An intelligent operation and maintenance method integrating multimodal data and active learning

Abstract

The present application relates to the field of intelligent operation and maintenance technology, and discloses an intelligent operation and maintenance method that integrates multimodal data and active learning. The method comprises: deploying a hierarchical Internet of Things device network in a target operation and maintenance area, collecting heterogeneous operation and maintenance data sets, and generating an operation and maintenance feature data set using a unified building operation and maintenance service framework coordinated by multiple MCPs; inputting the operation and maintenance feature data set and user feedback information into a deep learning intention recognition model for intent recognition and demand analysis, thereby obtaining structured user demand data and operation and maintenance task priority ranking; performing a real-time evaluation of the equipment operating status based on the operation and maintenance feature data set and the structured user demand data, and obtaining fault risk warning data; performing autonomous decision-making analysis through an agent-based artificial intelligence engine, and generating equipment maintenance plans and resource scheduling plans, thereby solving the problem in traditional operation and maintenance that a single fault prediction model has poor adaptability and prediction results are difficult to convert into effective decisions.

Description

Technical Field

[0001] The present application relates to the field of intelligent operation and maintenance technology, and in particular to an intelligent operation and maintenance method that integrates multimodal data and active learning. Background Art

[0002] In traditional operation and maintenance systems, equipment status monitoring usually relies on manual inspections or simple single-point sensors, making it difficult to detect equipment failures in a timely manner; maintenance processing processes are mostly passive responses, lack predictive capabilities, and are inefficient; user demand response systems are single and cannot accurately understand users' true intentions, resulting in untimely and inaccurate service responses; operation and maintenance decisions lack systematic data support, rely more on experience and judgment, and are not scientific enough; resource scheduling methods are fixed and cannot be dynamically optimized according to actual needs.

[0003] The continuous advancement of smart cities, smart campuses, and smart communities is placing higher demands on operations and maintenance management, but current technological implementation faces significant bottlenecks. While IoT devices are widely used, their deployment lacks systematic planning, leading to data redundancy and blind spots. Data collection frequencies are fixed and cannot be dynamically adjusted based on device status, resulting in wasted resources and insufficient monitoring. Information between devices is isolated, lacking a collaborative monitoring mechanism, leading to high false alarm rates. Operation and maintenance data is stored in a decentralized manner, hindering information flow and creating "data silos." Fault prediction models are limited in scope and unable to address complex and diverse fault types. Operation and maintenance decision-making lacks a closed-loop optimization mechanism, hindering continuous improvement. Summary of the Invention

[0004] This application provides an intelligent operation and maintenance method that integrates multimodal data and active learning, thereby solving the problem in traditional operation and maintenance that the single fault prediction model has poor adaptability and the prediction results are difficult to convert into effective decisions.

[0005] In a first aspect, the present application provides an intelligent operation and maintenance method that integrates multimodal data and active learning. The intelligent operation and maintenance method that integrates multimodal data and active learning includes:

[0006] Deploy a hierarchical IoT device network in the target operation and maintenance area and collect heterogeneous operation and maintenance data sets including environmental parameter data, device status data, resource consumption data, and security monitoring data;

[0007] Constructing a building operation and maintenance prompt word project for the heterogeneous operation and maintenance data set, and using a unified building operation and maintenance service framework coordinated by multiple MCPs to generate an operation and maintenance feature data set;

[0008] Input the operation and maintenance feature data set and user feedback information into the deep learning intention recognition model to perform intention recognition and demand analysis to obtain structured user demand data and operation and maintenance task priority ranking;

[0009] Based on the operation and maintenance feature data set and the structured user demand data, the equipment operation status is evaluated in real time to obtain fault risk warning data;

[0010] The fault risk warning data and the operation and maintenance task priority ranking are input into the agent-type artificial intelligence engine for autonomous decision-making analysis to generate equipment maintenance plans and resource scheduling plans.

[0011] Compared with the existing technology, the present application has the following beneficial effects: through hierarchical IoT device deployment and multi-level sampling strategies, differentiated monitoring schemes are implemented for different types of equipment, and the sampling frequency is dynamically adjusted in combination with adaptive sampling technology. While ensuring data quality, the system resource allocation is optimized, solving the data quality problems and resource waste problems caused by unreasonable sampling in traditional operation and maintenance. The multi-channel hybrid networking system intelligently allocates the optimal channel according to the communication needs of the equipment, and combines the model context protocol (MCP) to build a unified large language model service framework to achieve multi-model collaborative building operation and maintenance intelligent analysis. The three-layer encryption mechanism and priority transmission strategy ensure the security and timeliness of data transmission, solving the problems of unsafe, untimely and unintelligent data transmission in traditional operation and maintenance systems. The multimodal fusion intent recognition technology based on deep learning realizes efficient entity recognition through the selective state space layer of the Mamba model, and combines the LLM-Fine-Tuning technology for professional field intent classification. Compared with traditional methods, the entity recognition efficiency is improved by 5-10 times, and the intent understanding accuracy is improved by 15-20%. The innovative application of building operation and maintenance prompts transforms complex equipment status descriptions into structured operation and maintenance analysis results, enabling a precise understanding of user needs and addressing the inaccurate understanding and in-depth analysis inherent in traditional operation and maintenance. A graph neural network-based device topology is constructed to enable intelligent collaborative monitoring. Agent-based artificial intelligence (Agentic AI) technology is introduced to empower the system with autonomous decision-making capabilities. Through AI agents that perceive the environment, set goals, plan paths, and execute autonomously, the system achieves intelligent fault prediction and resource scheduling, addressing the high false alarm rates and inefficient resource allocation associated with isolated equipment monitoring in traditional operation and maintenance systems. A differentiated fault prediction system based on multi-model fusion selects the optimal algorithm combination for different fault characteristics, using dynamic ensemble learning and transfer learning techniques to adaptively optimize the prediction model. Digital twin technology is used to establish a bidirectional mapping between physical equipment and virtual models, providing a virtual experimental scenario for fault prediction. Blockchain technology enables tamper-proof storage of operation and maintenance data and multi-party collaboration, transforming fault predictions into actionable maintenance strategies. This addresses the challenges inherent in traditional operation and maintenance, such as the poor adaptability of single fault prediction models, the difficulty in translating prediction results into effective decisions, and low data credibility. Based on operation and maintenance feedback data and meta-learning algorithms, personalized optimization paths are customized for AI models to achieve continuous self-evolution of the system, enabling it to cope with new equipment, new failure modes and ever-changing user demand expressions, continuously maintain technological advancement, and fundamentally change the limitations of traditional operation and maintenance systems that are static and difficult to adapt to new scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0014] Figure 1 This is a flow chart of an intelligent operation and maintenance method that integrates multimodal data and active learning, provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0016] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0017] It should also be understood that the terms used in this specification are for the purpose of describing the target embodiments only and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0018] It should be further understood that the term "and / or" used in this specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 In the embodiments of the present application, an embodiment of an intelligent operation and maintenance method integrating multimodal data and active learning includes:

[0019] Step 100: Deploy a hierarchical IoT device network in the target operation and maintenance area, and collect a heterogeneous operation and maintenance data set including environmental parameter data, device status data, resource consumption data, and security monitoring data;

[0020] It is understandable that the execution subject of this application can be an intelligent operation and maintenance device based on the fusion of multimodal data and active learning, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0021] Specifically, all equipment in the target operation and maintenance area is classified into environmental monitoring equipment, equipment status monitoring equipment, resource consumption monitoring equipment and security monitoring equipment. Environmental monitoring equipment includes various sensors for collecting environmental parameters such as temperature, humidity, light, and smoke. Equipment status monitoring equipment covers special sensors such as vibration, temperature, current, and speed on key facilities such as water pumps, elevators, and air conditioners. Resource consumption monitoring equipment includes smart electricity meters, smart water meters, and gas meters. Security monitoring equipment includes cameras, infrared sensors, access control systems, and other terminals for security situation awareness. Based on the varying requirements for real-time data collection and accuracy across various devices, a multi-level data collection frequency control strategy is implemented. For environmental monitoring devices, a fixed sampling frequency of every five minutes is used. For status monitoring devices, the sampling frequency is dynamically adjusted based on the device's operating status. For example, the sampling frequency is automatically increased when operating parameters approach abnormalities or experience significant fluctuations, while the sampling frequency is reduced when the device's status is stable to conserve system resources. For resource consumption monitoring devices, a time-based, differentiated sampling scheme is implemented, increasing the sampling frequency to every ten minutes during peak energy consumption periods and extending it to once an hour during off-peak periods. For security monitoring devices, near-real-time data collection is required. Based on this multi-level data collection strategy, distributed sampling control is implemented across all IoT devices, centrally collecting heterogeneous raw data from multiple sources, at different frequencies, and in different physical locations to a local edge computing gateway. Data pre-screening is performed on the edge computing gateway, using rules such as value range checks, rate of change analysis, and data integrity checks to quickly identify and eliminate abnormal, redundant, or invalid data, improving data reliability and validity. Adaptively adjust the sampling frequency of pre-processed device data after preliminary screening. Based on parameters such as the real-time volatility of the currently collected data, historical operating status, and device health trends, the sampling frequency of each collection node is dynamically adjusted, allowing the collection frequency to be adaptively optimized based on the device's operating characteristics and environmental conditions. The dynamically optimized data collection results are associated with the device's unique identification code and integrated according to unified data encoding and structuring standards to obtain a heterogeneous operation and maintenance data set.

[0022] Step 200: Construct a building operation and maintenance prompt word project for heterogeneous operation and maintenance data sets, and use a unified building operation and maintenance service framework coordinated by multiple MCPs to generate an operation and maintenance feature data set;

[0023] Specifically, a multi-channel hybrid networking system is used to allocate communication channels to devices. IoT terminals are assigned to optimal communication channels based on parameters such as data flow, transmission frequency, communication distance, and power consumption requirements for different device types. For example, high-volume, real-time data (such as HD video streaming) is prioritized for 5G or wired broadband channels, while environmental parameter sensors with periodic, low-volume data are selected using low-power wide-area wireless protocols such as NB-IoT or LoRa. This approach dynamically configures a hybrid heterogeneous network architecture, forming a hybrid heterogeneous network configuration solution. A building operation and maintenance prompt project is constructed for heterogeneous operation and maintenance data. This project systematically converts semi-structured and unstructured data, such as building equipment status data, environmental parameters, and historical operation records, into carefully designed prompt templates, enabling large language models to accurately understand building operation and maintenance contexts. The prompt project includes four key components: scenario description, historical context, professional knowledge supplementation, and task guidance. It automatically selects the most appropriate prompt template for each operation and maintenance scenario, significantly improving the model's ability to understand and analyze operation and maintenance data. A unified building operation and maintenance service framework based on the Model Context Protocol (MCP) enables the collaboration of multiple large language models. As a new large language model application paradigm, MCP defines a unified input / output format, context management rules, and model call process, enabling multiple specialized large language models to work together to handle complex building operation and maintenance scenarios. This framework effectively addresses the limitations of a single model in terms of domain knowledge depth and reasoning capabilities. By transferring knowledge and sharing context between models, it enables more accurate equipment status assessment, fault diagnosis, and predictive maintenance analysis, generating high-value intelligent operation and maintenance data. Based on a hybrid heterogeneous network configuration and intelligent operation and maintenance data, a lightweight AES-128 encryption algorithm is applied at the device layer to ensure that every piece of raw data sent from IoT terminals is effectively encrypted at the source, preventing data hijacking or theft at the local device level. During data transmission, TLS 1.3 is used for end-to-end dynamic encryption and key negotiation, eliminating the risk of intermediary eavesdropping and tampering, and strengthening the confidentiality and integrity of the communication link. At the application layer, a comprehensive identity authentication and access control mechanism enables granular control of data access rights, ensuring that only authorized users and system processes can access, retrieve, and process critical data resources, preventing unauthorized access and data leakage. Furthermore, priority management is implemented for all transmitted data. Based on pre-set rules and real-time business scenarios, encrypted data with multiple layers of protection is prioritized. For example, security alarm data is assigned the highest priority, followed by device anomaly or fault status information, routine environmental monitoring data is given a lower priority, and historical archived data is given the lowest priority. All data streams are then sequentially input to the central server according to priority.All data arriving at the server is cleansed. Using various intelligent algorithms, including Z-score anomaly detection, Kalman filtering, and missing value imputation, we automatically identify and remove noisy data, duplicate data, and obviously erroneous data points, producing a cleansed dataset. Data standardization is then performed on the cleansed dataset. Using methods such as normalization or standard deviation normalization, data from different sources and dimensions is unified into a standardized data structure and distribution range, resulting in a standardized dataset. Feature extraction is performed on the standardized dataset, using various feature extraction algorithms, such as wavelet transform, fast Fourier transform, and principal component analysis, to analyze the data's time, frequency, and statistical characteristics, generating an operational feature dataset.

[0024] In this embodiment, a device communication parameter analysis is performed on the communication requirements of different devices in a heterogeneous operation and maintenance dataset. Key parameters such as data traffic characteristics, communication real-time requirements, sampling frequency, packet size, duty cycle, and deployment location are collected and statistically analyzed for each device. Based on the communication demand analysis results, all devices are classified according to metrics such as real-time performance, bandwidth, power consumption, and coverage. For example, high-definition video surveillance equipment with high real-time and high bandwidth requirements, environmental monitoring terminals with low power consumption and long duty cycles, and mobile security nodes with wide coverage requirements are categorized into different communication demand categories, resulting in a device communication demand classification result. Based on the device communication demand classification results, the transmission capacity of six communication technologies: NB-IoT, LoRa, Wi-Fi, ZigBee, 5G, and wired networks is evaluated. This evaluation covers parameters such as the theoretical and actual bandwidth limits, minimum and average communication delays, typical coverage, and energy consumption per unit of data transmitted for each communication technology, thereby constructing a multi-channel communication channel characteristic database. By conducting a multi-dimensional device-channel matching analysis based on the device communication requirement classification results and the aforementioned communication channel characteristic data, mathematical methods such as Euclidean distance and weighted scoring are used to calculate the matching scores between different device categories and each channel attribute. For example, devices requiring high bandwidth and low latency prioritize 5G and wired networks, while devices requiring low power and long distances prioritize NB-IoT or LoRa. Using this matching scoring mechanism, the optimal communication channel is assigned to each device category, initially forming a channel allocation plan that meets physical performance constraints. Network load analysis is then conducted on this initial channel allocation plan. By monitoring the load distribution of each communication channel in actual operation and maintenance scenarios, including metrics such as data traffic concentration, number of concurrent connections, communication conflict probability, and local bottleneck nodes, the load balancing status of each channel is dynamically monitored and evaluated in real time. When certain channel allocations are found to be congested or idle, the channel allocation of some devices is adjusted, migrating devices from overloaded channels to less loaded channels, optimizing overall resource allocation and forming a load-balanced channel allocation plan. A specialized building operation and maintenance prompt engineering system is being established for building operation and maintenance scenarios. The system incorporates a multi-layered prompt word construction strategy: the foundational layer converts sensor data, equipment status, and environmental parameters into standardized descriptions; the knowledge-enhanced layer incorporates building equipment expertise, failure mechanisms, and maintenance specifications; and the task-oriented layer customizes specialized instructions for specific O&M needs, such as fault diagnosis, energy efficiency optimization, and predictive maintenance. The system dynamically combines these three layers of prompt words based on different O&M scenarios to generate structured input suitable for large language model processing. For example, for air conditioning system fault diagnosis, prompt words might include multi-dimensional information such as "the system's historical five-day temperature fluctuation curve, a 20% drop in current cooling capacity, abnormal compressor sound, low refrigerant pressure readings, and data on similar fault cases," enabling the model to accurately grasp the essence of the problem. A unified large language model service framework is constructed based on the Model Context Protocol (MCP).MCP enables intelligent building operation and maintenance analysis through three core mechanisms: First, a context management mechanism maintains the continuity of critical operation and maintenance information, ensuring that the model understands the evolution of equipment status. Second, a model orchestration mechanism dynamically selects and combines expert and general models based on task complexity and expertise, forming an "expert collaboration network." Third, a feedback optimization mechanism verifies and corrects model outputs, enhancing the reliability of results through methods such as physical rule verification, historical case comparison, and multi-model cross-validation. This multi-model collaborative framework transforms complex building system states into clear operational and maintenance decision-making recommendations, such as high-value analytical results such as "Chiller Unit 3 has a 25% drop in cooling efficiency, likely due to heat exchanger scaling (83% confidence level). Cleaning and maintenance is recommended within 72 hours, with an estimated energy efficiency recovery of 18%." Differentiated communication strategies are developed based on the business needs and network characteristics of different devices. For example, high-priority, low-latency emergency data channels are enabled for core security equipment, batch reporting is scheduled for periodic monitoring equipment, and dynamic bandwidth adjustment is implemented for some temporarily high-traffic devices, enabling flexible response to changing application scenarios. The above differentiated communication strategies are integrated with the load balancing channel allocation scheme to generate a hybrid heterogeneous network configuration scheme.

[0025] Step 300: Input the operation and maintenance feature data set and user feedback information into the deep learning intention recognition model to perform intention recognition and demand analysis to obtain structured user demand data and operation and maintenance task priority ranking;

[0026] Specifically, demand data generated by repair requests and suggestions from mobile apps, human-computer interaction commands from voice assistants, manual recordings of customer service calls, and automatic transactions from the work order system are collected and integrated with multi-dimensional operational and maintenance feature datasets, including on-site equipment operating status, historical alarms, and environmental parameters, to construct a multimodal user demand dataset encompassing text, voice, and images. The textual data from this multimodal user demand dataset is fed into the Mamba model within a deep learning intent recognition model for entity recognition. Key elements such as device name, fault symptoms, location information, and operation time are extracted, converting the natural language content into highly structured entity annotations. This entity annotation data is then fed into the LLM-Fine-Tuning model within the deep learning intent recognition model for intent classification, resulting in the resulting intent classification results. This model, through pre-training a large language model and domain adaptation fine-tuning, combined with prompt word engineering, deeply understands the contextual semantics and domain knowledge of the text, distinguishing user requests as repair requests, inquiries, complaints, and suggestions, and then refining them down to specific service requests or root causes. Based on intent classification results and entity annotation data, automated knowledge graph construction technology is employed to structure all identified entities, their attributes, behaviors, and states according to pre-defined ontology and business logic, forming a multi-level, relatable, and queryable entity-relationship network. Based on this structured knowledge network and combined with historical user behavior data, including past repair frequency, problem resolution time, satisfaction ratings, and preferred request types, the system leverages the attention mechanism or memory enhancement mechanism within deep neural networks to analyze the implicit mapping between user expressions and actual needs, thereby refining and improving the accuracy of understanding personalized service requests. Through dynamic knowledge mapping, the system adapts to user expression habits and service scenarios to generate structured user demand data. Based on this structured demand data, an urgency score is calculated for each demand. This score is weighted based on multiple factors, including device importance, problem scope, duration, security risk, and user-specific needs. All demands are automatically sorted by urgency score to create a priority queue for maintenance tasks.

[0027] In this embodiment, the text data in the multimodal user demand dataset is segmented and cleaned, including pre-processing operations such as segmentation of Chinese sentences, stop word filtering, special character removal, and spelling normalization to obtain a processed text sequence. The processed text sequence is input into the selective state space layer of the Mamba model for sequence modeling. As a new architecture for sequence modeling, the Mamba model breaks through the limitations of traditional RNN and Transformer, and realizes long text processing with linear complexity through the structured state space model (SSM), while maintaining the ability to efficiently model long-distance dependencies. The model receives a text sequence and selectively encodes the input through a selective scanning mechanism, effectively capturing long-distance semantic dependencies in the text and obtaining a rich long-distance dependency feature representation. Based on the long-distance dependency feature representation, the Mamba model performs parallel sequence processing. Unlike traditional recurrent neural networks, Mamba converts sequence information into state representation through an innovative parallel computing method, and captures temporal patterns and key features in the text through a structured state space model. This processing approach enables the model to simultaneously consider global and local context, forming context-aware feature representations. This significantly improves the recognition of specialized terms such as device names, fault descriptions, and location information. A gating mechanism filters and enhances the context-aware feature representations. The Mamba model uses a complex gating network to dynamically adjust the importance of information at different locations, highlighting the feature representations of keywords and specialized terms while mitigating the influence of irrelevant information. This extracts key information from the text and forms a set of entity candidate features. This set of entity candidate features is then fed into a self-attention layer for global information integration. While Mamba itself already has the ability to capture long-range dependencies, introducing a local self-attention mechanism in entity recognition further enhances the model's ability to perceive potential entity locations. By calculating attention weights for each location in the sequence, it accurately identifies entity boundaries and generates a probability distribution for entity boundaries. Sequence labeling decisions are then made based on this probability distribution. The system combines the Mamba model's predictions with labeling transfer rules to construct a labeling score graph. The Viterbi algorithm then searches for the optimal labeling path, ensuring logical consistency across the entire sequence and avoiding unreasonable labeling combinations, ultimately achieving globally optimal entity recognition results. The globally optimal entity recognition results are post-processed and entity type normalized, including entity boundary adjustment, nested entity processing, synonym resolution, and entity type standardization, to obtain entity annotation data with a unified format and clear types, laying the foundation for subsequent intent analysis and demand understanding.

[0028] In this embodiment, the entity annotation data is subjected to text reconstruction and prompt word engineering optimization. The system first organizes the identified entities and their type information in a structured manner, and then designs an optimized prompt word template in combination with professional knowledge in the operation and maintenance field. Through technologies such as key information reinforcement, context supplementation, and task instruction clarification, the entity annotation data is converted into a more semantically expressive input sequence to form an enhanced semantic input sequence. The enhanced semantic input sequence is input into a pre-trained large language model. Utilizing a large language model pre-trained on massive general corpus and professional field texts, through its multi-layer Transformer architecture, a deep semantic analysis of the input text is performed, taking into account information at multiple levels such as vocabulary, syntax, and context, and extracting rich context-aware representation vectors. These vectors contain deep semantic information of the text intent. Domain-oriented fine-tuning strategies are designed based on specific operation and maintenance tasks. For the professional field of construction equipment operation and maintenance, efficient parameter fine-tuning technologies (such as LoRA, P-Tuning, etc.) are used to update only some key parameters of the large language model while retaining the model's general language understanding capabilities. During the fine-tuning process, domain expert knowledge and historical failure cases are incorporated, and a knowledge base for intent recognition in the O&M domain is constructed through gradient updates. This results in a domain-adapted fine-tuned model, significantly improving its ability to understand specialized terminology and complex expressions. The domain-adapted fine-tuned model performs multi-label classification on the input text. Unlike traditional single-intent recognition, the LLM, leveraging its powerful semantic understanding capabilities, can simultaneously identify both primary intent (e.g., "report a repair") and secondary intent (e.g., "query progress") in user expressions, generating a multi-dimensional intent probability distribution that more comprehensively captures the user's complex needs. Confidence thresholds are applied to the intent probability distribution, and intent conflict resolution is implemented. The system sets a dynamically adjusted confidence threshold to filter out low-probability intent predictions. Furthermore, it resolves potential intent conflicts by applying business rules and logical relationships between intents (e.g., "expedited processing" and "deferred processing" cannot both be true). Finally, intent normalization is performed based on O&M business rules, mapping various expressions to standardized business intent categories. This results in a final intent classification result, providing an accurate basis for subsequent task assignment and resource scheduling.

[0029] Step 400: Based on the operation and maintenance feature data set and the structured user demand data, the equipment operation status is evaluated in real time to obtain fault risk warning data;

[0030] Specifically, the operation and maintenance feature dataset is grouped according to system architecture and device function, organizing various IoT devices into several logical monitoring units with complementary functions. Each monitoring unit is composed of multiple types of sensors and monitoring devices. For example, a key device is equipped with different types of sensors, such as temperature, vibration, current, and noise, to provide a multi-dimensional reflection of the device's health status. Based on this, a comprehensive device status feature vector is constructed for each logical monitoring unit. This vector includes currently collected operating parameters (such as real-time temperature, pressure, and current), integrates environmental parameters (such as ambient humidity and air quality), and historical fault and maintenance records. The device status feature vector is input into a hybrid model consisting of a multilayer perceptron and a support vector machine for anomaly detection. The multilayer perceptron leverages the representation capabilities of deep neural networks to extract nonlinear and complex features from large-scale data and identify potential abnormal operating patterns. The support vector machine, with its excellent segmentation capabilities in high-dimensional feature spaces, improves the boundary clarity and generalization of anomaly identification. Working together, the two effectively filter out false positives and false negatives in device status and output preliminary abnormal status indicators. Based on preliminary abnormal state identification, a multi-dimensional device association topology based on physical space and functional relationships is constructed. This topology presents all devices in a graph structure, with nodes representing the devices themselves and edges indicating physical proximity, data collaboration, or task dependency strength. This generates a device association graph. Based on the device association graph and structured user demand data (such as user repair feedback and high-priority tasks in the target area), surrounding devices with associations exceeding a preset threshold are activated for collaborative monitoring. Peripheral associated devices simultaneously collect their own status and upload relevant data, generating multi-source verification information. This cross-device, multi-angle, and multi-dimensional data verification effectively filters out the randomness of isolated fault points and enhances the accuracy and reliability of anomaly detection. This multi-source verification information is input into a dynamic Bayesian network for probabilistic reasoning. The dynamic Bayesian network has the ability to model causal relationships and express uncertainty for time series data. By integrating historical evolution with current monitoring, it calculates the posterior probability distribution of the target device's multiple states (normal, warning, fault, etc.). Based on this posterior probability distribution, an intelligent assessment of the device's current risk state is automatically made, generating quantitative fault risk warning data.

[0031] Step 500: Input the fault risk warning data and operation and maintenance task priority ranking into the agent-based artificial intelligence engine for autonomous decision-making analysis to generate equipment maintenance plans and resource scheduling plans.

[0032] Specifically, fault type classification is automatically performed based on the abnormal status and risk probability provided by fault risk warning data, combined with multi-dimensional equipment operating information and historical O&M data. By analyzing monitoring data, it clearly distinguishes which equipment is currently experiencing gradual faults (such as continuous performance degradation or gradual deviation from normal ranges) and which equipment is experiencing sudden faults (such as instantaneous component failure or the sudden appearance of abnormal signals). These classification results are then standardized and output. An agent-based artificial intelligence (Agentic AI) multi-objective task model is constructed, enabling AI agents to perceive the environment, set goals, plan paths, and execute autonomously. This model optimizes fault repair rate, resource utilization efficiency, and time response speed. Through large-scale parameter learning and scenario adaptation, it develops an intelligent agent system capable of making independent decisions in different O&M scenarios. The model combines historical O&M experience with current scenario constraints to automatically calculate optimal decision parameters for subsequent maintenance plan generation and resource scheduling. Adaptive intelligent prediction algorithms are used for different fault categories. For gradual faults, a hybrid time series prediction model is constructed using a long-short-term memory network and an autoregressive integrated moving average model. Long-short-term memory networks (LSTMs) have strong representation capabilities for long-term historical data, complex time-series dependencies, and nonlinear trends, enabling them to capture subtle changes in equipment operating status over time and long-term degradation characteristics. The autoregressive integrated moving average model, with its efficient modeling capabilities for short-term trends and linear changes, enables adaptive smoothing and trend decomposition of sudden fluctuations. Combining these two models allows the system to provide highly accurate numerical predictions of the future development trends of gradual faults, determine how long it will take for equipment to enter a high-risk state, and infer the rate of fault deterioration and remaining lifespan. For sudden faults, a hybrid classification model of random forests and gradient boosted decision trees is used. Random forests, by constructing an ensemble of multiple decision trees, exploit high-dimensional features and nonlinear relationships between complex features to achieve highly robust classification. This model is suitable for capturing state changes before and after sudden events, abnormal signal distributions, and multi-source alarm patterns. The gradient boosted decision tree model, using an additive model and residual fitting mechanism, accurately identifies low-probability, high-impact, and targeted sudden fault patterns, enhancing the system's ability to detect atypical and hidden sudden faults. The two models work together to efficiently identify sudden equipment failures and predict their development trends. This effectively avoids false alarms and missed alerts in scenarios such as multi-source data integration, complex operating conditions, and weak fault signals. A failure risk index is calculated based on the failure trend prediction results and the prioritization of maintenance tasks. This prioritization is derived from preliminary intent identification, task diversion, and business impact assessment, taking into account factors such as the device's importance to the overall system, current business load, user demand urgency, and the scope of the failure. The system combines these indicators to quantitatively calculate the failure risk index for each fault point.The risk index model integrates multiple dimensions of data, including the current equipment status, future failure rate, historical failure probability, business weight, geographic impact, and repair difficulty. Using weighted aggregation or deep regression, it outputs a quantitative risk score between 0 and 1. A higher score indicates a greater risk to the equipment, a more urgent threat to the business, and a greater need for O&M resources. Based on the risk index and AI agent decision parameters, an optimal equipment maintenance plan is automatically generated. This plan includes details such as the maintenance target, estimated maintenance timeframe, recommended repair or replacement measures, required technician skill level, a list of necessary spare parts or specialized tools, and an ideal maintenance window (e.g., avoiding peak business hours). For gradual failures, the maintenance plan favors pre-planned inspections, proactive component replacement, and intensive monitoring of key indicators. For sudden failures, the maintenance plan prioritizes rapid response, emergency repair, fault isolation, and system reconfiguration. Based on the equipment maintenance plan and fault type classification, an agent-based autonomous resource allocation algorithm is executed to analyze maintenance resource requirements. This algorithm enables the AI agent to make autonomous decisions and generate the optimal resource scheduling plan without human intervention based on factors such as equipment status, maintenance priority, personnel skill matching, and spare parts inventory. This includes maintenance team selection, tool and equipment allocation, spare parts and material scheduling, and operation time scheduling, thereby achieving intelligent and precise allocation of operation and maintenance resources.

[0033] In this embodiment, after each equipment maintenance and resource scheduling plan is completed, operational feedback data from various dimensions is collected and organized. This includes user satisfaction ratings submitted through multiple channels, such as apps, phone calls, and service terminals, as well as comprehensive, structured operational process information, including the latest monitored values for equipment status, actual maintenance results (e.g., whether the fault has been completely eliminated and whether the repair time has been met), and resource utilization (e.g., the difference between actual and planned manpower, materials, and work hours). Evaluation indicators are calculated based on this operational feedback data for user satisfaction, equipment status, maintenance effectiveness, and resource utilization. User satisfaction is modeled in multiple dimensions using a combination of ratings, textual reviews, and complaints and suggestions. Equipment status data is compared based on changes in key indicators before and after maintenance. Maintenance effectiveness data is quantified using indicators such as task completion quality, repair success rate, and secondary repair rate. Resource utilization is measured based on actual consumption per task, resource allocation efficiency, and idle or redundant resources. Through automatic statistics, aggregation, and multi-indicator normalization, a comprehensive evaluation of operational effectiveness within the current cycle is obtained. Digital twin modeling is performed on the operational effectiveness evaluation results, creating a bidirectional mapping system between physical equipment and virtual models. The system integrates a physical rule engine, a finite element analysis module, and multi-physics coupled simulation tools to achieve high-precision digital simulation of equipment operating states, providing virtual experimental scenarios and predictive analysis capabilities. Furthermore, it leverages blockchain distributed ledger technology to immutably store and trace all O&M data and operation records. Smart contracts automatically trigger O&M task allocation and service quality assessment, providing reliable proof of O&M data. The current O&M strategy and plan serve as the initial state of a reinforcement learning environment. Within this environment, equipment stability (such as healthy operating days and mean time between failures) and user satisfaction (such as comprehensive evaluation scores and service response speed) are jointly defined as reward functions, reflecting the system's multi-faceted goals of safety, quality, and user experience. The allocation and scheduling of O&M resources (such as staffing, maintenance team scheduling, and spare parts warehouse management) is defined as an action space, making each policy adjustment or resource reorganization a selectable action. By sensing and recording the state changes and reward feedback resulting from different actions, the system gradually constructs an exploration space for intelligent decision-making. In this reinforcement learning environment, a dual temporal difference learning algorithm is employed to split the state-action value function used in traditional reinforcement learning into a value function and an advantage function. These functions capture the inherent value of the current state and the superiority of the target action, respectively, thereby improving the sensitivity and accuracy of strategy selection in complex scenarios. This dual temporal difference method effectively alleviates the overestimation problem in traditional Q-learning. By alternating independent target and evaluation networks, it enables the agent to converge to the optimal strategy more quickly when faced with large-scale, multi-dimensional, and complex spaces.After each round of strategy execution, the system continuously uses the new O&M effectiveness evaluation results as a reference for the next round of decision optimization. It adjusts strategy parameters based on reward signals, achieving a comprehensive shift from experience-driven to data-driven, adaptive, and self-evolving strategies. Through multiple rounds of iterative optimization using reinforcement learning, it develops targeted intelligent O&M strategies that adapt to different device types, business scenarios, and resource constraints.

[0034] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0035] The equipment in the target operation and maintenance area is divided into environmental monitoring equipment, equipment status monitoring equipment, resource consumption monitoring equipment, and security monitoring equipment;

[0036] Set a fixed sampling frequency for environmental monitoring equipment, set a dynamically adjusted sampling frequency for equipment status monitoring equipment, set differentiated sampling frequencies for resource consumption monitoring equipment, and set a real-time sampling frequency for security monitoring equipment to obtain a multi-level data collection strategy;

[0037] Based on a multi-level data collection strategy, IoT devices are sampled and controlled to obtain multi-source heterogeneous raw data. The multi-source heterogeneous raw data is then input into the edge computing gateway for data pre-screening to obtain pre-processed device data.

[0038] The sampling frequency of the pre-processed equipment data is adaptively adjusted to obtain dynamically optimized data collection results, and the dynamically optimized data collection results are associated with the unique identification code of the equipment to obtain a heterogeneous operation and maintenance data set.

[0039] Specifically, all equipment within the operation and maintenance area is functionally classified into environmental monitoring equipment, equipment status monitoring equipment, resource consumption monitoring equipment, and security monitoring equipment. For example, environmental monitoring equipment includes temperature and humidity sensors, light sensors, gas or smoke detectors, etc., designed to capture environmental changes within the area; equipment status monitoring focuses on key equipment such as water pumps, elevators, air conditioners, and transformers. Its sensors cover parameters such as temperature, vibration, current, voltage, and speed, which are related to equipment health and reliability; resource consumption monitoring equipment includes smart meters, water meters, and gas meters, which record energy consumption; security monitoring equipment includes cameras, infrared intrusion detectors, access control systems, etc., focusing on security risk perception and real-time capture of abnormal events. Differentiated data collection frequency strategies are formulated for different types of equipment based on their operating characteristics and business needs. Environmental monitoring equipment operates in a relatively stable environment, but with a long data change cycle, employs a fixed sampling frequency strategy, collecting data every 5 or 10 minutes to capture gradually changing environmental information such as temperature, humidity, and gas concentration. Equipment status monitoring equipment, on the other hand, requires high real-time anomaly detection and health assessment. Its operating parameters fluctuate dramatically under different operating conditions, so it employs a dynamic sampling frequency adjustment mechanism. This mechanism automatically increases the sampling frequency when the device status fluctuates, approaches anomaly thresholds, or has previously detected anomalies. During periods of long-term stable operation, the sampling frequency is appropriately reduced. Resource consumption monitoring equipment is suited to a time-based, differentiated sampling strategy, collecting data at a higher frequency during peak energy consumption periods and at a lower frequency during off-peak periods. Security monitoring equipment implements a near-real-time sampling frequency. Typical scenarios require continuous or high-frequency data uploads, such as cameras continuously streaming video or access control systems reporting every entry and exit event. Based on this multi-level data collection strategy, sampling parameters are distributed to IoT devices, centrally scheduling and controlling device collection behavior in real time, enabling distributed and synchronized data collection for all devices in the area. Large amounts of raw data from various devices, with varying collection frequencies and data types, flow through wireless or wired networks and are uploaded to edge computing gateway nodes. Within the edge computing gateway, this multi-source, heterogeneous raw data undergoes pre-screening, including verification of packet timestamps, source device IDs, and value validity. Data showing obvious anomalies, mutations, packet loss, or redundancy is cleaned. Basic algorithms such as sliding windows, interval checks, and threshold rules are used to filter out invalid data. The pre-processed device data is then output through aggregation, deduplication, and correction. Based on real-time operating status and historical data volatility, the sampling frequency of the pre-processed device data is adaptively adjusted to achieve dynamically optimized data collection results.Combining the data's temporal characteristics and business priorities, the system dynamically monitors equipment operating parameters, analyzing their fluctuations, abnormal trends, and the probability of emergencies. When it detects a change in equipment status, intensified data changes, or a signal that significantly deviates from the historical normal range, the system automatically increases the sampling frequency of that node, achieving more intensive status capture and rapid response. When the equipment status is stable for a long time and data fluctuations are small, the system proactively reduces the sampling frequency to maximize savings in transmission bandwidth and computing storage resources, achieving simultaneous optimization of data collection efficiency and energy utilization. The dynamically optimized data collection results are associated with the equipment's unique identification code, and all data is normalized and stored according to standard data formats and protocols to obtain a heterogeneous operation and maintenance data set.

[0040] Specifically, all equipment in the operation and maintenance area is functionally classified and divided into environmental monitoring equipment, equipment status monitoring equipment, resource consumption monitoring equipment, and security monitoring equipment. In the equipment deployment stage, a genetic algorithm is used in combination with geographic information system (GIS) technology to plan the three-dimensional deployment plan of IoT equipment. The algorithm takes maximizing monitoring coverage and minimizing equipment deployment costs as dual objective functions, encodes factors such as building structure, personnel flow, and equipment thermal distribution as constraints, and searches for the optimal deployment location through iterative evolution. The algorithm first generates a random initial population, each individual represents an equipment layout plan, and then continuously optimizes the layout plan through genetic operations such as fitness evaluation, selection, crossover, and mutation, and finally converges to the Pareto optimal frontier, achieving a 25% optimization of equipment deployment density and a 99% elimination rate of monitoring blind spots, greatly improving monitoring efficiency and resource utilization. In this embodiment, in the data preprocessing stage, an abnormal data repair algorithm based on generative adversarial networks (GAN) is innovatively introduced to intelligently repair missing, abnormal, or noisy data generated during the monitoring process. This algorithm consists of a generator and a discriminator. The generator generates possible missing data values based on contextual information, while the discriminator evaluates the rationality of the generated values. Through adversarial training, the restored data distribution approximates the true data distribution, improving data availability by 35%. During multimodal data fusion, a capsule neural network is used to extract key features from each modality, overcoming the drawback of traditional convolutional neural networks where pooling operations often lose spatial hierarchical relationships. A dynamic routing mechanism preserves geometric relationships between features, constructing a more complete multimodal feature representation. This improves feature extraction accuracy by 40%, laying a solid foundation for subsequent in-depth analysis.

[0041] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0042] According to the communication requirements of different devices in the heterogeneous operation and maintenance data set, the device communication channels are allocated through a multi-channel hybrid networking system to obtain a hybrid heterogeneous network configuration solution;

[0043] Build a building operation and maintenance prompt project based on heterogeneous operation and maintenance data, and use a unified building operation and maintenance service framework coordinated by multiple MCPs to generate intelligent operation and maintenance data;

[0044] Based on hybrid heterogeneous network configuration solutions and intelligent operation and maintenance data, the AES-128 encryption algorithm is applied at the device layer, the TLS 1.3 protocol is applied at the transport layer to achieve end-to-end encryption, and identity authentication and access control mechanisms are applied at the application layer to obtain encrypted data with multi-layer protection.

[0045] Perform data transmission priority analysis on the encrypted data with multiple layers of protection to obtain prioritized transmission data, and then input the prioritized transmission data into the server for data cleaning to obtain a purified data set;

[0046] Data standardization is performed on the purified data set to obtain a standardized data set, and feature extraction is performed on the standardized data set to obtain an operation and maintenance feature data set.

[0047] Specifically, a communication network architecture is planned based on the communication requirements of different devices in heterogeneous O&M datasets. Communication parameters are analyzed for each device type, and communication requirements are stratified based on their data volume, real-time requirements, power consumption constraints, and service priorities. For example, security cameras have extremely high requirements for high bandwidth and low latency, while environmental monitoring sensors prioritize stable connections with low power consumption, long distances, and low bandwidth. Based on this analysis, multi-channel hybrid networking technology is employed to integrate heterogeneous communication methods such as NB-IoT, LoRa, Wi-Fi, ZigBee, 5G, and wired networks. A hybrid network configuration with wide coverage, high bandwidth, low latency, and controllable energy consumption is dynamically constructed. This enables hierarchical scheduling of high-bandwidth, high-real-time data streams (such as video streams) and low-power periodic data collection. This allows for flexible response to changing O&M scenarios, resulting in adaptive, load-balanced network resource allocation strategies. The Model Context Protocol (MCP) is applied to standardize data encoding for heterogeneous O&M datasets, establishing a unified communication framework between devices. As a "common language" between industrial and IoT devices, the MCP protocol effectively resolves communication barriers between disparate devices, manufacturers, and protocols. Through standardized data format definitions and transmission rules, it enables seamless data transfer to a cloud-based time-series database, facilitating the construction of a unified data analysis platform and virtual-to-real mapping of device status, resulting in intelligent operations and maintenance data. Comprehensive security protection is provided throughout the data lifecycle. At the device level, each acquisition terminal encrypts raw data using the AES-128 encryption algorithm upon generation, ensuring high confidentiality and immutability even in the event of attacks on edge nodes, gateways, or transmission links. At the data transmission layer, all communications between devices and servers are end-to-end encrypted using the TLS 1.3 protocol. Leveraging dynamic key negotiation and high-strength encryption algorithms, this ensures data integrity, confidentiality, and resistance to man-in-the-middle attacks during transmission. At the application layer, fine-grained authentication and access control mechanisms are implemented. Users and system processes at different levels can only access, process, or forward designated data resources after passing authentication and authorization, eliminating the risk of internal unauthorized access and providing traceability throughout the entire data flow and operation process. Through this step, encrypted data with multi-layer protection is obtained. Multi-layer encrypted data streams are hierarchically managed and sorted according to business priorities. Different transmission priorities are assigned to each type of data stream based on device type, data content, business urgency, and operation and maintenance strategy requirements. For example, critical information such as security alarms and equipment anomalies has the highest priority, followed by routine environmental monitoring data, and historical archives and periodically collected data have the lowest priority. The priority scheduling mechanism is used to dynamically adjust the forwarding of data packets and bandwidth allocation to ensure that high-priority data can reach the central server with the lowest latency and highest reliability, thereby improving the operation and maintenance system's response capabilities to sudden and urgent events and reducing the probability of major risks.After the prioritized encrypted data arrives at the server, it undergoes intelligent data cleansing. This collaborative multi-algorithm cleansing process includes statistically based outlier detection (such as Z-score analysis), filtering and correction of time series fluctuation trends (such as Kalman filtering), missing value filling (such as interpolation algorithms and historical data inference), and redundant data deduplication. This automatically removes noise, errors, duplications, anomalies, or damaged data during transmission, resulting in a cleansed dataset. Data standardization is then performed on the cleansed dataset, including data normalization (such as unifying various physical quantities to a 0-1 range or a normal distribution), unit conversion (such as standardizing physical quantities like pressure, temperature, and current), and data structure adjustment (such as filling missing fields and unifying timestamp formats). This also includes standard coding and classification of core tags such as devices, collection points, and data types, resulting in a standardized dataset. Feature extraction is performed on the standardized dataset to improve the performance of subsequent intelligent analysis and decision-making. Feature extraction includes time-domain feature analysis (such as mean, variance, extreme value, and rate of change), frequency-domain feature analysis (such as Fourier transform to extract dominant frequency components and wavelet transform to capture signal mutation characteristics), spatial feature induction (such as geographic location and device topology), historical trend modeling (such as autoregressive coefficients and moving averages), and statistical anomaly indicator extraction (such as confidence intervals, skewness, and kurtosis). In line with actual O&M needs, we leverage business knowledge and AI algorithms to perform feature selection and dimensionality reduction on high-dimensional data, focusing on the most discriminative and representative features as the core content of the O&M feature dataset.

[0048] In this embodiment, during data transmission and processing, the innovative application of a federated learning algorithm breaks down data silos in IoT devices, enabling each device to jointly train models without sharing raw data, thus protecting data privacy and improving algorithm performance. In specific implementations, the edge computing gateway performs local model training and only uploads model parameters, not raw data, to the central server. The server aggregates the parameters of each node and distributes the updated model, enabling distributed collaborative learning. At the same time, quantum encryption technology is introduced to ensure data transmission security, and a quantum key distribution mechanism is used to ensure the absolute security of encryption keys. This ensures high security even under attacks with extremely powerful computing power, raising the data transmission security level to the level of quantum protection and providing cutting-edge protection for industrial-grade data security.

[0049] In a specific embodiment, the execution step allocates device communication channels through a multi-channel hybrid networking system based on the communication requirements of different devices in the heterogeneous operation and maintenance data set to obtain a hybrid heterogeneous network configuration solution, which may specifically include the following steps:

[0050] Analyze the communication parameters of different devices in the heterogeneous operation and maintenance data set to obtain the classification results of device communication requirements;

[0051] Based on the classification results of device communication requirements, the transmission capacity of six communication technologies, including NB-IoT, LoRa, Wi-Fi, ZigBee, 5G, and wired networks, was evaluated. Communication channel characteristic data including bandwidth parameters, delay parameters, coverage parameters, and energy consumption parameters were established through a multi-channel hybrid networking system.

[0052] Perform device channel matching on the device communication demand classification results and the communication channel characteristic data to obtain a matching score. Based on the matching score, the optimal communication channel is allocated to each type of device to obtain an initial channel allocation plan.

[0053] Perform network load analysis on the initial channel allocation plan to obtain a load-balanced channel allocation plan, and set differentiated communication strategies based on the load-balanced channel allocation plan;

[0054] Integrate differentiated communication strategies with load-balanced channel allocation schemes to generate hybrid heterogeneous network configuration solutions.

[0055] Specifically, a parameter analysis is conducted on the communication requirements of different devices in the heterogeneous operation and maintenance dataset. This analysis covers multiple dimensions, including the type of data generated by the devices, data volume, collection and reporting frequency, service priority, real-time and bandwidth requirements, energy consumption tolerance, and the distribution of device physical deployment. Through data-driven statistical modeling and hierarchical cluster analysis, all devices are classified into several categories based on their communication behavior and operation and maintenance scenarios, including high-real-time and high-bandwidth types, periodic low-data-volume types, ultra-low-power and long-distance types, and high-priority security alarm types. The resulting classification of device communication requirements is then output. Based on the communication requirements classification results, a multi-dimensional, quantitative assessment of the transmission capabilities of six major communication technologies: NB-IoT, LoRa, Wi-Fi, ZigBee, 5G, and wired networks is conducted. Each communication technology has unique performance parameters. For example, NB-IoT and LoRa prioritize low power consumption, wide coverage, and suitability for long-term, low-volume data transmission. Wi-Fi and ZigBee are suited to short- to medium-range, high-density scenarios. 5G offers ultra-high bandwidth and extremely low latency to meet real-time requirements for high-definition video streaming and critical control. Wired networks, with their maximum bandwidth and lowest packet loss, are suitable for critical infrastructure and data centers. Through experimental measurements and comparison with standard parameters, combined with actual deployment environments (such as signal strength distribution, interference levels, coverage gaps, power consumption, and battery life constraints), the actual capabilities of each channel in terms of bandwidth, latency, coverage, and energy consumption are collected and summarized. A communication channel characteristic database is established using structured data storage. Device-channel matching is performed by categorizing device communication requirements and matching the communication channel characteristic data. A weighted scoring method based on multi-dimensional attributes is used to quantify the compatibility between each device category and each channel, generating a matching score matrix. For example, for security video surveillance systems requiring high bandwidth and low latency, 5G and wired networks are the best match. For widely distributed, power-constrained environmental monitoring devices, LoRa and NB-IoT offer the best performance. For office automation sensors requiring flexible and low-cost networking, Wi-Fi or ZigBee are the right choice. Based on a compatibility score, the system automatically assigns the optimal communication channel to each device type, forming a preliminary channel allocation plan. This initial channel allocation plan undergoes network load analysis. During the load analysis phase, the system dynamically assesses the current load pressure on each channel by monitoring data traffic, number of connected devices, current throughput, latency jitter, and the frequency of abnormal events in real time, using techniques such as queueing theory, network simulation, and big data stream visualization. If a channel is overloaded (e.g., high 5G link utilization during peak video surveillance hours) or if certain channels have redundant resources, the system automatically adjusts the channel allocation of some devices, migrating some devices from overloaded channels to less-loaded channels to achieve global load balancing. Based on the load-balancing channel allocation results, differentiated communication strategies are designed to adapt to device service diversity and network dynamics.This includes setting different data packet priorities, bandwidth occupancy ratios, retransmission mechanisms, and sleep and wake-up strategies for different types of devices. For example, critical alarm and security event data can be assigned higher transmission priority, with instant push and redundant confirmation mechanisms enabled. Periodic data streams such as environmental monitoring that tolerate a certain degree of delay are then batched and reported in aggregate at regular intervals to reduce link occupancy and power consumption. For some high-value or special scenario equipment (such as power and critical medical infrastructure), redundant links and multi-channel concurrent backup strategies are established to prevent security risks caused by communication interruptions at critical moments. Dynamically adjust the modulation method, channel parameters, and device adaptive transmission power in response to network fluctuations and signal interference. The above-mentioned differentiated communication strategies are integrated with a channel allocation solution optimized for load balancing to generate a hybrid heterogeneous network configuration solution.

[0056] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0057] Collect user feedback from multiple channels, including mobile app feedback, voice assistant interactions, customer service call records, and work order system data. Combine this user feedback with the operation and maintenance feature dataset to obtain a multimodal user demand dataset.

[0058] Input the text data in the multimodal user demand dataset into the Mamba model in the deep learning intent recognition model for entity recognition to obtain entity annotation data;

[0059] Input the entity annotation data into the LLM-Fine-Tuning model in the deep learning intent recognition model for intent classification to obtain the intent classification results;

[0060] Build an entity relationship graph based on intent classification results and entity annotation data to obtain a structured knowledge network;

[0061] Based on the structured knowledge network and user historical behavior data, the mapping relationship between user expressions and actual needs is extracted to obtain structured user demand data. The urgency score is calculated based on the structured user demand data to generate the priority ranking of operation and maintenance tasks.

[0062] Specifically, a multi-faceted information collection system is constructed to continuously collect user feedback from multiple channels, including mobile app user feedback, voice assistant conversations, customer service call logs, and operational tasks flowing through the work order system. Each type of feedback data has its own unique information density, expression style, and business scenario. For example, app feedback often includes proactive repair requests or service reviews, while voice assistants tend to use natural language descriptions. Customer service call logs contain interaction details and user sentiment, and work order system data reflects historical problem resolution processes and outcomes. During the collection process, information from all channels is uniformly formatted, timestamped, source identified, and initially cleaned. Simultaneously, user feedback is deeply integrated with operational feature datasets collected on-site (such as equipment status, environmental parameters, and historical anomalies) to generate a multimodal user demand dataset covering text, voice, and structured task flows. The textual data from this multimodal user demand dataset is fed into the Mamba model within the deep learning intent recognition model for entity recognition. As a novel architecture for sequence modeling, the Mamba model transcends the limitations of traditional RNNs and Transformers, achieving linear complexity in long-text processing through a structured state-space model (SSM). The model first preprocesses the text, then performs sequence modeling using a selective state-space layer (S6), selectively encoding the input and effectively capturing long-range semantic dependencies. During the parallel sequence processing phase, Mamba converts sequence information into a state representation, capturing temporal patterns and key features through the structured state-space model to form a context-aware feature representation. Subsequently, the model performs feature filtering and enhancement through a complex gating network, dynamically adjusting the importance of information at different positions and extracting key information from the text to form a set of entity candidate features. A self-attention mechanism layer is introduced for global information integration, further enhancing the perception of potential entity locations and accurately identifying entity boundaries. Finally, the system combines the Mamba model's predictions with annotation transfer rules and uses the Viterbi algorithm to search for the optimal annotation path, ensuring overall annotation logic consistency and generating high-quality entity annotation data. This entity annotation data is then fed into the LLM-Fine-Tuning model for intent classification. The method first performs text reconstruction and prompt word engineering optimization on the entity annotation data to construct enhanced semantic input sequences. These sequences are then input into a pre-trained large language model, and deep semantic features are extracted through a multi-layer Transformer architecture to obtain context-aware representation vectors. Based on specific operation and maintenance tasks, the system designs a domain-oriented fine-tuning strategy, using efficient parameter fine-tuning techniques (such as LoRA and P-Tuning) to update only some key parameters of the large language model, retaining the model's general language understanding capabilities. During the fine-tuning process, domain expert knowledge and historical failure cases are introduced to build an intent recognition knowledge base in the operation and maintenance field.Unlike traditional single-intent recognition, the system uses multi-label classification to simultaneously identify both primary and secondary intents, more comprehensively capturing complex user needs. Finally, through confidence threshold screening and intent conflict resolution, combined with operational and maintenance business rules, intent normalization is performed to achieve accurate intent classification results. An entity relationship graph is constructed based on the intent classification results and entity annotation data. Utilizing automated knowledge extraction tools and ontology constraints, all identified entities, their attributes, event relationships, and spatial and temporal information are combined with pre-set business logic and relationship templates to construct a multi-level, queryable, and reasonable entity relationship graph. For example, a multi-dimensional association network is automatically constructed, such as "device-location-fault phenomenon-repair time-historical processing record-responsible person." Based on this, deep neural networks and graph mining algorithms are used to extract the mapping between user expressions and actual needs, combining structured knowledge networks with historical user behavior data. By modeling user information such as past repair frequency, need type, service satisfaction, response time, and historical resolution results, user profiles and behavioral preferences are formed, enabling personalized need understanding and service optimization. The model also clusters diverse user demand expressions for the same device and mines their urgency, automatically adapting to service decisions across multiple scenarios. Based on this structured user demand data, an urgency score is automatically calculated for each demand. The scoring model integrates multiple metrics, including device importance, fault impact, duration, security risk, historical priority, and user-specific attributes. Through weighted aggregation or multi-layer neural network regression, it converts each demand into a quantifiable urgency score. After all demand scores are ranked, a priority queue for maintenance tasks is output, enabling tiered scheduling and dispatching based on demand, risk, and impact. The system further integrates the Transformer-XL architecture with an active learning framework to enhance intent recognition performance. The Transformer-XL architecture expands the receptive field of contextual dependencies and can handle extremely long text sequences, effectively addressing the difficulty of understanding long texts such as service reports and maintenance records. Through inter-segment recursion and relative position encoding, it maintains coherent understanding of long texts. Working in conjunction with the Mamba model, intent recognition accuracy is improved by 15%. At the same time, an active learning framework based on uncertainty is constructed. The system automatically identifies samples with low confidence and fuzzy boundaries, actively requests labeling from human experts, and prioritizes learning the most informative and difficult examples, achieving a 300% increase in model update and iteration efficiency, enabling the system to continuously adapt to new equipment, new failure modes, and ever-changing ways of expressing user needs.

[0063] In a specific embodiment, the step of inputting text data in the multimodal user demand dataset into the Mamba model in the deep learning intent recognition model for entity recognition to obtain entity annotation data may specifically include the following steps:

[0064] Perform word segmentation and cleaning on the text data in the multimodal user demand dataset to obtain the processed text sequence;

[0065] The processed text sequence is input into the selective state space layer of the Mamba model for sequence modeling to obtain long-distance dependency feature representation;

[0066] Perform parallel sequence processing based on long-distance dependency feature representation, capture temporal patterns and key features in text through structured state space model, and obtain context-aware feature representation;

[0067] Context-aware feature representations are filtered and enhanced using a gating mechanism to extract key information from the text and obtain a candidate feature set for the entity.

[0068] The entity candidate feature set is input into the self-attention mechanism layer for global information integration, the potential entity location is accurately identified, and the entity boundary probability distribution is obtained;

[0069] Execute sequence labeling decisions based on the probability distribution of entity boundaries, calculate the optimal labeling path through the Viterbi algorithm, and obtain the globally optimal entity recognition result;

[0070] The globally optimal entity recognition results are post-processed and entity type normalized to obtain entity annotation data.

[0071] Specifically, the text data in the multimodal user demand dataset is segmented and cleaned. A segmentation tool based on a business dictionary and a general segmentation model is used to break the text data into its smallest semantic units. During this process, a custom domain vocabulary and device model dictionary are combined to identify common nouns, verbs, and adjectives, and to capture entities closely related to operation and maintenance scenarios, such as professional terminology, brand models, fault symptoms, geographic locations, and time expressions. The cleaning process removes stop words, punctuation, invalid numbers, and noise symbols, and normalizes and corrects common typos, mixed formats, and other anomalies. After segmentation and cleaning, a processed text sequence is obtained. This processed text sequence is input into the selective state space layer of the Mamba model for sequence modeling. As a next-generation architecture in the field of sequence modeling, the Mamba model breaks through the limitations of traditional RNNs, CNNs, and Transformers, achieving linear complexity long text processing through a selective state space model. The model accepts input sequences using a scanning mechanism and leverages an innovative selective state-space representation to dynamically weight the importance of input information at different positions, enabling the model to selectively retain and process key information while filtering out irrelevant content. This mechanism is particularly well-suited for processing lengthy texts such as equipment reports, maintenance records, and user feedback in the field of operations and maintenance. It efficiently extracts feature representations of long-range semantic dependencies and excels at capturing long-term information, such as fault descriptions spanning multiple paragraphs and changes in equipment status. Based on this long-range dependency feature representation, the model performs parallel sequence processing. Unlike the sequential computation of traditional recurrent neural networks, the Mamba model utilizes an innovative parallel computing architecture to simultaneously process the input sequence through state-space transformations, significantly improving computational efficiency. The structured state-space model treats each position in the sequence as a point in state space, capturing complex temporal dynamics through state transition functions and output mapping functions. This architecture excels at modeling temporal patterns and long-range dependencies in text, making it particularly well-suited for capturing temporal information such as the development of equipment failures and evolving user needs in operations and maintenance texts. It generates context-aware feature representations, enabling the model to gain a holistic understanding of the complete text context. A gating mechanism filters and enhances context-aware feature representations. The Mamba model integrates a complex gating network that dynamically adjusts the importance of features at each position through trainable gating parameters. This gating mechanism comprehensively considers the semantics and context of the current word, highlighting the representation of keywords and professional terms while suppressing the influence of secondary information. In maintenance scenarios, the system can accurately identify key fault descriptions such as "elevator vibration abnormality" and "temperature 5°C higher" while deemphasizing irrelevant modifiers and background information. This generates a highly focused set of entity candidate features, laying the foundation for subsequent accurate entity recognition. The entity candidate feature set is input into the self-attention mechanism layer for global information integration.While the Mamba model itself is capable of capturing long-range dependencies, introducing a lightweight self-attention mechanism can further enhance the model's sensitivity to entity boundaries in the delicate task of identifying entity boundaries. The self-attention layer calculates correlation scores between positions in the sequence and, through weighted aggregation of global information, enables the model to more accurately determine the start and end positions of entities. For example, in the description "Transformer main unit cooling fan anomaly," the self-attention mechanism accurately defines "transformer main unit" as the equipment entity, "cooling fan" as the component entity, and "anomaly" as the state entity, generating a probability distribution for entity boundaries and achieving precise entity boundary identification. Sequence labeling decisions are then made based on the entity boundary probability distribution. The system utilizes the combined output of the Mamba model and the self-attention layer to construct a sequence labeling decision graph, which contains the probability scores for each position under each label and the transition probabilities between adjacent labels. Using the Viterbi algorithm to dynamically program and search this decision graph, the system efficiently finds the globally optimal labeling path, ensuring the logical coherence and consistency of the labeling sequence. For example, avoid unreasonable combinations such as "B-device I-location" in the labeling rules, ensure the integrity and semantic consistency of each entity label sequence, and thus obtain the globally optimal entity recognition result, greatly improving the accuracy and robustness of entity recognition. The globally optimal entity recognition result is post-processed and entity type normalization is performed. The post-processing step includes steps such as fine-tuning entity boundaries, nested entity processing, homonym resolution, and entity attribute completion. Entity type normalization maps the various identified entities to a unified business ontology system. For example, different expressions such as "thermometer", "thermometer", and "temperature sensor" are uniformly classified into the "temperature sensor" entity type. This step ensures that downstream tasks can process entity data with a unified format and clear semantics, providing high-quality basic data input for subsequent intent classification and knowledge graph construction, forming a complete entity annotation dataset.

[0072] In a specific embodiment, the execution step of inputting the entity annotation data into the LLM-Fine-Tuning model in the deep learning intent recognition model for intent classification, and the process of obtaining the intent classification result may specifically include the following steps:

[0073] Perform text reconstruction and prompt word engineering optimization on entity annotation data to obtain enhanced semantic input sequences;

[0074] The enhanced semantic input sequence is fed into the LLM-Fine-Tuning model, and deep semantic features are extracted through a multi-layer Transformer architecture to obtain a context-aware representation vector.

[0075] Design a domain-specific fine-tuning strategy based on the target operation and maintenance tasks, perform gradient updates on the large language model parameters, build intent recognition knowledge in the operation and maintenance domain, and obtain a domain-adaptive fine-tuning model.

[0076] The domain-adapted fine-tuning model is used to perform multi-label classification on the input text, identifying the primary and secondary intents in the text and obtaining the intent probability distribution.

[0077] Confidence threshold screening and intent conflict resolution are applied to the intent probability distribution, and intent normalization is performed in combination with operation and maintenance business rules to obtain the intent classification result.

[0078] Specifically, the entity annotation data undergoes text reconstruction and prompt word engineering optimization. The system first structures the identified entities and their type information. For example, entity annotation data such as "Component: Water Pump," "Location: Building 3," and "Status: Abnormal Vibration" are reorganized according to professional domain templates. During the prompt word engineering phase, the system designs specific prompt word templates based on task characteristics, such as directive prompts like "Analyze the following equipment repair request and identify the user's core intent: [Entity Annotation Text]" or "This is an O&M requirement containing [Entity List]; please determine the user's primary and secondary intent." At the same time, based on historical cases and domain knowledge, the system injects key contextual information into the prompt words, such as auxiliary information such as common equipment failure types, standard O&M procedures, or common user expressions. This creates an information-rich and structured enhanced semantic input sequence, enabling the large language model to better understand the O&M needs of a professional domain. The enhanced semantic input sequence is input into the pre-trained large language model. Unlike traditional models such as BERT, modern large language models have a more powerful parameter scale and pre-training breadth, usually containing billions to hundreds of billions of parameters, and have been pre-trained on massive general corpora and multi-domain professional texts. When the input sequence enters the model, it is processed by a multi-layer Transformer architecture. Each layer can capture semantic features at different levels, from low-level lexical and syntactic information, to mid-level semantic associations, and then to high-level logical reasoning and domain knowledge. The model can automatically process very long texts without explicit positional encoding management, and has excellent understanding of complex multi-round conversations and detailed equipment descriptions. Through the self-attention mechanism, the model establishes a global association for each element in the input sequence, forming a context-aware representation vector that contains rich contextual information. These vectors retain the complete semantic information and implicit intent indications of operation and maintenance requirements. Design domain-oriented fine-tuning strategies based on specific operation and maintenance tasks. Considering the massive parameter size and rich general knowledge of large language models, the system employs parameter-efficient fine-tuning techniques (PEFT), such as low-rank adaptation (LoRA), prompt tuning (P-Tuning), or adapters. These methods update only a small number of key model parameters or introduce a small number of additional trainable parameters, while keeping most pre-trained parameters frozen. This approach not only significantly reduces computing resource requirements but also effectively prevents overfitting and preserves the model's general language understanding capabilities. During fine-tuning, the system uses a carefully labeled data set of operations and maintenance domain intent for supervised learning. It also incorporates domain expert knowledge and historical failure cases as auxiliary training signals, gradually building specialized intent recognition capabilities in the operations and maintenance domain through gradient updates. After fine-tuning, the model retains its original language understanding foundation while gaining a deep understanding of operations and maintenance terminology, industry processes, and equipment knowledge, resulting in a highly domain-adaptive fine-tuned model. This domain-adapted fine-tuned model performs multi-label classification on input text.Unlike traditional single-intent classification methods, the large language model, leveraging its powerful semantic understanding capabilities, can simultaneously identify multiple levels of intent within user expressions. For example, in a request such as "The elevator fault light is flashing, requiring urgent repair, and I also wish to know the last maintenance record," the system can simultaneously identify "report repair" (the primary intent) and "query history" (the secondary intent). The model uses a multi-head attention mechanism to focus on the importance of different semantic segments and leverages contextual understanding to capture implicit expressions. Ultimately, it assigns a probability value to each predefined intent category (such as report repair, consultation, complaint, suggestion, and emergency assistance), forming a complete intent probability distribution. This multi-label approach enables a more comprehensive understanding of complex user needs, providing a more comprehensive decision-making basis for subsequent task assignment and resource scheduling. The intent probability distribution is filtered and conflicting intents are resolved using a confidence threshold. The system sets a dynamically adjusted confidence threshold γ (typically in the range of 0.3-0.7), retaining only intent predictions with probabilities exceeding the threshold and filtering out low-probability intent guesses to ensure reliable results. The system also incorporates an intent conflict resolution mechanism, automatically identifying and resolving conflicting intent combinations using a predefined intent logical relationship matrix. For example, mutually exclusive intents such as "expedited processing" and "deferred processing," or "on-site repair" and "remote guidance" cannot exist simultaneously. The system uses logical rules or probabilistic advantages to retain the most reasonable intent combination. Finally, the system normalizes intents based on operational and maintenance business rules, mapping diverse user expressions into a standardized business intent classification system to ensure consistent understanding across downstream tasks. For example, expressions such as "broken," "faulty," and "not operating" are all normalized into the intent of "equipment fault repair." Through this series of processes, the system ultimately achieves accurate, comprehensive, and business-friendly intent classification results.

[0079] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0080] Based on the operation and maintenance feature dataset, IoT devices are organized into functionally complementary logical monitoring units. A device status feature vector containing operating parameters, environmental parameters, and historical fault information is constructed for each logical monitoring unit.

[0081] The device state feature vector is input into a hybrid model composed of a multi-layer perceptron and a support vector machine for anomaly detection to obtain a preliminary abnormal state identification;

[0082] Based on the preliminary abnormal state identification, a multi-dimensional device association topology of physical space and functional relationship is constructed to obtain a device association map;

[0083] Based on the device association map and structured user demand data, the surrounding devices with correlation exceeding the preset threshold are activated to perform collaborative monitoring and obtain multi-source verification information;

[0084] The multi-source verification information is input into the dynamic Bayesian network for probabilistic reasoning calculation to obtain the posterior probability distribution of the true state of the target equipment. Based on the posterior probability distribution, the equipment failure risk assessment is performed to obtain failure risk warning data.

[0085] Specifically, based on the operation and maintenance feature dataset, the large number of IoT devices within the entire operation and maintenance area are grouped and organized. By conducting a multi-dimensional analysis of each device's functional complementarity, business relevance, and data collaboration requirements, devices of different categories and uses, but with inherent synergies in actual operation and maintenance scenarios, are grouped into several logical monitoring units. A device state feature vector is constructed for each logical monitoring unit, containing operating parameters, environmental parameters, and historical fault information. Real-time operating parameters, such as current, voltage, temperature, pressure, and vibration frequency, reflect the device's current operating status and health. Environmental parameters, such as temperature and humidity, gas composition, noise levels, and ambient lighting in the monitored area, reveal the potential impact of external environmental changes on device operation. Historical fault and maintenance records include past device anomaly reports, repair times, fault types, repair results, part replacements, and maintenance cycles. Through data fusion, feature engineering, and historical information encoding, this data is integrated into a structured device state feature vector. This device state feature vector is input into a hybrid model consisting of a multilayer perceptron and a support vector machine for anomaly detection. The multi-layer perceptron uses a deep neural network structure to adaptively learn nonlinear, complex feature relationships, automatically extracting implicit abnormal feature patterns from large-scale, high-dimensional data. The support vector machine, with its high-dimensional spatial segmentation and outlier identification capabilities, models abnormal boundaries within device states. The two work together to balance complex feature representation and boundary identification, improving the accuracy of abnormal state detection under various operating conditions and effectively reducing false positives and negatives. After model processing, a preliminary abnormal state indicator is generated for each logical monitoring unit, clearly indicating which units or devices are experiencing operational anomalies, data fluctuations, or potential risks. Based on these preliminary abnormal state indicators, a multi-dimensional device association topology is constructed, reflecting physical and functional relationships. Based on factors such as the physical location of the current device, business processes, and functional relationships, a multi-dimensional device association topology is automatically constructed. This device association topology includes physical proximity (e.g., devices in the same room, circuit, or area are susceptible to each other) and functional and logical coupling (e.g., dependencies between front-end monitoring devices and back-end execution units, energy supply and consumption devices, and control and controlled devices). Using a graph modeling approach, devices are treated as nodes, and physical and functional dependencies as edges. Weight parameters are used to characterize the strength of associations between devices to obtain a device association graph. Based on the device association graph, combined with structured user demand data (such as high-priority areas, devices or workstations targeted by user repair reports, etc.), peripheral devices whose association with the initial abnormal device exceeds a preset threshold are automatically activated for collaborative monitoring. When a logical monitoring unit or device is detected to be abnormal, the system analyzes its first- and second-degree adjacent nodes based on the graph, and combines historical interactions and functional flows to select those device nodes that are closely related in space or function to simultaneously collect status information, environmental parameters, and abnormal signals to form multi-source, heterogeneous verification data. The multi-source verification information is input into a dynamic Bayesian network for probabilistic reasoning calculations.Dynamic Bayesian networks (DBNs) are capable of modeling time series, conditional probabilities, and causal relationships. Given partial observations, they infer the posterior probability distribution of a target device in various states (normal, warning, fault, etc.). By considering the current device's and neighboring device's state information, historical state transition probabilities, and anomaly development paths, the network enables multi-faceted probabilistic reasoning about the device's true health. Using the posterior probability distribution as the core, combined with comprehensive indicators such as historical failure rates, anomaly impact range, and service priority, it automatically generates a quantitative device failure risk assessment.

[0086] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0087] Perform fault type classification on the fault risk warning data to obtain fault type classification results, which include gradual faults and sudden faults;

[0088] Construct an agent-based artificial intelligence multi-objective task model, set fault repair rate, resource utilization efficiency, and time response speed as optimization goals, and obtain AI agent decision parameters;

[0089] For gradual faults, a long short-term memory network and autoregressive integrated moving average hybrid model are used for time series prediction. For sudden faults, a random forest and gradient boosting decision tree hybrid model are used for classification prediction to obtain fault development trend prediction results.

[0090] Calculate the fault risk index based on the fault development trend prediction results and the operation and maintenance task priority ranking to obtain a quantitative risk index. Then generate an equipment maintenance plan based on the quantitative risk index and AI agent decision parameters.

[0091] Based on the equipment maintenance plan and fault type classification results, an agent-based autonomous resource allocation algorithm is executed to analyze the maintenance resource demand and generate a resource scheduling plan.

[0092] Specifically, fault type classification is automatically performed based on the abnormal status and risk probability provided by fault risk warning data, combined with multi-dimensional equipment operating information and historical O&M data. By analyzing monitoring data, it clearly distinguishes which equipment is currently experiencing gradual faults (such as continuous performance degradation or gradual deviation from normal ranges) and which equipment is experiencing sudden faults (such as instantaneous component failure or the sudden appearance of abnormal signals). These classification results are then standardized and output. An agent-based artificial intelligence (Agentic AI) multi-objective task model is constructed, enabling AI agents to perceive the environment, set goals, plan paths, and execute autonomously. This model optimizes fault repair rate, resource utilization efficiency, and time response speed. Through large-scale parameter learning and scenario adaptation, it develops an intelligent agent system capable of making independent decisions in different O&M scenarios. The model combines historical O&M experience with current scenario constraints to automatically calculate optimal decision parameters for subsequent maintenance plan generation and resource scheduling. Adaptive intelligent prediction algorithms are used for different fault categories. For gradual faults, a hybrid time series prediction model is constructed using a long-short-term memory network and an autoregressive integrated moving average model. Long-short-term memory networks (LSTMs) have strong representation capabilities for long-term historical data, complex time-series dependencies, and nonlinear trends, enabling them to capture subtle changes in equipment operating status over time and long-term degradation characteristics. The autoregressive integrated moving average model, with its efficient modeling capabilities for short-term trends and linear changes, enables adaptive smoothing and trend decomposition of sudden fluctuations. Combining these two models allows the system to provide highly accurate numerical predictions of the future development trends of gradual faults, determine how long it will take for equipment to enter a high-risk state, and infer the rate of fault deterioration and remaining lifespan. For sudden faults, a hybrid classification model of random forests and gradient boosted decision trees is used. Random forests, by constructing an ensemble of multiple decision trees, exploit high-dimensional features and nonlinear relationships between complex features to achieve highly robust classification. This model is suitable for capturing state changes before and after sudden events, abnormal signal distributions, and multi-source alarm patterns. The gradient boosted decision tree model, using an additive model and residual fitting mechanism, accurately identifies low-probability, high-impact, and targeted sudden fault patterns, enhancing the system's ability to detect atypical and hidden sudden faults. The two models work together to efficiently identify sudden equipment failures and predict their development trends. This effectively avoids false alarms and missed alerts in scenarios such as multi-source data integration, complex operating conditions, and weak fault signals. A failure risk index is calculated based on the failure trend prediction results and the prioritization of maintenance tasks. This prioritization is derived from preliminary intent identification, task diversion, and business impact assessment, taking into account factors such as the device's importance to the overall system, current business load, user demand urgency, and the scope of the failure. The system combines these indicators to quantitatively calculate the failure risk index for each fault point.The risk index model integrates multiple dimensions of data, including the current equipment status, future failure rate, historical failure probability, business weight, geographic impact, and repair difficulty. Using weighted aggregation or deep regression, it outputs a quantitative risk score between 0 and 1. A higher score indicates a greater risk to the equipment, a more urgent threat to the business, and a greater need for O&M resources. Based on the risk index and AI agent decision parameters, an optimal equipment maintenance plan is automatically generated. This plan includes details such as the maintenance target, estimated maintenance timeframe, recommended repair or replacement measures, required technician skill level, a list of necessary spare parts or specialized tools, and an ideal maintenance window (e.g., avoiding peak business hours). For gradual failures, the maintenance plan favors pre-planned inspections, proactive component replacement, and intensive monitoring of key indicators. For sudden failures, the maintenance plan prioritizes rapid response, emergency repair, fault isolation, and system reconfiguration. Based on the equipment maintenance plan and fault type classification, an agent-based autonomous resource allocation algorithm is executed to analyze maintenance resource requirements. This algorithm enables the AI agent to make autonomous decisions and generate the optimal resource scheduling plan without human intervention based on factors such as equipment status, maintenance priority, personnel skill matching, and spare parts inventory. This includes maintenance team selection, tool and equipment allocation, spare parts and material scheduling, and operation time scheduling, thereby achieving intelligent and precise allocation of operation and maintenance resources.

[0093] In this embodiment, during the iterative process of system optimization, a meta-learning algorithm is used to improve the ability to "learn how to learn". Based on the rich operation and maintenance experience accumulated, this technology automatically customizes personalized optimization paths for various AI algorithms, enabling the system to quickly adapt to new equipment, new scenarios, and new failure modes, and achieve "fast learning with few samples". In specific implementation, the system regards different equipment categories, fault types, and operation and maintenance scenarios as different task domains, extracts common features between tasks and constructs a meta-knowledge base. When new scenarios appear, it can quickly migrate existing knowledge, greatly shortening the adaptation cycle. At the same time, the neural architecture search (NAS) technology is used to automatically explore the optimal deep neural network structure, dynamically generate and evaluate various network architecture combinations according to the characteristics and requirements of different operation and maintenance scenarios, and select the most efficient network structure, which speeds up the iteration speed of model performance by 2 times. The system continues to evolve and always maintains a technological leading edge.

[0094] In a specific embodiment, the intelligent operation and maintenance method for integrating multimodal data and active learning further includes the following steps:

[0095] Obtain operation and maintenance feedback data after the equipment maintenance plan and resource scheduling plan are executed;

[0096] Based on the operation and maintenance feedback data, the evaluation indicators of user satisfaction data, equipment status data, maintenance effect data and resource utilization data are calculated to obtain the operation and maintenance effect evaluation results;

[0097] Conduct digital twin modeling of the operation and maintenance effect evaluation results, establish a two-way mapping between physical equipment and virtual models, and obtain high-precision digital simulation data of the equipment's operating status;

[0098] High-precision digital simulation data of equipment operating status and operation and maintenance records are input into the blockchain system, and data encryption and distributed storage are performed to obtain credible proof of operation and maintenance data that cannot be tampered with;

[0099] Based on the O&M effectiveness evaluation results and the credible proof of unalterable O&M data, the current O&M strategy is used as the initial state, device stability and user satisfaction are used as the reward function, and O&M resource allocation is used as the action space to obtain a reinforcement learning environment.

[0100] In a reinforcement learning environment, a dual temporal difference learning algorithm is used to iteratively optimize the initial intelligent operation and maintenance strategy to obtain the target intelligent operation and maintenance strategy.

[0101] Specifically, after the equipment maintenance and resource scheduling plans are executed, operation and maintenance feedback data is collected. This includes proactive user satisfaction ratings (such as app service ratings, operation and maintenance follow-up questionnaires, and voice or text comments), as well as real-time equipment status data (such as key performance parameters after maintenance, continuous health monitoring data, and abnormality recurrence), maintenance effectiveness data (such as whether the task was completed in one go, whether rework occurred, whether the fault was completely eliminated, the work order closure time limit, and deviations from the planned maintenance plan), and resource utilization data (including actual man-hours consumed, actual consumption of spare parts and supplies, maintenance manpower arrangements and actual attendance, and on-site emergency response and deployment efficiency). Through automatic data aggregation, cleaning, and standardization, various types of feedback information are linked by core tags such as equipment ID, operation and maintenance order number, time node, and task category to form a structured operation and maintenance feedback data set. Based on the collected operation and maintenance feedback data, a systematic evaluation of the effectiveness of each round of operation and maintenance execution is conducted. For user satisfaction data, a multi-metric fusion model is employed. This model incorporates ratings, sentiment analysis of review text, historical user behavior, and the impact of the current outage. This model modifies and weights subjective evaluations based on single dimensions, enhancing the scientific nature and comparability of satisfaction assessments. For equipment status data, changes in key indicators before and after maintenance are used to assess the contribution of maintenance tasks to equipment health, stable operation, and performance improvement. Algorithms such as mean shift, variance comparison, and outlier reproduction are used to determine the durability and thoroughness of maintenance results. Regarding maintenance effectiveness data, metrics such as task completion rate, return rate, emergency response speed, timeliness of handling exceptions, and task deviation are analyzed to reflect the execution quality of maintenance plans from various perspectives. For resource utilization data, quantitative indicators such as planned and actual consumption of manpower and materials, task concurrency efficiency, scheduling response latency, and on-site resource redundancy or shortages are combined to assess the rationality and cost-effectiveness of resource allocation and scheduling. These evaluation results are summarized into a multi-dimensional O&M effectiveness report. Digital twin modeling is applied to the O&M effectiveness evaluation results, establishing a bidirectional mapping system between physical equipment and virtual models. The system integrates a physical rule engine, a finite element analysis module, and a multi-physics field coupling simulation tool. It can synchronize the state of the physical world in real time, perform high-precision digital simulations of equipment operating states, and simulate the potential impact of different O&M operations, providing virtual experimental scenarios and predictive analysis capabilities for O&M decision-making. High-precision digital simulation data of equipment operating states and O&M operation records are input into the blockchain system, and blockchain distributed ledger technology is used to provide tamper-proof, fully traceable, and reliable evidence for IoT device data collection, user demand data, and O&M operation records. Blockchain technology is used to achieve data encryption and distributed storage, while zero-knowledge proofs are used to protect data privacy while ensuring data authenticity and reliability. Smart contracts are used to automatically trigger processes such as O&M task allocation, service fee settlement, and service quality assessment, forming tamper-proof, credible proof of O&M data.The current O&M strategy (i.e., the maintenance and resource scheduling strategy actually implemented in this round) serves as the initial state of the reinforcement learning environment, with device stability and user satisfaction serving as the primary reward functions. Device stability is measured through metrics such as health score, troubleshooting time, and key performance maintenance period, while user satisfaction is reflected through normalized satisfaction scores, sentiment scores, and service suggestion adoption rates. The allocation of O&M resources constitutes the reinforcement learning action space, encompassing the various possible combinations of actions the system can choose from in each task assignment, personnel scheduling, material allocation, and priority setting. This approach not only simulates the multi-objective trade-offs inherent in real-world O&M decision-making but also encompasses complex game-plays at different levels and with varying resource constraints across a wide range of business scenarios. Within this reinforcement learning environment, deep reinforcement learning algorithms, such as Dueling Double DQN (dual decision + temporal difference), are employed to efficiently iteratively optimize the current intelligent O&M strategy. The Dueling network architecture decomposes the traditional Q-learning action-value function into a "state-value function" and an "action-advantage function," respectively capturing the long-term potential benefits of the current O&M strategy and the relative advantage of a specific action. This improves the learning process's sensitivity to sparse rewards, conflicting objectives, and strategy switching scenarios. The Double Q-learning mechanism, by alternately training independent objective and evaluation networks, effectively mitigates the overestimation of Q values in traditional reinforcement learning, improving the model's convergence speed and policy robustness in complex real-world environments. This approach automatically backtracks the performance of the current strategy after each round of O&M tasks, using the resulting O&M evaluation results as reward signals to continuously adjust policy parameters, optimizing maintenance dispatch, resource scheduling, and task prioritization, gradually approaching the theoretically optimal intelligent O&M strategy. During the iterative optimization process, the system proactively explores multiple O&M solution combinations. Leveraging the reinforcement learning model's self-learning and adaptive capabilities, it enables automatic policy migration and dynamic upgrades to address new equipment, unknown failures, sudden business changes, and resource constraints. At the same time, the reinforcement learning environment uses "offline playback" based on historical data to proactively verify, without actual risk, the potential for new strategies to improve device health and user experience, ensuring that each optimization update delivers real-world value. After multiple rounds of cyclical learning and strategy self-evolution, the system ultimately converges on a targeted intelligent O&M strategy tailored to the current business scenario and resource constraints. This strategy dynamically adapts to various fault types, device status, and user needs, balancing global device stability with local user satisfaction, maximizing O&M efficiency and resource utilization, and realizing a new model of intelligent O&M characterized by data-driven, intelligent self-optimization.

[0102] In this embodiment, a multi-objective optimization algorithm is applied to generate equipment maintenance plans and resource scheduling plans, taking into account multiple competing objectives such as cost, efficiency, and user experience. Pareto optimal solution and hierarchical analysis method are used to automatically generate a series of optimal trade-off solutions. A quantitative evaluation report is generated for each decision-making plan based on historical decision-making experience. The optimal solution hit rate of the decision-making plan is improved by more than 40% compared with the traditional one. At the same time, relying on deep reinforcement learning algorithms and digital twin simulation, the agent-type intelligent agent explores various possible operation and maintenance strategy combinations by interacting with the environment. Based on the experience replay mechanism and dual Q network architecture, it extracts high-value decision-making experience from massive historical data, avoids the overfitting and instability problems in traditional reinforcement learning, and balances exploration and utilization while searching for the global optimal solution, ensuring that the system can still provide high-quality solutions in the face of complex and dynamic decision-making environments.

[0103] In the embodiments of the present application, through hierarchical IoT device deployment and multi-level sampling strategies, differentiated monitoring schemes are implemented for different types of equipment, and the sampling frequency is dynamically adjusted in combination with adaptive sampling technology. While ensuring data quality, the system resource allocation is optimized, solving the data quality problems and resource waste problems caused by unreasonable sampling in traditional operation and maintenance. The multi-channel hybrid networking system intelligently allocates the optimal channel according to the communication needs of the equipment, and combines the model context protocol (MCP) to build a unified large language model service framework to achieve multi-model collaborative building operation and maintenance intelligent analysis. The three-layer encryption mechanism and priority transmission strategy ensure the security and timeliness of data transmission, solving the problems of unsafe, untimely and unintelligent data transmission in traditional operation and maintenance systems. The multimodal fusion intent recognition technology based on deep learning realizes efficient entity recognition through the selective state space layer of the Mamba model, and combines the LLM-Fine-Tuning technology to perform professional field intent classification. Compared with traditional methods, the entity recognition efficiency is improved by 5-10 times, and the intent understanding accuracy is improved by 15-20%. The innovative application of building operation and maintenance prompts transforms complex equipment status descriptions into structured operation and maintenance analysis results, enabling a precise understanding of user needs and addressing the inaccurate understanding and in-depth analysis inherent in traditional operation and maintenance. A graph neural network-based device topology is constructed to enable intelligent collaborative monitoring. Agent-based artificial intelligence (Agentic AI) technology is introduced to empower the system with autonomous decision-making capabilities. Through AI agents that perceive the environment, set goals, plan paths, and execute autonomously, the system achieves intelligent fault prediction and resource scheduling, addressing the high false alarm rates and inefficient resource allocation associated with isolated equipment monitoring in traditional operation and maintenance systems. A differentiated fault prediction system based on multi-model fusion selects the optimal algorithm combination for different fault characteristics, using dynamic ensemble learning and transfer learning techniques to adaptively optimize the prediction model. Digital twin technology is used to establish a bidirectional mapping between physical equipment and virtual models, providing a virtual experimental scenario for fault prediction. Blockchain technology enables tamper-proof storage of operation and maintenance data and multi-party collaboration, transforming fault predictions into actionable maintenance strategies. This addresses the challenges inherent in traditional operation and maintenance, such as the poor adaptability of single fault prediction models, the difficulty in translating prediction results into effective decisions, and low data credibility. Based on operation and maintenance feedback data and meta-learning algorithms, personalized optimization paths are customized for AI models to achieve continuous self-evolution of the system, enabling it to cope with new equipment, new failure modes and ever-changing user demand expressions, continuously maintain technological advancement, and fundamentally change the limitations of traditional operation and maintenance systems that are static and difficult to adapt to new scenarios.

[0104] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0105] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0106] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent operation and maintenance method integrating multimodal data and active learning, characterized in that: include: Deploy a hierarchical IoT device network in the target operation and maintenance area, and collect environmental parameter data, device status data, resource consumption data, and security monitoring data through differentiated collection frequencies. After pre-processing the collected data, adaptive sampling frequency adjustment is used to obtain dynamically optimized data collection results. The data collection results are associated with the unique device identification code to obtain a heterogeneous operation and maintenance data set. Constructing a building operation and maintenance prompt word project for the heterogeneous operation and maintenance data set, and using a unified building operation and maintenance service framework coordinated by multiple MCPs to generate an operation and maintenance feature data set; the prompt word project converts building equipment status data, environmental parameters, and historical operation record data into prompt word templates; The operation and maintenance feature data set and user feedback information are input into the deep learning intention recognition model for intent recognition and demand analysis to obtain structured user demand data and operation and maintenance task priority ranking; specifically including: collecting user feedback information from multiple channels including mobile phone APP feedback, voice assistant interaction, customer service phone records and work order system data, and combining the user feedback information with the operation and maintenance feature data set to obtain a multimodal user demand data set; inputting the text data in the multimodal user demand data set into the Mamba model in the deep learning intention recognition model for entity recognition to obtain entity annotation data; inputting the entity annotation data into the LLM-Fine-Tuning model in the deep learning intention recognition model for intent classification to obtain intent classification results; constructing an entity relationship graph based on the intent classification results and the entity annotation data to obtain a structured knowledge network; extracting the mapping relationship between user expression and actual needs based on the structured knowledge network and user historical behavior data to obtain structured user demand data, and performing urgency score calculation based on the structured user demand data to generate operation and maintenance task priority ranking; Based on the operation and maintenance feature data set and the structured user demand data, the equipment operation status is evaluated in real time to obtain fault risk warning data; The fault risk warning data and the operation and maintenance task priority ranking are input into the agent-type artificial intelligence engine for autonomous decision-making analysis to generate equipment maintenance plans and resource scheduling plans.

2. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1 is characterized in that: The hierarchical IoT device network is deployed in the target operation and maintenance area, and a heterogeneous operation and maintenance data set containing environmental parameter data, device status data, resource consumption data, and security monitoring data is collected, including: The equipment in the target operation and maintenance area is divided into environmental monitoring equipment, equipment status monitoring equipment, resource consumption monitoring equipment, and security monitoring equipment; A fixed sampling frequency is set for the environmental monitoring equipment, a dynamically adjusted sampling frequency is set for the equipment status monitoring equipment, a differentiated sampling frequency is set for the resource consumption monitoring equipment, and a real-time sampling frequency is set for the security monitoring equipment, thereby obtaining a multi-level data collection strategy; Sampling and controlling IoT devices based on the multi-level data acquisition strategy to obtain multi-source heterogeneous raw data, and inputting the multi-source heterogeneous raw data into the edge computing gateway for data pre-screening to obtain pre-processed device data; Adaptively adjust the sampling frequency of the pre-processed device data to obtain a dynamically optimized data collection result, and associate the dynamically optimized data collection result with a unique device identification code to obtain a heterogeneous operation and maintenance data set.

3. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1 is characterized in that: The said building operation and maintenance prompt word project for constructing the said heterogeneous operation and maintenance data set, and using a unified building operation and maintenance service framework coordinated by multiple MCPs to generate an operation and maintenance feature data set, includes: According to the communication requirements of different devices in the heterogeneous operation and maintenance data set, device communication channels are allocated through a multi-channel hybrid networking system to obtain a hybrid heterogeneous network configuration solution; Constructing a building operation and maintenance prompt word project based on the heterogeneous operation and maintenance data, and simultaneously using a unified building operation and maintenance service framework coordinated by multiple MCPs to generate intelligent operation and maintenance data; Based on the hybrid heterogeneous network configuration solution and the intelligent operation and maintenance data, the AES-128 encryption algorithm is applied at the device layer, the TLS 1.3 protocol is applied at the transport layer to achieve end-to-end encryption, and the identity authentication and access control mechanism is applied at the application layer to obtain encrypted data with multi-layer protection; Performing data transmission priority analysis on the multi-layer protected encrypted data to obtain prioritized transmission data, and inputting the prioritized transmission data into a server for data cleaning to obtain a cleansed data set; Data standardization is performed on the cleansed data set to obtain a standardized data set, and feature extraction is performed on the standardized data set to obtain an operation and maintenance feature data set.

4. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 3 is characterized in that: The method allocates device communication channels through a multi-channel hybrid networking system according to the communication requirements of different devices in the heterogeneous operation and maintenance data set to obtain a hybrid heterogeneous network configuration solution, including: Performing device communication parameter analysis on the communication requirements of different devices in the heterogeneous operation and maintenance data set to obtain a device communication requirement classification result; Based on the classification results of the communication requirements of the devices, the transmission capacity of six communication technologies, including NB-IoT, LoRa, Wi-Fi, ZigBee, 5G, and wired networks, was evaluated. Communication channel characteristic data including bandwidth parameters, delay parameters, coverage parameters, and energy consumption parameters were established through a multi-channel hybrid networking system. Performing device channel matching on the device communication requirement classification results and the communication channel characteristic data to obtain a matching score, and allocating an optimal communication channel to each type of device based on the matching score to obtain an initial channel allocation solution; Performing a network load analysis on the initial channel allocation scheme to obtain a load-balanced channel allocation scheme, and setting a differentiated communication strategy based on the load-balanced channel allocation scheme; The differentiated communication strategy is integrated with the load-balanced channel allocation solution to generate a hybrid heterogeneous network configuration solution.

5. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1 is characterized in that: The step of inputting the text data in the multimodal user demand dataset into the Mamba model in the deep learning intent recognition model for entity recognition to obtain entity annotation data includes: Performing word segmentation and cleaning processing on the text data in the multimodal user demand dataset to obtain a processed text sequence; Inputting the processed text sequence into the selective state space layer of the Mamba model for sequence modeling to obtain long-distance dependency feature representation; Performing parallel sequence processing based on the long-distance dependency feature representation, capturing temporal patterns and key features in the text through a structured state space model, and obtaining a context-aware feature representation; Performing gating mechanism filtering and feature enhancement on the context-aware feature representation to extract key information from the text and obtain an entity candidate feature set; The entity candidate feature set is input into the self-attention mechanism layer for global information integration, the potential entity position is accurately identified, and the entity boundary probability distribution is obtained; Execute sequence labeling decisions based on the entity boundary probability distribution, calculate the optimal labeling path through the Viterbi algorithm, and obtain the globally optimal entity recognition result; The globally optimal entity recognition result is post-processed and entity type normalized to obtain entity annotation data.

6. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1, characterized in that: Inputting the entity annotation data into the LLM-Fine-Tuning model in the deep learning intent recognition model to perform intent classification and obtain an intent classification result includes: Performing text reconstruction and prompt word engineering optimization on the entity annotation data to obtain an enhanced semantic input sequence; Inputting the enhanced semantic input sequence into the LLM-Fine-Tuning model, extracting deep semantic features through a multi-layer Transformer architecture, and obtaining a context-aware representation vector; Design a domain-specific fine-tuning strategy based on the target operation and maintenance tasks, perform gradient updates on the large language model parameters, build intent recognition knowledge in the operation and maintenance domain, and obtain a domain-adaptive fine-tuning model. Performing multi-label classification on the input text using the domain-adapted fine-tuning model to identify the primary and secondary intents in the text and obtain an intent probability distribution; Confidence threshold screening and intent conflict resolution are applied to the intent probability distribution, and intent normalization is performed in combination with operation and maintenance business rules to obtain intent classification results.

7. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1, characterized in that: The real-time evaluation of the equipment operation status based on the operation and maintenance feature data set and the structured user demand data to obtain fault risk warning data includes: Organizing IoT devices into functionally complementary logical monitoring units based on the operation and maintenance feature dataset, and constructing a device state feature vector containing operating parameters, environmental parameters, and historical fault information based on each logical monitoring unit; Inputting the device state feature vector into a hybrid model composed of a multilayer perceptron and a support vector machine for anomaly detection to obtain a preliminary abnormal state identification; Constructing a multi-dimensional device association topology of physical space and functional relationships based on the preliminary abnormal state identification to obtain a device association map; Based on the device association map and the structured user demand data, activate the peripheral devices whose association exceeds a preset threshold to perform collaborative monitoring to obtain multi-source verification information; The multi-source verification information is input into a dynamic Bayesian network for probabilistic reasoning calculation to obtain a posterior probability distribution of the true state of the target device, and an equipment failure risk assessment is performed based on the posterior probability distribution to obtain failure risk warning data.

8. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1, characterized in that: The inputting of the fault risk warning data and the operation and maintenance task priority ranking into the agent-based artificial intelligence engine for autonomous decision-making analysis to generate equipment maintenance plans and resource scheduling plans includes: Performing fault type classification on the fault risk warning data to obtain a fault type classification result, wherein the fault type classification result includes a gradual fault and a sudden fault; Construct an agent-based artificial intelligence multi-objective task model, set fault repair rate, resource utilization efficiency, and time response speed as optimization goals, and obtain AI agent decision parameters; For the gradual fault, a long short-term memory network and an autoregressive integrated moving average hybrid model are used for time series prediction. For the sudden fault, a random forest and gradient boosting decision tree hybrid model are used for classification prediction to obtain the fault development trend prediction results. Calculating a fault risk index based on the fault development trend prediction result and the operation and maintenance task priority ranking to obtain a quantitative risk index, and generating an equipment maintenance plan based on the quantitative risk index and the AI agent decision parameters; An agent-based autonomous resource allocation algorithm is executed based on the equipment maintenance plan and the fault type classification result to perform maintenance resource demand analysis and generate a resource scheduling plan.

9. The intelligent operation and maintenance method integrating multimodal data and active learning according to claim 1, characterized in that: The intelligent operation and maintenance method integrating multimodal data and active learning also includes: Obtaining operation and maintenance feedback data after the equipment maintenance plan and the resource scheduling plan are executed; Calculate evaluation indicators for user satisfaction data, equipment status data, maintenance effect data, and resource utilization data based on the operation and maintenance feedback data to obtain an operation and maintenance effect evaluation result; Conduct digital twin modeling on the operation and maintenance effect evaluation results, establish a two-way mapping between physical equipment and virtual models, and obtain high-precision digital simulation data of the equipment's operating status; Inputting high-precision digital simulation data of the equipment's operating status and operation and maintenance records into the blockchain system, performing data encryption and distributed storage, and obtaining credible proof of the operation and maintenance data that cannot be tampered with; Based on the operation and maintenance effect evaluation results and the tamper-proof operation and maintenance data credible proof, the current operation and maintenance strategy is used as the initial state, the device stability and user satisfaction are used as the reward function, and the operation and maintenance resource allocation is used as the action space to obtain a reinforcement learning environment; In the reinforcement learning environment, a dual temporal difference learning algorithm is used to iteratively optimize the initial intelligent operation and maintenance strategy to obtain the target intelligent operation and maintenance strategy.

Images