans in serious business scenarios.

[0053] In one example, after step S102, the method further includes: Step 1: Perform traffic monitoring, log recording, and security auditing on the general business capability interfaces; Step 2: Verify the identity and permissions of the received data access requests through the API gateway; Step 3: After successful verification, if the traffic level of the corresponding general business capability interface is within the preset threshold, the data access request is forwarded to the corresponding general business capability interface, and the access log of the corresponding general business capability interface is updated; if the traffic level of the corresponding general business capability interface exceeds the preset threshold, the data access request is intercepted and an error response is returned.

[0054] The API gateway is the sole entry point for all traffic into the system. It handles global, general, and business logic-independent management, including API routing, security, and monitoring. Based on the URL path of the request, the gateway accurately forwards requests to backend services. The security provided by the API gateway is coarse-grained; it verifies identity by checking tokens or certificates in API requests to prevent general network attacks, thus blocking unauthorized requests and attacks at the outermost layer and protecting vulnerable internal business services. Monitoring collects metrics such as the number of API calls, response time, and error rate. Access logs are generated for big data analysis, providing data support for system performance monitoring, troubleshooting, and capacity planning. The gateway limits the call frequency of each API to prevent a service from being overwhelmed by sudden traffic surges. It ensures system stability under high load and enables graceful degradation. Only requests passing through the API gateway enter the digital human service for subsequent permission verification and audit control. This leads to business-level, fine-grained, and context-sensitive permission control. For example, can user A perform operation C on resource B? Can traffic police officer Zhang San query the ID number of the owner of license plate '12345'? Dynamic judgments are made based on complex business rules such as "due to event, person, reason, and time." For example, Zhang San can query license plate information because he is handling an accident task, but this permission expires after the task is completed. This achieves core data security protection and prevents unauthorized access (e.g., users can only view their own data and cannot view others' private data). Complex permission logic is extracted from business code and centrally managed, making business development simpler and permission policies more consistent and easier to maintain.

[0055] As can be seen, the API gateway is the network traffic layer, responsible for global entry point control, identity authentication, and flood prevention and attack prevention. It is used to determine whether access to the service is possible without needing to understand the business logic. Authentication, on the other hand, is at the business application layer and is responsible for handling business logic such as permission authorization and whether the data can be manipulated. It requires a deep understanding of the business logic and context.

[0056] The collaboration process is summarized below. Step 1: Gateway. When a request arrives, the API gateway first checks if its token is valid (authentication) and then determines where to send it according to the routing. If the token is invalid or the traffic is too large, it is rejected directly.

[0057] The second step is authentication: the request carries a valid token to the specific digital human service. When the digital human needs to perform a sensitive operation, such as querying data, it will call the unified authentication interface.

[0058] The third step is to make a final decision based on your identity and the current work task and business context: Yes, you are allowed to see it; no, you do not have permission.

[0059] As can be seen, these are two indispensable and distinct lines of defense in the system's security architecture. The API gateway is the first line of defense, blocking unauthorized identities and network attacks. Authentication capabilities are the second line of defense, blocking unauthorized operations by authorized users and ensuring the security of business data. This design enables the system to possess both unified management and protection at the entry level and fine-grained, dynamic access control at the business level, together forming a defense-in-depth system.

[0060] In one example, the digital human model is trained in the following manner: Multi-source data with labeled information is input into the multimodal large model to be trained. The multimodal large model to be trained is trained based on the labeled information and the output of the multimodal large model to be trained to obtain a digital human model. The multimodal large model integrates a text generation model, an image recognition model, a speech recognition model and a deep thinking reasoning model.

[0061] Based on a pre-trained multimodal large model, fine-tuning is performed for the digital human construction task. Training algorithms such as contrastive learning and reinforcement learning are employed, with appropriate learning rates and training epochs set. During training, labeled multimodal data is used as input, and the desired digital human output is used as the target. The model parameters are continuously adjusted to enable the model to generate a digital human performance that meets the requirements based on the input data.

[0062] In one example, different general business capability interfaces are used to handle different business types, including: personal affairs category, vehicle service category, specific task category, material asset category, specific location category, organization category, event information category, and knowledge base category; wherein: The categories of personal affairs include personnel information management and personnel business information inquiry; The vehicle service categories include queries for basic vehicle information, transaction information, maintenance information, incident information, vehicle status and trajectory information; The specific task category includes the management of basic event information, change information, progress information, and workflow information; The material asset categories include basic item information, asset data, performance data, value, and transaction flow management; The specific location category includes basic information about the location, environmental information, information on the implementation of security measures, and visitor access management. The organization category includes information management of organization name, type, date of establishment, and place of registration; The event information categories include event time, geographical location, event participants, and event content information management; The knowledge base categories include knowledge updates and deep learning resource queries.

[0063] With the increasing popularity of digital human technology, its application scope has expanded from personal assistants to multiple fields such as enterprise services. It faces a wide variety of complex tasks, covering multiple professional fields and application scenarios, with needs being highly personalized. To enhance the intelligence and flexibility of digital humans, this invention modularizes common capabilities, strengthening their ability to handle complex tasks. Different groups and users can combine their needs and security authentication to flexibly schedule different general business capability interfaces, assembling different capabilities like building blocks, achieving a personalized digital human that best matches business requirements. Simultaneously, it optimizes the development process, improving the system's flexibility and scalability.

[0064] This invention primarily provides the following capabilities for encapsulation and construction: The "Personnel" category encompasses personal affairs management, including but not limited to personnel information management and personnel business information, to support users' daily work in querying and retrieving information.

[0065] Che focuses on vehicle-related services, such as basic vehicle information, transaction information, maintenance information, incident information, vehicle status, and trajectory information.

[0066] Events refer to the management of specific tasks or events, such as basic event information, event change information, event progress information, and event flow information, aiming to provide common event management capabilities.

[0067] The "things" category involves asset and material management, including basic information, management information, asset data, performance data, value, and transaction flow of items, aiming to provide the ability to query common data related to things.

[0068] Venue: Focuses on the safety and management of specific locations, such as basic venue information, management information, environmental quality monitoring, implementation of security measures, visitor access control, etc., to ensure the safety and comfort of the venue.

[0069] Organization: Organizational information typically includes several key attributes to comprehensively describe its characteristics and operations. These attributes mainly include: name, type, date of incorporation, place of registration, headquarters location, business scope, legal form, registered capital, number of employees, management information, contact information, organizational structure, financial status, shareholders or investors, social responsibility, certifications and qualifications, history, corporate culture, products and services, and market positioning. This detailed information allows for a better understanding and management of the organization's structure, functions, and its market positioning and development.

[0070] Event information includes several key attributes, such as: time information, geographical location, participants, event content, impact and consequences, response measures, media and public reaction, relevant documents and evidence, legal and regulatory information, economic data, technical data, social and cultural context, and predictive and preventative measures. This data helps in-depth analysis of all aspects of the event, supports decision-making, and provides a reference for future similar events.

[0071] Knowledge Base: As a storage and retrieval center for information and knowledge, it provides functions such as knowledge updates and deep learning material queries to support users' continuous learning and knowledge accumulation.

[0072] See Figure 2 This invention provides a system architecture diagram. A database is constructed through data acquisition and preprocessing. Services are built through service hooking and orchestration. The multimodal large model integrates text generation, image recognition, speech recognition, and deep reasoning models. It encapsulates eight categories of target data—people, vehicles, events, objects, organizations, places, incidents, and a knowledge base—into multiple building blocks. Furthermore, it includes security capabilities such as permission verification and audit control, and information push capabilities. When the digital human provides services, it performs intelligent agent orchestration (determining which models to use), multi-model invocation (determining which models to invoke), a building block capability supermarket (including the aforementioned encapsulated building blocks), plug-in capability expansion (expanding the knowledge base content), and automated task processing (proactively pushing messages to users).

[0073] In one example, to enable the digital human not only to possess fluent and natural language communication skills but also to provide customized services and support based on the user's specific needs, the digital human designed in this invention also possesses highly anthropomorphic basic attributes. These attributes include, but are not limited to: name, gender, date of birth, nationality, place of origin, identification number, height and weight, blood type, language ability, personality traits, and health status. These detailed attribute settings help the digital human to more realistically simulate human behavior patterns, enhancing the realism and interactivity of the user experience.

[0074] Secondly, the present invention provides a data processing device based on a modular digital human model, the device comprising: The acquisition module is used to acquire target data from multiple heterogeneous data sources and preprocess the target data to obtain structured target data; The construction module is used to build multiple independently callable general business capability interfaces based on the structured target data. The general business capability interfaces cover at least one type of business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. The input module is used to receive task requests input by the user, and to perform natural language understanding and generation processing on the task requests through a pre-trained digital human model to determine the user's intention and the task parameters required to generate the response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities. An assembly module is used to dynamically assemble one or more of the general business capability interfaces according to the user intent and task parameters determined by the digital human model, forming a task processing chain; The security module is used to perform permission verification and audit control on the task processing chain based on the model context protocol to ensure that only authorized users can access the corresponding interfaces. The output module is used to execute the task processing chain, generate processing results, and return the results to the user through at least one information push channel based on the user's preset push rules.

[0075] In one possible implementation, the acquisition module is specifically used for: When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

[0076] In one possible implementation, the acquisition module is further configured to: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data. Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described data processing methods.

[0077] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0078] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0079] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0080] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A data processing method based on a modular digital human model, characterized in that, The method includes: Target data from multiple heterogeneous data sources is acquired, and the target data is preprocessed to obtain structured target data; Based on the structured target data, multiple independent callable general business capability interfaces are constructed. The general business capability interfaces cover at least one business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. The system receives a task request input from a user and performs natural language understanding and generation processing on the task request using a pre-trained digital human model to determine the user's intent and the task parameters required to generate a response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities. Based on the user intent and task parameters determined by the digital human model, one or more of the general business capability interfaces are dynamically assembled to form a task processing chain. The task processing chain is subject to permission verification and audit control based on the model context protocol to ensure that only authorized users can access the corresponding interfaces. The task processing chain is executed to generate processing results, and the results are returned to the user through at least one information push channel based on the user's preset push rules. After the step of constructing multiple independently callable general business capability interfaces based on the structured target data, the method further includes: Traffic monitoring, logging, and security auditing are performed on the aforementioned general business capability interfaces; The API gateway is used to authenticate and verify the permissions of received data access requests. After successful verification, if the traffic level of the corresponding general business capability interface is within a preset threshold, the data access request is forwarded to the corresponding general business capability interface, and the access log of the corresponding general business capability interface is updated; if the traffic level of the corresponding general business capability interface exceeds the preset threshold, the data access request is intercepted and an error response is returned.

2. The method according to claim 1, characterized in that, The acquisition of target data from multiple heterogeneous data sources includes: When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

3. The method according to claim 1, characterized in that, The preprocessing of the target data includes: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data.

4. The method according to claim 1, characterized in that, The digital human model was trained in the following manner: Multi-source data with labeled information is input into the multimodal large model to be trained. The multimodal large model to be trained is trained based on the labeled information and the output of the multimodal large model to be trained to obtain a digital human model. The multimodal large model integrates a text generation model, an image recognition model, a speech recognition model and a deep thinking reasoning model.

5. The method according to claim 1, characterized in that, Different general business capability interfaces are used to handle different business processes. These business types include: personal affairs, vehicle services, specific tasks, material assets, specific locations, organizations, event information, and knowledge base. The categories of personal affairs include personnel information management and personnel business information inquiry; The vehicle service categories include queries for basic vehicle information, transaction information, maintenance information, incident information, vehicle status and trajectory information; The specific task category includes the management of basic event information, change information, progress information, and workflow information; The material asset categories include basic item information, asset data, performance data, value, and transaction flow management; The specific location category includes basic information about the location, environmental information, information on the implementation of security measures, and visitor access management. The organization category includes information management of organization name, type, date of establishment, and place of registration; The event information categories include event time, geographical location, event participants, and event content information management; The knowledge base categories include knowledge updates and deep learning resource queries.

6. A data processing device based on a modular digital human model, characterized in that, The device includes: The acquisition module is used to acquire target data from multiple heterogeneous data sources and preprocess the target data to obtain structured target data; The construction module is used to build multiple independently callable general business capability interfaces based on the structured target data. The general business capability interfaces cover at least one type of business theme among people, vehicles, events, things, organizations, places, events, and knowledge bases. The input module is used to receive task requests input by the user, and to perform natural language understanding and generation processing on the task requests through a pre-trained digital human model to determine the user's intention and the task parameters required to generate the response. The digital human model is obtained by fine-tuning a multimodal large model and integrates text generation, image recognition, speech recognition and deep reasoning capabilities. An assembly module is used to dynamically assemble one or more of the general business capability interfaces according to the user intent and task parameters determined by the digital human model, forming a task processing chain; The security module is used to perform permission verification and audit control on the task processing chain based on the model context protocol to ensure that only authorized users can access the corresponding interfaces. The output module is used to execute the task processing chain, generate processing results, and return the results to the user through at least one information push channel based on the user's preset push rules. The security module is specifically used for traffic monitoring, logging, and security auditing of the general business capability interface; it performs authentication and permission verification on received data access requests through the API gateway; after successful verification, if the traffic level of the corresponding general business capability interface is within a preset threshold, it forwards the data access request to the corresponding general business capability interface and updates the access log of the corresponding general business capability interface; if the traffic level of the corresponding general business capability interface exceeds the preset threshold, it intercepts the data access request and returns an error response.

7. The apparatus according to claim 6, characterized in that, The acquisition module is specifically used for: When the target data is a file, use a file reading tool or a custom script to periodically scan the file directory to obtain the file data. If the target data is a database type, use database connection tools or SQL queries to periodically pull the database and obtain the database data; When the target data is of an API type, use an HTTP request to call the API interface and retrieve the API data; When the target data is a message queue, the message queue consumer group is set up to consume messages in real time and obtain streaming data.

8. The apparatus according to claim 6, characterized in that, The acquisition module is also used for: Data cleaning of the target data includes: removing invalid data, filling in missing data, correcting erroneous data, and eliminating duplicate data; The cleaned target data is aligned to obtain standard target data. The standard target data is classified according to resource categories, including people, vehicles, events, objects, organizations, places, incidents, and knowledge bases. The classified standard target data is structured to obtain structured target data.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-5.

Images