Non-intrusive industrial agent system and method based on mcp and vertical domain large model

Abstract

The application belongs to the technical field of artificial intelligence and industrial software integration, and discloses a non-intrusive industrial intelligent agent system and method based on MCP and vertical domain large model, which comprises the following steps: target system investigation, data preparation and intelligent agent basic running environment deployment; constructing a large model decision center fine-tuned by a domain; defining an MCP protocol to eliminate the semantic gap between intelligent decision and front-end execution, and performing cross-system front-end adaptation; building a non-intrusive interaction channel between the intelligent agent and the target system, and deploying an instruction execution service; dynamic task arrangement, closed-loop feedback and system joint debugging optimization. The application integrates core technologies such as non-intrusive deployment, domain adaptive large model, standardized MCP protocol, intelligent resource scheduling and closed-loop optimization, realizes full-process automation of industrial operation and maintenance scenes, greatly shortens fault handling time and reduces the failure rate of manual operation, and solves the problems of fault handling delay and high cost of manual operation in the prior art.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and industrial software integration technology, and more specifically, to a non-intrusive industrial intelligent agent system and method based on MCP and vertical domain large model. Background Technology

[0002] Currently, industrial operation, monitoring, and management information systems generally suffer from the problem of "abundant data but weak intelligence." Although these systems can collect massive amounts of equipment status and alarm information, their intelligence level remains at the stage of rule matching or simple threshold judgment, lacking the ability to understand the context of complex faults, reason about causes and effects, and handle them dynamically.

[0003] To enhance intelligence levels, the industry is attempting to introduce large-scale models into industrial scenarios. However, directly integrating general-purpose large-scale models into industrial systems faces three core challenges: 1. Poor domain adaptability: General-purpose large-scale models lack a deep understanding of specific industrial terminology, procedures, equipment topologies, and failure modes, resulting in unreliable and unprofessional outputs. 2. Execution gaps and data processing bottlenecks: Even if large-scale models can output reasonable handling suggestions, it is difficult to effectively translate them into automatic retrieval and in-depth analysis of target system data, and further into actual operation of the front-end interface, failing to form a complete closed loop from "perception-decision-data analysis-execution". 3. High integration invasiveness: Traditional solutions often require deep coupling to the target system's back-end interfaces or databases, resulting in high modification costs, long cycles, and significant risks, making implementation particularly difficult in safety-sensitive critical infrastructure sectors such as rail transit and power.

[0004] Specifically, in practical applications, many industrial systems require complex fault diagnosis and predictive maintenance based on real-time status data. While existing automation tools (such as Selenium) can simulate user operations and complete some simple front-end tasks, they are highly dependent on page structure, and their scripts are fragile and difficult to maintain. More importantly, these tools lack the ability to interact dynamically with intelligent agents at the semantic level and driven by intent, failing to support complex, multi-step intelligent task flows, especially in scenarios requiring the analysis of internal system data for decision-making. Furthermore, traditional solutions typically require manually specifying APIs or directly accessing databases to obtain necessary data for analysis, which not only increases system complexity and development costs but may also introduce security risks. Ideally, intelligent agents should be able to autonomously extract the required data from the target system without interfering with the existing system architecture, combining this data to conduct in-depth analysis of faults, equipment status, etc., thereby providing more accurate and professional decision support.

[0005] The industry urgently needs an innovative solution that can leverage cutting-edge AI technology to enhance the intelligence of decision-making, while also achieving seamless integration with existing systems in a non-intrusive and low-cost manner to automate operations and data retrieval and analysis. The solution should significantly improve operational efficiency, reduce the cost of manual intervention, and ultimately achieve intelligent closed-loop management of operations and maintenance while ensuring system security.

[0006] In view of this, we propose a non-intrusive industrial intelligent agent system and method based on MCP and vertical domain large model. Summary of the Invention

[0007] 1. The technical problems to be solved.

[0008] The purpose of this application is to provide a non-intrusive industrial intelligent agent method and system based on MCP and vertical domain large model, which solves the technical problems mentioned in the background technology, and realizes the integration of core technologies such as non-intrusive deployment, domain adaptive large model, standardized MCP protocol, intelligent resource scheduling and closed-loop optimization, so as to realize full-process automation of industrial operation and maintenance scenarios, significantly shorten the fault handling time, reduce the error rate of human operation, and solve the problems of fault handling delay and high manual operation and maintenance costs in the prior art.

[0009] 2. Technical solution.

[0010] This application provides a non-intrusive industrial intelligent agent system based on MCP and a large vertical domain model, including a basic environment construction system, a dynamic task orchestration and closed-loop optimization system, a domain-adaptive intelligent decision kernel system, a standardized front-end control protocol system, and a non-intrusive plug-in execution architecture system.

[0011] Basic Environment Setup System: Completes the deployment of the basic system environment, data preparation, and security pre-configuration, providing support for the subsequent operation of core modules, including: target system adaptation module, domain data and knowledge base module, and basic operating environment module.

[0012] Domain-Adaptive Intelligent Decision-Making Kernel System: This system constructs a domain-fine-tuned large-model decision center to achieve semantic parsing, multi-hop reasoning, and structured instruction output for industrial events. It includes: a domain-specific large-model fine-tuning module, a decision kernel testing and optimization module, a RAG retrieval enhancement generation module, and a structured decision generation module.

[0013] Standardized front-end control protocol system: It realizes the definition of MCP protocol, mapping template configuration and execution engine development, bridging the semantic gap between intelligent decision-making and front-end execution, including protocol definition extension module, mapping template configuration module and execution engine module.

[0014] Non-intrusive plug-in execution architecture system: Builds a non-intrusive interaction channel between intelligent agents and target systems, deploys command execution services, and strengthens security isolation, including: command execution service and sandbox isolation module, non-intrusive data channel module and non-intrusive control channel module.

[0015] Dynamic Task Orchestration and Closed-Loop Optimization System: Enables complex task flow orchestration, constructs a fully autonomous closed loop of execution-feedback-optimization, and completes system integration and optimization, including: dynamic task orchestration module, closed-loop feedback autonomous optimization module, and system integration and iteration optimization module.

[0016] This invention provides a non-intrusive industrial intelligent agent method based on MCP and vertical domain large model, including the following steps.

[0017] Step S1, Preliminary Preparation and Basic Environment Setup: Complete the target system survey, data preparation, and deployment of the basic operating environment for the intelligent agent.

[0018] S11. Target system research, requirements analysis and compatibility pre-adaptation.

[0019] S12, Domain data collection, preprocessing, and intelligent construction of knowledge base.

[0020] S13. Deployment and security hardening of the basic operating environment for intelligent agents.

[0021] Step S2: Constructing a domain-adaptive intelligent decision-making kernel: Constructing a domain-fine-tuned large model decision center to achieve semantic parsing, multi-hop reasoning, and structured instruction output for industrial events, including the following steps.

[0022] S21, Vertical domain large model depth fine-tuning.

[0023] S22. The optimized configuration of the search enhancement generation RAG mechanism improves search accuracy and reasoning efficiency.

[0024] S23. Optimization of structured decision generation process.

[0025] S24. Decision kernel testing and intelligent optimization: Conduct multi-dimensional testing and adaptive optimization.

[0026] Step S3: Design and implementation of the standardized front-end control protocol MCP: Define the MCP protocol to bridge the semantic gap between intelligent decision-making and front-end execution, and achieve cross-system front-end adaptation.

[0027] S31, MCP protocol core definition and extension.

[0028] S32 and MCP mapping template intelligent configuration.

[0029] S33 and MCP execution engine development, testing, and optimization.

[0030] Step S4: Deployment of non-intrusive plug-in execution architecture: Establish a non-intrusive interaction channel between the intelligent agent and the target system, deploy instruction execution services, and ensure secure isolation and seamless integration.

[0031] S41. Data channel setup. S42. Control channel setup. S43. Command execution service deployment and sandbox enhancement.

[0032] Step S5, Dynamic Task Orchestration, Closed-Loop Feedback and System Integration Optimization: Implement complex task flow orchestration, build a continuous learning closed loop of execution-feedback-optimization, and complete system integration and iterative optimization.

[0033] S51, intelligent configuration of dynamic task orchestration.

[0034] S52. Enhanced closed-loop feedback mechanism: Achieve a fully autonomous closed loop of execution-feedback-optimization.

[0035] 3. Beneficial effects.

[0036] One or more technical solutions provided in this application have the following technical effects or advantages.

[0037] 1. This invention achieves complete isolation between the intelligent agent and the target system through loosely coupled runtime environment deployment, zero-intrusion event release adaptation, and lightweight script injection technologies. It does not require modification of the target system's source code, does not access the backend database, and does not occupy the target system's core resources. It solves the defects of high coupling, mutual interference, and strong integration intrusion between the intelligent agent and the target system. It is compatible with various industrial target systems such as rail transit signal operation and maintenance IOM systems and power equipment monitoring systems, and is compatible with various deployment environments such as physical machines, virtual machines, and containerization.

[0038] 2. The enhanced automated execution sandbox adopts a process, permission, and data layered isolation mechanism. Combined with a security rule engine and operation permission control, it strictly limits the scope of intelligent agent operation and completes interaction only at the front-end level, meeting the stringent requirements of rail transit SIL4 safety level and power industry data security standards. At the same time, it builds a full-process data security protection system through data encryption, two-way authentication, and anomaly filtering.

[0039] 3. Construct a comprehensive index model for data acquisition security and non-intrusiveness, and a quantitative model for sandbox security isolation levels to achieve accurate quantitative assessment of non-intrusiveness and security. This will overcome the limitations of existing technologies that cannot quantify non-intrusiveness and provide verifiable technical support for security adaptation in industrial scenarios.

[0040] 4. Adopting a three-level fine-tuning mechanism of domain knowledge internalization + instruction alignment + scenario adaptation, combined with efficient parameter fine-tuning technology, incremental model fine-tuning can be achieved with only a small amount of scenario-specific data, which greatly shortens the fine-tuning cycle and reduces computational overhead, solving the defects of existing technologies such as shallow fine-tuning of general large models, poor adaptability, long fine-tuning cycle and high cost.

[0041] 5. An optimized retrieval enhancement generation RAG mechanism, combined with a hierarchical local vectorized knowledge base and a multi-hop retrieval closed loop, improves the accuracy of knowledge retrieval and reasoning efficiency through dynamic parameter adjustment and weighted fusion of retrieval results; and constructs a multi-hop retrieval reasoning accuracy quantification model to achieve joint quantification of retrieval and reasoning effects.

[0042] 6. The structured decision generation process ensures that the output of instructions is standardized and parsable through dynamic adaptation of JSON Schema templates, semantic parsing optimization, enhanced reasoning interpretability design, and abnormal event handling mechanism, while also enabling effective handling of unknown faults.

[0043] 7. Standardized MCP protocol solves cross-system adaptation problems and improves versatility and reliability: It addresses the shortcomings of existing front-end control protocols, such as high coupling, poor adaptability, and weak scalability. It adapts to changes in front-end UI elements of the target system in real time, reducing template configuration and maintenance costs.

[0044] 8. The modular design and multi-level intelligent fault tolerance mechanism of the MCP execution engine, combined with the state prediction module and page crash recovery mechanism, greatly improve the success rate and continuity of operations; construct a quantitative model of the effectiveness of the fault tolerance mechanism to quantify the fault tolerance recovery capability and efficiency. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of a non-intrusive industrial intelligent agent system based on MCP and a vertical domain large model, as disclosed in a preferred embodiment of this application.

[0046] Figure 2 This is a timing diagram of the front-end automatic interaction of a non-intrusive industrial intelligent agent system based on MCP and a large vertical domain model, as disclosed in a preferred embodiment of this application. Detailed Implementation

[0047] The present application will be further described in detail below with reference to the accompanying drawings.

[0048] This invention aims to overcome the technical shortcomings of existing industrial target systems, such as "rich data but weak intelligence," poor adaptability to large-scale models, execution gaps, and strong integration intrusion. By constructing a loosely coupled intelligent agent runtime environment, it achieves full automation and intelligence in fault handling, while reducing integration costs and improving system security. Without modifying or with minimal modification to the core code of existing industrial target systems, it constructs an industrial intelligent agent with intelligent analysis, decision-making, and automatic front-end interaction capabilities through vertical domain fine-tuning of large models, the MCP protocol, and a non-intrusive add-on architecture, thus achieving full automation and intelligence in fault handling.

[0049] Reference Figure 1 and Figure 2 This invention provides a non-intrusive industrial intelligent agent system based on MCP and vertical domain large model, including...

[0050] Basic Environment Setup System: Completes the deployment of the basic system environment, data preparation and security pre-configuration, providing support for the operation of subsequent core modules, and solving the problems of poor compatibility between environment deployment and target system and lack of early prevention of security risks in existing technologies. It includes: target system adaptation module, domain data and knowledge base module and basic operating environment module.

[0051] Target System Adaptation Module: This module completes a comprehensive survey of the target system, outlines operational requirements, identifies security constraints, and performs pre-adaptation for compatibility, providing a basis for overall system adaptation. It includes a system survey unit, a requirements outlining unit, a security compliance unit, and a compatibility adaptation unit.

[0052] System Survey Unit: Collect the core functions, front-end technology stack, UI structure, event publishing mechanism, historical fault data and operation and maintenance pain points of the target system, and generate the "Target System Survey and Analysis Report".

[0053] Requirements Analysis Unit: Identify core operation and maintenance scenarios, including fault alarm handling and equipment status monitoring, and clarify system task priorities, input and output standards, and execution time limits.

[0054] Security and Compliance Unit: Investigate the security level, interface restrictions, and data encryption requirements of the target system, conduct security pre-assessment, and develop compliance adaptation plans.

[0055] Compatibility adaptation unit: Establishes a version-feature-adaptation strategy mapping table to adapt to different versions of the target system and different deployment environments, avoiding version incompatibility issues.

[0056] Domain Data and Knowledge Base Module: This module collects, cleans, and enhances multi-source domain data, constructs a hierarchical local vectorized knowledge base and knowledge graph, providing data support for subsequent large-scale model fine-tuning and inference. It includes a data acquisition unit, a data preprocessing unit, and a knowledge base construction unit.

[0057] Data acquisition unit: Collects data from multiple sources, including fault logs, maintenance manuals, technical standards, expert Q&A, and equipment lifecycle data, to ensure that the data volume meets the needs of fine-tuning, with ≥5000 unlabeled text entries and ≥1000 labeled instruction-response pairs.

[0058] Data preprocessing unit: Performs data cleaning and standardization, and enhances data through synonym replacement, fault scenario expansion, etc.; integrates a data enhancement effect quantification module to ensure a threshold ≥1.5. Data cleaning includes removing invalid, duplicate, and sensitive data. Standardization includes unifying technical terminology and fault code formats.

[0059] Knowledge base construction unit: An embedding model is used to encode text blocks into vectors, which are then stored in the FAISS vector retrieval engine to build a hierarchical knowledge base consisting of a base layer, a case layer, and a real-time layer. Knowledge relationships are mined using a graph neural network (GNN) to construct a fault-equipment-disposal knowledge graph. A dynamic update interface is set up to support real-time data entry and automatic removal of outdated data. A knowledge base retrieval accuracy quantification module is integrated to ensure a threshold ≥ 0.85. The embedding model is a bidirectional encoder-representation converter. The FAISS vector retrieval engine is used to improve retrieval efficiency, the bge-large-zh model is embedded to achieve text vectorization, knowledge relationships are built through GNN, and dynamic updates of the knowledge base and quantification of retrieval performance are supported.

[0060] Basic operating environment module: Deploys a loosely coupled intelligent agent operating environment, establishes a secure and stable communication channel, configures basic dependencies, and implements dynamic resource scheduling to ensure the safe and stable operation of the system. It includes an environment deployment unit, a basic dependency and resource scheduling unit, a communication channel unit, and a security hardening unit.

[0061] Environment Deployment Unit: Adopts Docker / K8s containerized deployment to isolate the intelligent agent's operating environment from the target system and avoid mutual interference.

[0062] Communication channel unit: Deploys Kafka / RabbitMQ message middleware - data channel and WebSocket service - control channel, adopts AES-256 encryption, breakpoint resume, heartbeat keep-alive, abnormal reconnection and other strategies to ensure data transmission security and real-time, with a latency of ≤100ms.

[0063] Basic Dependency and Resource Scheduling Unit: Configures basic dependencies such as embedded models, FAISS engine, computing resources, and audit databases; automatically allocates computing and storage resources based on task load.

[0064] Security hardening unit: Deploys firewalls and IDS intrusion detection systems to restrict access permissions; uses the national cryptographic SM4 algorithm to encrypt and store the audit database, and backs it up regularly; sets up a running status monitoring subunit to monitor CPU, memory, and network usage in real time, and triggers alarms and starts the backup environment when anomalies occur.

[0065] It implements containerized deployment based on Docker / K8s, adopts security technologies such as SSL / TLS two-way authentication, AES-256 encryption, and WSS encrypted communication, and integrates dynamic resource scheduling algorithms and runtime status monitoring tools.

[0066] Domain-Adaptive Intelligent Decision-Making Kernel System: This system constructs a domain-fine-tuned large-model decision center to achieve semantic parsing, multi-hop reasoning, and structured instruction output for industrial events. It includes: a domain-specific large-model fine-tuning module, a decision kernel testing and optimization module, a RAG retrieval enhancement generation module, and a structured decision generation module.

[0067] Vertical Domain Large Model Fine-Tuning Module: Through a triple fine-tuning mechanism, this module adapts a general-purpose large model to the needs of industrial applications, improving the accuracy and professionalism of instruction generation and overcoming the limitations of existing dual-training methods. It includes a domain knowledge internalization training unit, an instruction alignment training unit, and a scene-adaptive fine-tuning unit.

[0068] Domain knowledge internalization training unit: General large models such as LLaMA / ChatGLM are selected and optimized for lightweighting, including pruning redundant parameters and optimizing the attention mechanism; further pre-training is performed using over 5000 unlabeled domain texts, employing the Mini-Batch SGD algorithm with a learning rate of 1e. -5 Set up a pre-training verification sub-unit to verify the degree of knowledge mastery through a domain-specific test set with an accuracy of ≥90%. The test set includes 1000 professional questions.

[0069] Instruction Alignment Training Unit: Supervised fine-tuning of SFT based on more than 1000 labeled instruction-response pairs, and adjustment of training weights according to task type and urgency; fixed output format constraints through DPO preference optimization, setting up an output verification subunit to verify the legality and feasibility of instructions; setting up training verification and iteration subunits to ensure that instruction generation time is ≤500ms and accuracy is ≥95%.

[0070] Scene Adaptive Fine-Tuning Unit: Employing PEFT parameter fine-tuning technology, incremental fine-tuning is achieved using only 50-100 scene-specific data points, with a single fine-tuning cycle of ≤24 hours. A fine-tuning effect evaluation subunit is included to monitor scene adaptation accuracy, instruction generation speed, and resource utilization. Based on PEFT technology, it achieves efficient parameter fine-tuning, integrates training algorithms such as Mini-Batch SGD and DPO, and includes pre-training, fine-tuning verification, and effect quantification modules, supporting incremental adaptation across multiple scenes.

[0071] The RAG retrieval enhancement generation module associates a hierarchical local vectorized knowledge base with a fine-tuned large model. By optimizing the RAG mechanism, it improves retrieval accuracy and inference efficiency, achieving a multi-hop retrieval closed loop. It includes a RAG parameter dynamic adjustment unit, a multi-hop retrieval unit, and a retrieval result fusion unit.

[0072] The RAG parameter dynamic adjustment unit configures the retrieval similarity threshold and Top-N recall, automatically adjusting parameters based on retrieval result confidence and event urgency. The default retrieval similarity threshold is 0.75, and the default Top-N recall is Top-5.

[0073] Multi-hop retrieval unit: Implements a closed-loop logic of first-hop retrieval - second supplementary retrieval - third verification retrieval, sets up a retrieval cache sub-unit to cache high-frequency queries and results, with a default validity period of 24 hours; integrates a multi-hop retrieval inference accuracy quantification module to ensure that the threshold is ≥0.8.

[0074] Search result fusion unit: A weighted fusion algorithm is adopted to fuse multi-hop search results according to the hierarchical weight of the knowledge base: 0.6 for case layer, 0.3 for basic layer, and 0.1 for real-time layer, highlighting the role of high-value knowledge.

[0075] It connects the FAISS local vectorized knowledge base with the fine-tuned large model, integrates multi-hop retrieval logic, caching mechanism and weighted fusion algorithm, and supports dynamic adjustment of retrieval parameters and quantification of effect.

[0076] Structured Decision Generation Module: This module solidifies the structured decision generation process, dynamically adapts JSON Schema templates, generates machine-parseable structured instructions, handles exceptions, and addresses the shortcomings of existing technologies, such as inaccurate semantic parsing and infeasible instructions. It includes a template dynamic adaptation unit and a decision generation process unit.

[0077] Template Dynamic Adaptation Unit: Presets multiple JSON Schema templates, automatically matching them based on task type and target system type. Sets up a dynamic template update interface, supporting the addition and modification of templates without modifying the model code.

[0078] The decision generation process unit comprises three steps: solidified semantic parsing and query construction → multi-hop knowledge retrieval → enhanced reasoning and structured output. It employs NER+RE technology to optimize semantic parsing and generate accurate retrieval queries. A sub-unit for the interpretability of the reasoning process is included, outputting the reasoning logic chain. An exception handling sub-unit is also included, generating emergency alarms and manual intervention instructions for unknown faults, and storing the abnormal data in a training sample pool. It integrates NER and RE semantic parsing technologies, pre-sets multiple JSONSchema templates, supports dynamic template updates, and includes a corresponding reasoning interpretability module and exception handling logic.

[0079] Decision kernel testing and optimization module: Performs multi-dimensional testing on the decision kernel, achieving intelligent adaptive optimization to ensure the stability, accuracy, and efficiency of the decision kernel. Includes a multi-dimensional testing unit and an intelligent adaptive optimization unit.

[0080] Multi-dimensional test unit: Input scenario data such as common faults, complex faults, and unknown faults, and test the semantic parsing accuracy -≥98%, retrieval accuracy -≥95%, instruction generation accuracy -≥95%, response speed -≤500ms, etc.

[0081] Intelligent adaptive optimization unit: Automatically adjusts fine-tuning parameters, RAG configuration or schema template for problems encountered during testing; sets up a model performance monitoring subunit to monitor the running status in real time and automatically triggers incremental fine-tuning when performance degrades.

[0082] It integrates multi-dimensional testing tools, along with adaptive optimization algorithms and performance monitoring modules, supporting automatic parameter adjustment and incremental model fine-tuning.

[0083] A standardized front-end control protocol system: This system defines the MCP protocol, configures mapping templates, and develops the execution engine, bridging the semantic gap between intelligent decision-making and front-end execution, and overcoming the shortcomings of existing protocols such as high coupling and poor adaptability. It includes a protocol definition extension module, a mapping template configuration module, and an execution engine module.

[0084] Protocol Definition Extension Module: Defines a lightweight, declarative MCP protocol, employing operational primitives and extension mechanisms to achieve deep decoupling between business semantics and front-end technologies, enhancing protocol flexibility and scalability. It includes a protocol framework definition unit, a protocol decoupling unit, and a protocol extension unit.

[0085] Protocol framework definition unit: Defines the MCP protocol in JSON format, abstracts atomic operation primitives, including the original 4 types of primitives and the newly added 4 types of primitives capture, compare, delay, and call_api, and clarifies the JSON description format, positioning strategy and fault tolerance mechanism of each primitive.

[0086] Protocol decoupling unit: Strengthens the decoupling of "what to do" and "how to do it". The decision kernel generates semantic layer instructions, which are parsed into technical implementations by the MCP execution engine. Semantic tags are set to automatically map sub-units without the need for manual configuration of mapping templates.

[0087] Protocol extension unit: Sets up the MCP protocol extension interface, supports the addition of custom operation primitives, and adapts to different industrial scenarios and target systems; integrates the protocol extensibility quantization module to ensure that the threshold is ≥5.

[0088] The protocol is defined using JSON format, standardized operation primitives are encapsulated, an extension interface is designed, automatic semantic tag mapping is supported, and an extensible quantization module is provided.

[0089] Mapping Template Configuration Module: Enables fine-grained configuration, multi-system adaptation, and dynamic updates of MCP mapping templates, reducing template configuration costs and improving adaptation efficiency. It includes a fine-grained template configuration unit, a multi-system automatic mapping unit, and a dynamic template update unit.

[0090] Refined template configuration unit: Configure JSON format mapping templates for different front-end technology stacks to realize the mapping between business semantic tags and front-end technology implementation; adopt a layered design of basic template - scenario template - version template.

[0091] Multi-system automatic mapping unit: Reserves interfaces for multi-system adaptation, designs mapping templates that can be configured independently and loaded on demand; sets up a template automatic generation sub-unit, collects front-end UI elements through web crawling technology, automatically generates the initial mapping template, and only requires a small amount of manual verification.

[0092] Template dynamic update unit: Real-time monitoring of changes in front-end UI elements, automatic updating of mapped templates, support for template version management and historical retrospection; integrated module for quantifying the accuracy of mapped template adaptation.

[0093] It integrates web crawling tools and a UI element change monitoring module, supports hierarchical configuration, automatic generation and dynamic updates of mapping templates, and is equipped with a module for quantifying adaptation accuracy.

[0094] Execution Engine Module: Develops a modular MCP execution engine to implement instruction parsing, element location, action execution, status feedback, and intelligent fault tolerance, ensuring the accuracy, stability, and efficiency of instruction execution. It includes a modular execution unit, an intelligent fault tolerance unit, and a testing and performance optimization unit.

[0095] Modular Execution Unit: Includes an MCP parser, element locator, action executor, state feedback unit, and state prediction module. The MCP parser parses JSON command sequences and prioritizes them. The element locator performs multi-strategy positioning, intelligent waiting, and retries, achieving a positioning accuracy of ≥98%. The action executor simulates user interaction and adapts to different front-end frameworks. The state feedback unit collects operation results and sends back feedback data with a latency of ≤100ms. The state prediction module predicts operation results based on the LSTM algorithm, triggering backup strategies in advance.

[0096] Intelligent fault-tolerant unit: Implements a three-level fault-tolerant mechanism, including retrying element location failure, triggering backup operation sequences and abnormal alarms, setting up a page crash recovery subunit, automatically refreshing the page and re-establishing connections to ensure operational continuity; integrates a fault-tolerant mechanism effectiveness quantification module to ensure that the threshold is ≥8.

[0097] Testing and Performance Optimization Unit: Tests protocol compatibility, execution accuracy, and response speed; optimizes performance through methods such as batch instruction execution, parallel processing, and high-frequency operation caching.

[0098] It adopts a modular design, integrating LSTM state prediction algorithm, multi-strategy element localization logic, three-level fault tolerance mechanism, and supporting performance optimization and testing tools.

[0099] Non-intrusive add-on execution architecture system: This system establishes a non-intrusive interaction channel between the intelligent agent and the target system, deploys command execution services, strengthens security isolation, and overcomes the shortcomings of existing integrations that are highly invasive and lack sufficient security isolation. It includes: a command execution service and sandbox isolation module, a non-intrusive data channel module, and a non-intrusive control channel module.

[0100] Non-intrusive data channel module: Enables zero-intrusion acquisition of event data from the target system, ensuring secure and accurate data transmission and resolving the issue of some target systems being unable to publish events. It includes a data processing and security unit, a message middleware configuration unit, and a zero-intrusive event acquisition unit.

[0101] Message middleware configuration unit: Configure Kafka / RabbitMQ message middleware, adopt SSL / TLS two-way authentication, set parameters such as message partitioning and expiration time to avoid message backlog.

[0102] Zero-intrusion event collection unit: Coordinates the target system to publish events to a specified Topic without modifying the source code; for systems that do not have event publishing capabilities, deploys a lightweight, non-intrusive event collection plugin that only collects event data and does not access the backend database.

[0103] Data Processing and Security Unit: Receives event data, decrypts and deserializes it using AES-256; sets up a data anomaly filtering subunit to filter invalid data and desensitize it; sets up a data caching subunit to cache recent data (default 1 hour) to prevent data loss; integrates a non-intrusive and security quantification module for data acquisition to ensure a threshold ≥ 0.9.

[0104] Deploy a lightweight event collection plugin, employing AES-256 encryption and SSL / TLS two-way authentication, integrating data filtering, desensitization, and caching mechanisms, and equipped with a non-intrusive and security quantification module.

[0105] Non-intrusive control channel module: Establishes a non-intrusive control channel between the intelligent agent and the target system front-end, ensuring secure and stable command transmission and adaptability to different browsers and front-end technology stacks. Includes: lightweight script injection unit, communication unit,

[0106] Lightweight script injection unit: Injects JavaScript scripts with a size of ≤100KB, establishes a WebSocket long connection, has a self-protection mechanism to prevent tampering and deletion, and only has front-end operation permissions.

[0107] Communication unit: WSS encrypted communication is adopted, and the transmission format of instructions and status data is clearly defined; a heartbeat keep-alive and gradient reconnection strategy is set, and the number of reconnections is ≤5 times to ensure connection stability.

[0108] Compatibility and Load Optimization Unit: Optimizes WebSocket connection adaptation logic to adapt to different browsers and front-end technology stacks; sets up a channel load balancing subunit to allocate connection resources; sets up a command priority transmission subunit to ensure that urgent commands are transmitted first.

[0109] Channel monitoring unit: Tests channel stability, communication latency, and command transmission accuracy; monitors operational status in real time; triggers alarms and records logs when abnormalities occur.

[0110] Command Execution Service and Sandbox Isolation Module: Deploys command execution services, strengthens sandbox isolation, ensures operational safety, and implements status feedback and anomaly isolation to meet industrial system security level requirements. Includes a status feedback and anomaly isolation unit, a command execution service unit, and an enhanced sandbox unit.

[0111] Instruction Execution Service Unit: Deploys the MCP execution engine, associated with the decision kernel and WebSocket service, and adopts a microservice architecture; sets up a service health check subunit to monitor the module status in real time and start a backup module in case of failure.

[0112] Enhanced Sandbox Unit: Configures a layered sandbox with process isolation, permission isolation, and data isolation, with a built-in security rule engine to prohibit dangerous operations; sets up a sandbox logging sub-unit to record operation trajectories for security auditing and fault tracing.

[0113] Status feedback and anomaly isolation unit: Collects operation status, encapsulates feedback packets and sends them back to the decision kernel; sets up an anomaly isolation subunit to isolate abnormal operation sequences so as not to affect other operations; sets up an operation permission control subunit to restrict the scope of operations based on the permissions of maintenance personnel.

[0114] Dynamic Task Orchestration and Closed-Loop Optimization System: Enables complex task flow orchestration, constructs a fully autonomous closed loop of execution-feedback-optimization, completes system integration and optimization, and solves the shortcomings of existing technologies where task orchestration is simple and optimization relies on manual intervention. It includes: a dynamic task orchestration module, a closed-loop feedback autonomous optimization module, and a system integration and iteration optimization module.

[0115] Dynamic Task Orchestration Module: Enables intelligent orchestration, template adaptation, and dynamic adjustment of complex task flows, improving the flexibility and adaptability of task execution. It includes a task flow monitoring and adjustment unit, a complex task flow orchestration unit, and a task flow template unit.

[0116] Complex task flow orchestration unit: Configures task flow conditional branches, loops and parallel execution logic, supports task flow nesting; sets task flow priority management sub-unit, and sorts execution according to urgency and importance.

[0117] Task Flow Template Unit: Presets multiple task flow templates to automatically match different operation and maintenance scenarios; Sets up a template intelligent generation sub-unit to automatically generate new templates based on historical operation trajectories and expert experience, which can be used with minor manual adjustments; Sets up a template adaptive adjustment sub-unit to automatically adjust operation steps according to changes in the target system.

[0118] Task Flow Monitoring and Adjustment Unit: Monitors task flow execution progress and operation status in real time, and visualizes the execution trajectory; automatically switches to backup steps, adjusts the execution order, or triggers manual intervention in case of abnormalities.

[0119] Closed-loop feedback autonomous optimization module: Constructs a fully autonomous execution-feedback-optimization closed loop to achieve real-time feedback, audit traceability, and continuous learning, thereby enhancing the system's autonomous evolution capabilities. It includes a real-time feedback correction unit, an audit traceability unit, and a continuous learning autonomous optimization unit.

[0120] Real-time feedback correction unit: The front-end operation status is transmitted back to the decision kernel in real time. The feedback data analysis subunit is set up to identify the cause of failure and automatically trigger the correction strategy, adjusting the mapping template, RAG parameters, etc., without manual intervention.

[0121] Audit traceability unit: The operation trajectory is encrypted and stored in the audit database, and blockchain technology is used to ensure that it is tamper-proof; an audit report automatic generation sub-unit is set up to generate operation and maintenance reports on a regular basis; and multi-dimensional retrieval of audit data is supported.

[0122] Continuous learning and autonomous optimization units: Automatically filter high-frequency failures, unknown faults, and other cases, and import them into the training sample pool after anonymization and labeling; Set up automatic sample labeling sub-units to improve labeling efficiency; Regularly initiate large-scale model incremental fine-tuning, knowledge base updates, and template optimization to achieve autonomous evolution of the agent.

[0123] System integration and iterative optimization module: Completes system-wide integration testing, troubleshoots interface issues between modules, implements system iterative optimization, and ensures the system meets the actual needs of industrial scenarios. Includes a system-wide integration testing unit and an iterative optimization unit.

[0124] System-wide integration unit: This unit coordinates all the aforementioned modules to test inter-module interface compatibility, data transmission accuracy, and command execution consistency, and to identify and resolve issues during the integration process.

[0125] Iterative optimization unit: Based on the joint debugging results and actual operation data, optimize the parameters, algorithms and configurations of each module, improve innovative functions, ensure continuous improvement of system performance, and adapt to dynamic changes in industrial scenarios.

[0126] This invention provides a non-intrusive industrial intelligent agent method based on MCP and vertical domain large model, including the following steps.

[0127] Step S1, Preliminary Preparation and Basic Environment Setup: Complete the target system survey, data preparation, and deployment of the basic operating environment for the intelligent agent.

[0128] S11. Target system research, requirements analysis and compatibility pre-adaptation.

[0129] (1) Comprehensive Survey of the Target System: Identify the core functional modules, front-end technology stack, UI structure characteristics, and event publishing mechanism of the target system. Simultaneously, survey the system's historical fault data and operational pain points, such as delayed fault handling and high human error rates, to generate a "Target System Survey and Analysis Report," clarifying the key areas for intelligent agent adaptation. Target systems include rail transit signal maintenance IOM systems and power equipment monitoring systems. The front-end technology stack includes React / Vue / jQuery, etc. The event publishing mechanism includes log push protocols, event reporting frequency, and data formats.

[0130] (2) Refined analysis of operation and maintenance scenarios and tasks: The core operation and maintenance scenarios of the target system are analyzed, including fault alarm handling, real-time equipment status monitoring, predictive maintenance, routine inspection, fault review and analysis, etc. The specific tasks that the intelligent agent needs to perform are clarified, the task priorities are refined, such as the priority of emergency fault handling is higher than that of routine inspection, and the input and output standards and execution time limits of each task are defined.

[0131] (3) Security Constraints and Compliance Path Investigation: Investigate the target system's security level, interface opening restrictions, and data encryption requirements to determine the compliant paths for data acquisition and front-end interaction; set up a security pre-assessment stage to detect security vulnerabilities in the target system, such as cross-domain access restrictions and permission control deficiencies, through simulated data interaction and front-end operations; develop avoidance plans in advance to ensure that the agent's operations do not touch the target system's back-end logic and database permissions, while complying with industry security standards, such as the security standards for rail transit signaling systems and the data security standards for the power industry. For example, in the rail transit field, the security level must meet SIL4.

[0132] (4) Compatibility pre-adaptation: For different versions of the target system and different deployment environments, such as physical machines, virtual machines and containerized deployments, a compatibility adaptation module is set up to collect the UI element features and event publishing formats of each version of the system in advance, and establish a version-feature-adaptation strategy mapping table to avoid version incompatibility issues during subsequent intelligent agent deployment.

[0133] S12, Domain data collection, preprocessing, and intelligent construction of knowledge base.

[0134] (1) Comprehensive Collection of Data from Multiple Sources: Collect relevant corpora from the target domain, including fault logs, maintenance manuals, technical standards, expert Q&A, historical fault handling cases, equipment topology data, real-time operation data, etc., with a total amount sufficient for fine-tuning needs, at least 5,000 unlabeled text entries and 1,000 labeled instruction-response pairs. Utilize equipment lifecycle data, including equipment factory parameters, maintenance records, and aging data, to provide data support for predictive maintenance. Target domains include rail transit signaling and power equipment. Fault logs include fault codes, fault phenomena, handling processes, and handling results. Real-time operation data includes equipment voltage, current, and temperature data.

[0135] (2) Data Cleaning and Enhancement: Invalid, duplicate, and sensitive data are removed; text is segmented, denoised, and standardized, such as unifying professional terminology, fault code formats, and equipment naming rules; labeled data is verified to ensure the accuracy of the mapping logic between input events and handling instructions; data enhancement steps are set up to expand the amount of labeled data through synonym replacement, fault scenario expansion, and instruction variant generation, such as expanding "FAM mode loss" to similar scenarios such as "FAM mode abnormally missing" and "FAM signal interruption," generating corresponding instruction-response pairs to solve the problems of insufficient labeled data and poor model generalization ability in existing technologies. A quantitative model is constructed to quantify the data enhancement effect; the quantitative model is as follows.

[0136] .in: , These are the weighting coefficients; To increase the amount of post-annotated data; This represents the original amount of labeled data. To enhance the average semantic similarity between the data and the original data, a value of 0-1 is used; The data augmentation effect evaluation index has a threshold of ≥1.5. If the value is lower than the threshold, the data needs to be expanded. This quantitative model can accurately quantify the effectiveness of data augmentation, which is different from the lack of quantitative standards in existing technologies. It is a data augmentation effectiveness evaluation metric. It is used to accurately quantify the effectiveness of data augmentation steps, unlike existing technologies that lack quantitative standards.

[0137] (3) Intelligent Construction and Layering of Local Vectorized Knowledge Base: The cleaned maintenance manuals, historical cases, equipment topology data, and other texts are divided into blocks and encoded, and converted into vectors, such as bge-large-zh, through an embedding model, and stored in the local vectorized knowledge base constructed by FAISS; a layered knowledge base design is adopted, divided into a basic layer, a case layer, and a real-time layer, with each layer having an independent retrieval weight, and the case layer having a higher weight than the basic layer. At the same time, a dynamic update interface for the knowledge base is set up to support the real-time entry of new data and the automatic elimination of outdated data, ensuring the timeliness of the knowledge base; in addition, a knowledge association mining module is set up to mine the association between different faults, equipment, and handling instructions through graph neural networks (GNN), and construct a fault-equipment-handling knowledge graph to provide support for subsequent multi-hop reasoning, solving the defects of the single knowledge base structure and insufficient knowledge association in the existing technology. The basic layer includes technical standards and equipment manuals; the case layer includes historical fault handling cases; and the real-time layer includes real-time operating data of the target system and the latest fault cases.

[0138] A knowledge base retrieval quantification model is constructed.

[0139] ;in: To determine the number of knowledge fragments retrieved; The relevance between the i-th knowledge fragment and the query is represented by a value between 0 and 1, which is verified by the knowledge graph. The weight of the knowledge base level where the i-th knowledge fragment is located is 0.6 for the case layer, 0.3 for the basic layer, and 0.1 for the real-time layer. To ensure search accuracy, with a threshold of ≥0.85, this knowledge base retrieval quantification model enables quantifiable verification of search results, overcoming the limitation of existing technologies where search results cannot be accurately evaluated.

[0140] S13. Deployment and security hardening of the basic operating environment for intelligent agents.

[0141] (1) Deployment of loosely coupled runtime environment: The runtime environment of the intelligent agent is deployed using independent processes or containers, such as Docker and K8s, to ensure loose coupling with the target system and not occupy the core resources of the target system; an environment isolation mechanism is adopted to isolate the runtime environment of the intelligent agent from the target system environment through virtualization technology, so as to avoid the abnormal operation of the intelligent agent from affecting the normal operation of the target system, and solve the problem of high coupling and mutual interference between the intelligent agent and the target system in the existing technology.

[0142] (2) Communication Channel Deployment and Optimization: Deploy message middleware, such as Kafka / RabbitMQ, to build a data channel between the agent and the target system. Optimize the message transmission mechanism by using partitioned storage, message encryption, and breakpoint resume technology to ensure the security and real-time performance of data transmission, with a latency of ≤100ms. Deploy WebSocket service to build a control channel between the two systems, using a connection keep-alive mechanism and an abnormal reconnection strategy. The reconnection interval is gradually increased to avoid frequent reconnections consuming resources and improve the stability of the control channel. Message encryption uses AES-256 encryption.

[0143] (3) Basic Dependency Configuration and Dynamic Resource Scheduling: A dynamic resource scheduling module is set up to configure basic dependencies, including embedded models, the FAISS vector retrieval engine, computing resources required for fine-tuning large vertical models, and an audit database. Based on the agent's task load, such as peak fault handling and model fine-tuning tasks, computing and storage resources are automatically allocated to avoid resource waste and ensure the agent's response speed under high load scenarios, thus addressing the shortcomings of fixed resource allocation and low utilization in existing technologies. An embedded model such as bge-large-zh is used. The audit database is used to store operation trajectories and feedback data.

[0144] A resource scheduling efficiency quantification model is constructed, and the resource scheduling efficiency quantification model is as follows.

[0145] .in, The average response time of tasks under a fixed resource allocation mode; The average response time of tasks under dynamic resource scheduling mode; The threshold for improving resource scheduling efficiency is ≥30%, and the resource scheduling efficiency quantification model can quantify the advantages of dynamic scheduling.

[0146] (4) Security hardening of the operating environment: Set up an operating status monitoring module, deploy firewalls and intrusion detection systems (IDS), restrict access permissions of the intelligent agent's operating environment, and only open necessary ports for interaction with the target system; encrypt and store the audit database, using the national cryptographic algorithm SM4, and perform regular data backups; monitor the CPU, memory, and network usage of the intelligent agent's operating environment in real time through the operating status monitoring module, and automatically trigger alarms and start the backup environment when an anomaly occurs to ensure the continuous and stable operation of the intelligent agent and meet the 7×24-hour operating requirements of the industrial system.

[0147] Step S2: Constructing a domain-adaptive intelligent decision-making kernel: Constructing a domain-fine-tuned large model decision center to achieve semantic parsing, multi-hop reasoning, and structured instruction output for industrial events. This differs from the shortcomings of existing technologies that directly apply general large models, do not fine-tune them deeply, and have simple reasoning logic. It includes the following steps.

[0148] S21, Vertical domain large model depth fine-tuning.

[0149] (1) Training of domain knowledge internalization layer, including basic model selection and optimization and training target verification.

[0150] Base Model Selection and Optimization: A general-purpose large model was selected as the base model. For industrial scenarios, such as LLaMA and ChatGLM, the base model was lightweighted and optimized by pruning redundant parameters, optimizing the attention mechanism, and reducing model runtime overhead to ensure the agent can be deployed on industrial edge devices. The base model was further pre-trained using over 5000 cleaned unlabeled domain texts, including equipment manuals, maintenance logs, and equipment topology data. Training employed the Mini-Batch SGD algorithm with a learning rate of 1e. -5 The number of iterations is dynamically adjusted based on the pre-training effect, typically 5-10 rounds.

[0151] Training objective validation: This involves deeply adapting the model parameters to the target domain's terminology, fault logic chains, and handling standards, allowing the model to internalize core domain knowledge, such as the FAM mode of rail transit signaling systems, the working principle of RMCR networks, and the fault propagation patterns of power equipment. This avoids non-specialized issues found in general-purpose models. A pre-training validation phase is employed, using a domain-specific test set containing 1000 domain-specific questions to verify the model's mastery of domain knowledge. Accuracy must reach over 90%. If this target is not met, the amount of pre-training data is increased, and training parameters are adjusted until the requirements are met.

[0152] (2) Training of instruction alignment layer.

[0153] Supervised fine-tuning of SFT optimization: Based on over 1000 manually labeled instruction-response pairs, such as "FAM mode lost → check RMCR network + reset AOM" and "equipment temperature too high → reduce load + check cooling system," supervised fine-tuning is performed on the pre-trained model. Instruction classification training is employed, categorizing instructions by task type and urgency, and adjusting training weights accordingly. Emergency fault handling instructions have higher training weights than regular instructions to ensure the model's accuracy in responding to urgent tasks. Task types include fault handling, status monitoring, and inspection. Urgency levels include urgent, regular, and low priority.

[0154] Preference-based DPO optimization and output constraints: By optimizing the DPO, the model learns the mapping logic of "input event → structured operation instruction", solidifies the output format constraint capability, and ensures that the output conforms to the preset JSON Schema template. An output validation mechanism is adopted. After the model outputs, the rule engine verifies the legality and feasibility of the instruction, such as checking whether the device ID in the instruction exists and whether the operation parameters conform to the device specifications. Invalid instructions are returned to the model for regeneration, which solves the defect of the model output instructions being infeasible in the existing technology.

[0155] Training, Validation, and Iteration: Input typical fault events, including emergency, routine, and complex faults, to validate the accuracy and professionalism of the model's output instructions and whether they conform to domain handling standards. Instruction generation time should be ≤500ms, and accuracy ≥95%. If the standards are not met, iteratively adjust the training samples and supplement similar fault cases. Optimize training parameters, including adjusting the learning rate and number of iterations, until they meet actual requirements. Employ cross-validation to validate the model using fault data from different target systems to ensure its universality.

[0156] A quantitative model for fine-tuning effect is constructed to evaluate the fine-tuning effect. The quantitative model for fine-tuning effect is as follows.

[0157] .in, These are the weighting coefficients; An accuracy rate is generated for the instruction, with a value between 0 and 1; This refers to the average instruction generation time. To evaluate the fine-tuning effect, the threshold is ≥0.9. The fine-tuning effect quantification model comprehensively quantifies the accuracy and response speed, breaking through the limitations of single-indicator evaluation in existing technologies and highlighting the superiority of the fine-tuning mechanism.

[0158] (3) Scene adaptive fine-tuning breaks through the limitations of existing fine-tuning.

[0159] Scenario-adaptive fine-tuning mechanism: For different industrial scenarios, such as rail transit signal fault handling and predictive maintenance of power equipment, and different target system versions, a scenario-adaptive fine-tuning module is designed. It does not require full fine-tuning of the model. Only incremental fine-tuning with a small amount of scenario-specific data, such as 50-100 data points, is required to adapt the model to a specific scenario. For example, for power equipment monitoring systems, by supplementing power equipment fault data and handling specifications, incremental fine-tuning enables the model to quickly adapt to the professional needs of the power field.

[0160] Optimization of fine-tuning efficiency: The PEFT technique for efficient parameter fine-tuning is adopted, which only fine-tunes some parameters of the model, such as LoRA adapter parameters, reducing the computational overhead of fine-tuning and shortening the fine-tuning cycle. The time for single scene adaptive fine-tuning is ≤24 hours, which solves the defects of long model fine-tuning cycle and high cost in existing technologies. Fine-tuning effect evaluation indicators are adopted, including scene adaptation accuracy, instruction generation speed and resource utilization, to ensure that the performance of the model does not degrade after fine-tuning.

[0161] Construct a scenario-adaptive fine-tuning efficiency quantification model to quantify the efficiency of scenario-adaptive fine-tuning. The scenario-adaptive fine-tuning efficiency quantification model is as follows: .in, The amount of data required for full fine-tuning; This is the full-scale fine-tuning cycle; The amount of data required for scene-adaptive fine-tuning; For scene-adaptive fine-tuning of the cycle; To fine-tune the efficiency improvement factor, the threshold is ≥20. This scenario-adaptive fine-tuning efficiency quantification model can accurately quantify the efficiency advantage of scenario-adaptive fine-tuning compared to full fine-tuning.

[0162] S22. The optimized configuration of the search enhancement generation RAG mechanism improves search accuracy and reasoning efficiency.

[0163] RAG Core Parameter Optimization and Dynamic Adjustment: The dynamic parameter adjustment module links the hierarchical local vectorized knowledge base with the fine-tuned large model, configuring RAG core parameters, including the retrieval similarity threshold and Top-N recall. Based on the confidence level of the retrieval results and the urgency of the event, the module automatically adjusts the similarity threshold and recall. For example, when the confidence level of the first-hop retrieval result is below 0.75, the threshold is automatically lowered to 0.70, and the recall is increased to Top-8 to ensure sufficient relevant knowledge is retrieved. The default retrieval similarity threshold is 0.75, which can be dynamically adjusted; the default Top-N recall is Top-5, but is adjusted to Top-3 in urgent fault scenarios to improve retrieval speed.

[0164] Multi-hop retrieval logic optimization: Pre-set trigger conditions for first-hop retrieval, second-hop supplementary retrieval, and third-hop verification retrieval to form a multi-hop retrieval closed loop of "retrieval-verification-supplementation"; First hop: The query is converted into a vector through the embedding model, and Top-N relevant fragments are retrieved in the local vectorized knowledge base; Second hop: If the confidence of the first-hop result is insufficient, such as similarity <0.75, the model generates a supplementary query based on the first-hop content, such as "the impact of RMCR network disconnection on FAM mode", triggering a second retrieval to extend the causal reasoning chain; Third hop: The results of the first-hop and second-hop retrievals are merged, and the relevance of the retrieval results is verified through the knowledge graph, and irrelevant fragments are removed to ensure the accuracy of the retrieved knowledge; At the same time, a retrieval caching mechanism is adopted to cache high-frequency retrieval queries and results, shortening the retrieval time. The cache validity period can be dynamically set, with a default of 24 hours.

[0165] A multi-hop retrieval inference accuracy quantification model is constructed to accurately evaluate multi-hop retrieval. The multi-hop retrieval inference accuracy quantification model is as follows: ,in, To improve the accuracy of multi-hop retrieval; The accuracy of reasoning based on retrieved knowledge is denoted by a value of 0-1 and is obtained by comparing the reasoning results with the expert handling plan. To achieve a comprehensive accuracy of multi-hop retrieval and reasoning, with a threshold of ≥0.8, this multi-hop retrieval and reasoning accuracy quantification model achieves joint quantification of multi-hop retrieval and reasoning performance, which differs from the limitations of existing technologies that only quantify retrieval or only quantify reasoning.

[0166] Search result fusion optimization: A search result fusion module is set up, which adopts a weighted fusion algorithm. According to the hierarchical weight of the knowledge base, the case layer has a weight of 0.6, the basic layer has a weight of 0.3, and the real-time layer has a weight of 0.1. Multi-hop search results are weighted and fused to highlight the role of high-value knowledge, such as historical failure cases, and provide more accurate contextual support for subsequent reasoning, thus solving the defects of messy search results and poor relevance in existing technologies.

[0167] S23. Optimization of structured decision generation process.

[0168] (1) Dynamic Adaptation of JSON Schema Templates: Multiple JSON Schema templates are preset, and the corresponding template is automatically matched according to the task type and target system type, defining the field constraints of the instruction output; for example, the fault handling instruction template includes fields such as fault ID, handling steps, execution priority, and number of retries, while the status monitoring instruction template includes fields such as monitoring indicators, threshold range, and alarm triggering conditions; a dynamic template update interface is set up to support adding and modifying templates according to business needs without modifying the model code, thus improving system scalability. Task types include fault handling, status monitoring, and inspection. Field constraints include action / params fields, execution time limits, and fault tolerance parameters.

[0169] The three-step process for structured decision generation has been solidified and optimized: The solidified process consists of three steps: semantic parsing and query construction → multi-hop knowledge retrieval → enhanced reasoning and structured output. The execution logic and input / output formats for each step are clearly defined, and optimization mechanisms have been added, including semantic parsing optimization, enhanced reasoning optimization, and exception handling.

[0170] Semantic parsing optimization: Named Entity Recognition (NER) and Relationship Extraction (RE) technologies are adopted to extract key entities from the original event, including device ID, fault code, timestamp, and fault level. At the same time, the relationships between entities are extracted, such as the association between "device ID=XXX" and "fault code=XXX", to generate accurate semantic search queries, which solves the defects of inaccurate semantic parsing and unreasonable query generation in the existing technology; the original event includes alarm JSON.

[0171] Enhanced reasoning optimization: The original events, multi-hop recalled knowledge fragments, preset JSON Schema templates, and knowledge graph relationships are combined to construct prompt words, which are then input into the fine-tuning model. After the model undergoes domain knowledge fusion reasoning, it strictly follows the Schema to output machine-parseable structured instructions. At the same time, a new interpretability module for the reasoning process is added, which outputs the reasoning logic chain, such as "due to the loss of FAM pattern → the RMCR network is found to be abnormal, which will cause the loss of FAM pattern → therefore, the instruction to check the RMCR network + reset AOM is generated", which meets the interpretability requirements of industrial systems.

[0172] Abnormal event handling: For unidentified fault events, such as unknown fault codes and abnormal event formats, an abnormal handling branch is set up to automatically generate an "emergency alarm + manual intervention" instruction. At the same time, the abnormal event and related data are stored in the training sample pool for subsequent incremental fine-tuning of the model, which solves the deficiency of the existing technology in handling unknown faults.

[0173] S24. Decision kernel testing and intelligent optimization: Conduct multi-dimensional testing and adaptive optimization.

[0174] Multi-dimensional testing: Input typical raw event data of the target system, such as alarm JSON and device status data, covering scenarios such as common faults, complex faults, unknown faults, and abnormal data. Test the decision kernel's parsing ability, retrieval accuracy, reasoning logic, structured output effect, and response speed. Construct a test indicator system, including semantic parsing accuracy ≥98%, retrieval accuracy ≥95%, command generation accuracy ≥95%, response speed ≤500ms, and abnormal event handling accuracy ≥90%.

[0175] Intelligent optimization: Through the adaptive optimization module, problems encountered during testing, such as retrieval bias, non-standard instructions, and sloppy reasoning, are automatically adjusted by the module to fine-tune parameters, RAG configuration, or schema templates without manual intervention. For example, if the retrieval bias rate exceeds 5%, the retrieval similarity threshold and knowledge base weights are automatically adjusted; if the instructions are non-standard, the schema template constraints are automatically optimized. At the same time, by setting up a model performance monitoring module, the running status of the decision kernel is monitored in real time. When performance degradation occurs, incremental fine-tuning of the model is automatically triggered to ensure the continuous stability and accuracy of the decision kernel.

[0176] Step S3: Design and Implementation of Standardized Front-End Control Protocol (MCP): Define the MCP protocol to bridge the semantic gap between intelligent decision-making and front-end execution, achieving cross-system front-end adaptation, and distinguishing it from the shortcomings of existing technologies such as "high coupling, poor adaptability, and weak fault tolerance of front-end control protocols." Through protocol extension mechanisms, intelligent fault tolerance, and cross-framework adaptation, the protocol's universality and reliability are improved.

[0177] S31, MCP protocol core definition and extension.

[0178] Lightweight, declarative MCP protocol framework definition: Defines a lightweight, declarative Model Control Protocol (MCP). The protocol uses JSON format for transmission, is small in size, has high transmission efficiency, and is suitable for the low bandwidth requirements of industrial scenarios. It abstracts front-end interactions into a set of atomic, composable operation primitives, including navigate(url) page navigation, click(selector) click operation, fill(selector, value) input operation, assert(condition) assertion judgment, and also includes the following operation primitives.

[0179] capture(selector, type): Screenshot / element screenshot, used for fault review and operation tracing. The type parameter supports full-screen screenshot and specified element screenshot.

[0180] compare(selector, baseline): Element state comparison, used to verify the effect of operation, such as comparing whether the device state changes before and after the operation.

[0181] delay(time): Delay execution, used to adapt to dynamic front-end loading scenarios. The time parameter supports dynamic calculation, such as automatically adjusting the delay time according to the page loading speed.

[0182] call_api(api_url, params): Lightweight API call, used to call the lightweight interface exposed by the front end of the target system without touching the back end, to obtain supplementary data.

[0183] Standardized design of operation primitives: The JSON description format for each operation primitive is clearly defined, binding element location strategies and fault tolerance mechanisms. Simultaneously, an error code system for operation primitives is defined, clarifying the meaning and handling method for each error. For example, error code 1001 indicates element location failure, triggering a retry mechanism; error code 1002 indicates operation timeout, triggering a backup operation sequence. Element location strategies support CSS selectors, XPath, and semantic tags, using device ID location to adapt to the specific characteristics of industrial equipment front-end interfaces. Fault tolerance mechanisms include dynamically configurable retries and adaptively adjusted timeouts based on operation type; the timeout for emergency operations is shortened to 5 seconds, while the timeout for regular operations is set to 10 seconds.

[0184] Deep decoupling of "what to do" and "how to do it": Further strengthening the decoupling principle, the instructions generated by the intelligent agent decision kernel only contain business semantic layer operations, such as {"action": "click", "target": "fault confirmation button", "priority": "high"}, without involving any front-end technical details, such as selectors and frame types; the MCP execution engine is responsible for parsing semantic instructions into specific technical implementations, and adopts an automatic semantic tag mapping mechanism to achieve automatic association between semantic tags and front-end elements without the need for manual configuration of mapping templates, thereby improving adaptation efficiency.

[0185] Protocol Extension Mechanism: An MCP protocol extension interface is designed to support custom operation primitives for different industrial scenarios and target systems without modifying the core protocol framework. For example, for rail transit signaling systems, the "signal_check(channel, params)" operation primitive is set for signal channel status checking; for power systems, the "power_monitor(device_id, index)" operation primitive is set for power equipment parameter monitoring, addressing the shortcomings of existing protocols in poor extensibility and inability to adapt to multiple scenarios. An MCP protocol extensibility quantification model is constructed for protocol extensibility evaluation. The protocol extensibility quantification model is as follows: .in: The number of custom operation primitives that can be added; The number of target system types that can be adapted; The number of basic operation primitives in the protocol; This is a protocol scalability metric with a threshold of ≥5, which quantifies the protocol's scalability using a protocol scalability quantification model.

[0186] S32 and MCP mapping template intelligent configuration.

[0187] Fine-grained configuration of mapping templates: For the target system's front-end technology stack, configure JSON-formatted mapping templates to map business semantic tags to specific front-end technology implementations, such as {"selector": ".btn-confirm", "framework": "Vue", "retry": 2, "timeout": 5000}. A layered template design is adopted, divided into basic templates (general operation semantic mappings), scenario templates (specific scenario operation semantic mappings), and version templates (semantic mappings for different system versions), improving template adaptability. Business semantic tags include "fault confirmation button" and "device status query box". The front-end technology stack includes Vue, React, and jQuery.

[0188] Multi-system adaptation and automatic mapping: Reserving interfaces for multi-system adaptation, setting up independently configurable and on-demand loading mapping templates for target systems with different technology stacks and UI structures, ensuring that the same set of semantic instructions can be reused across systems; setting up an automatic mapping template generation module, collecting front-end UI elements of the target system through web crawling technology, automatically generating the mapping relationship between semantic tags and front-end elements, generating an initial mapping template, which can be used by humans with only a small amount of verification, reducing template configuration costs.

[0189] Dynamic updating of mapping templates: A dynamic updating mechanism for mapping templates is adopted. Through the front-end element change monitoring module, changes to the front-end UI elements of the target system are monitored in real time, such as modifications to element selectors and changes to button names. When changes are detected, the mapping template is automatically updated, and an alarm is sent to the operation and maintenance personnel for verification. In addition, template version management is supported, and historical templates can be traced back to avoid operation failures caused by template update errors. This solves the defects of existing technologies where mapping templates cannot adapt to front-end changes and have high maintenance costs.

[0190] Construct a quantitative model for mapping template adaptation accuracy, and evaluate the mapping template adaptation accuracy. The quantitative model for mapping template adaptation accuracy is as follows: ,in: This represents the number of successful mapping operations. This represents the total number of operations. This is the template update latency rate, calculated as the template update latency time divided by the UI element change time, with a value between 0 and 1. To achieve a mapping template adaptation accuracy of ≥0.98, this mapping template adaptation accuracy quantification model comprehensively quantifies both mapping success rate and update timeliness, overcoming the limitation of existing technologies that only focus on mapping success.

[0191] S33 and MCP execution engine development, testing, and optimization.

[0192] Modular Development of MCP Execution Engine: Develop the MCP execution engine, which includes five core modules.

[0193] MCP parser: Parses JSON instruction sequences and converts them into operation objects with context. It supports batch parsing of instructions and priority sorting, with urgent instructions being executed first.

[0194] Element locator: Based on the positioning strategy specified in the instruction, such as CSS / XPath / semantic tags / device ID, combined with intelligent waiting and retry mechanisms, such as polling to detect element visibility, handling dynamically loaded components, and automatically switching positioning methods when element positioning fails, it accurately locates the target DOM node.

[0195] Action executor: Simulates user interaction behaviors, including clicking, input, scrolling, and screenshotting, and adapts event triggering logic for the features of frameworks such as Vue / React, such as manually triggering input events to update reactive data, and for special interactions of industrial equipment front-ends, such as keyboard shortcut operations.

[0196] Status Feedback Unit: Collects operation results in real time, including success indicators, exceptions, time consumption, snapshots of key element status, and error codes. These are encapsulated into structured feedback packets and sent back to the decision kernel via WebSocket. The feedback latency is ≤100ms.

[0197] Status prediction module: Based on historical operation data and front-end element status data, it uses machine learning algorithms to predict operation results, such as using LSTM to predict whether element positioning will fail or whether the operation will time out, triggering backup operation strategies in advance, improving the success rate of operations, and solving the shortcomings of existing technologies that only take remedial measures after operation failure and are inefficient.

[0198] Intelligent fault tolerance and exception handling optimization: A multi-level fault tolerance mechanism is designed to address common exceptions in front-end operations, such as element location failure, operation timeout, and page crash.

[0199] Level 1 fault tolerance: When element positioning fails, automatically switch the positioning method, such as switching from CSS selector to XPath, and retry a specified number of times, with a default of 3 times.

[0200] Level 2 fault tolerance: If a retry fails, a backup operation sequence is triggered. If the "fault confirmation button" cannot be clicked, a confirmation operation is triggered via a shortcut key.

[0201] Level 3 fault tolerance: If the backup operation sequence fails, an abnormal alarm is automatically generated and sent back to the decision kernel. At the same time, abnormal scenario data is recorded for subsequent model optimization.

[0202] In addition, a page crash recovery mechanism is adopted. When a crash is detected in the front-end page of the target system, the page is automatically refreshed, the WebSocket connection is re-established, and the unfinished operation is resumed to ensure the continuity of operation.

[0203] Construct a quantitative model for the effectiveness of the fault tolerance mechanism to evaluate its effectiveness. The quantitative model for the effectiveness of the fault tolerance mechanism is as follows: ,in: This refers to the number of operations that successfully recovered via the fault-tolerant mechanism. This represents the total number of abnormal operations. Mean time for fault-tolerant recovery; This represents the average recovery time after manual intervention. The threshold value is ≥8, which is used as an indicator of the effectiveness of the fault tolerance mechanism. This fault tolerance mechanism effectiveness quantification model quantifies the recovery capability and efficiency of the fault tolerance mechanism.

[0204] Multi-dimensional testing and performance optimization: Testing the compatibility, execution accuracy, and response speed of the MCP protocol, with single-step operation execution time ≤ 1 second; setting up performance optimization steps to improve execution efficiency through batch execution of instructions and parallel processing of operations, such as executing multiple unrelated operations simultaneously; employing an operation caching mechanism for high-frequency operations to shorten operation execution time; testing fault tolerance capabilities under abnormal scenarios to ensure the timeliness and effectiveness of exception handling, meeting the high reliability requirements of industrial scenarios. The MCP protocol is compatible with different front-end technology stacks and different system versions.

[0205] Step S4: Deployment of non-intrusive plug-in execution architecture: Establish a non-intrusive interaction channel between the intelligent agent and the target system, deploy instruction execution services, and ensure secure isolation and seamless integration.

[0206] S41. Data channel setup.

[0207] (1) Message middleware configuration and optimization: Configure message middleware connection parameters, obtain the target system's Broker address, event publishing Topic, and authentication credentials, and adopt a two-way authentication mechanism (SSL / TLS) to ensure that the agent can securely subscribe to the target system's event stream as a consumer; optimize the message middleware configuration, set message partitioning strategies and message expiration times to avoid message backlog. Message middleware such as Kafka / RabbitMQ.

[0208] (2) Zero-Intrusion Event Publishing Adaptation: Coordinate with the target system to publish its internal events to a designated Topic in the message middleware according to a preset JSON format. This does not require modification of the target system's source code or database permissions, relying only on its existing event publishing capabilities, such as the log push module. For target systems that do not have event publishing capabilities, a lightweight event collection plugin is used. This plugin is embedded into the front end of the target system in a non-intrusive manner, only collecting event data without modifying the system code or accessing the backend database. The data is then encrypted and pushed to the message middleware, solving the defects of some target systems in the existing technology that cannot publish events and cannot achieve zero-intrusion data acquisition. Internal events include alarms and device status changes, etc.

[0209] Data reception, deserialization, and anomaly filtering: The agent subscribes to event streams through a message middleware, receives event data in real time, decrypts the data using the AES-256 encryption algorithm, and then deserializes it into a format recognizable by the decision kernel. A data anomaly filtering module is set up to filter invalid and abnormal data through a rule engine. Abnormal data includes format errors, missing data, and outliers. At the same time, the data is anonymized to hide sensitive information, such as device passwords and maintenance personnel information, to ensure data security and accuracy. In addition, a data caching mechanism is used to cache recently received event data.

[0210] A comprehensive index model for data acquisition security and non-intrusiveness is constructed to conduct a quantitative assessment of the non-intrusiveness and security of data acquisition. The comprehensive index model for data acquisition security and non-intrusiveness is as follows.

[0211] ;in: This is an intrusive indicator, including the number of lines of code modified in the target system / the total number of lines of code in the target system, with a value of 0-1, and 0 for zero intrusion; For encryption security, AES-256 uses a value of 0.95, and the Chinese national standard SM4 uses a value of 0.98. This represents the accuracy of filtering outlier data, with a value between 0 and 1. The data acquisition security and non-invasiveness comprehensive index has a threshold of ≥0.9. This data acquisition security and non-invasiveness comprehensive index model quantifies non-invasiveness and security, breaking through the limitation of existing technologies that cannot quantify non-invasiveness.

[0212] S42. Establish a control channel.

[0213] (1) Lightweight JavaScript script optimization: A lightweight JavaScript script with a size of ≤100KB is injected into the front end of the target system. This script does not affect the running speed of the target system. The script is responsible for establishing a long connection with the WebSocket service of the agent and continuously listening to the connection status. A script self-protection mechanism is adopted to prevent the script from being tampered with or deleted. If the script is detected to be abnormal, it will be automatically re-injected and an alarm will be sent to notify the operation and maintenance personnel. The script only has front-end operation permissions and does not have permissions to access the back-end interface or database of the target system to ensure system security.

[0214] (2) WebSocket Communication Rule Optimization and Encryption: Configure WebSocket communication rules, clarify the transmission format of MCP rendering instructions and MCP execution instruction sequences, and the return format of page status information; adopt the WebSocket encrypted communication - WSS protocol to encrypt the transmitted instructions and status data end-to-end to prevent data theft and tampering; adopt a communication heartbeat mechanism, sending a heartbeat packet every 30 seconds to check the connection status. If a connection is detected to be broken, reconnection is automatically triggered, with the number of reconnections ≤ 5, and the reconnection interval gradually increasing, such as 1 second, 3 seconds, 5 seconds, 10 seconds, and 15 seconds, to ensure the stability of the control channel. MCP rendering instructions dynamically generate "Intelligent Analysis" buttons and "Fault Handling" pop-ups. Page status information includes DOM changes and operation results.

[0215] A connection stability index model is constructed to quantitatively evaluate the stability of the control channel. The connection stability index model is as follows: ,in: The number of connection drops per unit time; Average recovery time for a single disconnection, in seconds; Total running time, in seconds; The connection stability index has a threshold of ≥0.999. The stability of the control channel is precisely quantified through the connection stability index model.

[0216] (3) Control Channel Compatibility and Load Optimization: WebSocket connection adaptation logic is optimized for different browsers and front-end technology stacks to ensure stable operation of the control channel in various environments. A channel load balancing mechanism is adopted to automatically allocate connection resources when multiple agents simultaneously establish connections with the target system, avoiding communication delays caused by excessive load on a single channel. Furthermore, a command priority transmission mechanism is used, prioritizing the transmission of urgent commands to ensure timely execution of emergency tasks. Different browsers include Chrome, Firefox, and Edge. Emergency commands include fault handling commands.

[0217] (4) Control channel testing and monitoring: Test the stability of the control channel, communication delay ≤100ms, and accuracy of command transmission; set up a channel operation status monitoring module to monitor the number of connections, command transmission rate, and number of abnormalities in real time. When an abnormality occurs, such as connection disconnection and command transmission failure, an alarm will be automatically triggered and an abnormality log will be recorded for subsequent troubleshooting and optimization.

[0218] S43, Command Execution Service Deployment and Sandbox Enhancement.

[0219] (1) Deployment and association of instruction execution service: The MCP execution engine is deployed to the intelligent agent runtime environment and associated with the decision kernel and WebSocket service. A microservice architecture is adopted to realize the independent deployment and flexible expansion of each module. A new service health check mechanism is added to monitor the running status of each module in real time. If a module fails, a backup module is automatically started to ensure the continuity of instruction execution service. The MCP execution engine includes a parser, locator, executor, status feedback unit and status prediction module.

[0220] (2) Enhanced Automated Execution Sandbox: Configure an enhanced automated execution sandbox, adopting a layered isolation mechanism, process isolation, permission isolation, and data isolation, limiting the scope of instruction execution, and ensuring that all operations are completed at the browser front-end level without touching the target system's back-end logic and database; the sandbox has a built-in security rule engine to prohibit dangerous operations, such as modifying front-end code, accessing sensitive interfaces, and downloading malicious files. If a dangerous operation is detected, execution will be terminated immediately and an alarm will be triggered; a sandbox logging module is set up to record all operation trajectories within the sandbox for security auditing and fault tracing, meeting the stringent security isolation requirements of industrial systems, such as the SIL4 safety level of rail transit and the safety specifications of the power industry.

[0221] Construct a quantitative model for the security isolation level of the sandbox, and evaluate the security isolation level of the sandbox. The quantitative model for the security isolation level of the sandbox is as follows: ,in: , , These are the weighting coefficients; This represents the process isolation level, with a value between 0 and 1. This represents the access control isolation level, with a value between 0 and 1. The data isolation level, with a value between 0 and 1; The sandbox safety isolation level has a threshold of ≥0.95. This sandbox safety isolation level quantification model quantifies the sandbox isolation effect to meet industrial safety level requirements.

[0222] Status Feedback and Anomaly Isolation: A status feedback module is deployed to collect the status of each operation in real time, including success / failure / anomaly details, operation time, and element status snapshots. This data is encapsulated into a structured feedback package and sent back to the decision kernel via WebSocket. An anomaly isolation mechanism is employed. When an operation anomaly is detected, such as triggering sandbox security rules or excessive operation failures, the operation sequence is automatically isolated without affecting the execution of other operations. The anomaly data is also stored in an audit database for subsequent analysis and optimization. An operation permission control module is set up to restrict the scope of the intelligent agent's operations based on the permission level of the operations personnel. For example, ordinary operations personnel are only allowed to perform routine operations, while administrators are allowed to perform advanced operations, thus improving system security.

[0223] Step S5, Dynamic Task Orchestration, Closed-Loop Feedback and System Integration Optimization: Achieve complex task flow orchestration, construct a continuous learning closed loop of execution-feedback-optimization, complete system integration and iterative optimization, and solve the defects of existing technologies such as "simple task orchestration, imperfect feedback mechanism, and reliance on manual optimization".

[0224] S51, intelligent configuration of dynamic task orchestration.

[0225] (1) Complex task flow orchestration and conditional branch optimization: Configure conditional branches, loop logic and parallel execution logic of MCP instruction sequence to support dynamic adjustment of multi-step and complex task flows; for example, when an abnormal pop-up is detected, it automatically jumps to the pop-up closing branch, and continues to execute the original task after closing; supports task flow nesting, such as nesting device status detection tasks in fault handling task flow, to improve the flexibility of task orchestration; adopts task flow priority management, sets task flow priority according to the urgency and importance of the task, such as high, medium and low, and the agent executes the task according to the priority to ensure that urgent tasks are handled first.

[0226] (2) Intelligent generation and adaptation of task flow templates: Multiple task flow templates are preset, such as fault alarm handling templates, equipment inspection templates and predictive maintenance templates, and the corresponding templates are automatically matched for different operation and maintenance scenarios; a task flow template intelligent generation module is set up to automatically generate task flow templates for new scenarios based on historical operation trajectories and expert experience, and only minor adjustments are required for manual use; at the same time, a template adaptive adjustment mechanism is adopted to automatically adjust the operation steps in the task flow template according to the dynamic changes of the target system, such as front-end UI updates and operation and maintenance process changes, to ensure the adaptability of the template.

[0227] (3) Task flow execution monitoring and dynamic adjustment: The task flow execution monitoring module monitors the execution progress and status of each step of the task flow in real time and displays the execution trajectory of the task flow in a visual way. When an abnormality is detected in the task flow execution, such as the failure or timeout of a certain step, the dynamic adjustment mechanism is automatically triggered. According to the type of abnormality, the backup operation step is switched, the execution order is adjusted, or the task flow is paused and manual intervention is triggered to ensure the smooth execution of the task flow. For example, when the equipment status detection step fails, the detection method is automatically switched. If it still fails, the task flow is paused and the operation and maintenance personnel are notified.

[0228] Construct a task flow execution efficiency model and evaluate its efficiency. The task flow execution efficiency model is as follows: ,in: The number of task flows that were successfully completed; This represents the total number of task flows; This refers to the standard task flow execution time; This refers to the actual execution time of the task flow; To determine the task flow execution efficiency, a threshold of ≥80% is set, and the completion rate and execution efficiency of the task flow are quantified through a task flow execution efficiency model.

[0229] S52. Enhanced closed-loop feedback mechanism: Achieve a fully autonomous closed loop of execution-feedback-optimization.

[0230] (1) Real-time feedback and immediate correction optimization: Configure real-time feedback logic to transmit the status feedback of the front-end operation back to the decision kernel in real time for immediate correction of the current task flow; add a feedback data analysis module to analyze the feedback data in real time, identify the reasons for operation failure, such as element location failure, non-standard instructions, and front-end page abnormalities, and automatically trigger preset correction strategies, such as adjusting the MCP mapping template, optimizing RAG retrieval parameters, and switching operation modes, without manual intervention, solving the defects of the existing technology where feedback and correction are disconnected and rely on manual intervention. The status feedback of the front-end operation includes success / failure / abnormality, error code, and operation time.

[0231] (2) Enhanced Audit and Traceability Mechanism: An audit and traceability module is deployed to encrypt and store the complete operation trajectory in the audit database. Blockchain technology is used to ensure that the operation trajectory is tamper-proof, supporting full-link traceability and compliance review. Through the automatic audit report generation module, operation and maintenance audit reports are generated regularly, including daily, weekly, and monthly reports, displaying indicators such as the intelligent agent's operating status, operation success rate, and fault handling efficiency, providing data support for operation and maintenance management. At the same time, it supports searching audit data by device, fault type, time range, and other dimensions, facilitating fault review and accountability. The complete operation trajectory includes instruction sequence, execution result, environmental context, feedback data, and sandbox logs.

[0232] An audit traceability reliability index model is constructed to conduct a quantitative assessment of audit traceability reliability. The audit traceability reliability index model is as follows: ,in: To determine the accuracy of the traceability, a value between 0 and 1 is used. The data tampering rate is approximately 0 under blockchain technology. The reliability index for audit traceability is set at a threshold of ≥0.99. The reliability of audit traceability is quantified through the audit traceability reliability index model, highlighting the application value of blockchain technology.

[0233] Continuous Learning and Autonomous Optimization Mechanism: A fully autonomous continuous learning closed loop is constructed, automatically filtering high-frequency failure cases, unknown fault cases, inefficient operation cases, and their associated contexts, including original events, generated instructions, page structure fragments, and feedback data. After desensitization and annotation, these are incorporated into the training sample pool. An automatic sample annotation module is set up to improve sample annotation efficiency through model-assisted annotation and manual verification. Incremental fine-tuning of the vertical domain large model and dynamic updates of the RAG knowledge base are initiated regularly. At the same time, the MCP mapping template and task flow template are optimized, enabling the decision kernel to continuously optimize the accuracy and robustness of instruction generation, and the MCP execution engine to continuously improve the success rate of operations. This achieves the autonomous evolution of the agent and solves the defects of existing technologies where agents cannot autonomously optimize and performance is difficult to continuously improve.

[0234] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model, characterized in that, Includes the following steps: Preliminary preparation and basic environment setup: Conduct target system research, data preparation, and deployment of the basic operating environment for intelligent agents; Constructing a domain-adaptive intelligent decision-making kernel: Building a large model decision center finely tuned by the domain to achieve semantic parsing, multi-hop reasoning, and structured instruction output for industrial events; Design and implementation of standardized front-end control protocol MCP: Define the MCP protocol to eliminate the semantic gap between intelligent decision-making and front-end execution, and perform cross-system front-end adaptation; Non-intrusive plug-in execution architecture deployment: Establish a non-intrusive interaction channel between the intelligent agent and the target system, deploy instruction execution services, and achieve secure isolation and seamless integration between the intelligent agent and the target system; Dynamic task orchestration, closed-loop feedback and system integration optimization: Perform complex task flow orchestration, build a continuous learning closed loop of execution-feedback-optimization, and complete system integration and iterative optimization; The deployment of the basic operating environment for intelligent agents includes the following steps: Comprehensive collection of data from multiple sources and domains; Data cleaning and enhancement: Remove invalid, duplicate, and sensitive data; perform word segmentation, noise reduction, and standardization on the text. Intelligent construction and layering of local vectorized knowledge base: The cleaned text data is encoded and converted into computer-recognizable vectors through a bidirectional encoder-representation converter. In the intelligent construction and hierarchical steps of the local vectorized knowledge base: A knowledge base retrieval quantification model is constructed, which is as follows: ;in: To retrieve the number of knowledge fragments; The relevance between the i-th knowledge fragment and the query; The weight of the knowledge base level where the i-th knowledge fragment is located; To improve search accuracy; Let i be the knowledge fragment retrieved by the retrieval.

2. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 1, characterized in that, Building a domain-adaptive intelligent decision-making kernel includes the following steps: Deep fine-tuning of large vertical models; The search enhancement generation mechanism has been optimized and configured, including parameter optimization and dynamic adjustment, multi-hop search logic optimization, and search result fusion optimization. Optimization of the structured decision generation process, including semantic parsing optimization, enhanced reasoning optimization, and abnormal event handling; Decision kernel testing and intelligent optimization: Conduct multi-dimensional testing and adaptive optimization.

3. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 1, characterized in that, Cross-system front-end adaptation includes the following steps: MCP protocol core definition and extensions; Intelligent configuration of MCP mapping templates; MCP execution engine development, testing, and optimization.

4. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 3, characterized in that, In the intelligent configuration step of MCP mapping template, a dynamic update mechanism for mapping template is adopted; at the same time, a quantitative model of mapping template adaptation accuracy is constructed to evaluate the mapping template adaptation accuracy and comprehensively judge the adaptation effect of mapping template.

5. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 2, characterized in that, In the vertical domain large model deep fine-tuning step, a scene adaptive fine-tuning efficiency quantification model is constructed to quantify the scene adaptive fine-tuning efficiency. The scene adaptive fine-tuning efficiency quantification model is as follows: ,in, These are the weighting coefficients; Generate accuracy for the instructions; This refers to the average instruction generation time. To fine-tune the performance evaluation indicators.

6. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 1, characterized in that, Deploying a non-intrusive add-on execution architecture involves the following steps: Data channel setup: including message middleware configuration and optimization, and zero-intrusion event publishing adaptation; Building a control channel includes optimizing lightweight JavaScript, optimizing WebSocket communication rules, optimizing encryption and control channel compatibility and load, and testing and monitoring the control channel. Command execution service deployment and sandbox enhancement: This includes command execution service deployment and association, as well as automated execution sandbox enhancement.

7. The non-intrusive industrial intelligent agent construction method based on MCP and vertical domain large model according to claim 1, characterized in that, System integration and iterative optimization include the following steps: Intelligent configuration of dynamic task orchestration: including complex task flow orchestration and conditional branch optimization, intelligent generation and adaptation of task flow templates, and monitoring and dynamic adjustment of task flow execution; Strengthened closed-loop feedback mechanism: including real-time feedback and immediate correction optimization, and enhanced audit and traceability mechanism.

8. A non-intrusive industrial intelligent agent system based on MCP and a vertical domain large model, employing the method described in any one of claims 1-7, comprising: The system comprises a basic environment setup system, a dynamic task orchestration and closed-loop optimization system, a domain-adaptive intelligent decision-making kernel system, a standardized front-end control protocol system, and a non-intrusive plug-in execution architecture system, characterized by: Basic environment setup system: Completes the deployment of the basic system environment, data preparation and security pre-configuration, including: target system adaptation module, domain data and knowledge base module and basic operating environment module; Domain-Adaptive Intelligent Decision-Making Kernel System: Constructs a domain-fine-tuned large model decision center to realize semantic parsing, multi-hop reasoning, and structured instruction output for industrial events, including: a vertical domain large model fine-tuning module, a decision kernel testing and optimization module, a RAG retrieval enhancement generation module, and a structured decision generation module; Standardized front-end control protocol system: It realizes the definition of MCP protocol, mapping template configuration and execution engine development, bridging the semantic gap between intelligent decision-making and front-end execution, including protocol definition extension module, mapping template configuration module and execution engine module; Non-intrusive plug-in execution architecture system: Builds a non-intrusive interaction channel between intelligent agents and target systems, deploys command execution services, and strengthens security isolation, including: command execution service and sandbox isolation module, non-intrusive data channel module and non-intrusive control channel module; Dynamic Task Orchestration and Closed-Loop Optimization System: Enables complex task flow orchestration, constructs a fully autonomous closed loop of execution-feedback-optimization, and completes system integration and optimization, including: dynamic task orchestration module, closed-loop feedback autonomous optimization module, and system integration and iteration optimization module.

Images