Intelligent operation and maintenance double-track transition system and instruction self-evolution method for small and medium-sized enterprises

Abstract

This invention relates to a dual-track transition system and instruction self-evolution method for intelligent operation and maintenance (O&M) in small and medium-sized enterprises (SMEs), belonging to the field of industrial IoT intelligent O&M technology. The system includes: seamless access to legacy O&M platforms via an API gateway and IoT message queue, providing data source semantic query services and supporting the discovery of callable interfaces using natural language; an optional multimodal perception layer generating structured semantic events; a human-machine collaborative interaction platform comprising five windows; instructions generated manually or by a lightweight decision-making model based on multi-event, multi-indicator context, with effectiveness feedback and context jointly constituting training samples for optimizing the decision-making model; automatic triggering of replay record generation at key nodes, supporting explicit user activation of events to capture high-value scenarios, achieving sample correction and instruction strategy self-evolution. This invention is suitable for the gradual intelligent upgrade of resource-constrained SMEs, solving the problem that SMEs cannot deploy advanced AI model-based intelligent operation and maintenance systems.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of industrial Internet of Things (IoT), and involves IoT operation and maintenance and artificial intelligence application technology. Specifically, it relates to an intelligent operation and maintenance dual-track transition system and instruction self-evolution method suitable for resource-constrained small and medium-sized enterprises. Background Technology

[0002] Operational maintenance of enterprise facilities and application software systems is an essential technical means for enterprises to ensure the safe and stable operation of their businesses. Current operation and maintenance management primarily relies on Computerized Maintenance Management Systems (CMMS) and Supervisory Control and Data Acquisition (SCADA) systems. CMMS helps enterprises / organizations automate and improve their core maintenance operations, record activities, and optimize workflows. SCADA is a computer-based automated production process control and scheduling system that can monitor and control operating equipment on-site. Operational maintenance operations are typically stored in a work order database. Each work order typically records detailed information such as work order number, creation time, work order status, priority, assigned personnel, and solutions.

[0003] Currently, operations and maintenance (O&M) scenarios face core challenges such as data fragmentation, delayed fault location, frequent false positives and false negatives, and difficulty in reusing expert knowledge. For enterprise system O&M lacking sufficient funding and technical team support, the technical means become outdated as the enterprise's equipment and technology are updated after system establishment. Multiple platforms operate independently, leading to data fragmentation. The independent operation of each system, requiring collaboration among multiple personnel and platforms, further delays fault location. Differences in platform integration, varying personnel skills, and differing understandings of the connections and potential impacts between events result in frequent false positives and false negatives. The degree of standardization of O&M procedures and incident handling plans within an enterprise also affects operational flexibility and the possibility of omissions. While personnel involved in project construction and acceptance often possess the most comprehensive system knowledge, with standardization and personnel changes, without a robust knowledge system, expert knowledge often gradually disappears during organizational operations, leaving the organization reactive in emergency responses.

[0004] Intelligent operation and maintenance (O&M) systems based on large models have been applied in multiple industries. Through RAG retrieval-enhanced generation, multi-agent collaboration, and large language models (LLM), massive amounts of messy logs are transformed into structured knowledge, enabling cross-dimensional logical reasoning. However, while the multi-reward native multimodal reasoning models currently used in intelligent O&M systems possess native cross-modal fusion and multi-dimensional reward optimization capabilities, their effective deployment heavily relies on large-scale, high-quality, multimodal, and labeled O&M datasets. However, most SMEs lack both the technical capabilities to build such datasets and the financial and operational costs of replacing outdated CMMS / SCADA platforms, hindering the widespread adoption of advanced AI technologies.

[0005] Small and medium-sized enterprises (SMEs) are unable to deploy advanced AI model-based intelligent operation and maintenance systems due to data scarcity, outdated platforms, and limited resources. Specifically, the following problems exist: (1) If the existing operation and maintenance management system is completely replaced by an intelligent operation and maintenance system based on a large model, there will be high costs, long cycles, and high risks of business interruption; (2) The old system interface has no business semantics, chaotic naming, and missing documentation, making it difficult for non-technical personnel and AI to understand its purpose; (3) AI agents cannot discover data sources independently, requiring a large amount of hard coding for adaptation, making it difficult to generalize; (4) Most systems confuse "information display" with "action recommendation," mistaking document click behavior as a model optimization signal; (5) There is a lack of structured modeling and feedback mechanisms for "executable operations" such as querying, subscribing, and controlling; (6) There is a lack of collaborative architecture between advanced multimodal perception capabilities and lightweight decision execution; (7) The existing feedback mechanism is crude, unable to trace the operation context, and difficult to correct noise labels. Therefore, there is an urgent need for a transitional architecture that is low-threshold, supports semantic interfaces, focuses on action recommendation, can generate high-quality training data on its own, is compatible with multimodal perception evolution, and has the ability to replay operations and correct samples, in order to solve the problem that small and medium-sized enterprises cannot deploy AI model-based intelligent operation and maintenance systems. Summary of the Invention

[0006] The purpose of this invention is to provide a dual-track transition system for intelligent operation and maintenance for small and medium-sized enterprises (SMEs) and a self-evolving instruction method, which solves the problem that SMEs are unable to deploy advanced AI model-based intelligent operation and maintenance systems due to data scarcity, outdated platforms, and limited resources.

[0007] The dual-track transition system for intelligent operation and maintenance for small and medium-sized enterprises provided by this invention includes: a multimodal perception layer, an old platform interface layer, a human-machine collaborative interaction platform, a lightweight decision model, and an operation playback and sample correction module.

[0008] The legacy platform interface layer includes an API gateway, an IoT message queue, and a data source semantic query service. The system accesses the enterprise's current operations and maintenance management system, monitoring system, and work order database through the API gateway to obtain operations and maintenance data. The system subscribes to device sensor data, events, and control commands in real time through the IoT message queue. Events generated by the device monitoring system are injected into the IoT message queue. The data source semantic query service semantically transforms the interfaces of the enterprise's operations and maintenance management system and monitoring system through a registration-indexing-query process. This includes: receiving registered API interface information or message topic information, obtaining the business semantic description of the API interface, notifying the API interface or message channel of successful registration, and storing the API interface information or message topic information and the corresponding business semantic description in a semantic index library. The API interface information or message topic information includes the interface URL, request method, parameter definition, return format, and authentication method. The semantic index library supports semantic queries based on natural language or keywords, returning a list of relevant API interfaces and message topics.

[0009] The human-machine collaborative interaction platform includes an IoT event window, an information window, an emergency response plan window, an instruction input window, and an instruction execution and feedback window. The IoT event window displays real-time events in a scrolling fashion, with alarm-level events pinned to the top. Clicking on an event activates associated operations. The information window dynamically displays event query results and execution feedback for manually confirmed instructions, and it interacts bidirectionally with the IoT event window. The emergency response plan window displays structured technical documents related to alarm events to assist human decision-making. The instruction input window receives executable operation instructions, which are either manually input or generated by a lightweight decision-making model. Operation instructions include at least one of the following: querying statistical indicators, subscribing to device messages, and remotely controlling devices. The instruction execution and feedback window displays confirmed executed instructions and their execution status, interacts with the information window, and mandates user feedback on the validity of instructions. It constructs a triplet training sample of decision context-executed instruction-validity feedback to optimize the lightweight decision-making model. The decision context includes the user-selected event and statistical indicators from real-time or historical reports.

[0010] The lightweight decision model generates executable operation instructions based on the decision context. When generating instructions, it calls the data source semantic query service to query the API interfaces and / or message topics that need to be called.

[0011] When the operation replay and sample correction module is triggered, it generates a replay record; the replay record is a structured record containing the triggering event, decision context, execution instructions, feedback results and related timestamps.

[0012] The multimodal perception layer is an optional module. It is initially disabled and re-enabled after acquiring more than a set amount of multimodal operation and maintenance data. The multimodal perception layer deploys a large-scale native multimodal inference model with multiple rewards. This model receives raw multimodal data, including device infrared images, visible light images, acoustic recordings, and sensor time series, and outputs structured semantic events. The multimodal perception layer then injects these structured semantic events into the IoT message queue.

[0013] The instruction self-evolution method based on the present invention for a dual-track intelligent operation and maintenance system for small and medium-sized enterprises includes:

[0014] (1) Receive one or more events, associate indicators with users, and form a decision context consisting of the events and indicators; based on the context, generate executable operation instructions by a lightweight decision model or manually input them; record the actual execution results of the instructions and the effectiveness feedback provided by the user; generate formatted samples;

[0015] The formatted sample is expressed as {context:{events[],metrics[]},action,reward,timestamp}, where context is the decision context, events[] is an array of events, metrics[] is an array of metrics, action is the executed instruction, reward is the digitized instruction feedback result, and timestamp is the time stamp;

[0016] (2) Regularly review the playback records manually, correct the instruction feedback results, and regularly export formatted samples of instruction validity feedback;

[0017] (3) When a user manually marks a high-value scene on the human-computer collaborative interaction platform, a replay record is generated, the event type is the scene the user is interested in, the state of the five windows of the human-computer collaborative interaction platform at that time is saved, and a formatted sample is generated.

[0018] (4) Regularly obtain formatted samples and use the triplet of decision context-execution instruction-validity feedback to train or optimize the lightweight decision model.

[0019] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0020] (1) The system of the present invention transforms the technical interface into business language through the data source semantic query service, realizes the semanticization of the interface, and enables non-technical personnel and AI to discover and use the interface data of the old system independently.

[0021] (2) The system of the present invention supports AI autonomous connection, and the lightweight decision model can dynamically query semantic services to realize automatic mapping of “instruction → API” without hard coding; it has a layered decoupled architecture, and the multi-reward native multimodal reasoning large model used focuses on perception, while the lightweight decision model focuses on decision-making and each performs its own function.

[0022] (3) The system of the present invention has a flexible evolution path. Small and medium-sized enterprises can first use the traditional event + lightweight decision model, and then add multimodal perception in the later stage for smooth upgrade; it has the advantages of low cost and high compatibility, adapts to the heterogeneous old systems of small and medium-sized enterprises, and can ensure business continuity.

[0023] (4) The system and instruction self-evolution method of the present invention focus on action recommendation, and the feedback mechanism directly serves to improve the intelligent recommendation capability from "decision context to executable instruction"; based on real decision modeling, it supports decision context with multiple events and multiple indicators, and can accurately reflect the complex fault handling logic in industrial sites; it has the advantage of high-quality sample governance, and through the playback and correction mechanism, it filters out noise labels and improves the signal-to-noise ratio of training data.

[0024] (5) The system and instruction self-evolution method of the present invention are traceable throughout the entire chain, meet the requirements of industrial audit and compliance, and enhance the credibility of intelligent operation and maintenance system; support active scene capture, operation and maintenance experts can manually mark high-value or complex interaction scenarios, make up for the blind spots of automatic triggering, and improve sample diversity and teaching value. Attached Figure Description

[0025] Figure 1 This is an overall architecture diagram of the dual-track intelligent operation and maintenance transition system for small and medium-sized enterprises according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart of the data source semantic query service workflow according to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of human-computer collaborative five-window interaction and playback triggering according to an embodiment of the present invention;

[0028] Figure 4 This is a flowchart illustrating the automated workflow from playback recording to training samples in an embodiment of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] This invention implements a dual-track intelligent operation and maintenance transition system suitable for resource-constrained SMEs. It achieves seamless access to legacy operation and maintenance platforms through an API gateway and IoT message queue, and provides data source semantic query services to lower the barrier to entry. It can optionally integrate a multi-reward native multimodal reasoning model as the underlying perception engine to generate high-order semantic events. It combines a rule engine and Retrieval Enhancement Generation (RAG) to provide knowledge assistance. Through a human-computer collaborative interaction interface, it collects structured samples of "multi-event + multi-indicator context → executable instructions → validity feedback" for training a lightweight action recommendation model. Simultaneously, it supports the automatic generation of timestamped "replay records" at key interaction nodes for post-event review, sample correction, and high-quality data governance, supporting the gradual implementation of intelligent operation and maintenance capabilities.

[0031] The main technical ideas of this invention to realize a dual-track intelligent operation and maintenance system for small and medium-sized enterprises include:

[0032] (1) Achieve dual-track parallelism + semantic interface: Connect the old platform with the IoT message queue through the API gateway and provide data source semantic query service;

[0033] (2) Implement optional multimodal perception layer: Support the deployment of multi-reward native multimodal reasoning large models, which are only used as the underlying event generator;

[0034] (3) Achieve knowledge-action decoupling: RAG is only used to display technical documents, does not generate instructions, and does not participate in training;

[0035] (4) Implement a lightweight decision-making model focused on action recommendation: recommend executable instructions based on multi-event + multi-index context;

[0036] (5) Implement feedback-driven strategy evolution: User feedback on the effectiveness of actual execution instructions is used to optimize the lightweight decision model;

[0037] (6) Implement operation replay and sample correction mechanism: The system automatically triggers the generation of replay records at key interaction nodes, and can optionally support users to explicitly activate a window event as an additional trigger source; the replay records can be used for post-event review, sample quality verification or manual correction to ensure the accuracy of training data and business alignment.

[0038] like Figure 1 As shown, the dual-track transition system for intelligent operation and maintenance for small and medium-sized enterprises implemented in this embodiment of the invention mainly includes an optional multimodal perception layer, an old platform interface layer, a human-machine collaborative interaction platform, a lightweight decision model, and an operation playback and sample correction module.

[0039] The multimodal perception layer is optional and is used to run a large-scale native multimodal inference model with multiple rewards, transforming raw multimodal data into structured semantic events. This model receives raw multimodal data such as infrared images, visible light images, voiceprint recordings, and sensor time series data from devices, and outputs structured semantic events, such as the event "abnormal temperature rise in the bearing area of ​​water pump B," which is then injected into a unified IoT message queue. This model does not directly generate operational instructions; it serves only as a high-quality event source for upper-layer decision-making.

[0040] The primary function of the multimodal perception layer is to improve the quality of input data, thereby reducing the complexity of lightweight decision-making models and improving the accuracy and usability of the system. For example, after data is semantically encoded, the amount of training data for lightweight decision-making models is significantly compressed. Semantizing IoT data through the multimodal perception layer also allows operations and maintenance personnel to quickly grasp the status of IoT systems. The multimodal perception layer can also semantically and contextually encode data and associate relevant parameters according to user needs. However, the effective deployment of large-scale, multimodal inference models with multiple rewards heavily relies on large-scale, high-quality, multimodal, and labeled operations and maintenance datasets. Various system operations and maintenance (IoT deployments) typically lack universally applicable models that can be directly used. Small and medium-sized enterprises (SMEs) may initially lack large-scale, high-quality operations and maintenance data, making it difficult to effectively utilize large-scale inference models. Therefore, for SMEs, this layer can be initially disabled, and they can continue to use traditional monitoring systems to inject events generated by monitoring data into the IoT message queue. After obtaining a certain amount of high-quality operations and maintenance data, a large-scale multimodal perception inference model can then be added to achieve a smooth upgrade.

[0041] The legacy platform interface layer includes an API gateway, an IoT message queue, and a data source semantic query service. It connects to existing maintenance management systems (CMMS), supervisory control and monitoring (SCADA) systems, and work order databases through a standardized API gateway to obtain multimodal maintenance data. The IoT message queue subscribes to device sensor data, alarm events, and control commands in real time. IoT messaging systems include lightweight IoT communication protocols such as MQTT and high-throughput stream processing platforms such as Kafka. The data source semantic query service receives registered interface information and associates it with business semantic descriptions, supporting queries using natural language or keywords to return callable data source addresses and parameters.

[0042] Events generated by the device monitoring system and / or the multimodal sensing layer are injected as messages into the IoT message queue. Message topics are divided into system-level messages and alarm event-level messages. System-level messages include event channels and alarm channels. Alarm event-level messages establish an independent channel for each alarm event, such as opening a message channel for an abnormal event of excessive current sensor usage. Control commands and feedback results from each subsystem are broadcast through message channels. Taking the event log service as an example, by listening to messages on the corresponding channel, the monitored data is recorded to a file, forming a complete log stream for that event.

[0043] Enterprise administrators register the API interfaces and message topics related to the operation and maintenance management system, monitoring system, and work order database into the system of this invention through the data source semantic query service. The data source semantic query service implements the following functions: a) receiving registered API interfaces or message queue topic information, including technical metadata such as URL, request method, parameter definition, return format, and authentication method; b) allowing administrators or AI-assisted tools to annotate each API interface with a business semantic description, such as "get real-time current of water pump B" or "subscribe to motor temperature alarm"; c) notifying the API interface or message channel of successful registration after successful registration, automatically building and maintaining a built-in semantic index library, which stores metadata and corresponding business semantic descriptions of registered API interface information and message topic information, and supports semantic queries based on natural language or keywords, such as inputting "water pump vibration data" to return a list of relevant API interfaces and message topics; d) returning a list of matching data source metadata to the upper-layer application, including interface address, parameter specifications, authentication method, and business semantic description; the upper-layer application includes a human-computer interaction platform and a lightweight decision model; e) enabling non-technical personnel to discover data sources autonomously through natural language, while enabling the lightweight decision model to dynamically parse the underlying interfaces to be called when generating instructions, without the need for hard-coded adaptation.

[0044] like Figure 2 As shown, the data source semantic query service achieves semanticization and discoverability of legacy APIs through a three-step process of registration, indexing, and querying. The data source semantic query service receives registered API interface information and its business semantic description, and stores the API interface information and business semantic description in a semantic index library, such as... Figure 2 The example API interface semantic description is to obtain vibration data of a water pump; after successful registration, the API interface is notified of successful registration. The data source semantic query service receives query requests from the upper-layer application, such as a query request to obtain vibration data of water pump B, performs a semantic query in the semantic index, obtains a list of API interfaces, and returns it to the upper-layer application. The metadata in the list includes the interface address, parameter specifications, and authentication method.

[0045] The human-computer collaborative interaction platform includes five interconnected windows, such as... Figure 3As shown, the IoT event window, information window, emergency response plan window, command input window, and command execution and feedback window are respectively: The IoT event window displays real-time events in a scrolling manner, with alarm-level events pinned to the top; events originate from traditional monitoring systems and / or multimodal sensing layers; clicking on any event automatically activates related operations, such as viewing device statistics, topology relationships, and status curves. The information window dynamically displays event query results and the execution feedback of manually confirmed commands, and is interactively linked with the event window. The emergency response plan window displays structured technical document content to assist human decision-making, but does not generate executable commands.

[0046] The emergency response plan window serves as a knowledge support layer, comprising a rule matching unit and a Retrieval Enhancement Generation (RAG) unit. Events are first retrieved through the rule matching unit, which pre-deterministic handling rules for high-risk events. When a match is found, the standard emergency response plan for that event is directly returned. High-risk events include overpressure, overheating, and cascading shutdowns. The RAG unit semantically segments enterprise standard operating procedures (SOPs), manuals, and work orders, using domain-adaptive embedding models such as BGE-M3 to generate vectors, which are then stored in a vector database. For events where no rules are matched, the RAG unit filters and retrieves based on metadata such as equipment type and system module, returning Top-K relevant document fragments. It automatically extracts key content such as possible causes, typical handling steps, and safety warnings, and displays them in a structured format in the window for quick reference by operations and maintenance personnel, without requiring forced expansion of the original text links. The output of the RAG unit is only auxiliary information and does not constitute executable instructions, nor does it directly participate in model training or feedback optimization.

[0047] The command input window receives executable operation commands, which can be manually input or generated by a lightweight decision model. These commands include at least one of the following: querying statistical indicators, subscribing to device messages, and remotely controlling devices. As the action decision layer, the command input window receives user-inputted commands, supporting both voice and text input. The lightweight decision model can also generate executable operation suggestions based on the current decision context (selected event + indicators) and knowledge provided by RAG. For example, the command input window displays executable operation commands as follows:

[0048] - "Query the RMS trend of pump B's vibration over the past 1 hour"

[0049] - "Subscribe to the motor temperature MQTT topic / line3 / motor / temp"

[0050] - "Remote Reset Control Cabinet Circuit Breaker QF5"

[0051] Operators can click on instructions to select them. The selected instructions will appear in the instruction input window, and the instructions can be edited and confirmed for execution.

[0052] The system mandates that all commands must be explicitly bound to the current decision context, i.e., one or more events or one or more reporting metrics selected by the user. When generating commands, the lightweight decision model can invoke the data source semantic query service to automatically parse which API or message topic and parameters should be called.

[0053] The Command Execution and Feedback window uses user feedback on the validity of executed commands, along with the corresponding decision context, to form training samples for optimizing the lightweight decision-making model. This window displays confirmed commands and their execution status, such as executing, completed, or failed. The Command Execution and Feedback window is linked to the information window; when a command is completed and the result is returned, this window updates the command status and simultaneously displays the structured result in the information window. When a user clicks on any completed command or its associated result item, the system automatically highlights or jumps to the corresponding physical device, indicator chart, or topology node, enabling three-way traceability of operation, data, and assets. The Command Execution and Feedback window requires users to provide operation validity feedback, such as "valid / invalid / partially valid." This feedback, along with the corresponding "decision context + command," forms a training sample represented by the triple "decision context - executed command - validity feedback," used to optimize the lightweight decision-making model.

[0054] The lightweight decision model generates executable operation instructions based on context. When generating instructions, it can call the data source semantic query service to determine the specific API or message topic to be called.

[0055] The lightweight decision-making model optimizes through an instruction recommendation strategy optimization mechanism, learning which executable operation with the most practical value should be recommended in a specific context. The system records a triplet of "decision context → execution instruction → effectiveness feedback" formed in each human-machine collaborative operation. The decision context includes one or more IoT events actively selected by the user, and one or more statistical indicators from real-time / historical reports. The execution instruction is the actual confirmed and executed operation command, regardless of whether it originates from AI suggestions or human input. The effectiveness feedback is provided by the user after instruction execution and is converted into a numerical reward, for example, a reward of 1.0 / 0.5 / 0.0 corresponding to effective / partially effective / ineffective.

[0056] This invention provides an evolution interface for lightweight decision-making models, reserves standardized data export and model training interfaces, and supports the periodic use of accumulated structured samples for fine-tuning of lightweight decision-making models. In this embodiment, samples are uniformly formatted as JSON Schema, as shown in the example below. This sample structure naturally supports future expansion of multimodal contexts, such as adding an image_snapshot_id field, to provide high-quality supervision signals for more advanced models.

[0057]

[0058] The decision context of this example sample is recorded in the context, including IoT events recorded in the event set "events", statistical indicators recorded in the indicator set "metrics", execution instructions recorded in "action", and effectiveness feedback recorded in "reward". A timestamp is also added to the sample. In the example, EVT-20251202-102258-087 and MM-EVT-20251202-102259-001 correspond to event numbers, the indicator pump_B_vibration_rms-20251202-102259-224 refers to the report ID of the pump B vibration RMS trend, and motor_temp_C-20251202-102259-502 refers to the report ID of the motor temperature.

[0059] The operation replay and sample correction module is used to automatically generate replay records when triggered in the following situations: (1) an event arrives, including new alarms, device status updates from the IoT message queue, or multimodal sensing events; (2) a report message returned after the instruction is executed arrives, and the returned report message carries the instruction ID and associated event ID. For example, when the user issues the instruction "query the current of pump B in the past 24 hours", the system receives the returned report through the message queue. The message carries the command ID, the original trigger event ID, the query result and other related information; (3) the user completes the multi-event / multi-index context selection; (4) the user confirms the execution of an instruction; (5) the user submits validity feedback, that is, feedback is valid, invalid or partially valid; (6) the user responds to the user's explicit activation operation of an event in any window and uses that event as a trigger event to generate a replay record.

[0060] Users can explicitly activate a window through actions such as collapsing or highlighting it, which will then be used as a trigger event for the playback recording. The trigger event type is "User Attention Event," even if the event does not trigger any subsequent instructions or feedback. Users can also explicitly activate a window through actions such as clicking "Save this Scene," which will be used as a trigger event for the playback recording. The trigger event type is "User Annotation Event," even if the event does not trigger any subsequent instructions or feedback.

[0061] The replay records generated by this invention contain structured records including trigger event names and IDs, decision context, a list of recommended execution instruction IDs, a list of executed instruction IDs, feedback results, and relevant timestamps. They also support sample correction via the replay interface. The decision context includes a list of selected event IDs and a list of indicator IDs. The corrected replay records generate training samples for optimizing the lightweight decision-making model.

[0062] The operation replay and sample correction module provides a replay interface that supports searching and replaying the operation process by trigger event ID, instruction ID, or time range. Experts are allowed to correct or supplement the original feedback with fine-grained annotations during the replay process, and the corrected samples are given priority for model retraining.

[0063] The instruction self-evolution method for the dual-track transition system of intelligent operation and maintenance for SMEs based on the present invention mainly refers to using a sample set of high-quality, high-value scenarios accumulated over time to continuously fine-tune and update the lightweight decision-making model offline. First, one or more events are received, and user-associated reports / statistical indicators are used to form a decision context composed of the events and indicators. Based on the context, executable operation instructions are generated by the lightweight decision-making model or manually input. The actual execution results of the instructions and the effectiveness feedback provided by the user are recorded. Formatted samples are generated. In this embodiment, the sample format is {context:{events[],metrics[]},action,reward,timestamp}, where context is the decision context, events[] is the selected event array, metrics[] is the statistical indicator array, action is the executed instruction, reward is the digitized instruction feedback result, and timestamp is the timestamp.

[0064] Secondly, the playback records are manually reviewed periodically to correct command feedback results, and formatted samples of command validity feedback are exported periodically. The automated flow from playback records to training samples is as follows: Figure 4 As shown, users periodically review playback records and correct feedback results for operation commands, revise error feedback labels, and periodically export standard format data of command validity feedback. By periodically correcting labels and exporting valid feedback samples, a high-quality sample set is obtained.

[0065] By having users manually mark high-value or complex interaction scenarios on the human-computer collaborative interaction platform, playback records are generated. The trigger event type of the playback record is the user's focus scenario. The five window states of the human-computer collaborative interaction platform at that time are saved, and a formatted sample is generated.

[0066] Regularly obtain formatted samples and use the decision context-execution instruction-validity feedback triples to train or optimize lightweight decision models.

[0067] Example 1: Traditional monitoring event processing and playback correction.

[0068] Taking a water pump malfunction in a manufacturing enterprise as an example, the following is observed: (1) The IoT message queue receives the event: "Water pump B vibration exceeds limit", with event ID EVT-20250802-102155-087; (2) The system automatically generates a playback record REPLAY-20250802-102156-001, with the trigger type being "new event"; (3) The maintenance personnel selects the event in the event window, triggering information changes in the linkage window: (31) Based on the RAG knowledge system, the solution window displays multiple entries and file links related to "water pump" and "vibration exceeds limit", including "... To further view the historical records and check the power quality, the operator can select a text with the mouse and select "Create Command" from the pop-up command options; (32) The lightweight decision model suggests in the command input window: "Query the vibration trend of pump B in the past 24 hours", and the command ID (command_id) is CMD-20250802-102201-045; (4) The operator confirms the command in the command input window, and the system calls the historical database through the API gateway; (5) The system automatically generates the playback record REPLAY-20250802-102258 -002, where command_id is CMD-20250802-102201-045, and trigger type is "new command"; (6) The query result is returned through the message queue, and the message carries command_id:CMD-20250802-102201-045 and event_id:EVT-20251202-102345-277; (7) The information window displays the trend chart; (8) The system automatically generates a replay record REPLAY-20250802-102346-302, with trigger type "new information". (9) The maintenance personnel marked the feedback as "valid"; (10) The system automatically generated a replay record REPLAY-20250802-103018-003, with the trigger type being "instruction validity confirmation"; (11) A week later, the experts reviewed the results and found that the judgment was wrong, so it was corrected to "partially valid" because the root cause of the problem was the power stability issue, and a note was added; (12) The system automatically updated the replay record REPLAY-20250802-103018-003, and the validity was changed from "valid" to "partially valid"; (13) The corrected sample was used for the next training of the lightweight decision model.

[0069] Example 2: Multimodal sensing event integration.

[0070] This invention's system is deployed on a production line equipped with an edge AI box. An infrared camera uploads a thermal image of water pump B to the object storage service every 5 seconds, while an acoustic sensor simultaneously pushes a 10-second audio clip to a message queue. A large-scale, multi-reward native multimodal inference model runs as a microservice, subscribing to the aforementioned data streams. After fusion analysis, it outputs structured events, which are then pushed to the alarm event channel of the message queue. An example of a structured event is as follows:

[0071]

[0072] This example records the event ID, event type, occurring device, conclusion confidence level, data source modality, and storage path. The event type confidence level is based on the data source modality ["thermal imaging", "vibration spectrum"]. When the IoT event window service starts, it registers to listen for alarm event channels. Upon receiving the pushed event, it displays the event using the event type template for maintenance personnel to make a comprehensive judgment.

[0073] Example 3: Data source semantic registration and query.

[0074] The administrator registers an API interface through the management interface: "Interface ( / pump / getData), Request Method (GET), Request Parameters ([id, mode], id is the pump ID in integer form, mode is a fixed value raw), Return Format (CSV), Character Encoding (GBK), Response Code Mapping (200 Success, 503 Device Offline)". The system guides the administrator to annotate the business semantics: "Get the current raw vibration signal of the pump by pump ID". Subsequently, when the lightweight decision model needs to "get the raw vibration data of pump B", it calls the data source semantic query service, passing in the natural language query "get the raw vibration data of pump B", and the service returns the registration information and call template, achieving zero-code integration.

[0075] Example 4: Automated transfer of playback records to training samples.

[0076] The system of this invention performs a sample extraction task every morning at dawn: it iterates through all playback records, filters out records whose trigger events are of the "instruction validity confirmation" type, converts them into standard training sample format, such as {"context": {...}, "action": "...", "reward": 1.0}, and writes them to the sample repository. The training pipeline of the lightweight decision model pulls the latest sample set daily for incremental fine-tuning to ensure that the recommendation strategy continuously aligns with best practices in the field. If further analysis of user behavior is needed, records whose trigger events are of the "user attention event" type can be selected.

[0077] Example 5: Users actively activate high-value scenarios.

[0078] During an alarm handling process, the operations engineer discovered a rare "vibration spectrum harmonic anomaly." Although no immediate instructions were given, sensing its importance, the engineer clicked the "Save this scenario" button next to "Query Vibration Spectrum Report" in the information window. The system immediately generated a playback record (REPLAY-20250802-102558-031) with the trigger event type "User-Concerned Scenario," fully saving the state of the five windows at that time. This record was later used for subsequent processing, internal training, and building a long-tail fault training set.

[0079] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.

Claims

1. A dual-track intelligent operation and maintenance transition system for small and medium-sized enterprises, characterized in that: The system includes: an interface layer for legacy platforms, a human-computer collaborative interaction platform, a lightweight decision-making model, and an operation playback and sample correction module; The legacy platform interface layer includes an API gateway, an IoT message queue, and a data source semantic query service. The system accesses the enterprise's current operations and maintenance management system, monitoring system, and work order database through the API gateway to obtain operations and maintenance data. The system subscribes to device sensor data, events, and control commands in real time through the IoT message queue. Events generated by the device monitoring system are injected into the IoT message queue. Enterprise administrators register the API interfaces and message topics related to the operations and maintenance management system, monitoring system, and work order database into the system through the data source semantic query service. The data source semantic query service semantically transforms the interfaces of the enterprise's operations and maintenance management system and monitoring system through a registration-indexing-query process, including: receiving registered API interface information or message topic information, obtaining the business semantic description of the API interface or message topic, and notifying the API interface or message channel of successful registration; storing the API interface or message topic information and business semantic description in a semantic index library. The API interface or message topic information includes the interface address, parameter specifications, request method, return format, and authentication method. The semantic index library supports semantic queries based on natural language or keywords, returning a list of relevant API interfaces and message topics. The human-machine collaborative interaction platform includes an IoT event window, an information window, an emergency response plan window, an instruction input window, and an instruction execution and feedback window. The IoT event window displays real-time events in a scrolling fashion, with alarm-level events prioritized. Clicking on an event activates associated operations. The information window dynamically displays event query results and execution feedback for manually confirmed instructions, and it interacts bidirectionally with the IoT event window. The emergency response plan window displays structured technical documents related to alarm events to assist human decision-making. The instruction input window receives executable operation instructions, which are either manually input or generated by a lightweight decision-making model. Operation instructions include at least one of the following: querying statistical indicators, subscribing to device messages, and remotely controlling devices. The instruction execution and feedback window displays confirmed executed instructions and their execution status, interacts with the information window, and mandates user feedback on the validity of the instructions. It constructs a triplet training sample of decision context-execution instruction-validity feedback to optimize the lightweight decision-making model. The decision context includes the user-selected event and statistical indicators reported in real-time or historically. The lightweight decision model generates executable operation instructions based on the decision context. When generating instructions, it calls the data source semantic query service to query the API interfaces and / or message topics that the instructions need to call. When the operation replay and sample correction module is triggered, it generates a replay record; the replay record is a structured record containing the triggering event, decision context, execution instructions, feedback results and timestamps; API stands for Application Programming Interface, and URL stands for Uniform Resource Locator.

2. The system according to claim 1, characterized in that, The system also includes a multimodal perception layer, which deploys a large multimodal inference model. The large model is used to receive raw multimodal data including device infrared images, visible light images, voiceprint recordings, and sensor timing data, and output structured semantic events. The multimodal perception layer injects the structured semantic events into the IoT message queue. The multimodal perception layer is an optional module. It is initially turned off and then turned on after obtaining more than a set amount of multimodal operation and maintenance data.

3. The system according to claim 1, characterized in that, The emergency response plan window includes a rule matching unit and a search enhancement generation (RAG) unit. Events are first searched through the rule matching unit, which pre-sets deterministic handling rules for high-risk events. When a match is found, the standard emergency response plan for that event is directly returned. The RAG unit semantically segments enterprise standard operating procedures, manuals, and work orders, uses an embedding model to generate vectors, and stores them in a vector database. For events not found in the rule matching unit's search, the RAG unit filters the search based on the device type and system module that sent the event, returning Top-K relevant document fragments. It automatically extracts the event cause, typical handling steps, and safety warnings from these document fragments for user reference.

4. The system according to claim 1, characterized in that, The triggering events for the operation playback and sample correction module include the following types: (1) New events arrive, including alarms, device status updates from IoT message queues, or multimodal sensing events; (2) The report message returned after the instruction is executed arrives, and the returned report message carries the instruction ID and the associated event ID; (3) The user completes the decision context selection, instruction confirmation, or feedback submission; (4) In response to the user’s explicit activation of an event in any window, the event is used as the trigger event to generate a playback record.

5. The system according to claim 1, characterized in that, The operation replay and sample correction module provides a replay interface that supports retrieving replay records by trigger event ID, instruction ID, or time range, and replaying the instruction operation process. It supports replaying the recordings in the replay interface and generating training samples from the corrected replays to optimize the lightweight decision model.

6. The system according to any one of claims 1-5, characterized in that, The instruction self-evolution method implemented based on this system includes: (1) Receive one or more events, associate indicators with users, and form a decision context consisting of the events and indicators; based on the context, generate executable operation instructions by a lightweight decision model or manually input them; record the actual execution results of the instructions and the effectiveness feedback provided by the user; generate formatted samples; The formatted sample is expressed as {context:{events[],metrics[]},action,reward,timestamp}, where context is the decision context, events[] is an array of events, metrics[] is an array of metrics, action is the executed instruction, reward is the digitized instruction feedback result, and timestamp is the time stamp; (2) Regularly review the playback records manually, correct the instruction feedback results, and regularly export formatted samples of instruction validity feedback; (3) When a user manually marks a high-value scene on the human-computer collaborative interaction platform, a replay record is generated, the event type is the scene the user is interested in, the state of the five windows of the human-computer collaborative interaction platform at that time is saved, and a formatted sample is generated. (4) Regularly obtain formatted samples and use the triplet of decision context-execution instruction-validity feedback to train or optimize the lightweight decision model.

Images