An intelligent agent service system based on a domestic operating system

Abstract

The present invention discloses an intelligent agent service system based on a domestic operating system, which belongs to the technical field of the integration of operating systems and artificial intelligence. The technical problem to be solved by the present invention is the poor behavioral perception sensitivity, delayed response to preference changes, and low decision-making interpretability of conventional intelligent agent systems in a domestic ecosystem. The technical solution adopted is: the system integrates operation behavior perception and reinforcement learning technology, and forms a complete technical closed loop of perception-modeling-optimization-interpretation through an operation perception engine, a human feedback reinforcement learning center, an MCP protocol adapter, and a thinking chain visual designer, thereby realizing the paradigm transition of intelligent agents from passive execution to active collaboration; wherein, the operation perception engine captures and analyzes user behavior data in real time to obtain operation sequence data, and the operation sequence data is input as features into the reinforcement learning center, and the reinforcement learning center outputs operation instructions.

Description

Technical Field

[0001] The present invention relates to the technical field of the integration of operating systems and artificial intelligence, and specifically to an intelligent body service system based on a domestically produced operating system. Background Art

[0002] As intelligent agent technology rapidly penetrates the desktop, the shortcomings of existing technology systems in terms of local adaptation, security and trustworthiness, and ecological collaboration have seriously restricted my country's progress in information technology independence. These shortcomings are manifested in three core bottlenecks:

[0003] 1. Fragmented Protocol Ecosystem: Currently, intelligent agent tools generally utilize proprietary communication protocols (such as MCP's JSON-RPC and ANP's semantic network model), leading to data silos and skyrocketing development costs. For example, traditional approaches to enterprise ERP system integration require developing independent interfaces for each tool, extending integration cycles by 3-8 times and increasing maintenance costs by more than 400%. This fragmentation extends beyond technical architecture to security mechanisms—MCP's ACL whitelisting and ANP's zero-knowledge proofs struggle to interoperate, potentially leading to privacy risks.

[0004] Second, in-depth reconstruction of behavior perception: Traditional systems can only capture surface-level events like clicks and scrolling, and their understanding of GUI operation semantics is inaccurate. For example, in WPS office scenarios, the collaborative editing needs of users frequently switching between "revision modes" are often overlooked, and the correlation analysis capabilities of multi-window overlay operations (such as Excel and PowerPoint) are weak.

[0005] 3. Lag in responding to preference changes: Conventional reinforcement learning with human feedback (RLHF) processes are too long. Traditional thought processes rely on manually labeled data, and RLHF training cycles typically exceed 72 hours, making it difficult to respond to changes in user preferences in real time. When users adjust document formatting standards, static models require full parameter retraining, which can easily lead to catastrophic forgetting.

[0006] Therefore, how to overcome the defects of conventional intelligent systems in the domestic ecosystem, such as poor behavioral perception sensitivity, delayed response to preference changes, and low decision-making interpretability, is a technical problem that needs to be solved urgently. Summary of the Invention

[0007] The technical task of the present invention is to provide an intelligent agent service system based on a domestic operating system to solve the problems of poor behavioral perception sensitivity, delayed response to preference changes, and low decision-making interpretability existing in conventional intelligent agent systems in the domestic ecology.

[0008] The technical task of the present invention is achieved in the following manner: an intelligent agent service system based on a domestically produced operating system, which integrates operational behavior perception and reinforcement learning technology. Through an operational perception engine, a human feedback reinforcement learning hub, an MCP protocol adapter, and a thinking chain visual designer, a complete closed-loop technology of perception-modeling-optimization-interpretation is formed, thus achieving a paradigm shift from passive execution to active collaboration of intelligent agents.

[0009] Among them, the operation perception engine captures and analyzes user behavior data in real time, obtains operation sequence data, and inputs the operation sequence data as features into the reinforcement learning center. The reinforcement learning center outputs operation instructions. The MCP protocol adapter interacts with different data sources and services through standardized interfaces for the operation instructions recommended by the reinforcement learning center to obtain the results corresponding to the operation instructions, and passes the feedback from the external system (such as user ratings and eye tracking data) back to the human feedback reinforcement learning center, continuously optimizing the strategy network of the human feedback reinforcement learning center; the thinking chain visualization designer converts the complex decision-making process and data relationship of the human feedback reinforcement learning center into an intuitive visualization view, helping users understand the decision logic and behavior patterns of the intelligent behavior analysis device, improving the system's interpretability and user trust, and guiding the adjustment and optimization of the human feedback reinforcement learning center to form a closed-loop optimization process.

[0010] Preferably, the operation perception engine includes:

[0011] The operation sequence data acquisition module is used to capture user operation behaviors in real time based on the kernel-level hook mechanism of the domestic operating system to form operation sequence data; user operation behaviors include GUI operation events and file system access traces; GUI operation events include window focus switching, control clicks, and shortcut key triggering; file system access traces include create, read, write, and delete operations;

[0012] The user multi-dimensional portrait construction module is used to build a multi-dimensional user behavior portrait through event tracing technology and realize the contextual relationship analysis of operational semantics.

[0013] Preferably, the operation perception engine also has the following functions:

[0014] ①Supports multi-device synchronous perception: unified capture of mobile, desktop, and cloud operation behaviors;

[0015] ②Add abnormal behavior detection function: real-time alarm for atypical operation modes;

[0016] ③ Provide encrypted storage and transmission of behavioral data to ensure data security;

[0017] ④Support plug-in extensions, allowing third-party developers to access customized behavior perception rules.

[0018] Preferably, the reinforcement learning hub includes:

[0019] A model training module is used to train a behavior-document multimodal joint probability model. This model combines behavior data with document data (such as text and images) to model multimodal associations using a probabilistic framework. This model uses probabilistic graphical models (such as Bayesian networks and Markov random fields) or deep learning frameworks (such as variational autoencoders and generative adversarial networks) to achieve joint modeling of multimodal data.

[0020] The optimization engine building module is used to build a dual-channel feedback-driven policy optimization engine to optimize the behavior-document multimodal joint probability model;

[0021] The privacy protection and security module is used to implement privacy protection in the reinforcement learning hub using privacy protection and security mechanisms.

[0022] Preferably, the model training module works as follows:

[0023] (1) Calculate the corresponding characteristic values of the three dimensions of operation frequency, duration and path complexity for each operation behavior; among them, operation frequency refers to the number of times the user performs each operation behavior in a specific time period, reflecting the frequency of users' use of different operation behaviors; duration refers to the time the user spends on each operation behavior, reflecting the degree of attention and time invested by the user in different operations; path complexity refers to analyzing the complexity of the path when the user performs the operation, such as the directory depth of the accessed file, the number of jumps, etc., to measure the complexity of the user's operation path;

[0024] (2) Considering different user operations as words and a series of user operations as documents, the TF-IDF algorithm is used to quantify the user's preference weight W for different types of operations. The specific formula is as follows;

[0025] ;

[0026] in, Indicates the number of times the word t appears in the document Di; Represents a document The number of words in all the vocabulary; N represents the total number of documents; Represents a document Whether the word t is included, if included, it is 1, if not included, it is 0;

[0027] TF-IDF (Term Frequency-Inverse Document Frequency) is a commonly used weighting technique for text retrieval and text mining. TF-IDF is a statistical method used to assess the importance of a word to a document set or a document in a corpus. The importance of a word increases proportionally with the number of times it appears in a document, but decreases inversely with its frequency in the corpus. This effectively reduces the influence of common words on keywords and improves the relevance between keywords and articles.

[0028] (3) Using the user’s preference weights and three-dimensional feature values for different types of operation behaviors, all operation sequence data are converted into structured feature vectors as input for training the behavior-file multimodal joint probability model;

[0029] (4) The joint probability distribution of operation behavior and file access behavior is calculated through a dynamic Bayesian network, and the user's operation sequence and file access behavior at different time points are modeled as a conditional probability distribution to capture the causal relationship between operation behavior and file access;

[0030] (5) The structured feature vector is input into a multi-layer neural network, and the output of the multi-layer neural network is the probability distribution of the operation suggestion. A supervised learning method is used to train the behavior-file joint probability model through historical behavior data, construct the spatiotemporal association between operation sequence and file access, and dynamically update the probability distribution to reflect the temporal nature and contextual dependence of user behavior;

[0031] (6) Verify the performance of the behavior-document multimodal joint probability model through cross-validation and indicator evaluation (such as accuracy, recall rate, F1 score) to ensure the generalization ability of the model.

[0032] Preferably, the working process of the optimization engine establishment module is as follows:

[0033] (1) Design a feedback channel, which includes an explicit feedback channel and an implicit feedback channel. The explicit channel receives the user's star rating of the intelligent suggestion through the user interface, with a rating range of 1-5. The implicit channel records the user's gaze point, scanning path, and pupil changes during the operation through eye tracking, and records the user's dwell time on a specific operation or interface element. The cognitive load index, such as the number of gazes and average dwell time, is calculated based on the eye movement and dwell time data. The rating and cognitive load index are then converted into numerical feedback signals as input to the reward function of reinforcement learning.

[0034] (2) Designing a reward function by combining explicit and implicit feedback signals, where explicit feedback directly serves as the reward value, and implicit feedback indirectly affects the reward through cognitive load indicators;

[0035] (3) The PPO algorithm is used to calculate the gradient update of the behavior-file multimodal joint probability model policy network, and the stability of training is ensured by truncating the policy update. Among them, the PPO algorithm adopts multi-objective optimization to optimize the user operation path by minimizing the path entropy, thereby improving the operation efficiency, and by maximizing the confusion of sensitive operations, it improves the unrecognizableness of sensitive operations and protects user privacy.

[0036] Preferably, the privacy protection and security module works as follows:

[0037] (1) During the back-propagation process, gradient masking technology is used to mask gradients involving sensitive data to ensure that private data is not leaked;

[0038] (2) By adding noise or transforming feature vectors, the behavioral characteristics of sensitive operations are obfuscated, thereby reducing the identifiability of sensitive operations;

[0039] (3) Regularly evaluate the privacy protection effect of the system to ensure the effectiveness of the privacy protection mechanism.

[0040] Preferably, the MCP protocol adapter includes a protocol gateway deployment module and a dynamic service discovery module;

[0041] The protocol network management deployment module is used to deploy the MCP protocol gateway using a client-server architecture. The client is used to receive the operation strategy generated by the reinforcement learning center, and the server is an external data source and tool. The client and server negotiate functions to determine the functions and services provided by the client and server to each other. The protocol network management deployment module integrates the JSON-RPC 2.0 standard protocol and supports two communication modes, as follows:

[0042] ① Local pipeline (stdio) mode: achieves low latency response of <10ms, suitable for processing local operation behavior data;

[0043] ② Network Stream (SSE) mode: supports high-concurrency calls and is suitable for processing behavioral data in distributed systems;

[0044] The dynamic service discovery module is used to identify available MCP servers through an automatic scanning mechanism. Available MCP servers include local IDE plug-ins, enterprise ERP system interfaces, and cloud-based AI services (such as the Claude inference engine). The client sends a request to the server based on the user request or the needs of the AI model. The server processes the user request and may interact with local or remote resources. After the operation is completed, the server returns the processing result to the client, and the client then passes the information back to the host application. URI dynamic templates are then used to implement parameterized resource positioning, supporting the JSON-RPC 2.0 standard protocol to ensure the flexibility and dynamism of service discovery.

[0045] Preferably, the MCP protocol adapter has the following functions:

[0046] ①Support cross-platform compatibility: Windows, Linux, macOS, and mobile terminals;

[0047] ②Support protocol version management function and support seamless switching of different versions of MCP protocols;

[0048] ③ Provide service health monitoring function to monitor the availability of MCP servers in real time;

[0049] ④Support service circuit breaker mechanism to avoid system crash due to single point failure.

[0050] Preferably, the Thought Chain Visual Designer includes:

[0051] The decision traceability model construction module is used to build a decision traceability model based on the multi-head attention mechanism, integrating time-series behavior data with system state characteristics, tracking and recording the formation process of each decision generated by the reinforcement learning center, and building a complete decision chain. The formation process of each decision generated by the reinforcement learning center includes the key influencing factors of the decision and the contextual information at the time of decision-making;

[0052] The decision path reconstruction module is used to reconstruct the decision path using an LSTM network weighted by a time decay factor. It models and analyzes the time series data in the decision-making process, highlights the decision trends and patterns that change over time, and helps understand the evolution of decisions.

[0053] The causal relationship analysis module is used to combine knowledge graph technology to generate an explainable view of causal relationships. It associates various events and operations in the decision-making process with their resulting results to form a causal relationship diagram, revealing the logic and motivation behind the decision;

[0054] The visualization output module is used to obtain behavior heat maps, file relationship networks, and strategy evolution timelines. The behavior heat map is used to display the distribution of operation modes, showing the frequency and patterns of users' operations at different times and in different scenarios, helping to identify high-frequency operation areas and user behavior habits. The file association network is used to reveal implicit knowledge structures, showing the direct references, indirect associations, and content similarity between files, helping to discover potential knowledge structures and information flows. The strategy evolution timeline is used to display the learning process, showing the evolution and optimization process of the reinforcement learning strategy over time, helping to evaluate the learning effect and the convergence of the strategy. The evolution and optimization process of the reinforcement learning strategy over time includes the key nodes of strategy adjustment and the changing trends of performance indicators.

[0055] The feedback and optimization module is used to feed back the visualization results to the reinforcement learning center to provide a reference for further optimization of the reinforcement learning strategy. By analyzing the visualization output, it can discover potential problems and improvement space in the reinforcement learning strategy, guide the adjustment and optimization of the reinforcement learning algorithm, and form a closed-loop optimization process.

[0056] The intelligent agent service system based on the domestic operating system of the present invention has the following advantages:

[0057] (1) This invention innovatively integrates MCP protocol access, dynamic behavior modeling, and human-machine collaborative evolution mechanisms through multimodal behavior perception, behavior-file multimodal reinforcement learning optimization, and lightweight thought chain generation technology. This solves the problems of poor behavior perception sensitivity, delayed response to preference changes, and low decision-making interpretability that exist in conventional intelligent agent systems in a domestic ecosystem. This enables autonomous control and efficient reasoning of intelligent agent services in a domestic environment.

[0058] (2) This invention uses a context-sensitive protocol (MCP) to achieve dynamic access to cross-platform intelligent agent tools, combining operational behavior perception, file semantic understanding, and human feedback reinforcement learning algorithms to build a user-personalized thought chain system;

[0059] (3) This invention breaks through the core bottleneck of the application of intelligent agent technology in domestic operating systems through three major technical paths: localization adaptation, security enhancement, and ecological collaboration;

[0060] (IV) The present invention aims to construct an intelligent agent service system and device based on a domestic operating system, and realize the paradigm shift of intelligent agents from passive execution to active collaboration by building a technical closed loop of "operation-feedback-optimization". Specifically, it uses the MCP protocol to build a unified tool access framework, realizes seamless integration of heterogeneous tools based on the client-server architecture, and realizes the integration of heterogeneous tools through JSON-RPC. The 2.0 protocol encapsulates tool call requests, and the dynamic service discovery mechanism can automatically match local tools or server-side APIs. Combined with fine-grained permission control, it ensures that only verified devices and users can access system resources. A three-dimensional behavior modeling system of user operation-file-feedback is established. Through bidirectional LSTM analysis of window focus trajectory and gesture operation sequence, a multimodal feedback interface is designed to convert user ratings into incentive signals for reinforcement learning. A thinking chain optimization engine driven by a hierarchical PPO algorithm is deployed. In this engine, the meta-policy network integrates operation timing characteristics and file TF-IDF vector optimization long-term goals, and the task policy network achieves rapid decision updates through real-time data pipelines. Through protocol ecosystem standardization, in-depth behavior modeling and dynamic optimization of thinking chains, this invention constructs a new intelligent collaborative paradigm of "operation is training, feedback is evolution", providing a safe and reliable digital foundation for the integration of AI technology into human work.

[0061] (5) The present invention ensures data sovereignty through domestic kernel-level monitoring and implements transmission encryption using a national secret algorithm. It also constructs a dynamic behavior-file association model to break through the static limitations of traditional log analysis. At the same time, it innovatively integrates human feedback mechanisms and explainable AI technology to make the intelligent agent's decision-making process both evolvable and transparent. Experiments have shown that the present invention can improve the operational efficiency of common office scenarios by 37% and reduce the error rate by 62%, while providing privacy protection capabilities that comply with the GB / T 35273 standard. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention will be further described below with reference to the accompanying drawings.

[0063] Attachment Figure 1 This is the structural block diagram of the intelligent agent service system based on the domestic operating system. DETAILED DESCRIPTION

[0064] The following is a detailed description of an intelligent body service system based on a domestic operating system of the present invention with reference to the accompanying drawings and specific embodiments.

[0065] Example:

[0066] As attached Figure 1 As shown, this embodiment provides an intelligent agent service system based on a domestic operating system. The system integrates operation behavior perception and reinforcement learning technology. Through the operation perception engine, human feedback reinforcement learning center, MCP protocol adapter and thinking chain visual designer, it forms a complete technical closed loop of perception-modeling-optimization-interpretation, realizing the paradigm transition of intelligent agents from passive execution to active collaboration.

[0067] Among them, the operation perception engine captures and analyzes user behavior data in real time, obtains operation sequence data, and inputs the operation sequence data as features into the reinforcement learning center. The reinforcement learning center outputs operation instructions. The MCP protocol adapter interacts with different data sources and services through standardized interfaces for the operation instructions recommended by the reinforcement learning center to obtain the results corresponding to the operation instructions, and passes the feedback from the external system (such as user ratings and eye tracking data) back to the human feedback reinforcement learning center, continuously optimizing the strategy network of the human feedback reinforcement learning center; the thinking chain visualization designer converts the complex decision-making process and data relationship of the human feedback reinforcement learning center into an intuitive visualization view, helping users understand the decision logic and behavior patterns of the intelligent behavior analysis device, improving the system's interpretability and user trust, and guiding the adjustment and optimization of the human feedback reinforcement learning center to form a closed-loop optimization process.

[0068] The operation perception engine in this embodiment includes:

[0069] The operation sequence data acquisition module is used to capture user operation behaviors in real time based on the kernel-level hook mechanism of the domestic operating system to form operation sequence data; user operation behaviors include GUI operation events and file system access traces; GUI operation events include window focus switching, control clicks, and shortcut key triggering; file system access traces include create, read, write, and delete operations;

[0070] The user multi-dimensional portrait construction module is used to build a multi-dimensional user behavior portrait through event tracing technology and realize the contextual relationship analysis of operational semantics.

[0071] The operation perception engine in this embodiment also has the following functions:

[0072] ①Supports multi-device synchronous perception: unified capture of mobile, desktop, and cloud operation behaviors;

[0073] ②Add abnormal behavior detection function: real-time alarm for atypical operation modes;

[0074] ③ Provide encrypted storage and transmission of behavioral data to ensure data security;

[0075] ④Support plug-in extensions, allowing third-party developers to access customized behavior perception rules.

[0076] The reinforcement learning hub in this embodiment includes:

[0077] A model training module is used to train a behavior-document multimodal joint probability model. This model combines behavior data with document data (such as text and images) to model multimodal associations using a probabilistic framework. This model uses probabilistic graphical models (such as Bayesian networks and Markov random fields) or deep learning frameworks (such as variational autoencoders and generative adversarial networks) to achieve joint modeling of multimodal data.

[0078] The optimization engine building module is used to build a dual-channel feedback-driven policy optimization engine to optimize the behavior-document multimodal joint probability model;

[0079] The privacy protection and security module is used to implement privacy protection in the reinforcement learning hub using privacy protection and security mechanisms.

[0080] The working process of the model training module in this embodiment is as follows:

[0081] (1) Calculate the corresponding characteristic values of the three dimensions of operation frequency, duration and path complexity for each operation behavior; among them, operation frequency refers to the number of times the user performs each operation behavior in a specific time period, reflecting the frequency of users' use of different operation behaviors; duration refers to the time the user spends on each operation behavior, reflecting the degree of attention and time invested by the user in different operations; path complexity refers to analyzing the complexity of the path when the user performs the operation, such as the directory depth of the accessed file, the number of jumps, etc., to measure the complexity of the user's operation path;

[0082] (2) Considering different user operations as words and a series of user operations as documents, the TF-IDF algorithm is used to quantify the user's preference weight W for different types of operations. The specific formula is as follows;

[0083] ;

[0084] in, Indicates the number of times the word t appears in the document Di; Represents a document The number of words in all the vocabulary; N represents the total number of documents; Represents a document Whether the word t is included, if included, it is 1, if not included, it is 0;

[0085] (3) Using the user’s preference weights and three-dimensional feature values for different types of operation behaviors, all operation sequence data are converted into structured feature vectors as input for training the behavior-file multimodal joint probability model;

[0086] (4) The joint probability distribution of operation behavior and file access behavior is calculated through a dynamic Bayesian network, and the user's operation sequence and file access behavior at different time points are modeled as a conditional probability distribution to capture the causal relationship between operation behavior and file access;

[0087] (5) The structured feature vector is input into a multi-layer neural network, and the output of the multi-layer neural network is the probability distribution of the operation suggestion. A supervised learning method is used to train the behavior-file joint probability model through historical behavior data, construct the spatiotemporal association between operation sequence and file access, and dynamically update the probability distribution to reflect the temporal nature and contextual dependence of user behavior;

[0088] (6) Verify the performance of the behavior-document multimodal joint probability model through cross-validation and indicator evaluation (such as accuracy, recall rate, F1 score) to ensure the generalization ability of the model.

[0089] The working process of the optimization engine establishment module in this embodiment is as follows:

[0090] (1) Design a feedback channel, which includes an explicit feedback channel and an implicit feedback channel. The explicit channel receives the user's star rating of the intelligent suggestion through the user interface, with a rating range of 1-5. The implicit channel records the user's gaze point, scanning path, and pupil changes during the operation through eye tracking, and records the user's dwell time on a specific operation or interface element. The cognitive load index, such as the number of gazes and average dwell time, is calculated based on the eye movement and dwell time data. The rating and cognitive load index are then converted into numerical feedback signals as input to the reward function of reinforcement learning.

[0091] (2) Designing a reward function by combining explicit and implicit feedback signals, where explicit feedback directly serves as the reward value, and implicit feedback indirectly affects the reward through cognitive load indicators;

[0092] (3) The PPO algorithm is used to calculate the gradient update of the behavior-file multimodal joint probability model policy network, and the stability of training is ensured by truncating the policy update. Among them, the PPO algorithm adopts multi-objective optimization to optimize the user operation path by minimizing the path entropy, thereby improving the operation efficiency, and by maximizing the confusion of sensitive operations, it improves the unrecognizableness of sensitive operations and protects user privacy.

[0093] The working process of the privacy protection and security module in this embodiment is as follows:

[0094] (1) During the back-propagation process, gradient masking technology is used to mask gradients involving sensitive data to ensure that private data is not leaked;

[0095] (2) By adding noise or transforming feature vectors, the behavioral characteristics of sensitive operations are obfuscated, thereby reducing the identifiability of sensitive operations;

[0096] (3) Regularly evaluate the privacy protection effect of the system to ensure the effectiveness of the privacy protection mechanism.

[0097] The MCP protocol adapter in this embodiment includes a protocol gateway deployment module and a dynamic service discovery module;

[0098] The protocol network management deployment module is used to deploy the MCP protocol gateway using a client-server architecture. The client is used to receive the operation strategy generated by the reinforcement learning center, and the server is an external data source and tool. The client and server negotiate functions to determine the functions and services provided by the client and server to each other. The protocol network management deployment module integrates the JSON-RPC 2.0 standard protocol and supports two communication modes, as follows:

[0099] ① Local pipeline (stdio) mode: achieves low latency response of <10ms, suitable for processing local operation behavior data;

[0100] ② Network Stream (SSE) mode: supports high-concurrency calls and is suitable for processing behavioral data in distributed systems;

[0101] The dynamic service discovery module is used to identify available MCP servers through an automatic scanning mechanism. Available MCP servers include local IDE plug-ins, enterprise ERP system interfaces, and cloud-based AI services (such as the Claude inference engine). The client sends a request to the server based on the user request or the needs of the AI model. The server processes the user request and may interact with local or remote resources. After the operation is completed, the server returns the processing result to the client, and the client then passes the information back to the host application. URI dynamic templates are then used to implement parameterized resource positioning, supporting the JSON-RPC 2.0 standard protocol to ensure the flexibility and dynamism of service discovery.

[0102] The MCP protocol adapter in this embodiment has the following functions:

[0103] ①Support cross-platform compatibility: Windows, Linux, macOS, and mobile terminals;

[0104] ②Support protocol version management function and support seamless switching of different versions of MCP protocols;

[0105] ③ Provide service health monitoring function to monitor the availability of MCP servers in real time;

[0106] ④Support service circuit breaker mechanism to avoid system crash due to single point failure.

[0107] The thinking chain visual designer in this embodiment includes:

[0108] The decision traceability model construction module is used to build a decision traceability model based on the multi-head attention mechanism, integrating time-series behavior data with system state characteristics, tracking and recording the formation process of each decision generated by the reinforcement learning center, and building a complete decision chain. The formation process of each decision generated by the reinforcement learning center includes the key influencing factors of the decision and the contextual information at the time of decision-making;

[0109] The decision path reconstruction module is used to reconstruct the decision path using an LSTM network weighted by a time decay factor. It models and analyzes the time series data in the decision-making process, highlights the decision trends and patterns that change over time, and helps understand the evolution of decisions.

[0110] The causal relationship analysis module is used to combine knowledge graph technology to generate an explainable view of causal relationships. It associates various events and operations in the decision-making process with their resulting results to form a causal relationship diagram, revealing the logic and motivation behind the decision;

[0111] The visualization output module is used to obtain behavior heat maps, file relationship networks, and strategy evolution timelines. The behavior heat map is used to display the distribution of operation modes, showing the frequency and patterns of users' operations at different times and in different scenarios, helping to identify high-frequency operation areas and user behavior habits. The file association network is used to reveal implicit knowledge structures, showing the direct references, indirect associations, and content similarity between files, helping to discover potential knowledge structures and information flows. The strategy evolution timeline is used to display the learning process, showing the evolution and optimization process of the reinforcement learning strategy over time, helping to evaluate the learning effect and the convergence of the strategy. The evolution and optimization process of the reinforcement learning strategy over time includes the key nodes of strategy adjustment and the changing trends of performance indicators.

[0112] The feedback and optimization module is used to feed back the visualization results to the reinforcement learning center to provide a reference for further optimization of the reinforcement learning strategy. By analyzing the visualization output, it can discover potential problems and improvement space in the reinforcement learning strategy, guide the adjustment and optimization of the reinforcement learning algorithm, and form a closed-loop optimization process.

[0113] The working process of this embodiment is as follows:

[0114] S1. Capture user operation behaviors in real time based on the operation perception engine; user operation behaviors are the input basis for subsequent modules (such as the reinforcement learning hub and the MCP protocol adapter). Based on the kernel-level hook mechanism of the domestic operating system, user operation behaviors are captured in real time, including GUI operation events (such as window focus switching, control clicks, shortcut key triggering) and file system access traces (such as create / read / write / delete operations), forming operation sequence data.

[0115] S2. Input the user operation behavior data captured by the operation perception engine as features into the reinforcement learning center to train the behavior-file multimodal joint probability model. The behavior-file multimodal joint probability model is used to receive user task instructions and output the probability distribution of recommended operation instructions, as follows:

[0116] S201. For each operation, calculate the characteristic values of the three dimensions of operation frequency, duration, and path complexity. Operation frequency refers to the number of times a user performs each operation within a specific time period, reflecting the frequency of use of different operations by the user. Duration refers to the time a user spends on each operation, reflecting the user's attention and time investment in different operations. Path complexity refers to analyzing the complexity of the path taken by the user when performing the operation, such as the directory depth of the file accessed and the number of jumps, to measure the complexity of the user's operation path.

[0117] S202: Consider different user operations as "vocabulary" and a series of user operations as "documents". Use the TF-IDF algorithm to quantify the user's preference weight W for different types of operations. The specific formula is as follows:

[0118] ;

[0119] in, Indicates the number of times the word t appears in the document Di; Represents a document The number of words in all the vocabulary; N represents the total number of documents; Represents a document Whether the word t is included, if included, it is 1, if not included, it is 0;

[0120] S203. Using the calculated preference weights and the extracted three-dimensional features, all operation sequence data are converted into structured feature vectors as input for model training;

[0121] S204. Calculate the joint probability distribution of the operation behavior and the file access behavior using a dynamic Bayesian network, model the user's operation sequence and file access behavior at different time points as a conditional probability distribution, and capture the causal relationship between the operation behavior and file access.

[0122] S205. Construct a multi-layer neural network with structured feature vectors as input and probability distribution of action suggestions as output. Use supervised learning to train a behavior-file joint probability model using historical behavior data, build a spatiotemporal association between action sequences and file accesses, and dynamically update the probability distribution to reflect the temporal order and contextual dependency of user behavior.

[0123] S206. Verify model performance through cross-validation and indicator evaluation (such as accuracy, recall, and F1 score) to ensure the generalization ability of the model;

[0124] S3. Receive the user's task instructions, obtain the optimal operation instructions based on the trained behavior-document multimodal joint probability model, call external data sources and services through the standardized interface provided by the MCP protocol adapter, and execute the corresponding instructions to obtain the operation results corresponding to the operation instructions; among them, the operation strategy generated by the reinforcement learning center is passed to the external system (such as ERP, AI service) through the MCP protocol adapter to automatically execute the corresponding operation; the construction of the MCP protocol adaptation framework includes two parts: protocol gateway deployment and dynamic service discovery. The protocol network management deployment adopts a client-server architecture to deploy the MCP protocol gateway. The client is used to receive the operation strategy generated by the reinforcement learning center, and the server is an external data source and tool. The client and server negotiate functions to determine which functions and services they can provide to each other; integrate the JSON-RPC 2.0 standard protocol and support two communication modes:

[0125] ① Local pipeline (stdio) mode achieves a low latency response of <10ms, suitable for processing local operation behavior data;

[0126] ② The network stream (SSE) mode supports high-concurrency calls and is suitable for processing behavioral data in distributed systems;

[0127] Dynamic service discovery uses an automatic scanning mechanism to identify available MCP servers, including but not limited to local IDE plug-ins, enterprise ERP system interfaces, and cloud-based AI services (such as the Claude inference engine). Based on user requests or AI model needs, the client sends requests to the server, which processes these requests and may interact with local or remote resources. After the operation is completed, the server returns the processing results to the client, which then passes the information back to the host application. URI dynamic templates are used to implement parameterized resource location, and the JSON-RPC 2.0 standard protocol is supported to ensure the flexibility and dynamism of service discovery.

[0128] S4: A dual-channel feedback-driven policy optimization engine in the reinforcement learning hub collects user feedback on operation results (such as user ratings and eye tracking data) to further optimize the policy, improve model operation efficiency and privacy protection capabilities; the details are as follows:

[0129] S401. Design feedback channels, including explicit and implicit feedback channels. The explicit channel receives the user's star rating of the intelligent suggestions through the user interface, with a rating range of 1-5. The implicit channel records the user's gaze point, scanning path, and pupil changes during the operation through eye tracking, and records the user's dwell time on specific operations or interface elements. Cognitive load indicators, such as the number of fixations and average dwell time, are calculated based on eye movement and dwell time data. The rating and cognitive load indicators are converted into numerical feedback signals as input to the reward function of reinforcement learning.

[0130] S402. Design a reward function by combining explicit and implicit feedback signals, where explicit feedback directly serves as the reward value, and implicit feedback indirectly affects the reward through the cognitive load indicator;

[0131] S403. Using the PPO algorithm to calculate the gradient update of the behavior-file multimodal joint probability model policy network, and ensuring the stability of training by truncating the policy update; wherein the PPO algorithm uses multi-objective optimization to optimize the user operation path by minimizing path entropy to improve operation efficiency, and to maximize the obfuscation of sensitive operations to increase the unrecognizableness of sensitive operations and protect user privacy;

[0132] S5. Use the Thinking Chain Visual Designer to transform the decision-making process and data relationships of the behavior-document multimodal joint probability model into an intuitive visualization. This helps users understand the system's decision logic and behavior patterns, improves the system's interpretability and user trust, and guides the adjustment and optimization of the reinforcement learning algorithm, forming a closed-loop optimization process. The details are as follows:

[0133] S501. Decision traceability model construction: Build a decision traceability model based on a multi-head attention mechanism, integrating time-series behavior data with system state characteristics. This model tracks and records the formation process of each decision generated by the reinforcement learning hub, including key influencing factors and contextual information at the time of decision-making, to build a complete decision chain.

[0134] S502, Decision Path Reconstruction: Use a time-decay factor-weighted LSTM network to reconstruct the decision path, model and analyze the time series data in the decision-making process, highlight the decision trends and patterns that change over time, and help understand the evolution of the decision process;

[0135] S503, Causal Association Analysis: Combined with knowledge graph technology, this technology generates an interpretable view of causal relationships, linking various events and operations in the decision-making process with their resulting outcomes to form a causal relationship diagram, revealing the logic and motivation behind the decision;

[0136] S504, visual output, as follows:

[0137] S50401, Behavior Heat Map: This shows the distribution of operation modes, displaying the frequency and patterns of users' operations at different times and in different scenarios, helping to identify high-frequency operation areas and user behavior habits.

[0138] S50402, Document Association Network: Reveals implicit knowledge structures and displays associations between documents, including direct citations, indirect associations, and content similarities, helping to discover potential knowledge structures and information flows.

[0139] S50403, Strategy Evolution Timeline: Shows the learning process, presenting the evolution and optimization process of reinforcement learning strategies over time, including key points of strategy adjustment and trends in performance indicators, to help evaluate learning effects and strategy convergence.

[0140] S505, Feedback and Optimization: Feedback the visualization results to the reinforcement learning center to provide a reference for further optimization of the strategy; by analyzing the visualization output, potential problems and improvement space in the strategy are discovered, guiding the adjustment and optimization of the reinforcement learning algorithm, forming a closed-loop optimization process.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent agent service system based on a domestic operating system, characterized in that: The system integrates operational behavior perception and reinforcement learning technologies. Through the operational perception engine, human feedback reinforcement learning center, MCP protocol adapter, and thinking chain visual designer, it forms a complete technical closed loop of perception-modeling-optimization-interpretation, realizing the paradigm shift of intelligent agents from passive execution to active collaboration. Among them, the operation perception engine captures and analyzes user behavior data in real time, obtains operation sequence data, and inputs the operation sequence data as features into the reinforcement learning center. The reinforcement learning center outputs operation instructions. The MCP protocol adapter interacts with different data sources and services through standardized interfaces for the operation instructions recommended by the reinforcement learning center to obtain the results corresponding to the operation instructions, and transmits the feedback of the external system back to the human feedback reinforcement learning center, continuously optimizing the strategy network of the human feedback reinforcement learning center; the thinking chain visual designer converts the complex decision-making process and data relationship of the human feedback reinforcement learning center into an intuitive visual view, and guides the adjustment and optimization of the human feedback reinforcement learning center to form a closed-loop optimization process; The reinforcement learning hub includes: Model training module, used for behavior-document multimodal joint probability model training; The optimization engine building module is used to build a dual-channel feedback-driven policy optimization engine to optimize the behavior-document multimodal joint probability model; Privacy and security module, which is used to implement privacy protection in reinforcement learning hubs using privacy and security mechanisms; The working process of the model training module is as follows: (1) Calculate the corresponding characteristic values of the three dimensions of operation frequency, duration and path complexity for each operation behavior; among them, operation frequency refers to the number of times the user performs each operation behavior in a specific time period, reflecting the frequency of users' use of different operation behaviors; duration refers to the time the user spends on each operation behavior, reflecting the degree of attention and time invested by the user in different operations; path complexity refers to analyzing the complexity of the path when the user performs the operation, measuring the complexity of the user's operation path; (2) Considering different user operations as words and a series of user operations as documents, the TF-IDF algorithm is used to quantify the user's preference weight W for different types of operations. The specific formula is as follows; ； in, Indicates the number of times the word t appears in the document Di; Represents a document The number of words in all the vocabulary; N represents the total number of documents; Represents a document Whether the word t is included, if included, it is 1, if not included, it is 0; (3) Using the user’s preference weights and three-dimensional feature values for different types of operation behaviors, all operation sequence data are converted into structured feature vectors as input for training the behavior-file multimodal joint probability model; (4) The joint probability distribution of operation behavior and file access behavior is calculated through a dynamic Bayesian network, and the user's operation sequence and file access behavior at different time points are modeled as a conditional probability distribution to capture the causal relationship between operation behavior and file access; (5) The structured feature vector is input into a multi-layer neural network, and the output of the multi-layer neural network is the probability distribution of the operation suggestion. A supervised learning method is used to train the behavior-file joint probability model through historical behavior data, construct the spatiotemporal association between operation sequence and file access, and dynamically update the probability distribution to reflect the temporal nature and contextual dependence of user behavior; (6) Verify the performance of the behavior-document multimodal joint probability model through cross-validation and indicator evaluation to ensure the generalization ability of the model; The working process of the optimization engine establishment module is as follows: (1) Design a feedback channel, which includes an explicit feedback channel and an implicit feedback channel. The explicit channel receives the user's star rating of the intelligent suggestion through the user interface, with a rating range of 1-5. The implicit channel records the user's gaze point, scanning path, and pupil changes during the operation through eye tracking, and records the user's dwell time on a specific operation or interface element. The cognitive load index is calculated based on the eye movement and dwell time data. The rating and cognitive load index are then converted into numerical feedback signals as the input of the reward function of reinforcement learning. (2) Designing a reward function by combining explicit and implicit feedback signals, where explicit feedback directly serves as the reward value, and implicit feedback indirectly affects the reward through cognitive load indicators; (3) The PPO algorithm is used to calculate the gradient update of the behavior-file multimodal joint probability model policy network, and the stability of training is ensured by truncating the policy update. Among them, the PPO algorithm adopts multi-objective optimization to optimize the user operation path by minimizing the path entropy, and to improve the unrecognizableness of sensitive operations by maximizing the confusion of sensitive operations, thereby protecting user privacy.

2. The intelligent agent service system based on the domestic operating system according to claim 1 is characterized in that: The Operation Awareness Engine includes: The operation sequence data acquisition module is used to capture user operation behaviors in real time based on the kernel-level hook mechanism of the domestic operating system to form operation sequence data; user operation behaviors include GUI operation events and file system access traces; GUI operation events include window focus switching, control clicks, and shortcut key triggering; file system access traces include create, read, write, and delete operations; The user multi-dimensional portrait construction module is used to build a multi-dimensional user behavior portrait through event tracing technology and realize the contextual relationship analysis of operational semantics.

3. The intelligent agent service system based on a domestic operating system according to claim 1 or 2, characterized in that: The operation perception engine also has the following functions: ①Supports multi-device synchronous perception: unified capture of mobile, desktop, and cloud operation behaviors; ②Add abnormal behavior detection function: real-time alarm for atypical operation modes; ③ Provide encrypted storage and transmission of behavioral data to ensure data security; ④Support plug-in extensions, allowing third-party developers to access customized behavior perception rules.

4. The intelligent agent service system based on the domestic operating system according to claim 1 is characterized in that: The working process of the privacy protection and security module is as follows: (1) During the back-propagation process, gradient masking technology is used to mask gradients involving sensitive data to ensure that private data is not leaked; (2) By adding noise or transforming feature vectors, the behavioral characteristics of sensitive operations are obfuscated, thereby reducing the identifiability of sensitive operations; (3) Regularly evaluate the privacy protection effect of the system to ensure the effectiveness of the privacy protection mechanism.

5. The intelligent agent service system based on the domestic operating system according to claim 1 is characterized in that: The MCP protocol adapter includes a protocol gateway deployment module and a dynamic service discovery module; The protocol network management deployment module is used to deploy the MCP protocol gateway using a client-server architecture. The client is used to receive the operation strategy generated by the reinforcement learning center, and the server is an external data source and tool. The client and server negotiate functions to determine the functions and services provided by the client and server to each other. The protocol network management deployment module integrates the JSON-RPC 2.0 standard protocol and supports two communication modes, as follows: ① Local pipeline mode: achieves low latency response of <10ms, suitable for processing local operation behavior data; ② Network flow mode: supports high-concurrency calls and is suitable for processing behavioral data in distributed systems; The dynamic service discovery module is used to identify available MCP servers through an automatic scanning mechanism. Available MCP servers include local IDE plug-ins, enterprise ERP system interfaces, and cloud-based AI services. The client sends a request to the server based on the user request or the needs of the AI model. The server processes the user request and may interact with local or remote resources. After the operation is completed, the server returns the processing result to the client, and the client then passes the information back to the host application. URI dynamic templates are then used to implement parameterized resource positioning, supporting the JSON-RPC 2.0 standard protocol to ensure the flexibility and dynamism of service discovery.

6. The intelligent agent service system based on the domestic operating system according to claim 1 is characterized in that: The MCP protocol adapter has the following functions: ①Support cross-platform compatibility: Windows, Linux, macOS, and mobile terminals; ②Support protocol version management function and support seamless switching of different versions of MCP protocols; ③ Provide service health monitoring function to monitor the availability of MCP servers in real time; ④Support service circuit breaker mechanism to avoid system crash due to single point failure.

7. The intelligent agent service system based on the domestic operating system according to claim 1 is characterized in that: The Thought Chain Visual Designer includes: The decision traceability model construction module is used to build a decision traceability model based on the multi-head attention mechanism, integrating time-series behavior data with system state characteristics, tracking and recording the formation process of each decision generated by the reinforcement learning center, and building a complete decision chain. The formation process of each decision generated by the reinforcement learning center includes the key influencing factors of the decision and the contextual information at the time of decision-making; The decision path reconstruction module is used to reconstruct the decision path using an LSTM network weighted by a time decay factor. It models and analyzes the time series data in the decision-making process, highlights the decision trends and patterns that change over time, and helps understand the evolution of decisions. The causal relationship analysis module is used to combine knowledge graph technology to generate an explainable view of causal relationships. It associates various events and operations in the decision-making process with their resulting results to form a causal relationship diagram, revealing the logic and motivation behind the decision; The visualization output module is used to obtain behavior heat maps, file relationship networks, and strategy evolution timelines. The behavior heat map is used to display the distribution of operation modes, showing the frequency and patterns of users' operations at different times and in different scenarios, helping to identify high-frequency operation areas and user behavior habits. The file association network is used to reveal implicit knowledge structures, showing the direct references, indirect associations, and content similarity between files, helping to discover potential knowledge structures and information flows. The strategy evolution timeline is used to display the learning process, showing the evolution and optimization process of the reinforcement learning strategy over time, helping to evaluate the learning effect and the convergence of the strategy. The evolution and optimization process of the reinforcement learning strategy over time includes the key nodes of strategy adjustment and the changing trends of performance indicators. The feedback and optimization module is used to feed back the visualization results to the reinforcement learning center to provide a reference for further optimization of the reinforcement learning strategy. By analyzing the visualization output, it can discover potential problems and improvement space in the reinforcement learning strategy, guide the adjustment and optimization of the reinforcement learning algorithm, and form a closed-loop optimization process.

Images