Intelligent knowledge management system for clean energy project

Abstract

The invention discloses an intelligent knowledge management system for a clean energy project, and the system comprises the steps: collecting multi-source heterogeneous data from the clean energy project, and converting the multi-source heterogeneous data into structured knowledge input data; performing knowledge extraction on the structured knowledge input data, and aligning and fusing knowledge extracted from different data sources to construct a structured knowledge base; performing data processing on the structured knowledge base through a time-aware graph attention network T-GAT and a graph structure causal model SCM to construct a dynamic time sequence knowledge graph; according to user requirements and scenes, a decision result and a knowledge recommendation result are obtained through an integrated causal-driven multi-objective optimization algorithm and a dynamic time sequence knowledge graph, and user feedback data are output; and processing the user feedback data by fusing Temporal Graph Network and an online near-end strategy optimization Online PPO algorithm, and carrying out dynamic optimization and adaptive adjustment on the knowledge graph, the rule base and the decision model according to the user feedback data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge management, and in particular to an intelligent knowledge management system for clean energy projects. BACKGROUND

[0002] With the increasing global demand for renewable energy, the complexity and scale of clean energy projects continue to expand. However, traditional project management methods face many challenges. In the prior art, traditional knowledge management techniques have problems such as knowledge fragmentation, loss of tacit knowledge, poor dynamic adaptability, and limited decision support. The patent with publication number CN116384615A fails to better solve the problems of tacit knowledge mining and dynamic updating. The optimization algorithm mentioned in the patent with publication number CN118277593 lacks deep dependence on knowledge graphs, and the decision logic is not transparent. Other patent technologies (such as the patent with publication number CN117093674A) have some shortcomings in cross-modal data alignment and timeliness modeling. In knowledge management, there is little integration of explicit and tacit knowledge, little application of artificial intelligence technology in knowledge collection, and difficulty in forming interaction with users and displaying real-time data in knowledge application. Knowledge application efficiency is low, and knowledge management is difficult to develop continuously, which cannot meet the needs of efficient and reliable knowledge management. SUMMARY

[0003] The present application aims to at least partially solve one of the technical problems in the related art.

[0004] To this end, the present application proposes an intelligent knowledge management system for clean energy projects, aiming to improve the efficiency and decision quality of clean energy project knowledge management.

[0005] To achieve the above purpose, the present application proposes an intelligent knowledge management system for clean energy projects, comprising a data collection and preprocessing module, a knowledge extraction and integration module, a knowledge graph construction module, a knowledge application and decision module, and a feedback and self-evolution module. Wherein, The data collection and preprocessing module integrates text, image, time series data and unstructured records through multi-source heterogeneous data fusion technology, extracts key experience rules using implicit knowledge pre-labeling technology, and realizes real-time data cleaning and anomaly repair combined with stream data processing, thereby providing high-quality, structured knowledge input for clean energy projects, realizing deep fusion of multi-modal data, improving real-time performance and reliability, efficiently transforming tacit knowledge, and laying a foundation for subsequent knowledge extraction.

[0006] The knowledge extraction and integration module realizes deep mining of clean energy project data through fine multi-modal data processing technology, including entity recognition, relation extraction, event tagging and implicit knowledge analysis, and constructs a time-effective structured knowledge base through cross-modal alignment and dynamic fusion technology, thereby providing efficient and accurate core input for knowledge graph construction and intelligent decision-making, innovatively reducing the implicit knowledge conversion cost, and ensuring the dynamic updating and quality control of knowledge management, significantly improving the knowledge completeness and decision support capability of the system.

[0007] The knowledge graph construction module realizes the construction of dynamic time sequence knowledge graph by integrating time perception graph attention network (T-GAT) and graph structure causal model (SCM), supports knowledge semantic association, time-effectiveness modeling and intelligent reasoning of clean energy project throughout its life cycle, and adopts a hybrid storage architecture of Neo4j and Apache AGE to ensure efficient query and distributed storage of nodes, thereby providing accurate and explainable causal chain analysis for project decision-making, and significantly improving the dynamic, explainability and intelligence level of knowledge management.

[0008] The knowledge application and decision-making module realizes intelligent decision support for clean energy projects by integrating causal-driven multi-objective optimization algorithm and dynamic knowledge graph, covering scenario-based knowledge recommendation, multi-objective optimization decision-making, decision explainability display and dynamic strategy adjustment, wherein the combination of proximal policy optimization (PPO) and graph structure causal model (SCM) ensures the scientificity and explainability of the decision, and the incremental learning and federated learning mechanism enhances the dynamic adaptability and cross-project knowledge sharing capability of the system, thereby significantly improving the intelligence level of project management and the transparency of decision-making.

[0009] The feedback and self-evolution module realizes dynamic optimization and self-evolution of the knowledge base and decision-making model by integrating multi-source feedback data, implementing incremental knowledge update and online model learning, and combining quality verification and risk control mechanism. This module innovatively integrates Temporal Graph Network (TGN) and Online Proximal Policy Optimization (Online PPO) algorithm to ensure real-time updating of the knowledge graph and adaptability of the decision-making model, and A / B testing and rollback mechanism ensure system stability and safety of the updating process, thereby providing continuous optimization and self-adaptive intelligent management support for clean energy projects.

[0010] The beneficial effects of the present application include: Knowledge fusion depth and timeliness are improved, a dynamic knowledge graph is constructed through a T-GAT time graph attention network, time dimension modeling of knowledge elements is realized, historical technical evolution analysis and future trend prediction (such as patent technology elimination early warning) are supported, and semantic alignment of text description, equipment image and sensor time sequence data is completed in combination with an ERNIE-ViL multi-modal pre-training model, so that the problem of multi-source data island existing in a traditional system is effectively solved.

[0011] Implicit knowledge conversion and decision credibility enhancement, a Few-shot Learning small sample learning framework is adopted, only 10-20 expert annotations are needed to extract domain experience rules (annotation cost is reduced by 80%), and a visual causal chain of 'equipment upgrade -> energy efficiency improvement -> revenue growth' is constructed in cooperation with an SCM structure causal model, so that the explainability of decision suggestions is improved by 80%.

[0012] System adaptation and cross-scene generalization capability, a distributed knowledge base architecture based on federated learning is used, knowledge migration of different scenes such as photovoltaic and wind energy is realized through a meta-learning framework, and a Neo4j+Apache AGE hybrid storage engine is used to keep the query response speed < 200ms under the condition of a billion node scale, and the performance is improved by 5 times compared with a single graph database.

[0013] Optimization of whole life cycle management efficiency, the implicit knowledge mining (expert dialogue record analysis) and the explicit knowledge management (project document structuring) are innovatively combined to build a whole-process knowledge system covering technical research and development, equipment operation and maintenance, and resource scheduling.

[0014] The clean energy project intelligent knowledge management system of the embodiment of the application integrates the above modules to build a clean energy project intelligent knowledge management system, which can effectively improve the intelligent level of project management, reduce operation cost, shorten project cycle, enhance the sustainability and competitiveness of the project, and has stronger professionalism and landing value through dynamic knowledge graph and implicit knowledge mining technology.

[0015] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter in the description of the application. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and / or additional aspects and advantages of the application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which: Figure 1 The structural diagram of the clean energy project intelligent knowledge management system of the embodiment of the application.

[0017] Figure 2The diagram below shows the overall framework of an intelligent knowledge management system for clean energy projects based on dynamic knowledge graphs and tacit knowledge mining, provided as an embodiment of the present invention.

[0018] Figure 3 The flowchart of the data acquisition and preprocessing module provided in the embodiments of the present invention is shown.

[0019] Figure 4 This is a knowledge extraction and integration module architecture diagram provided for embodiments of the present invention.

[0020] Figure 5 This is a schematic diagram of a knowledge graph construction module provided in an embodiment of the present invention.

[0021] Figure 6 The flowchart illustrates the knowledge application and decision-making module provided in the embodiments of the present invention.

[0022] Figure 7 This is a schematic diagram of the feedback and self-evolution module provided in an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes an intelligent knowledge management system for clean energy projects based on embodiments of the present invention.

[0026] This embodiment uses a hydropower project as an example, and analyzes in detail the application of the system in a real-world project, combining the relevant modules and technologies of this invention. Through data acquisition and preprocessing, knowledge extraction and integration, knowledge graph construction, knowledge application and decision-making, and feedback and self-evolution modules, the system of this invention can provide knowledge management and decision support for hydropower projects, optimize resource allocation, improve project execution efficiency, and reduce risks.

[0027] The water power generation project is located in a reservoir area of the Yangtze River Basin, and a medium-sized hydropower station with a installed capacity of 100 MW is planned to be constructed, which involves water resource scheduling, equipment selection, fund investment, construction management, operation and maintenance and other aspects. In the implementation process of the project, various challenges such as water flow change, equipment failure and fund scheduling are faced, and fine management and optimization decision of each link of the project are needed.

[0028] Figure 1 The structure diagram of the intelligent knowledge management system of the clean energy project according to the embodiment of the application is shown in Figure 1 The intelligent knowledge management system of the clean energy project comprises a data acquisition and preprocessing module, a knowledge extraction and integration module, a knowledge graph construction module, a knowledge application and decision module and a feedback and self-evolution module. The data acquisition and preprocessing module is used for acquiring multi-source heterogeneous data from the clean energy project, and converting the multi-source heterogeneous data into structured knowledge input data through intelligent preprocessing technology. The knowledge extraction and integration module is used for knowledge extraction on the structured knowledge input data, and aligning and fusing the knowledge extracted from different data sources through multi-modal alignment and dynamic fusion technology to construct a structured knowledge base. The knowledge graph construction module is used for data processing on the structured knowledge base through a time-aware graph attention network (T-GAT) and a graph structure causal model (SCM) to construct a dynamic time series knowledge graph. The knowledge application and decision module is used for obtaining decision results and knowledge recommendation results according to user needs and scenarios, and through integration of a causal-driven multi-objective optimization algorithm and a dynamic time series knowledge graph, and outputting user feedback data. The feedback and self-evolution module is used for processing the user feedback data by fusing a Temporal Graph Network and an online proximal policy optimization (Online PPO) algorithm, and dynamically optimizing and adaptively adjusting the knowledge graph, the rule base and the decision model according to the user feedback data, so as to feed back the optimized results to the knowledge extraction and integration module, the knowledge graph construction module and the knowledge application and decision module.

[0029] As shown in Figure 2 The overall framework diagram of the intelligent knowledge management system of the clean energy project based on the dynamic knowledge graph and the implicit knowledge mining provided by the embodiment of the application.

[0030] In an embodiment of the application, as shown in Figure 3As shown, the data acquisition and preprocessing module includes a data source and acquisition layer and a data preprocessing layer, which includes a structured data source unit, an unstructured data source unit, a multi-modal data interface unit, and a data cleaning and standardization unit, an implicit knowledge pre-structuring unit, and a streaming data processing unit.

[0031] The data source layer realizes distributed acquisition and privacy protection integration of multi-source heterogeneous data through API interface, sensors, and federated learning framework.

[0032] The structured data source unit automatically captures patent specifications, technical standards, and papers through API interface or document parsing tools, and collects sensor data and external database information in real time to form a multi-dimensional structured data pool.

[0033] The unstructured data source unit uses speech recognition, image analysis, and event extraction technology to convert expert experience, operation logs, and technical drawings into structured knowledge items.

[0034] The multi-modal data interface unit realizes cross-source data privacy protection integration through the federated learning framework and edge computing nodes, reduces the central load, and improves the processing timeliness.

[0035] The data processing layer uses TFT model, Whisper model, and Flink engine to realize data anomaly repair, rule extraction, and real-time feature extraction with a delay of ≤500 milliseconds.

[0036] The data cleaning and standardization unit uses TFT model, BERT model, and OpenCV technology to realize time series anomaly repair, text error correction, and image denoising, and unifies the data format.

[0037] The implicit knowledge pre-structuring unit extracts IF-THEN rules and constructs event-attribute triples through the Whisper model and Few-shot Learning technology.

[0038] The streaming data processing unit builds a real-time pipeline based on Kafka and Flink to realize data cleaning, window aggregation, and edge feature extraction.

[0039] Specifically, the data acquisition and preprocessing module of the embodiment of the present application is realized through the following sub-modules: The data acquisition and preprocessing module is responsible for acquiring multi-source heterogeneous data from the clean energy project, and converting it into high-quality, structured knowledge input through intelligent preprocessing technology, providing a foundation for subsequent knowledge extraction and graph construction.

[0040] The data source and acquisition layer includes a structured data source unit, an unstructured data source unit, and a multi-modal data interface unit, and uses a distributed data collector, each unit is equipped with an independent processor and a data buffer.

[0041] The structured data source unit realizes efficient data integration through multi-dimensional collection and automated processing technology, specifically including: Through API interface or document parsing tools (such as Apache Tika), automatically capture and analyze patent specifications, technical standards, scientific papers and other document data, extract standardized technical parameters and related information.

[0042] Deploy Internet of Things terminals such as water level sensors and vibration monitoring equipment to collect physical parameters such as water flow, equipment temperature and pressure in real time at a sampling frequency of no less than 10Hz, ensuring data timeliness and continuity.

[0043] Connect to the open API of the meteorological bureau to obtain environmental data such as temperature, precipitation and wind speed, and access project progress and fund flow information from the energy management platform, forming a multi-dimensional structured data pool covering technology, environment and operation, providing high-quality input for knowledge extraction and graph construction.

[0044] Real-time collection of the operating status of water turbines, generators and other equipment, real-time monitoring of generator output power, temperature, pressure and other parameters, and timely detection of equipment failure or abnormal operation.

[0045] The unstructured data source unit realizes the efficient transformation of implicit knowledge through multi-modal collection and intelligent analysis technology, specifically including: Use voice recording equipment to collect audio data such as clean energy field expert meeting discussions and on-site guidance recordings, use voice recognition models (such as Whisper) to convert them into text and perform semantic segmentation, and extract key experience rules, such as the rule "IF running time ≥ 720 hours THEN trigger lubrication maintenance".

[0046] Extract unstructured operation and maintenance logs (such as equipment failure descriptions and maintenance records) from the project management system, identify event types, participating entities and time attributes based on event extraction models (such as DEGREE), and build event-attribute triples.

[0047] Integrate image data such as technical drawings and CAD design drawings, use visual recognition models (such as YOLOv7) to label key component coordinates and function descriptions, and use cross-modal pre-training models (such as ERNIE-ViL) to realize semantic alignment between image entities and text descriptions, solve the semantic gap between heterogeneous data, and finally convert unstructured data into structured knowledge items, providing multi-modal input for dynamic knowledge graph construction.

[0048] The multi-modal data interface unit realizes efficient integration of cross-source data through distributed architecture and privacy protection technology, specifically including: Deploy lightweight agents on edge devices (such as RTU controllers) to achieve real-time caching and preliminary cleaning of local data (including redundant data filtering and format standardization), significantly reducing the load on the central server and improving data processing timeliness.

[0049] Build an encrypted channel using the federated learning framework to aggregate multi-project data (such as operation and maintenance records of multiple hydropower stations and equipment status time series data). Use homomorphic encryption and secure multi-party computation technology to ensure data privacy, support cross-institutional knowledge sharing and model collaborative training, and achieve dynamic optimization of global knowledge base and generalization of decision-making capabilities while protecting sensitive information. Provide a safe and efficient multi-modal data fusion foundation for clean energy projects.

[0050] The data preprocessing layer uses data cleaning and standardization units, implicit knowledge pre-structuring units, and streaming data processing units to efficiently collect, deeply clean, and convert implicit knowledge from multi-source data in clean energy projects, providing high-quality input for dynamic knowledge graph construction, which can significantly improve the real-time and intelligent level of the knowledge management system.

[0051] The data cleaning and standardization unit uses multi-modal data intelligent processing technology to achieve comprehensive optimization of data quality, including: For time series data (such as water flow and equipment vibration parameters), use the Temporal Fusion Transformer (TFT) model for real-time anomaly detection, and use the sliding window interpolation method to repair missing values and noise interference.

[0052] For text data (such as technical documents and patent specifications), use rule engines (regular expressions) and BERT models for spelling correction and term standardization (such as unifying "turbine" and "water turbine" to "HT-3000 water turbine").

[0053] For image data (such as technical drawings), use OpenCV libraries for denoising and binarization, and use the YOLOv7 model to identify key components (such as valves and pipelines) and label coordinate information with an error of less than or equal to 2 pixels.

[0054] Use format unification technology to convert text data into JSON-LD format with embedded semantic labels (such as "technical parameters" and "fault description"), convert image data into vector graphics (SVG) and dynamically associate with text metadata, and store time series data as Parquet columnar files after unifying timestamps according to ISO8601 standards, forming a standardized and highly consistent multi-modal data pool, providing a high-quality input foundation for subsequent knowledge extraction and graph construction.

[0055] The implicit knowledge pre-structuring unit realizes the conversion of unstructured implicit knowledge into structured rules through intelligent analysis technology, specifically including: The expert recording is converted into text and semantically segmented by using the Whisper model to extract independent semantic fragments (such as "rainy season needs to increase the inspection frequency of flood discharge gate"), and IF-THEN type decision rules (for example, "IF water level exceeds warning line THEN start emergency flood discharge program") are extracted from the text based on few-shot learning and prompt engineering.

[0056] The DEGREE event extraction model is used to identify event types (such as "equipment failure", "maintenance completed"), participating entities (device number, responsible person) and time attributes from operation and maintenance logs, build event-attribute triples, and finally convert expert experience and log records into structured knowledge entries that can be stored and reasoned, providing implicit knowledge input for dynamic knowledge graph.

[0057] The streaming data processing unit realizes efficient real-time data management through hierarchical architecture and edge computing optimization technology, specifically including: A real-time data pipeline is built by using Apache Kafka, data types are partitioned (such as " / sensor / water_flow / " and "sensor / temperature"), sensor data streams are received, and Flink stream processing engine is used to perform window aggregation (such as 5-minute sliding window calculation of water flow mean and maximum), dynamic threshold alarm (based on historical data to adjust device temperature threshold and trigger real-time notification) and data persistence operation, and the cleaned data is written into the hybrid storage architecture of time series database (such as InfluxDB) and graph database (Neo4j) for synchronization; A lightweight TinyML model is deployed on the sensor node to realize local data filtering (such as removing transient noise) and feature extraction (such as Fourier transform analysis of device vibration spectrum), and only valid features are uploaded to the central server, reducing bandwidth occupation by more than 70%, ensuring the real-time, efficiency and resource utilization of data processing.

[0058] In an embodiment of the present application, as shown in Figure 4 The knowledge extraction and integration module includes a multi-modal knowledge extraction layer, a cross-modal knowledge integration layer, and a dynamic update and quality control layer, including a text knowledge extraction unit, an image knowledge extraction unit, a time series knowledge extraction unit, an implicit knowledge extraction unit, a multi-modal entity alignment unit, a dynamic knowledge fusion unit, a knowledge storage optimization unit, an incremental knowledge update unit, an online learning unit, an automatic verification unit, and a manual review interface unit.

[0059] The multi-modal knowledge extraction layer extracts entities, events, and IF-THEN rules through the RoBERTa-LSTM-CRF model, YOLOv7-CLIP alignment technology, Fourier transform, and Few-shot Learning, significantly improving accuracy.

[0060] The text knowledge extraction unit identifies entities through the RoBERTa-LSTM-CRF model and extracts semantic relationships based on GCN, with an accuracy of ≥95%.

[0061] The image knowledge extraction unit achieves image component recognition and image-text semantic alignment through the YOLOv7 and CLIP models, addressing the cross-modal semantic gap.

[0062] The temporal knowledge extraction unit correlates temporal events and device logs through Fourier transform, TFT model, and GNN, constructing a temporal-event subgraph.

[0063] The tacit knowledge extraction unit quantifies expert rule weights using RoBERTa sentiment analysis and confidence scores, enhancing decision-making credibility.

[0064] The cross-modal knowledge integration layer implements cross-modal semantic correlation, timeliness modeling, and distributed storage of billion-level nodes based on ERNIE-ViL, T-GAT network, and Neo4j+Apache AGE architecture.

[0065] The multi-modal entity alignment unit achieves cross-modal semantic correlation and spatio-temporal alignment of text, image, and temporal data based on ERNIE-ViL and GNN technology.

[0066] The dynamic knowledge fusion unit adds timeliness weights through the T-GAT network and resolves multi-source data conflicts based on a logic engine, optimizing knowledge consistency.

[0067] The knowledge storage optimization unit uses a Neo4j+Apache AGE hybrid architecture and RDF / SWRL rule base, supporting millisecond-level queries and logical reasoning for billion-level nodes.

[0068] The dynamic updating and quality control layer forms a closed-loop quality optimization process through real-time updating triggered by Kafka, incremental BERT model, SWRL rule engine, and visual annotation tools.

[0069] The incremental knowledge updating unit triggers graph updates in real time through Kafka, with an end-to-end delay of ≤1 second, supporting dynamic response.

[0070] The online learning unit dynamically optimizes entity recognition accuracy ≥5% and decision model adaptability using incremental BERT and online PPO algorithms.

[0071] The automatic verification unit verifies the logic consistency through a SWRL rule engine, freezes conflicting entries, and marks time-sensitive invalid knowledge.

[0072] The artificial review interface unit generates correction instructions through visual marking and a contrast learning algorithm, forming a "human-model" closed-loop optimization link.

[0073] Specifically, the knowledge extraction and integration module of the embodiment of the present application is implemented through the following sub-modules: The knowledge extraction and integration module includes a multi-modal knowledge extraction layer, a cross-modal knowledge integration layer, and a dynamic updating and quality control layer. This module extracts explicit and implicit knowledge from multi-source heterogeneous data of clean energy projects, and constructs a structured knowledge base through multi-modal alignment and dynamic fusion technology, realizes precise extraction, deep integration and dynamic updating of multi-source knowledge, and provides core data support for the construction of knowledge graph and intelligent decision-making.

[0074] The multi-modal knowledge extraction layer includes a text knowledge extraction unit, an image knowledge extraction unit, a time sequence knowledge extraction unit, and an implicit knowledge extraction unit, which extracts valuable knowledge elements from a large amount of text, image, time sequence, and implicit knowledge data to form structured data that can be used for knowledge graph.

[0075] The text knowledge extraction unit realizes high-precision extraction of structured knowledge through multi-level semantic parsing technology, specifically including: A pre-trained model RoBERTa combined with a bidirectional LSTM-CRF is used to identify technical terms (such as "water turbine efficiency"), equipment models (such as "HT-3000"), and expert names from patent documents and technical reports, with an accuracy of not less than 95%.

[0076] Based on graph convolution network (GCN) and dependency syntax analysis, the semantic relationship between entities (such as "water turbine - efficiency - improvement") is extracted, and a technology association network is constructed to support deep analysis of domain knowledge.

[0077] The image knowledge extraction unit realizes deep semantic association of image-text data through cross-modal analysis technology, specifically including: A YOLOv7 model is used to analyze technical drawings, identify and label the coordinate positions and function descriptions of key equipment components (such as turbine blades and floodgate), and form a structured component information library The CLIP model is used to calculate the semantic similarity between image regions and text descriptions, establish a cross-modal mapping relationship (such as dynamically associating the turbine structure in the drawing with the "high-efficiency fluid dynamics design" description in the patent text), solve the semantic gap problem of image-text heterogeneous data, realize precise alignment of image entities and text knowledge, and provide a multi-modal fusion semantic basis for dynamic knowledge graph.

[0078] The time sequence knowledge extraction unit realizes the semantic conversion of time sequence data through multi-dimensional signal analysis and event modeling technology, specifically including: Performing Fourier transform and wavelet analysis on the time sequence data collected by sensors (such as water flow, equipment vibration frequency), extracting periodic characteristics and abnormal patterns (such as low-frequency vibration representing equipment wear trend).

[0079] Detecting key events (such as "flow sudden drop") in time sequence data through Temporal Fusion Transformer (TFT) model, and dynamically associating the detected events with device status logs based on Graph Neural Network (GNN) (such as associating to "floodgate failure" record), constructing time sequence-event knowledge subgraph, realizing accurate mapping from raw data to structured semantics, and associating to device status logs.

[0080] The implicit knowledge extraction unit realizes the digital conversion of expert experience through multi-modal semantic analysis and intelligent rule generation technology, specifically including: Based on Few-shot Learning and Prompt Engineering technology, automatically extract IF-THEN type executable decision rules (such as "IF vibration amplitude exceeds threshold THEN trigger shutdown check") from text, realize the conversion of unstructured experience to structured rules.

[0081] Using RoBERTa model to perform sentiment analysis (positive / negative) and confidence score (0-1) on expert statements, providing quantitative basis for rule weight allocation in knowledge base, ensuring accurate expression and reasonable application of expert implicit knowledge in knowledge graph.

[0082] The knowledge integration layer includes multi-modal entity alignment unit, dynamic knowledge fusion unit, and knowledge storage optimization unit. This module aims to effectively integrate the knowledge extracted from different data sources (such as real-time data, literature, reports, databases, expert interviews, etc.) to provide complete and consistent knowledge information for the next step of knowledge graph construction.

[0083] The multi-modal entity alignment unit realizes the deep integration of heterogeneous data through cross-modal semantic mapping and spatio-temporal correlation technology, specifically including: Using pre-trained model ERNIE-ViL to map text entities (such as "turbine efficiency parameters") and image entities (such as turbine structure diagram in technical drawings) to the same vector space, establishing cross-modal semantic correlation through cosine similarity calculation (threshold > 0.85), solving the semantic gap problem of heterogeneous data between text and image.

[0084] The key events (such as "traffic sudden drop") in the sensor timing data are spatio-temporally aligned with the records (such as "floodgate failure") in the operation and maintenance logs, and the "event-impact" relationship chain is constructed based on graph neural network (GNN), forming a multi-dimensional data linkage structured knowledge network, providing accurate cross-modal semantic basis for dynamic updating and intelligent reasoning of knowledge graph, and enhancing the global correlation and decision support capability of clean energy project data.

[0085] The dynamic knowledge fusion unit realizes the continuous optimization and consistency maintenance of the knowledge base through timeliness modeling and multi-source conflict resolution technology, including: The time-aware graph attention network (T-GAT) is used to add timestamp attributes (such as "a certain technology patent is valid until 2030") to the nodes (such as technology patents, equipment models) and relationships (such as "technology association", "patent citation") in the knowledge graph, supporting historical state backtracking (such as technology evolution path analysis) and future trend prediction (such as patent expiration warning).

[0086] Based on the logical reasoning engine and expert feedback mechanism, the contradictory descriptions in multi-source data (such as the definition differences of the same technical parameter in different documents) are detected and resolved, and through the confidence evaluation model (such as weighted scoring based on data source authority and update frequency), high-confidence data is preferentially adopted, and the knowledge graph is dynamically corrected combined with expert review results, ensuring the timeliness, consistency and decision reliability of the knowledge base, providing accurate and traceable knowledge support for clean energy projects.

[0087] The knowledge storage optimization unit realizes the efficient management and application of massive knowledge through the mixed architecture and standardized rule construction technology, including: A hybrid architecture of Neo4j and Apache AGE is used to build a distributed graph database, supporting storage of billions of nodes and millisecond-level multi-hop queries (such as "retrieve all maintenance solutions and technical patents associated with a faulty device"), and improving query efficiency in complex correlation scenarios through parallel computing and index optimization.

[0088] Implicit knowledge (such as expert experience rules) is converted into RDF triples (subject-predicate-object) and SWRL semantic rules (such as "IF device vibration exceeds threshold THEN trigger shutdown instruction"), a standardized rule base is constructed and integrated into the dynamic knowledge base, realizing the logical storage and programmable invocation of implicit knowledge, forming a knowledge management system that balances efficient storage and intelligent reasoning, and providing high availability and high scalability knowledge service foundation for the whole life cycle management of clean energy projects.

[0089] The dynamic updating and quality control layer includes an incremental knowledge updating unit, an online learning unit, an automated verification unit, and a manual review interface unit. Through Kafka, it triggers real-time updates of the incremental BERT model, the SWRL rule engine, and the visual annotation tool, forming a closed-loop quality optimization process.

[0090] The incremental knowledge updating unit drives the dynamic expansion and updating of the knowledge base through real-time data streams. Specifically, it includes: A real-time data pipeline is built using Apache Kafka to receive new data streams from sensors, patent databases, etc. (such as the latest patent files and device state time series data), triggering knowledge extraction and graph update processes. Based on a stream processing engine (such as Flink), the entire link automation of knowledge extraction, entity association, and graph update is realized, ensuring that the end-to-end delay from data reception to knowledge storage does not exceed 1 second. This supports the millisecond-level dynamic response capability of the knowledge base, meeting the high timeliness requirements of clean energy projects for real-time decision-making.

[0091] The online learning unit continuously optimizes the performance of the knowledge extraction model through incremental learning techniques. Specifically, it includes: Incremental BERT models are used to fine-tune new text data (such as technical documents and expert reports) online, dynamically adjusting the model parameters for entity recognition and relationship extraction. Combined with domain adaptation training, the model's generalization ability for new scenario data is improved, with entity recognition accuracy increasing by ≥5% after incremental updates. This ensures that the knowledge extraction process continuously optimizes as data evolves, adapting to the technical iteration and diversified management needs of clean energy projects.

[0092] The automated verification unit ensures the accuracy and consistency of the knowledge base through rule engines and logical reasoning techniques. Specifically, it includes: The SWRL rule engine is used to verify the logical consistency of knowledge, such as detecting whether the "device maintenance period" is less than the "maintenance time" and automatically freezing conflicting data.

[0093] A timeliness conflict detection model is built to identify outdated knowledge (such as invalid patents and obsolete technologies) and trigger alerts or automatic invalidation markers. Through multi-dimensional verification mechanisms, the static logical rationality and dynamic timeliness of the knowledge base are ensured, reducing decision-making risks.

[0094] The manual review interface unit realizes double protection of knowledge quality through human-machine collaboration mechanisms. Specifically, it includes: It provides a visual review interface, allowing experts to annotate and comment on erroneous nodes in the knowledge graph (such as incorrectly associated "technology-patent" relationships). The system uses a contrastive learning algorithm to extract features from the differences between human feedback and the existing knowledge base, automatically generate model correction strategies, and update the knowledge graph.

[0095] The expert review results are fed back into the incremental learning model, optimizing the accuracy of subsequent automated update processes and forming a closed-loop feedback chain of "manual annotation - model correction - knowledge optimization," which significantly improves the credibility of the knowledge base and the reliability of decision-making.

[0096] In one embodiment of the present invention, such as Figure 5 As shown, the knowledge graph construction module includes a knowledge structuring layer, a dynamic temporal knowledge graph construction layer, a distributed storage and query optimization layer, and a causal reasoning engine unit, including entity node units, relation edge units, time-aware graph embedding units, event-driven graph update units, index optimization units, and hybrid storage architecture units.

[0097] The knowledge structuring layer constructs a structured knowledge storage format through attribute annotation (technical parameters, project status) and semantic relationship definition (technical associations, participation relationships).

[0098] Each node in the entity node unit represents a single entity and comes with rich attribute information.

[0099] Each edge in the relation edge unit represents a semantic relationship between entities, enabling querying and reasoning within the graph.

[0100] The dynamic temporal knowledge graph construction layer adds timestamp attributes through the T-GAT network and real-time response event streams through the TGN framework, supporting technology evolution retrospective and millisecond-level graph updates.

[0101] The time-aware graph embedding unit adds timestamp attributes to node relationships through the T-GAT network, supporting historical backtracking and technology trend prediction.

[0102] The event-driven graph update unit responds to event streams in real time based on the TGN framework, dynamically adjusts the graph topology, and achieves millisecond-level semantic adaptation.

[0103] The distributed storage and query optimization layer achieves query latency ≤50ms and cache hit rate ≥90% through spatiotemporal composite indexes, subgraph caching, and a hybrid engine of Neo4j+InfluxDB+PostgreSQL.

[0104] The index optimization unit improves the high-frequency query response hit rate to ≥90% through spatiotemporal composite indexes and subgraph caching technology.

[0105] The hybrid storage architecture unit integrates the Neo4j graph database, the InfluxDB time-series database, and distributed PostgreSQL to ensure query latency ≤50ms.

[0106] The causal reasoning engine unit constructs a causal DAG based on the SCM model, supports counterfactual reasoning and root cause localization, and improves the interpretability of decisions.

[0107] Specifically, the knowledge graph construction module in this embodiment of the invention is implemented through the following sub-modules: The knowledge graph construction module includes a knowledge structuring layer, a dynamic temporal knowledge graph construction layer, a distributed storage and query optimization layer, and a causal reasoning engine unit. It integrates various types of knowledge through the knowledge graph and provides it to project managers through the intelligent reasoning engine. For example, if the system identifies a fault warning signal in equipment, it can quickly link it to similar historical fault cases and solutions, and provide optimal maintenance decision recommendations. Based on structured knowledge extracted from multiple sources, a dynamic temporal knowledge graph (DT-KG) is constructed to achieve semantic association, timeliness modeling, and intelligent reasoning of knowledge throughout the entire lifecycle of clean energy projects.

[0108] The knowledge structuring layer includes entity node units and relation edge units. Through the aforementioned knowledge extraction and integration techniques, the extracted knowledge information is structured and stored in a format suitable for subsequent applications, specifically including: Each entity node in the unit represents an entity such as technology, data, project, patent, person, or organization. Each entity node is accompanied by rich attribute information, such as technical parameters, project status, and patent application date.

[0109] Each edge in the relation edge unit represents a semantic relationship between entities, such as "technology A is associated with technology B" or "expert X participates in project Y". Each relation edge is accompanied by relation type and weight information to facilitate querying and reasoning in the graph.

[0110] The dynamic temporal knowledge graph construction layer includes a time-aware graph embedding unit and an event-driven graph update unit. By integrating timeliness modeling and streaming event response mechanisms, it achieves multi-dimensional dynamic evolution and precise adaptation of the knowledge graph, breaking through the static limitations of traditional knowledge bases. It constructs a dynamic knowledge foundation that combines temporal depth and real-time response capabilities, providing precise support for the technology lifecycle management, root cause localization, and inter-period decision optimization of clean energy projects.

[0111] The Time-Aware Graph Embedding Unit utilizes a Time-Aware Graph Attention Network (T-GAT) to achieve dynamic timeliness modeling and historical state management of the knowledge graph, specifically including: Add timestamp attributes (valid_from, valid_to) to nodes (such as technology patents, equipment models) and relationships (such as "technology A improves efficiency") in the knowledge graph. For example, mark "Patent P001 validity period: 2020-2030". Dynamically calculate the weight of the relationship between nodes within different time windows based on the graph attention mechanism (GAT) (such as "the association strength of a certain technology in 2023 is 0.92") to quantify the trend of technological evolution and the timeliness of association. For example, construct a subgraph of "HT-3000 turbine technology chain", associate patents (such as patent CN202310001), maintenance cases (such as the bearing overheating failure solution in 2023), and expert experience rules (confidence level 0.92), and mark "patent CN202310001 validity period until 2035" through the T-GAT network, and automatically trigger the association weight update when a similar new technology patent (CN202410002) is detected in 2024.

[0112] By using versioned storage technology to preserve historical snapshots of the knowledge graph, it is possible to support the tracing of the technology development path along the timeline (such as "the path of improving the efficiency of water turbines from 2015 to 2023"), forming a dynamic time-series graph that combines time dimension and semantic depth, providing high-precision knowledge support for the technology life cycle management, trend prediction and historical tracing of clean energy projects.

[0113] The event-driven graph update unit achieves dynamic, real-time updates of the knowledge graph through an event-driven mechanism and streaming graph neural network technology, specifically including: By receiving real-time event streams such as device fault alarms and patent expiration notifications through Apache Kafka, the graph incremental update process is triggered (such as adding "faulty equipment-maintenance solution" associations). Based on the Streaming Graph Neural Network (StreamGNN) framework Temporal Graph Network (TGN), the continuous time-series data is modeled, and the graph structure is dynamically adjusted (such as merging similar technology nodes and optimizing redundant association edges). Through time-sensitive node embedding and relation weight calculation, the knowledge graph structure and semantics are adapted in real time, ensuring that the graph can accurately reflect the latest status and technological evolution trends of clean energy projects, and providing highly timely and reliable knowledge support for scenarios such as fault diagnosis and resource scheduling.

[0114] The distributed storage and query optimization layer includes an index optimization unit and a hybrid storage architecture unit. Through spatiotemporal index acceleration and multimodal storage fusion technology, it builds a high-throughput, low-latency data service system, solves the storage fragmentation and query efficiency bottlenecks of multi-source heterogeneous data in the clean energy field, significantly improves the engineering availability and cross-scenario scalability of large-scale knowledge graphs, and empowers clean energy projects to evolve towards intelligence and high responsiveness.

[0115] The index optimization unit achieves efficient retrieval and performance optimization of knowledge graphs through spatiotemporal composite indexing and subgraph caching technology, specifically including: Establish a composite index structure based on time and space dimensions for nodes in the knowledge graph (such as "the operating status of a certain device in September 2023"), and accelerate multi-condition combined queries (such as historical data analysis and regional device status statistics) through spatiotemporal joint key values.

[0116] Memory caching optimization is performed on frequently accessed subgraphs (such as the "technology chain of a certain type of water turbine"), combined with the LRU (Least Recently Used) replacement strategy and preloading mechanism, so that the cache hit rate is no less than 90%, significantly reducing the response latency of complex association queries, providing millisecond-level retrieval support for real-time monitoring, fault tracing and decision analysis of clean energy projects, and improving the availability of knowledge services and the overall performance of the system.

[0117] The hybrid storage architecture unit enables efficient management and collaborative analysis of multimodal data through a hybrid storage architecture, specifically including: The Neo4j graph database is used to store the topology and attribute information of the knowledge graph, and the Cypher query language is supported to realize multi-hop association retrieval (such as "retrieve all technical patents and maintenance records associated with a certain device").

[0118] The InfluxDB time-series database stores time-series data such as water flow and equipment vibration frequency collected by sensors, and dynamically associates them with graph nodes (e.g., "reservoir node ← association → real-time water flow time-series curve") to achieve semantic fusion of structured knowledge and time-series signals.

[0119] By integrating Apache AGE to extend PostgreSQL, a distributed storage and computing architecture is built, supporting horizontal scaling and parallel queries on billions of nodes, ensuring query latency of ≤50 milliseconds in complex correlation scenarios, and providing high-throughput, low-latency multimodal data service capabilities for clean energy projects.

[0120] The causal reasoning engine unit is independent of the three levels of this module. It uses a graph-structured causal model (SCM) to model causal relationships and provide intelligent decision support for clean energy projects. Specifically, it includes: Based on conditional independence tests (such as PC algorithms) and domain knowledge, causal directed acyclic graphs (DAGs) can be constructed, for example, establishing a causal relationship chain of "water level drop → power generation reduction".

[0121] By using counterfactual reasoning to simulate the impact of intervention measures (such as changing turbine materials) on project performance (such as predicting the improvement in power generation efficiency), a quantitative basis for decision-making pre-assessment can be provided.

[0122] When abnormal power generation efficiency is detected, the source of the fault can be quickly located based on causal path tracing (such as "efficiency decline ← abnormal vibration ← blade corrosion"), forming a closed-loop reasoning link from causal discovery to root cause diagnosis, which significantly improves the efficiency of fault investigation and the scientific nature of decision-making, and provides explainable and verifiable intelligent support for risk control and optimization management of clean energy projects.

[0123] In one embodiment of the present invention, such as Figure 6 As shown, the knowledge application and decision-making module includes an intelligent knowledge recommendation layer and a multi-objective optimization decision-making layer, which includes a user profile modeling unit, a real-time triggering mechanism unit, an optimization model construction unit, a reinforcement learning algorithm unit, and a hybrid optimization strategy unit.

[0124] The intelligent knowledge recommendation layer is based on clustering algorithms, attention ranking networks, and GNN fault association retrieval to achieve scenario-based recommendations and emergency responses.

[0125] The user profile modeling unit uses clustering algorithms and dynamic priority models to achieve scenario-driven weight adjustment for knowledge recommendation.

[0126] The real-time triggering mechanism unit retrieves the fault correlation chain through GNN and pushes the emergency plan within 1 second, forming a closed-loop alarm response.

[0127] The multi-objective optimization decision layer generates multi-constraint optimal decision schemes through Pareto front analysis, PPO-MADDPG algorithm and genetic-reinforcement learning fusion.

[0128] The optimization model building unit defines multi-objective functions and constraints, and uses Pareto front analysis to generate non-dominated solution sets.

[0129] The reinforcement learning algorithm unit realizes digital twin simulation and multi-agent collaborative decision optimization through PPO and MADDPG algorithms.

[0130] The hybrid optimization strategy unit combines genetic algorithms and reinforcement learning to improve the efficiency of solving discrete / continuous variables and global convergence.

[0131] Specifically, the knowledge application and decision-making module is implemented through the following sub-modules: The knowledge application and decision-making module includes an intelligent knowledge recommendation layer and a multi-objective optimization decision-making layer. This module enables intelligent management of clean energy projects from knowledge acquisition to decision execution, significantly improving management efficiency, decision quality, and the ability to cope with complex environments, thus providing technical support for the industry's intelligent and sustainable development.

[0132] The intelligent knowledge recommendation layer includes a user profile modeling unit and a real-time triggering mechanism unit. Through dynamic profile analysis and event-driven response technology, it achieves accurate push of scenario-based knowledge, realizing full-scenario coverage from personalized recommendations to emergency response. It solves the problems of lagging recommendations and fragmented emergency response in traditional systems, and provides real-time and accurate decision support for clean energy projects.

[0133] The user profiling modeling unit provides personalized knowledge services through multi-dimensional user behavior analysis and dynamic scenario adaptation technology, specifically including: User profiles are built based on historical behavioral data (such as query records and operational preferences) and role definitions (such as operations engineers and project managers). User types are categorized and behavioral characteristics are quantified through clustering algorithms and feature engineering. For example, for the operations engineer role profile, based on historical operational preferences (such as frequently querying "bearing maintenance"), similar equipment failure cases are prioritized for recommendation.

[0134] By combining real-time environmental data (such as rainstorm warnings and equipment failure alarms) with project phases (planning, construction, and operation and maintenance), and using dynamic priority models (such as recommendation ranking networks based on attention mechanisms) to adjust knowledge recommendation weights, we can achieve scenario-driven adaptive services. For example, in the case of equipment failure, we can prioritize pushing relevant maintenance solutions, providing accurate and real-time knowledge support for different roles and scenarios, and significantly improving the efficiency of clean energy project management and user decision-making experience.

[0135] The real-time triggering mechanism unit achieves intelligent handling of abnormal scenarios through event-driven real-time response technology, specifically including: When sensors detect real-time signals such as excessive equipment vibration amplitude or abnormal water flow, the dynamic association mechanism of the knowledge graph is automatically triggered. Based on the graph neural network (GNN), the "fault-solution" chain is retrieved (such as matching historical fault cases, maintenance plans and related technical documents), and the emergency plan (such as shutdown inspection procedures and spare parts replacement guidelines) is pushed to the user terminal (such as mobile devices and monitoring screens) in real time through the message queue (such as RabbitMQ). This ensures that the end-to-end delay from anomaly detection to decision response is ≤1 second, forming a closed-loop alarm and handling process, which significantly improves the emergency efficiency and operational standardization of clean energy projects in the event of sudden failures.

[0136] The multi-objective optimization decision layer includes an optimization model building unit, a reinforcement learning algorithm unit, and a hybrid optimization strategy unit. Through the organic combination and progressive layering of these three components, the system achieves precise optimization of clean energy projects in complex decision-making environments, improves the overall benefits and sustainability of projects, significantly enhances decision convergence speed and multi-objective collaborative capabilities, and provides strong technical support for intelligent management in the field of clean energy.

[0137] The optimization model building unit maximizes the comprehensive benefits of clean energy projects through multi-objective optimization and constraint modeling techniques, specifically including: A multi-objective optimization problem is constructed with the objectives of maximizing power generation efficiency, minimizing maintenance costs, and minimizing ecological impact. Pareto front analysis is used to balance the conflicts between objectives and generate a non-dominated solution set.

[0138] By integrating technical feasibility constraints (such as equipment load limits and technical parameter ranges), resource limitations (budget and manpower allocation), and policy and regulatory requirements (such as carbon emission standards and environmental compliance), the optimal decision-making scheme that satisfies multiple constraints is found through hybrid optimization strategies (such as the fusion of genetic algorithms and reinforcement learning). This provides a quantitative basis for resource allocation, technology selection, and risk management of clean energy projects, and significantly improves the overall benefits and sustainability of the projects.

[0139] The reinforcement learning algorithm unit achieves dynamic decision optimization for clean energy projects through deep reinforcement learning and multi-agent collaborative technology, specifically including: A virtual decision-making environment (digital twin) is constructed based on the Proximal Policy Optimization (PPO) algorithm to simulate the long-term impact of scheduling schemes such as water level regulation and equipment maintenance. The policy network is trained through a reward function (such as a comprehensive score of power generation gain and cost weight) to generate decision-making strategies that take into account both efficiency and sustainability.

[0140] For complex scenarios such as multi-power station coordinated scheduling and regional energy allocation, the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm is adopted. Through the competition-cooperation mechanism among distributed agents, the decision-making of each subsystem is coordinated (such as the dynamic balance between reservoir water release and downstream power station power generation load). Combined with global state sharing and local policy optimization, it realizes efficient allocation of cross-system resources and risk collaborative control, significantly improves the decision convergence speed and multi-objective collaborative capability of clean energy projects, and provides efficient and reliable algorithm support for intelligent management in complex scenarios.

[0141] The hybrid optimization strategy unit achieves efficient solutions to complex decision-making scenarios through a fusion strategy of genetic algorithms and reinforcement learning, specifically including: For discrete decision variables such as equipment selection and technology route selection, a genetic algorithm is used to generate a pool of candidate solutions. The search space is expanded through crossover and mutation operations, and individuals with high fitness (such as low maintenance cost and high power generation efficiency solutions) are retained. At the same time, a reinforcement learning policy network is combined to perform gradient optimization on continuous decision variables (such as water level adjustment range and maintenance cycle), forming a hybrid optimization framework.

[0142] By combining the global search capability of genetic algorithms with the policy fine-tuning advantages of reinforcement learning through a dynamic weight allocation mechanism, the problem effectively avoids local optima traps, significantly improves the solution efficiency and global convergence of multi-objective optimization problems, and provides decision support with both breadth and precision for technology selection and resource allocation in clean energy projects.

[0143] In one embodiment of the present invention, such as Figure 7 As shown, the multi-source feedback acquisition layer and self-evolution layer include explicit feedback units, implicit feedback units, real-time data units, knowledge graph dynamic correction units, rule base optimization units, and model online learning units.

[0144] The multi-source feedback acquisition layer generates standardized feedback streams to drive system optimization through user ratings, behavior clustering, and sensor anomaly detection.

[0145] The explicit feedback unit verifies the consistency between user ratings and decision-making logic based on the rule engine, and corrects the knowledge graph in real time.

[0146] Implicit feedback units locate decision-making blind spots through behavioral clustering and causal models, generating implicit signals to drive a self-evolutionary mechanism.

[0147] The real-time data unit builds a high-throughput data pipeline using Apache Kafka, receives user ratings and device alarm signals by topic partitioning, and uses the Flink engine to perform data cleaning, feature extraction and window aggregation, ensuring end-to-end latency ≤500 milliseconds.

[0148] The self-evolutionary layer achieves a cold start time reduction of ≥70% and knowledge base self-optimization through GNN conflict resolution, Bayesian rule weight adjustment, and MAML cross-scene migration.

[0149] The knowledge graph dynamic correction unit identifies contradictory nodes based on GNN and TGN technologies, corrects the association weights, and automatically triggers the replacement of old knowledge through time decay weights to ensure the real-time performance and consistency of the graph.

[0150] The rule base optimization unit uses a Bayesian confidence model to dynamically adjust the weight of IF-THEN rules, and combines it with the SWRL engine to verify logical conflicts and freeze violation entries, forming a closed-loop rule optimization process.

[0151] The online learning unit of the model uses the MAML framework to achieve cross-scene transfer, reducing cold start time by ≥70%.

[0152] Specifically, the feedback and self-evolution module is implemented through the following sub-modules: The feedback and self-evolution module includes a multi-source feedback acquisition layer and a self-evolution layer. Through explicit / implicit feedback fusion and dynamic optimization closed-loop technology, it realizes the continuous evolution of knowledge system and decision-making model, solves the problems of static rigidity and response lag in traditional systems, provides a highly robust and flexible intelligent management base for clean energy projects, and significantly enhances the reliability of decision-making and the ability of technology iteration in complex environments.

[0153] The multi-source feedback acquisition layer includes explicit feedback units, implicit feedback units, and real-time data units. Through direct correction by explicit feedback and in-depth mining by implicit feedback, a two-way optimization closed loop of the knowledge management system is realized, providing highly responsive and robust intelligent support for clean energy projects.

[0154] The explicit feedback unit achieves efficient utilization of user explicit feedback through structured feedback collection and dynamic knowledge base correction technology, specifically including: The system collects rating data (e.g., 1-5 star ratings), adoption / rejection markers for decision-making solutions, and manually annotated error node information (e.g., "Technology A is incorrectly associated with Patent B"). Based on a rule engine and logical reasoning mechanism, the system verifies the rationality of the feedback, such as detecting the logical consistency between the rating and the decision results. Furthermore, through a contrastive learning algorithm, the system extracts features from the differences between the manually annotated error relationships and the existing knowledge graph, generating knowledge correction instructions (e.g., deleting incorrect associations and updating node attributes). This triggers real-time dynamic updates to the knowledge graph, ensuring closed-loop processing of explicit feedback and continuous optimization of the knowledge base.

[0155] The implicit feedback unit mines the value of users' implicit feedback through behavioral analysis and anomaly detection techniques, specifically including: Collect user interaction behavior data (such as frequently searched terms and skipped recommended items) and model potential user needs through clustering algorithms and attention mechanisms to identify blind spots in knowledge recommendation (such as long-tail technical documents that are not covered).

[0156] Analyze the correlation between abnormal sensor data (such as sudden increase in equipment vibration without warning) and knowledge graph, locate the missing response of decision-making module through time-series causal model, and generate implicit feedback signals (such as "fault warning rules need to be expanded").

[0157] By combining the aforementioned implicit feedback, the weights of the recommendation model are dynamically adjusted and the decision-making logic is optimized, forming a self-evolutionary mechanism that links "behavior-anomaly-knowledge," which significantly improves the system's scenario adaptability and the accuracy of knowledge services.

[0158] The real-time data unit achieves efficient processing of multi-source feedback data through a layered architecture and streaming computing technology, specifically including: A high-throughput data acquisition layer is built using Apache Kafka to receive real-time feedback streams (such as user ratings and device alarm signals) from user terminals, sensors, and project management systems. Data is then classified and transmitted in parallel by topic partitioning (such as " / feedback / user_rating" and " / feedback / sensor_alert").

[0159] The Flink stream processing engine is used to perform real-time cleaning (such as removing invalid fields and correcting format errors) and feature extraction on raw data, including normalization of user ratings and confidence marking of abnormal data (such as calculating the probability of anomalies based on historical distribution). Window aggregation and state management ensure low latency (end-to-end latency ≤ 500 milliseconds) and high reliability of data processing, forming a standardized, structured, high-quality data stream, providing a real-time and reliable input foundation for knowledge updates and decision optimization.

[0160] The self-evolutionary layer includes a knowledge graph dynamic correction unit, a rule base optimization unit, and a model online learning unit. It enables continuous optimization and adaptive evolution of the knowledge system and decision-making model of clean energy projects, ensuring the real-time nature of the knowledge graph, the reliability of the rule base, and the cross-scenario generalization ability of the model. This significantly improves the intelligence level of clean energy management and its ability to cope with complex environmental changes.

[0161] The knowledge graph dynamic correction unit achieves dynamic optimization of the knowledge graph through multi-dimensional feedback-driven and time-sensitive modeling techniques, specifically including: The matching degree between user feedback data and knowledge graph is calculated based on graph neural networks (GNNs). Inconsistent nodes (such as the conflict between the user-annotated "incorrect association" and the existing relationship confidence) are identified. The association weights are corrected or invalid relationships are deleted through a confidence fusion model (such as Bayesian weighting).

[0162] Add time decay weights to nodes and relationships. When patent expiration or technical standard updates are detected, the priority of old knowledge is automatically reduced and a replacement process is triggered (such as associating with the latest technical documents).

[0163] To address sudden feedback such as equipment failure alarms, a Time Sequence Graph Network (TGN) is used to update the knowledge graph in real time, constructing a "faulty equipment - root cause - solution" relationship chain. Streaming processing enables millisecond-level response times, ensuring the real-time nature, consistency, and decision reliability of the knowledge graph, providing highly robust dynamic knowledge support for clean energy projects. For example, by using TGN to update the "gate corrosion - coating failure - leakage" event chain in real time, the average fault repair time was reduced from 8 hours to 3 hours. The rule base optimization unit achieves continuous optimization and reliability assurance of the rule base through dynamic rule correction and logical consistency verification technologies, specifically including: Based on user adoption / rejection records and execution result verification data (such as the actual benefits of decision-making schemes), a Bayesian confidence update model is used to dynamically adjust the weight of IF-THEN rules (such as triggering a manual review process after a rule is rejected 3 times in a row). Combined with the SWRL rule engine, new knowledge is logically consistent (such as verifying the compliance of "maintenance cycle ≥ minimum maintenance cycle"). Detected logically conflicting items (such as maintenance cycle set too short) are automatically frozen and alarm notifications are generated.

[0164] Through rule versioning management and historical retrospective mechanisms, experts can modify or roll back conflicting rules, forming a closed-loop optimization process of "dynamic adjustment - logical verification - manual review" to ensure the scientific nature, compliance, and scenario adaptability of the rule base, providing highly reliable decision-making logic support for clean energy projects.

[0165] The online learning unit for models achieves adaptive optimization and rapid cross-scene transfer of models through incremental learning and meta-learning techniques, specifically including: The Online Proximal Policy Optimization (Online PPO) framework is adopted to dynamically adjust the policy network parameters based on real-time feedback data (such as the actual benefits of the decision scheme and the user adoption rate). The training stability is ensured by advantage function pruning and policy entropy constraints, so that the decision model can adapt to environmental changes (such as equipment aging and weather fluctuations) in real time.

[0166] Based on the incremental BERT model, online fine-tuning is performed on newly added text data (such as the latest technical documents and expert reports). Combined with domain adaptation training, the accuracy of entity recognition and relation extraction is improved (e.g., the accuracy of new term recognition is improved by ≥8%). For cross-scenario migration requirements (such as from hydropower station to wind farm), the Model Independent Meta-Learning (MAML) framework is adopted. The pre-trained model is quickly adapted using a small amount of new scenario data (such as wind farm operation and maintenance records). Through inner loop task optimization and outer loop meta-update mechanism, the cold start time is reduced by more than 70%, which significantly improves the model's generalization ability and deployment efficiency in multi-energy scenarios, and provides highly elastic algorithm support for the dynamic management and technology iteration of clean energy projects.

[0167] This invention's intelligent knowledge management system for clean energy projects constructs a full lifecycle knowledge management system for clean energy projects through the collaborative innovation of dynamic knowledge graphs, multimodal data fusion, and intelligent decision-making technologies. The system achieves efficient integration of multi-source heterogeneous data based on real-time data acquisition and streaming processing technologies. Combined with cross-modal entity alignment, event-driven updates, and a causal reasoning engine, it completes a deep transformation from explicit knowledge structuring to implicit experience rule-making. Furthermore, this invention generates dynamic decision-making schemes that balance efficiency, cost, and ecological protection through reinforcement learning and multi-objective optimization algorithms, and achieves transparent display of the decision-making logic through user interaction and visualization modules. Simultaneously, this invention continuously optimizes the knowledge graph and decision-making model through feedback-driven incremental learning and a federated knowledge-sharing mechanism, forming a self-evolving closed loop of "data—knowledge—decision—feedback." The implementation of this invention significantly improves the level of intelligent project management and the management efficiency of clean energy projects, reduces equipment failure risks and operating costs, optimizes the rationality of resource allocation and the scientific nature of technology selection, and has the ability to quickly migrate across scenarios (such as hydropower, wind power, and photovoltaic scenarios). It provides a high-precision, highly robust, and scalable intelligent management solution for the clean energy field, and promotes the industry towards high-quality development in the direction of digitalization and sustainability.

[0168] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0169] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A smart knowledge management system for clean energy projects, characterized in that, It includes modules for data acquisition and preprocessing, knowledge extraction and integration, knowledge graph construction, knowledge application and decision-making, and feedback and self-evolution; among which, The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from clean energy projects and transform the multi-source heterogeneous data into structured knowledge input data through intelligent preprocessing technology. The knowledge extraction and integration module is used to extract knowledge from structured knowledge input data, and to align and fuse knowledge extracted from different data sources through multimodal alignment and dynamic fusion technology to build a structured knowledge base. The knowledge graph construction module is used to process data from a structured knowledge base using the Time-Aware Graph Attention Network (T-GAT) and the Graph Structure Causal Model (SCM) to construct a dynamic temporal knowledge graph. The knowledge application and decision-making module is used to obtain decision results and knowledge recommendation results based on user needs and scenarios, and to output user feedback data by integrating causal-driven multi-objective optimization algorithms and dynamic temporal knowledge graphs. The feedback and self-evolution module is used to process user feedback data by integrating Temporal Graph Network and Online Proximal Policy Optimization (OnlinePPO) algorithm, and to dynamically optimize and adaptively adjust the knowledge graph, rule base and decision model based on the user feedback data, so as to feed the optimized results back to the knowledge extraction and integration module, the knowledge graph construction module and the knowledge application and decision module.

2. The system according to claim 1, characterized in that, The data acquisition and preprocessing module includes a data source and acquisition layer and a data preprocessing layer; the knowledge extraction and integration module includes a multimodal knowledge extraction layer, a cross-modal knowledge integration layer, and a dynamic update and quality control layer. The knowledge graph construction module includes a knowledge structuring layer, a dynamic temporal knowledge graph construction layer, a distributed storage and query optimization layer, and a causal reasoning engine unit; The knowledge application and decision-making module includes an intelligent knowledge recommendation layer and a multi-objective optimization decision-making layer; the feedback and self-evolution module includes a multi-source feedback collection layer and a self-evolution layer.

3. The system according to claim 2, characterized in that, The data source and acquisition layer includes structured data source units, unstructured data source units, and multimodal data interface units. The structured data source unit is used to acquire structured data through multi-dimensional collection and automated processing technologies. The structured data includes standardized technical parameters and related information extracted from document-type data through API interfaces, real-time collected physical parameters, a constructed multi-dimensional structured data pool, collected equipment operating status, and real-time monitored generator count. The unstructured data source unit is used to collect audio data using a voice recording device, and to convert the audio data into text and perform semantic segmentation using a speech recognition model to extract key experience rules; to extract unstructured operation and maintenance logs from the project management system, and to identify event types, participating entities and time attributes based on an event extraction model to construct event-attribute triples; and to integrate image data, use a visual recognition model to annotate the coordinates and functional description information of key components, and use a cross-modal pre-trained model to perform semantic alignment between image entities and text descriptions, outputting processed unstructured data. The multimodal data interface unit is used to deploy lightweight agent programs on edge devices to perform real-time caching and preliminary cleaning of local data; it uses a federated learning framework to build an encrypted channel to aggregate multi-project data, and uses homomorphic encryption and secure multi-party computation technology to encrypt the data and output multimodal data.

4. The system according to claim 2, characterized in that, The data preprocessing layer includes data cleaning and standardization units, tacit knowledge pre-structuring units, and streaming data processing units. The data cleaning and standardization unit is used to perform real-time anomaly detection on time-series data in structured data using the Temporal FusionTransformer model, and to repair missing values ​​and noise interference using the sliding window interpolation method; to perform spelling correction and terminology standardization on text data based on a rule engine and BERT model; to perform denoising and binarization on image data using the OpenCV library, and to identify key components and label coordinate information using the YOLOv7 model; and to convert text data into JSON-LD format with embedded semantic tags using format unification technology, to convert image data into vector graphics and dynamically associate them with text metadata, and to store time-series data as Parquet columnar files after unifying timestamps according to the ISO8601 standard, thus obtaining cleaned and standardized data. The tacit knowledge pre-structured unit is used to input processed unstructured data and cleaned and standardized data. The Whisper model is used to convert expert recordings into text and perform semantic segmentation to extract independent semantic fragments. Based on few-shot learning and prompting engineering, IF-THEN type decision rules are extracted from the text. The DEGREE event extraction model is used to identify event types, participating entities and time attributes from operation and maintenance logs, and event-attribute triples are constructed to output structured tacit knowledge. The streaming data processing unit receives multimodal data, as well as cleaned and standardized data and structured tacit knowledge. It builds a real-time data pipeline using Apache Kafka, receives sensor data streams by data type partitioning, and uses the Flink streaming engine to perform window aggregation, dynamic threshold alarms, and data persistence operations. The cleaned data is synchronously written to a hybrid storage architecture of time-series database and graph database. A lightweight TinyML model is deployed on the sensor nodes to achieve local data filtering and feature extraction to output the streaming data processing results.

5. The system according to claim 2, characterized in that, The multimodal knowledge extraction layer includes text knowledge extraction units, image knowledge extraction units, temporal knowledge extraction units, and implicit knowledge extraction units; among them, The text knowledge extraction unit is used to input cleaned and standardized text data. It uses the pre-trained model RoBERTa combined with bidirectional LSTM-CRF to identify entities, and extracts semantic relationships between entities based on graph convolutional network GCN and dependency parsing to extract text knowledge. The image knowledge extraction unit is used to input cleaned and standardized image data, and uses the YOLOv7 model for parsing to identify and label the coordinate positions and functional descriptions of key equipment components, forming a structured component information database; the CLIP model is used to calculate the semantic similarity between image regions and text descriptions, and to establish cross-modal mapping relationships to extract image knowledge; The time-series knowledge extraction unit is used to input cleaned and standardized time-series data, perform Fourier transform and wavelet analysis to extract periodic features and abnormal patterns, detect key events in the time-series data through the Temporal Fusion Transformer model, and dynamically associate the detected events with the device status log based on the graph neural network to construct a time-series-event knowledge subgraph to extract time-series knowledge. The implicit knowledge extraction unit is used to automatically extract IF-THEN type executable decision rules from text based on few-shot learning and prompting engineering techniques; and to extract implicit knowledge by performing sentiment analysis and confidence scoring on expert statements using the RoBERTa model.

6. The system according to claim 2, characterized in that, The cross-modal knowledge integration layer includes a multimodal entity alignment unit, a dynamic knowledge fusion unit, and a knowledge storage optimization unit. The multimodal entity alignment unit receives knowledge elements from the text knowledge extraction unit, image knowledge extraction unit, and temporal knowledge extraction unit. It uses the pre-trained model ERNIE-ViL to map text entities and image entities to the same vector space and establishes cross-modal semantic associations through cosine similarity calculation. It performs spatiotemporal alignment of key events in sensor time-series data with records in operation and maintenance logs, constructs event-influence relationship chains based on graph neural networks, forms a structured knowledge network with multidimensional data linkage, and outputs aligned multimodal knowledge. The dynamic knowledge fusion unit receives alignment knowledge from the multimodal entity alignment unit and implicit knowledge from the implicit knowledge extraction unit. It uses a time-aware graph attention network to add timestamp attributes to nodes and relationships in the knowledge graph. Based on a logical reasoning engine and an expert feedback mechanism, it performs conflict detection and resolution on contradictory descriptions in multi-source data. It prioritizes high-confidence data through a confidence evaluation model and dynamically corrects the knowledge graph based on expert review results, outputting the fused knowledge. The knowledge storage optimization unit receives fused knowledge from the dynamic knowledge fusion unit and uses Neo4j and ApacheAGE to build a distributed graph database hybrid architecture. It transforms implicit knowledge into RDF triples and SWRL semantic rules, builds a standardized rule base and integrates it into the dynamic knowledge base, and outputs the optimized knowledge.

7. The system according to claim 2, characterized in that, The dynamic update and quality control layer includes an incremental knowledge update unit, an online learning unit, an automated verification unit, and a manual review interface unit. The incremental knowledge update unit is used to receive new data streams from the data acquisition and preprocessing module through a real-time data pipeline built by Apache Kafka, triggering the knowledge extraction and graph update process. Based on the stream processing engine, it realizes full-link automation of knowledge extraction, entity association and graph update, so as to update the knowledge graph with new knowledge. The online learning unit is used to receive new knowledge from the incremental knowledge update unit. It uses the incremental BERT model to fine-tune the new text data online, dynamically adjusts the model parameters for entity recognition and relation extraction, and improves the model's generalization ability to new scenario data through domain-adaptive training. An automated verification unit receives new knowledge from the incremental knowledge update unit and verifies the logical consistency of the knowledge based on the SWRL rule engine; it constructs a timeliness conflict detection model to identify expired knowledge and trigger alarms or automatic expiration markers, and ensures the static logical rationality and dynamic timeliness of the knowledge base through a multi-dimensional verification mechanism. The manual review interface unit receives the verification results from the automated verification unit, supports experts in annotating and commenting on erroneous nodes in the knowledge graph, extracts features from the differences between manual feedback and the existing knowledge base through a comparative learning algorithm, automatically generates model correction strategies and updates the knowledge graph; and feeds back the expert review results to the incremental learning model to optimize the accuracy of subsequent automated update processes, forming a closed-loop feedback chain.

8. The system according to claim 2, characterized in that, The knowledge structuring layer includes entity node units and relation edge units, among which... Entity node units are used to receive fused knowledge from the knowledge extraction and integration module. Each node represents an entity, and each entity node is accompanied by attribute information, which is stored in a structured manner as nodes in the knowledge graph. The relation edge unit is used to receive fused knowledge from the knowledge extraction and integration module. Each edge represents the semantic relationship between entities, and each relation edge is accompanied by relation type and weight information. Each relation and its type and weight information are structured and stored as edges in the knowledge graph.

9. The system according to claim 2, characterized in that, The dynamic temporal knowledge graph construction layer includes a time-aware graph embedding unit and an event-driven graph update unit; among which, The time-aware graph embedding unit receives structured knowledge from entity node units and relation edge units, adds timestamp attributes to nodes and relations in the knowledge graph, dynamically calculates the weights of relations between nodes within different time windows based on the graph attention mechanism, quantifies the trend of technological evolution and the timeliness of association; retains historical state snapshots of the knowledge graph through versioned storage technology, supports backtracking the technological development path along the timeline, and forms a dynamic time-series graph that combines time dimension and semantic depth; The event-driven graph update unit receives real-time event streams from the data acquisition and preprocessing module, triggers the incremental graph update process, and models continuous time-series data based on the Temporal Graph Network framework to dynamically adjust the graph structure. Through time-sensitive node embedding and relation weight calculation, it achieves real-time adaptation of the knowledge graph structure and semantics.

10. The system according to claim 2, characterized in that, The distributed storage and query optimization layer includes an index optimization unit and a hybrid storage architecture unit. The index optimization unit receives dynamic knowledge graph structures from the time-aware graph embedding unit and the event-driven graph update unit, establishes a composite index structure based on time and space dimensions for nodes in the knowledge graph, accelerates multi-condition combined queries through spatiotemporal joint key-value pairs, optimizes memory caching for frequently accessed subgraphs, and outputs optimized index information by combining LRU replacement strategy and preloading mechanism. The hybrid storage architecture unit receives data from entity node units, relation edge units, time-aware graph embedding units, event-driven graph update units, and index optimization units. It uses the Neo4j graph database to store the topology and attribute information of the knowledge graph, and uses the InfluxDB time-series database to store time-series data, dynamically associating it with graph nodes to achieve semantic fusion of structured knowledge and time-series signals. It integrates Apache AGE to extend PostgreSQL and builds a distributed storage and computing architecture.

11. The system according to claim 2, characterized in that, The causal reasoning engine unit is used for: It receives dynamic knowledge graph structures from the event-driven graph update unit and knowledge graph data stored from the hybrid storage architecture unit, and constructs a causal directed acyclic graph based on conditional independence tests and domain knowledge. Simulate the impact of intervention measures on project performance through counterfactual reasoning; When abnormal power generation efficiency is detected, the source of the fault is quickly located based on causal path tracing, forming a closed-loop reasoning link from causal discovery to root cause diagnosis.

12. The system according to claim 2, characterized in that, The intelligent knowledge recommendation layer includes a user profile modeling unit and a real-time triggering mechanism unit, among which... The user profile modeling unit is used to build user profiles based on historical behavior data and role definitions. It classifies user types and quantifies behavioral characteristics through clustering algorithms and feature engineering. Combining real-time environmental data and project stages, it uses a dynamic priority model to adjust knowledge recommendation weights and realize scenario-driven adaptive services. The real-time triggering mechanism unit receives user profiles and personalized knowledge recommendation weights from the user profile modeling unit, as well as real-time sensor data from the data acquisition and preprocessing module. When the sensor detects an abnormal real-time signal, it automatically triggers the dynamic association mechanism of the knowledge graph, retrieves the fault-solution chain based on the graph neural network, and pushes the emergency plan to the user terminal in real time through the message queue.

13. The system according to claim 2, characterized in that, The multi-objective optimization decision layer includes an optimization model building unit, a reinforcement learning algorithm unit, and a hybrid optimization strategy unit, among which... The optimization model building unit receives dynamic knowledge graph data from the knowledge graph building module, constructs a multi-objective optimization problem with the objective functions of maximizing power generation efficiency, minimizing maintenance costs, and minimizing ecological impact, and uses Pareto front analysis to balance the conflicts between objectives and generate a non-dominated solution set; it integrates technical feasibility constraints, resource limitations, and policy and regulatory requirements, and solves the optimal decision scheme that satisfies multiple constraints through a hybrid optimization strategy; The reinforcement learning algorithm unit receives multi-objective optimization models and preliminary decision schemes from the optimization model construction unit, as well as real-time data from the data acquisition and preprocessing module. It constructs a virtual decision-making environment based on the proximal policy optimization algorithm to simulate the long-term impact of scheduling schemes. It trains the policy network through the reward function to generate decision policies. The multi-agent deep deterministic policy gradient algorithm coordinates the decisions of each subsystem through the competition-cooperation mechanism among distributed agents. It combines global state sharing and local policy optimization to achieve efficient allocation of cross-system resources and collaborative risk control. The hybrid optimization strategy unit receives optimization decision strategies from the reinforcement learning algorithm unit and multi-objective optimization models from the optimization model construction unit. For discrete decision variables, a genetic algorithm is used to generate a candidate solution pool. Crossover and mutation operations are used to expand the search space and retain high-fitness individuals. At the same time, the reinforcement learning policy network is combined to perform gradient optimization on continuous decision variables, forming a hybrid optimization framework. The global search capability of the genetic algorithm is combined with the policy fine-tuning advantage of reinforcement learning through a dynamic weight allocation mechanism.

14. The system according to claim 2, characterized in that, The multi-source feedback acquisition layer includes explicit feedback units, implicit feedback units, and real-time data units. The explicit feedback unit is used to collect scoring data of knowledge recommendation results, adoption / rejection tags of decision schemes, and manually labeled error node information through the user interface. It verifies the rationality of the feedback based on the rule engine and logical reasoning mechanism, and extracts features from the differences between manually labeled error relationships and existing knowledge graphs through a contrastive learning algorithm to generate knowledge correction instructions. The implicit feedback unit is used to collect user interaction behavior data and model potential user needs through clustering algorithms and attention mechanisms to identify blind spots in knowledge recommendation; analyze the correlation between abnormal sensor data and knowledge graphs, locate the response deficiencies of the decision-making module through time-series causal models, and generate implicit feedback signals; based on implicit feedback, dynamically adjust the weights of the recommendation model and optimize the decision-making logic to form a linked self-evolution mechanism. The real-time data unit is used to build a high-throughput data acquisition layer using Apache Kafka, receiving real-time feedback streams from user terminals, sensors, and project management systems. It classifies and transmits data in parallel by topic partitioning; it uses the Flink stream processing engine to perform real-time cleaning and feature extraction on the raw data, including normalization of user ratings and confidence marking of outlier data. Low latency in data processing is ensured through window aggregation and state management.

15. The system according to claim 2, characterized in that, The self-evolutionary layer includes a knowledge graph dynamic correction unit, a rule base optimization unit, and a model online learning unit. The knowledge graph dynamic correction unit receives explicit feedback data and knowledge correction instructions from the explicit feedback unit, implicit feedback signals from the implicit feedback unit, and real-time feedback data from the real-time data unit. It calculates the matching degree between user feedback data and the knowledge graph based on a graph neural network, identifies contradictory nodes, corrects association weights or deletes invalid relationships through a confidence fusion model, adds time decay weights to nodes and relationships, and updates the graph in real time using a temporal graph network to construct relationship chains for sudden feedback. The rule base optimization unit receives corrected knowledge graph data from the knowledge graph dynamic correction unit and decision execution result verification data from the system. It dynamically adjusts the weights of IF-THEN rules using a Bayesian confidence update model and combines the SWRL rule engine to perform logical consistency verification on new knowledge. It automatically freezes and generates alarm notifications for detected logically conflicting entries. Through rule versioning management and historical backtracking mechanisms, it supports experts in correcting or rolling back conflicting rules, forming a closed-loop optimization process. The online learning unit receives corrected knowledge graph data from the dynamic knowledge graph correction unit, optimized rule base data from the rule base optimization unit, and real-time feedback data from the real-time data unit. It employs an online near-end policy optimization framework to dynamically adjust policy network parameters based on real-time feedback data, ensuring training stability through advantage function pruning and policy entropy constraints. Based on the incremental BERT model, it performs online fine-tuning of newly added text data, combining domain-adaptive training to improve the accuracy of entity recognition and relation extraction. For cross-scene transfer requirements, it adopts a model-independent meta-learning framework, adapting the pre-trained model using a small amount of new scene data, through inner-loop task optimization and outer-loop meta-update mechanisms.

Images