Multi-agent cooperative processing system and processing method for wind power fault questions and answers

Abstract

The invention discloses a multi-agent cooperative processing system for wind power fault questions and answers. The system comprises a knowledge construction agent module used for extracting a knowledge triple from an operation and maintenance document and constructing a wind power fault knowledge graph; the task planning agent module is used for analyzing a problem input by a user and generating a task execution plan; and the task execution agent module sequentially calls information extraction, protection analysis and parameter analysis agents to execute tasks, and inputs a result into a language model to generate a natural language answer with a reasoning path. According to the method, cooperative processing of knowledge integration, task decomposition and data analysis can be realized, the problems of knowledge fragmentation, term inconsistency and inexplanatory reasoning in the prior art are effectively solved, and the accuracy, reliability and traceability of wind turbine generator fault diagnosis are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the technical field of wind turbine fault diagnosis and handling methods, specifically relating to a multi-agent collaborative processing system for wind power fault question answering, and also to the aforementioned multi-agent collaborative processing method for wind power fault question answering. Background Technology

[0002] Wind turbines are prone to various electrical and mechanical faults due to their complex structure, variable operating conditions, and harsh operating environment. Current operation and maintenance (O&M) knowledge primarily comes from heterogeneous materials from multiple sources, including equipment manufacturer manuals, protection setting and commissioning records, duty defect tickets, maintenance reports, and historical operating data. To improve fault handling efficiency, the industry typically employs intelligent question-answering solutions, such as Retrieval Enhanced Generation (RAG) and Large Language Modeling (LLM), attempting to combine document retrieval with answer generation to provide reference conclusions for frontline personnel. However, in actual wind farm applications, existing technologies still have significant shortcomings: 1. Fragmented documentation and inconsistent terminology: Documentation from different models and manufacturers lacks consistency in terminology, with multiple naming conventions or abbreviations for the same concept, leading to unstable retrieval and difficulties in cross-document reasoning. In particular, the correspondence between "alarm / fault codes—trigger logic—handling measures" is scattered across multiple documents, making it difficult to directly connect and utilize them.

[0003] 2. Difficulty in Explicitly Defining Cross-Source Causality and Rules: Fault diagnosis often requires linking protection settings (such as current / voltage / frequency thresholds and delays) with operating parameters before the fault (such as wind speed, pitch angle, generator current, and temperature rise). Existing RAGs mostly remain at the level of piecing together text fragments of evidence, lacking a mechanism for structured expression and joint reasoning of setting rules, cascading trigger relationships, and time series patterns, making it difficult to meet the requirements of the power and energy industries for safety, interpretability, and auditability.

[0004] 3. LLM's Insufficient Capabilities in Numerical and Time Series Computation: The actual handling of wind turbine faults often relies on comparing actual operational measurements from a period prior to the fault with the protection / limit / logic rules in the turbine configuration. This is a typical structured computation and time-series discrimination task. However, general-purpose LLMs lack the precise computational or time-series analysis capabilities for such tasks. Simply feeding raw data into an LLM is unlikely to yield stable and consistent conclusions. Furthermore, tasks such as protection setting verification, threshold trigger determination, trend / abrupt change detection, and interval statistics require robust time-series computation and interpretable output. Relying solely on LLM's generative inference of text can easily lead to calculation errors, misjudgments, and unverifiable issues.

[0005] 4. The knowledge graph construction is not systematic: Most existing work focuses on extracting knowledge of general equipment or single models in a piecemeal manner, lacking an integrated graph model and conflict detection mechanism for "fault diagnosis - protection setting - operating parameters" (such as multiple sets of codes corresponding to the same fault across documents, the same fault being marked as "resettable / non-resettable", etc.), which makes it difficult to reuse knowledge and continuously iterate.

[0006] In summary, existing technologies cannot simultaneously meet the comprehensive requirements of "unified cross-source knowledge, calculable and verifiable protection settings, interpretable and analyzable operating parameters, and traceable and verifiable Q&A results" in wind power scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent collaborative processing system for wind power fault question answering, which can structure multi-source heterogeneous knowledge, perform unified modeling and evaluation of protection settings and operating parameters, and decompose and traceably manage the question answering process, thereby improving the accuracy, efficiency and engineering usability of wind power fault diagnosis and handling.

[0008] Another objective of this invention is a multi-agent collaborative processing method for wind power fault response.

[0009] The technical solution adopted in this invention is a multi-agent collaborative processing system for wind power fault response, comprising the following modules: The knowledge construction agent module is used to extract knowledge triples from wind turbine operation and maintenance documents and construct a wind power fault knowledge graph. The knowledge graph includes at least the following node types: fault node, trigger condition node, protection setting node, operating parameter node, and handling measure node. The task planning agent module is used to perform semantic parsing on the input natural language questions, determine the type of information involved in the questions, and generate a task execution plan that includes three parts: basic fault information extraction, protection setting analysis, and operating parameter analysis. The task execution agent module is used to sequentially call the information extraction sub-agent, protection analysis sub-agent, parameter analysis sub-agent, and answer generation sub-agent to execute the task according to the task execution plan.

[0010] The invention is further characterized in that, The knowledge building agent module includes: The system includes a text parsing unit, a triplet extraction unit, and a conflict detection unit. The conflict detection unit is used to detect situations where multiple codes correspond to the same fault name or where there are contradictions in the same alarm attribute, and outputs a conflict identifier.

[0011] The task planning agent module performs semantic parsing on the input natural language question by combining template matching and semantic vector similarity calculation, with a semantic vector similarity threshold of 0.85 to 0.95.

[0012] The information extraction sub-agent uses a recursive knowledge graph retrieval algorithm to obtain the set of all triples related to the target node from the knowledge graph; The protection analysis sub-agent reads the unit's operating data within a preset time window before a fault occurs and calculates whether each protection is triggered based on the protection threshold rules. The parameter analysis sub-agent performs time series analysis on specified operating parameters to identify trends, ranges, and abrupt changes. The answer generation sub-agent integrates the aforementioned analysis results and prompts into the language model, and outputs a natural language answer that includes the reasoning path and knowledge source.

[0013] The parameter analysis sub-agent detects abrupt changes using the sliding window variance method, with a window width of 10–60 seconds.

[0014] Another technical solution adopted in this invention is a multi-agent collaborative processing method for wind power fault question answering, which uses the above-mentioned multi-agent collaborative processing system for wind power fault question answering, and the steps are as follows: S1. Receive natural language questions input by the user and perform semantic standardization processing; S2. The task planning agent module performs semantic parsing on the problem and generates a task planning table containing flags for basic fault information extraction, protection setting analysis, and operating parameter analysis. S3. The information extraction sub-agent calls the knowledge graph interface provided by the knowledge construction agent module to obtain the complete set of triples related to the target node according to the recursive knowledge graph retrieval algorithm. S4. The protection analysis sub-agent extracts the running data within the preset time window and calculates the trigger status, parameter range and trend according to the protection setting rules. S5. The parameter analysis sub-agent performs time series analysis on the specified parameters to detect rising, falling, fluctuating or abrupt features and generate physical interpretations. S6. The answer generation sub-agent integrates the above results and prompts, inputs them into the language model to generate a natural language answer with a reasoning path, and outputs it.

[0015] Another feature of the present invention is that, In S2, semantic parsing employs an algorithm that combines template matching with semantic vector retrieval.

[0016] In S3, the recursive knowledge graph retrieval algorithm includes the following process: 1) Initialize the access table and read the seed node set; 2) Call the RetrieveTriples function to retrieve triples associated with the current node; 3) Add the new node to the access table and recursively call the next level of retrieval; 4) Return the complete set of triples.

[0017] In S4, the data sampling period is 1 to 10 seconds, and the time window length is 3 to 15 minutes.

[0018] The final output of S6 includes natural language responses, trigger rule numbers, referenced document identifiers, and inference path information.

[0019] The beneficial effects of this invention are: Compared with existing single RAG or pure LLM question-answering technologies, the present invention has the following advantages: 1) Upgrading from document retrieval to knowledge reasoning: By transforming fragmented documents into structured graphs through WindKG-Agent, this invention can answer questions without relying on whether the user happens to mention the correct keywords. Instead, it can extend all contextual knowledge related to the fault downstream based on the semantic relationship of the graph, realizing reasoning-based question answering across documents and sources, and solving the problem of information isolation in the prior art. 2) Upgrading from black-box generation to interpretable generation: Since the Task-Planning Agent has broken down the analysis to be performed step by step in the pre-stage, and the Task-Action Agent has explicitly output the structured analysis results of each step, the final natural language answer can be traced back to "which protection rule was used", "which segment of operational data was viewed", and "which graph triplet was referenced", which greatly improves the interpretability and traceability that are most valued in the wind power field; 3) Traditional LLM has difficulty performing numerical comparisons, threshold judgments, and time series trend identification stably. This invention achieves a closed loop of rules / data → structured results → language model interpretation by delegating these computational tasks to specialized protection analysis agents and parameter analysis agents, and then feeding the results back to LLM. 4) Conflicts are perceptible and versions are manageable: With the help of conflict detection and expert verification, this invention can proactively discover situations that are easy to mislead large models, such as "different codes with the same name" and "mutually exclusive processing suggestions for different documents", and explicitly expose them in the final answer, avoiding the risk of traditional systems returning conflict information mixed together, leading to misjudgment by operation and maintenance personnel. 5) Facilitates the construction of wind power fault Q&A benchmark sets and automated evaluation Since the task planning and execution of this invention are structured, recordable, and replayable, it is also very suitable for building a wind power fault diagnosis question-and-answer benchmark: each question can fix the spectrum segment it retrieves and the protection / parameter analysis path it adopts. Subsequently, only the LLM or the inference strategy needs to be changed to make repeatable comparisons in the same context. Attached Figure Description

[0020] Figure 1 This is a block diagram of the multi-agent collaborative processing system for wind power fault response according to the present invention; Figure 2 This is a flowchart of the task execution process for the multi-agent collaborative processing method for wind power fault response in this invention. Figure 3 This is a flowchart of the knowledge graph recursive retrieval algorithm in the multi-agent collaborative processing method of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 The multi-agent collaborative processing system (WindServe-Agent) for wind power fault question answering of this invention consists of three core modules: a knowledge construction layer, a task planning layer, and a task execution layer. These three modules collaborate through a unified data interaction and context passing mechanism. Figure 1 As shown, the module relationships and data flow of the knowledge building agent module WindKG-Agent, the task planning agent module Task-Planning Agent, and the task execution agent module Task-Action Agent are illustrated.

[0023] This system can be viewed as an intelligent assembly, consisting of approximately the following modules based on the system's calculated load percentage: WindKG-Agent module: accounts for 30% to 40% of the system's computing resources; Task-Planning Agent module: occupies 10% to 20% of the system's computing resources; Task-Action Agent module: accounts for 40% to 60% of the system's computing resources; The Task-Action Agent contains four sub-modules, with the following system resource allocations: information extraction agent 20%–25%, protection analysis agent 20%–30%, parameter analysis agent 25%–35%, and answer generation agent 15%–20%.

[0024] Preferably, the multi-agent collaborative processing system for wind power fault Q&A of the present invention is deployed on a server with a GPU or AI accelerator and can be called through a Python or C++ interface; it can also be integrated into a cloud platform, operation and maintenance monitoring terminal or mobile device, or edge computing node and can be called by operation and maintenance personnel, algorithm system or host computer through API or graphical terminal.

[0025] The knowledge layer problem that this invention needs to solve is: how to transform unstructured knowledge that is originally scattered in different operation and maintenance documents, control strategy descriptions, unit protection tables, and manufacturer manuals into a unified wind power fault knowledge graph that can be automatically retrieved by programs and stably referenced by models, and how to automatically discover and label problems such as "different codes with the same name", "attribute conflicts", and "inconsistent handling measures" during the graph construction process, thereby ensuring that the knowledge called upon in the subsequent question-and-answer stage is structured, conflict-free, and traceable.

[0026] The WindKG-Agent knowledge construction agent module of this invention is used to extract entities and relationships related to wind power faults from multiple sources of operation and maintenance documents, such as original operation and maintenance documents, vendor manuals, and fault handling records, to transform unstructured data into a computable knowledge graph, and to detect and label potential conflicts.

[0027] This module includes a text parsing unit, a triple extraction unit, and a conflict detection unit. The text parsing unit parses the text using an LLM (Large Language Model) and extracts knowledge triples in the form of "subject-relation-object".

[0028] To ensure the consistency of knowledge, this invention sets up a conflict detection unit: when the same fault name corresponds to multiple codes or unified alarm attributes, the system performs similarity calculation and rule comparison (similarity threshold 0.8~0.95), and outputs conflict triplet tags for expert verification.

[0029] The knowledge graph output by WindKG-Agent includes at least the following node types: Fault node, Trigger node, Protection Setting node, Parameter node, and Action node, with semantic relationships including "Triggering Reason", "Protection Constraint", and "Recommended Action".

[0030] By transforming fragmented documents into structured graphs using WindKG-Agent, questions can be answered without relying on whether the user happens to mention the correct keywords. Instead, the graph's semantic relationships can be used to extend downstream all contextual knowledge related to the problem, enabling reasoning-based question answering across documents and sources. This solves the problem of information silos in existing technologies.

[0031] The reasoning execution problem to be solved by this invention is: when a user's question may involve multiple dimensions such as "What is the fault?", "Can this alarm be reset?", "What were the wind speed, pitch, pitch hydraulic pressure, and grid connection status at the time?", "Is the protection malfunctioning?", and "Should the power be cut off?", how can the system automatically decompose a natural language question into multiple executable tasks (e.g., checking basic fault information → checking protection settings → performing operational parameter analysis → generating an answer), and orderly schedule these tasks to different agents for execution, and finally unify the natural language answer, structured results, and trigger rule descriptions, avoiding the redundancy and instability caused by dumping all documents onto the LLM at once.

[0032] Based on this, the Task-Planning Agent module of the present invention is to parse the semantics of the natural language question input by the user, determine the knowledge type and analysis type to be invoked in this question-and-answer session, and generate a structured task execution plan. The task planning agent employs an algorithm combining template matching and semantic vector similarity retrieval to perform semantic parsing and classification of the input question. The semantic vector similarity threshold is 0.85~0.95, identifying whether it involves three types of information: 1) Basic fault information (e.g., "What is the cause of the XXX fault"); 2) Protection setting analysis (e.g., "Did this protection activate?"); 3) Analysis of operating parameters (e.g., "How do wind speed and blade pitch angle change?").

[0033] Each type of information is controlled by a Boolean flag (True / False), which forms a task planning table to guide the scheduling of subsequent task execution layers.

[0034] The Task-Action Agent module is the core of the execution layer. It sequentially calls sub-agents such as information extraction, protection analysis, parameter analysis, and answer generation according to the task planning table to complete specific analysis actions and integrate the analysis results with the language model generation results. Information-Extraction Agent: Employs a recursive knowledge graph retrieval algorithm, such as... Figure 3 As shown, the set of all triples related to the target node is obtained from the knowledge graph to realize multi-level knowledge retrieval; Protection Analysis Sub-Agent: Within a preset time window Δt_pre (3 min to 15 min) before the fault occurs, read the unit's SCADA data and calculate whether each protection is triggered according to the protection threshold rules; if the measured value meets the threshold condition, it is determined to be triggered, and the trigger status and change trend are output. Parametric Analysis Sub-Agent: Performs time series analysis on selected operating parameters to identify increases, decreases, fluctuations, or abrupt changes. The parametric analysis sub-agent detects abrupt change points using the sliding window variance method, with a window width of 10–60 seconds, calculates the rate of change range of 0.1–10%, and generates a physical interpretation. Answer generation sub-agent: This sub-agent integrates the structured data, text descriptions, trigger rules, time series, parameter trends, and other information generated from the aforementioned analysis, along with prompts, and organizes them into multiple prompt segments. These segments are then input into the Language Model (LLM) prompt to generate a natural language answer that includes the reasoning path and cited knowledge sources. After obtaining the LLM output, this answer, along with the source information, is returned to the user.

[0035] Because the Task-Planning Agent breaks down the analysis to be performed step by step in the pre-processing stage, and the Task-Action Agent explicitly outputs the structured analysis results of each step, the final natural language answer can be traced back to "which protection rule was used", "which segment of operational data was viewed", and "which graph triplet was referenced", which greatly improves the interpretability and traceability that are most valued in the wind power field.

[0036] The advantages of the multi-agent collaborative processing system for wind power fault response of the present invention are as follows: A wind turbine fault diagnosis question-and-answer framework is constructed, based on a knowledge graph, centered on task planning, and implemented by multiple agents. This framework enables the system to first identify the problem type, then consult relevant knowledge, examine setpoints and curves, and finally provide evidence-based textual conclusions and sources. This achieves a highly accurate, consistent, and interpretable intelligent question-and-answer service for wind turbine faults. It also enables collaborative processing of knowledge integration, task decomposition, and data analysis, effectively addressing the problems of fragmented knowledge, inconsistent terminology, and uninterpretable reasoning in existing technologies. This significantly improves the accuracy, reliability, and traceability of wind turbine fault diagnosis.

[0037] Through a multi-agent collaborative mechanism, a three-layer structure of knowledge graph, task planning and task execution was introduced for the first time in the intelligent question-answering system for wind turbines, enabling interpretable reasoning for complex problems. A recursive knowledge graph retrieval algorithm is proposed, which can efficiently extract related knowledge without losing context. A unified protection setting analysis function model was designed to realize threshold judgment and trend analysis across different models; The generated answers carry natural language responses, trigger rule numbers, referenced document identifiers, and reasoning path information to achieve traceability and interpretability of the results, meeting the safety and traceability requirements in the energy sector.

[0038] Example 2 Furthermore, the knowledge building agent module WindKG-Agent can use any model with information extraction capabilities, such as GPT-4, Gemini, Claude, DeepSeek-V, etc.

[0039] Protection rules can be stored in relational databases or graph databases, such as Neo4j and OrientDB.

[0040] Parametric analysis can be replaced by trend detection models based on machine learning, such as LSTM and Prophet.

[0041] System interfaces can be implemented via RESTful API or gRPC.

[0042] Example 3 The multi-agent collaborative question-answering system of the present invention includes: WindKG-Agent (Wind Power Knowledge Graph Agent) is used to extract entities and relationships related to wind power faults from multi-source operation and maintenance documents, build a unified knowledge graph, and detect and label potential conflicts. The Task-Planning Agent is used to perform semantic parsing on the user's natural language questions, determine the type of knowledge and analysis required for this question-and-answer session, and generate a structured task execution plan. The Task-Action Agent is used to call sub-agents such as information extraction, protection analysis, parameter analysis, and answer generation according to the task execution plan to complete specific analysis actions. It also organizes the structured data, text descriptions, trigger rules, time strings, parameter trends, and other information generated by the aforementioned agents into a Prompt that can be input into an LLM, and returns it to the user along with the source information after obtaining the LLM output.

[0043] The system can be deployed on servers, cloud platforms, or edge computing nodes, and can be invoked by operations and maintenance personnel, algorithm systems, or host computers via API or graphical terminals.

[0044] The knowledge construction layer, WindKG-Agent, employs a clear three-stage design for building the wind turbine fault knowledge graph. This ensures that the graph is not manually entered all at once, but rather follows a closed-loop process of "LLM extraction → rule / statistical assistance → conflict detection → manual confirmation," guaranteeing high coverage, high consistency, and maintainability of the final knowledge. Its specific technical solutions include: Knowledge graph pattern definition stage By systematically reviewing wind turbine alarm code tables, PLC logic specifications, pitch / current / yaw subsystem manuals, fault handling cards, historical maintenance records, and other relevant documents, at least the following types of nodes are defined: Fault / Alarm node; Fault Code Node; TriggerLogic; Protection / Setting Node; Action / Measure node; Operating parameter nodes (e.g., wind speed, wind direction, pitch angle, nacelle wind speed, bus voltage, converter temperature, etc.); Equipment / component nodes (such as pitch system, yaw system, main bearing, etc.).

[0045] Several standardized semantic relationships are defined for the above nodes, such as “trigger cause”, “has code”, “recommended action”, “protected by”, “depends on”, and “applicable to”, which constitute the graph skeleton.

[0046] LLM-based knowledge extraction stage Input the above pattern along with the original document into the LLM, and use the few-shot hint method to make the LLM output multiple triples in the form of "(subject, relation, object)"; The generated triples are structured and merged to form a set of candidate triples to be added to the database; For documents with hierarchical information, such as the three-level structure of fault → cause → handling steps in a manufacturer's maintenance manual, LLM can be used to extract information hierarchically through recursive template prompts, ensuring the semantic integrity of fault nodes and their upstream and downstream action nodes.

[0047] Conflict detection and expert verification phase The system constructs a conflict detection rule chain for situations where the same subject, such as the same fault name, is assigned different objects in different documents (e.g., different fault codes, different "resettable / non-resettable" attributes), including but not limited to: Same name, different code detection; Detection of different descriptions for the same code; Detection of inconsistent priorities in multiple handling measures for the same fault; Inconsistent threshold settings for the same protection point across different documents will be detected.

[0048] Detected conflicting triples will be marked as "pending confirmation" or "requires manual review," and can be accompanied by metadata such as source document, page number, and timestamp, facilitating one-time confirmation by domain experts in the system backend.

[0049] Once confirmed, the triples are entered into the "highly reliable knowledge subgraph" and are the content that is prioritized for retrieval by default during the question-and-answer phase. Triples that are not confirmed or have multiple versions can be presented to the LLM in a "multiple candidate" manner during the question-and-answer phase, and the LLM will make a secondary decision based on the context.

[0050] Through the above steps, WindKG-Agent can transform the originally loose wind power operation and maintenance knowledge into a strong semantic, recursive, and conflict-detectable graph, providing a computable knowledge foundation for subsequent task planning and execution.

[0051] The key design principle of the Task-Planning Agent layer is that not all user problems require all analyses; rather, the decision to "check graphs," "examine constant values," or "analyze curves" should be based on the problem itself. To this end, a planning method based on "problem semantics → three types of tasks" is proposed, and its technical solution is as follows: Problem semantic recognition The input natural language questions are subjected to structured semantic parsing to identify whether keywords, semantic roles or hidden intentions related to "fault name / code", "protection / setting value / threshold" or "operating parameters / fluctuation / trend / wind speed / pitch / grid side measurement" are present. By combining alarm data from wind farms / generators with domain-specific fine-tuning models or prompting engineering, the recognition rate of industry abbreviations (such as "yaw fault", "grid connection failure", "pitch tracking error", "wind speed anomaly suppression") can be improved.

[0052] Task mapping of three types of information If it is identified that the user mainly wants to ask "What is this fault / how to deal with it / can it be reset", then mark "basic fault information extraction" as required in the execution plan; If the identified problem involves "why the protection trips / is the setting too low / the machine shuts down when a certain value is exceeded", then mark "protection setting analysis" as required; If the problem is identified as involving "whether there are abnormal trends in wind speed / power / pitch / oil pressure / converter temperature at the time", then mark "operating parameter analysis" as required and attach a parameter list.

[0053] Structured output of the execution plan The Task-Planning Agent module outputs a standardized plan object, which must include at least: Do you need to query fault information: Yes / No; Is protection setting analysis required: Yes / No; Parameter analysis required: Yes / No; The knowledge graph node to be queried is: [fault name, fault code, fault cause, ...]. List of parameters to be analyzed: [Wind speed, blade angle, , grid voltage, ...] This structured plan will be passed to the task execution layer as is, ensuring that the tasks received by each sub-Agent are consistent, replayable, and recordable.

[0054] Through this planning, the present invention can significantly reduce the noise caused by irrelevant information input into LLM, and make the reasoning path of each question and answer predictable and reconfigurable, which is suitable for later automated regression testing and the construction of fault question and answer benchmark sets.

[0055] The Task-Action Agent in the task execution layer executes specific analysis tasks according to the plan. It includes the following sub-modules, each of which can be viewed as a "small agent targeting a specific sub-problem": The Information-Extraction Agent recursively searches the knowledge graph based on the "knowledge graph nodes to be queried" provided by the Task-Planning Agent, using a "depth-first + visited marker" approach.

[0056] Each time a triple is retrieved, it is added to the result set, and the process continues from the target node of that triple as a new starting point until all necessary relations have been retrieved or nodes have been marked as visited.

[0057] This recursive strategy ensures that not only "the name and code of the fault" are obtained, but also upstream and downstream information such as "the logic that triggered it", "which protection restrictions it is subject to", "recommended processing steps", and "related operating parameter requirements", thus providing context-complete knowledge blocks for subsequent natural language generation.

[0058] The Protection-Analysis Agent extracts a data sequence from the historical operation database or unit SCADA data for a period of time prior to the fault, based on the fault name and the time window to be analyzed in the task plan. Extract the protection rules related to this fault / machine model / component from the knowledge graph. The rules include parameter names, comparison methods (greater than, less than, absolute value exceeding the limit, rate of change exceeding the limit), set value thresholds, and durations, etc. Establish a unified calculation function for each rule, and output whether the protection is triggered, the parameter range before and after triggering, and the trend of change (increase, sudden drop, fluctuation, periodic jitter) before the fault critical point. The evaluation results of all rules are summarized into a structured protection setting analysis conclusion, which is then used in the final answer.

[0059] This approach transforms "complex, manufacturer-defined protection logic" into a unified, programmable analysis model, which is something that traditional RAG and pure LLM cannot achieve.

[0060] The Parameter-Analysis Agent retrieves data for the same time window for each of the key parameters specified by the Task-Planning Agent (such as wind speed, wind direction, pitch angle, nacelle wind speed, main bearing temperature, generator outlet current, grid-connected power, etc.). For parameters with clear physical meaning in the wind power field, priority should be given to calling "dedicated analysis functions" (e.g., whether wind speed and pitch match, whether wind speed exceeds the cut-out wind speed, whether there is hysteresis in pitch control, whether there is a sudden drop in load but the wind speed is normal, etc.). For the remaining general parameters, use general time series analysis: detect the maximum value, minimum value, mean, standard deviation, whether there are abrupt changes greater than the threshold, whether there are segments that are consistently higher than the normal range, and convert these results into "human-readable parameter descriptions"; The final output of the protection setting analysis conclusions provides sufficient context for generating explanatory text for the large model.

[0061] The Answer-Generation Agent organizes the outputs of the three sub-agents above (graph knowledge, protection decisions, and parameter analysis) along with the task planning description into a hierarchical Prompt, for example: Problem restatement; Standard definitions and handling procedures for related faults; The actual operating data and protection triggering status of this fault; Possible triggering causes and troubleshooting order; Referenced knowledge sources (document name / graph node / rule name); Then it is handed over to LLM to generate the final natural language answer; To improve traceability, the final answer includes explanatory information such as "This conclusion references rule R23 and operating parameters wind direction and blade angle" and "The fault has a name conflict in the graph, and two versions have been returned. Please confirm manually."

[0062] Example 4 Based on the basic technical solution of Embodiment 3, the present invention may further include one or more of the following optional solutions to improve the applicability and engineering implementation capability of the system: Option 1: Multi-version knowledge coexistence mechanism When the handling steps for the same fault are not synchronized in the documentation of different manufacturers, the system allows multiple handling solutions to be included in the returned results in a "parallel candidate" manner, and the source and document date are marked in the metadata, so that on-site maintenance personnel can choose the latest or most applicable version.

[0063] Option 2: Knowledge filtering at the aircraft / site level During knowledge retrieval, irrelevant nodes in the graph are filtered based on context such as site, model, installed capacity, and controller version, reducing the LLM input length and improving the accuracy of the response.

[0064] Option 3: Debug / Audit Mode The system can provide an expert-oriented "interpretive view" that returns not only the final answer but also the intermediate plan for this task planning, the execution logs of each sub-agent, the identified conflict triples, and the context fragments ultimately adopted by the LLM, facilitating system acceptance and continuous improvement by industry experts.

[0065] Option 4: Time synchronization mechanism for integration with real-time SCADA / historical database When a fault occurs from an external alarm system, the system can automatically request data segments from the running data source for the same time period and align them according to a preset sampling period for protection and parameter analysis, ensuring the timeliness and accuracy of the analysis conclusions.

[0066] Example 5 The multi-agent collaborative processing method for wind power fault question answering of the present invention adopts the multi-agent collaborative processing system for wind power fault question answering of Embodiment 1, and the process is as follows: Figure 2 As shown, please follow these steps: Step 1: Input Phase Users enter natural language questions in the question-and-answer interface; the system receives the input text and performs semantic standardization processing, including word segmentation, synonym replacement, and stop word removal.

[0067] Step 2: Task Planning Phase The Task-Planning Agent module performs semantic parsing on the problem. The semantic parsing uses an algorithm that combines template matching and semantic vector retrieval to extract keyword vectors. The vectors are then compared with pre-trained semantic templates to calculate cosine similarity (threshold 0.85-0.95) to determine the type of task to be executed and generate a task planning table that includes flags for basic fault information extraction, protection setting analysis, and operating parameter analysis.

[0068] Step 3: Knowledge Retrieval Stage The Information-Extraction Agent receives the planning table and calls the knowledge graph interface in WindKG-Agent, such as... Figure 3 As shown, according to the recursive knowledge graph retrieval algorithm: Initialize the access table and read the seed node set S0 (consisting of fault name and device type); Call the RetrieveTriples function to query the set of triples E0 related to S0 from the knowledge graph; Call HarvestDown(v) for each target node v∈E0 until all visited nodes are marked as traversed; The complete set of triplets T is obtained.

[0069] Step 4: Protection Setting Analysis Stage The Protection-Analysis Agent extracts runtime data within a preset time window and calculates trigger status, parameter range, and trends based on protection setting rules. The data sampling period is 1 to 10 seconds, and the time window length is 3 to 15 minutes.

[0070] Here, the real-time unit operation data sequence D_pre within the time window Δt_pre = 5 min is calculated using the formula: Triggering the decision function:

[0071] in The current value of the parameter. For the corresponding threshold, The allowable fluctuation range is 0.5% to 2%. The output includes the trigger status, parameter range, and trend type.

[0072] Step 5: Parameter Analysis Phase The Parameter-Analysis Agent performs time series analysis on specified parameters. Calculate the trend slope If |k| ≥ k0 (threshold 0.05~0.5), it is marked as "significant change"; The sliding window variance method was used to detect abrupt change points, with a window width of 10–60 s.

[0073] Explanation of the output parameter variation patterns and their physical meaning.

[0074] Step 6: Answer Generation Stage The Answer Generation Agent integrates the three types of results and task planning information to construct a prompt template, including the question background, analysis summary, reasoning chain, and cited documents. An example template is shown below: "Based on the triggering logic and operational data analysis in the graph, the wind speed parameter continuously increased (+15%) within 3 minutes before the fault, exceeding the protection threshold of 12.5 m / s, thus triggering the wind speed power limiting protection. According to rule R23, this fault can be reset." Step 7: Output Stage The system output includes natural language responses, trigger rule numbers, cited document identifiers, and inference path information to ensure the traceability and interpretability of the results. The results are visualized on the terminal interface and archived for review.

[0075] The principle of the multi-agent collaborative processing method for wind power fault response in this invention is as follows: Step 2 aims to reduce irrelevant information input and improve the stability of the model's response through semantic planning; its principle is to use high-dimensional semantic similarity matching to replace keyword retrieval.

[0076] Step 3 employs a recursive knowledge graph retrieval algorithm, which can maintain the semantic connectivity of the knowledge graph and avoid obtaining only partial information from local nodes.

[0077] Steps 4–5 make the numerical analysis of traditional protection logic and operating parameters explicit, so that LLM does not have to be calculated directly, thereby significantly improving consistency and interpretability.

[0078] The advantage of step 6 is that it integrates structured results with natural language descriptions to form a traceable and verifiable answer.

[0079] To verify the effectiveness of the present invention, the performance of the traditional RAG-LLM question-and-answer system and the multi-agent collaborative processing system for wind power fault question-and-answer in Embodiment 1 of the present invention were compared.

[0080] The test results on a dataset containing 200 typical wind power fault problems are shown in Table 1 below: Table 1 Performance Test Comparison

[0081] The results show that the system of the present invention improves accuracy by approximately 20 percentage points and is significantly superior to the comparative system in terms of interpretability. For example... Figure 2 As shown, task planning and multi-agent collaborative execution mechanisms significantly improve the overall system performance.

[0082] Example 6 Wind farm A was selected, with a 1.5 MW doubly-fed induction generator (DFIG) as the turbine type.

[0083] System configuration parameters: Δt_pre = 5 min, protection threshold wind speed is 12.5 m / s.

[0084] When the user enters the question "What are the reasons for the wind speed protection tripping?", the system identifies the task type as "protection analysis + parameter analysis" and outputs an explanatory answer that is consistent with the actual fault record.

[0085] Example 7 A 2 MW direct-drive turbine was selected for testing the fault "yaw tracking error". The system retrieved the relevant triggering logic and repair measures from the fault map, and identified the yaw delay caused by sudden changes in wind direction through parameter analysis. The results show that the system of this invention can correctly identify complex multi-factor fault chains.

[0086] Example 8 For offshore wind farm turbines, an alarm is triggered due to abnormal ambient temperature. Using the method described in this invention, the system simultaneously considers three parameters: temperature, wind speed, and blade pitch, to analyze the combined cause of insufficient heat dissipation and low wind speed leading to protection activation.

[0087] As can be seen from the above embodiments, the multi-agent collaborative processing system for wind power fault question answering of the present invention can effectively integrate knowledge, data and reasoning, and significantly improve the accuracy, transparency and reliability of wind power fault diagnosis.

Claims

1. A multi-agent collaborative processing system for wind power fault response, characterized in that, Includes the following modules: The knowledge construction agent module is used to extract knowledge triples from wind turbine operation and maintenance documents and construct a wind power fault knowledge graph. The knowledge graph includes at least the following node types: fault node, trigger condition node, protection setting node, operating parameter node, and handling measure node. The task planning agent module is used to perform semantic parsing on the input natural language questions, determine the type of information involved in the questions, and generate a task execution plan that includes three parts: basic fault information extraction, protection setting analysis, and operating parameter analysis. The task execution agent module is used to sequentially call the information extraction sub-agent, protection analysis sub-agent, parameter analysis sub-agent, and answer generation sub-agent to execute the task according to the task execution plan.

2. The multi-agent collaborative processing system for wind power fault response according to claim 1, characterized in that, The knowledge construction agent module includes: The system includes a text parsing unit, a triple extraction unit, and a conflict detection unit. The conflict detection unit is used to detect situations where multiple codes correspond to the same fault name or where there are contradictions in the same alarm attribute, and outputs a conflict identifier.

3. The multi-agent collaborative processing system for wind power fault response according to claim 1, characterized in that, The task planning agent module performs semantic parsing on the input natural language question by combining template matching and semantic vector similarity calculation, with a semantic vector similarity threshold of 0.85 to 0.

95.

4. The multi-agent collaborative processing system for wind power fault response according to claim 1, characterized in that, The information extraction sub-agent uses a recursive knowledge graph retrieval algorithm to obtain the set of all triples related to the target node from the knowledge graph; The protection analysis sub-agent reads the unit's operating data within a preset time window before the fault occurs, and calculates whether each protection is triggered according to the protection threshold rules. The parameter analysis sub-agent performs time series analysis on specified operating parameters to identify trends, ranges, and abrupt changes. The answer generation sub-agent integrates the aforementioned analysis results and prompts into the language model, and outputs a natural language answer that includes the reasoning path and knowledge source.

5. The multi-agent collaborative processing system for wind power fault response according to claim 4, characterized in that, The parameter analysis sub-agent detects abrupt changes using the sliding window variance method, with a window width of 10–60 seconds.

6. The processing method of the multi-agent collaborative processing system for wind power fault response as described in any one of claims 1-5, characterized in that, The steps are as follows: S1. Receive natural language questions input by the user and perform semantic standardization processing; S2. The task planning agent module performs semantic parsing on the problem and generates a task planning table containing flags for basic fault information extraction, protection setting analysis, and operating parameter analysis. S3. The information extraction sub-agent calls the knowledge graph interface provided by the knowledge construction agent module to obtain the complete set of triples related to the target node according to the recursive knowledge graph retrieval algorithm. S4. The protection analysis sub-agent extracts the running data within the preset time window and calculates the trigger status, parameter range and trend according to the protection setting rules. S5. The parameter analysis sub-agent performs time series analysis on the specified parameters to detect trend slope or abrupt change characteristics and generate physical interpretations. S6. The answer generation sub-agent integrates the above results and prompts, inputs them into the language model to generate a natural language answer with a reasoning path, and outputs it.

7. The multi-agent collaborative processing method for wind power fault response according to claim 6, characterized in that, In S2, the semantic parsing adopts an algorithm based on a combination of template matching and semantic vector retrieval.

8. The multi-agent collaborative processing method for wind power fault response according to claim 6, characterized in that, In S3, the recursive knowledge graph retrieval algorithm includes the following process: 1) Initialize the access table and read the seed node set; 2) Call the RetrieveTriples function to retrieve triples associated with the current node; 3) Add the new node to the access table and recursively call the next level of retrieval; 4) Return the complete set of triples.

9. The multi-agent collaborative processing method for wind power fault response according to claim 6, characterized in that, In S4, the sampling period for the running data is 1 to 10 seconds, and the time window length is 3 to 15 minutes.

10. The multi-agent collaborative processing method for wind power fault response according to claim 6, characterized in that, The final output of S6 includes natural language response, trigger rule number, referenced document identifier, and inference path information.

Images