A wind farm fault processing method, device, equipment and storage medium

Abstract

The application discloses a wind farm fault processing method and device, equipment and storage medium, relates to the wind power generation operation and maintenance technical field, and includes: acquiring multi-source heterogeneous data of a wind farm, determining an initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology; determine the target knowledge graph library, retrieve the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault related subgraph; input the target fault related subgraph into a preset fault troubleshooting large model to infer the initial diagnosis scheme of the initial fault problem object based on the target fault related subgraph by using the preset fault troubleshooting large model; the initial diagnosis scheme is converted into a structured target fault work order through a preset entity extraction and slot filling operation, and a corresponding target fault troubleshooting tree is generated based on the target fault work order, so that the user end can troubleshoot the wind farm through the target fault troubleshooting tree. The fault processing efficiency of the wind farm can be improved.

Description

Technical Field

[0001] This invention relates to the field of wind power operation and maintenance technology, and in particular to a method, apparatus, equipment and storage medium for handling wind farm faults. Background Technology

[0002] In existing technologies, fault handling relies on human experience, and maintenance plans lack standardized and intelligent handling suggestions, resulting in inconsistent solutions for the same type of fault handled by different personnel. Static and fixed on-site troubleshooting paths cannot cope with complex operating conditions, requiring repeated telephone consultations or consulting paper manuals, increasing downtime and causing power generation losses. Work order entry involves a large amount of repetitive work, and historical high-value experience is difficult to effectively accumulate and reuse. Traditional diagnostic systems are mostly based on fixed rules or static fault trees, unable to understand complex on-site phenomena described in natural language, and struggling to provide dynamic and in-depth reasoning suggestions.

[0003] In conclusion, improving the efficiency of fault handling in wind farms is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and storage medium for handling wind farm faults, which can improve the efficiency of wind farm fault handling. The specific solution is as follows: Firstly, this application provides a method for handling wind farm faults, including: Acquire multi-source heterogeneous data of the wind farm, and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology; Determine the target knowledge graph base, and retrieve the initial fault problem object in the target knowledge graph base to obtain the corresponding target fault-related subgraph; The target fault correlation subgraph is input into a preset fault diagnosis model so that the preset fault diagnosis model can be used to deduce the initial diagnosis scheme for the initial fault problem object based on the target fault correlation subgraph. The initial diagnostic scheme is transformed into a structured target fault work order through preset entity extraction and slot filling operations. Based on the target fault work order, a corresponding target fault investigation tree is generated so that the user terminal can use the target fault investigation tree to investigate the faults of the wind farm.

[0005] Optionally, the multi-source heterogeneous data includes time-series waveform data, alarm logs, and natural language descriptions from the wind farm monitoring system; Accordingly, the step of determining the initial fault problem object of the wind farm based on the multi-source heterogeneous data using a preset multi-modal alignment technique includes: Receive fault alarm notifications from the wind farm; A time window is constructed based on the time-series waveform data, and a preset sliding window technique is used to extract a preset length of the time-series waveform data before and after the time of the fault alarm notification according to the time window and perform normalization processing to obtain normalized time-series waveform data. The normalized timing waveform data is input into a preset timing encoder to obtain a timing feature vector; The alarm log and the natural language description are input into a pre-trained language model to perform semantic extraction to generate text feature vectors; The association weight matrix between the temporal feature vector and the text feature vector is determined by a cross-modal attention mechanism. Based on the weight matrix, the time series feature vectors are weighted and aggregated to obtain the aggregated time series feature vectors. The aggregated temporal feature vector and the text feature vector are mapped to a preset dimension through a multi-layer perceptron mechanism to obtain an aligned feature representation; A preset problem identification operation is performed on the aligned feature representation to determine the initial fault problem object; the initial fault problem object includes any one or more of the following: the station number, the fault code, and the ambient wind speed at the time of the fault.

[0006] Optionally, determining the target knowledge graph base includes: Identify the individual components of the wind turbine and the connections between them; Each component of the wind turbine is mapped as a physical node, and the connection edges between the physical nodes are determined based on the connection relationships between the components; wherein, the connection edges include any one or more of electrical connections, mechanical connections, signal control, and hydraulic pipeline relationships; The target knowledge graph base is determined based on the entity nodes and the connecting edges.

[0007] Optionally, the step of retrieving the initial fault problem object from the target knowledge graph database to obtain the corresponding target fault-related subgraph includes: The initial fault problem object is queried in the target knowledge graph to determine the target entity node and target connection relationship corresponding to the initial fault problem object. Based on graph retrieval enhancement generation technology, target fault-related subgraphs are extracted from the target knowledge graph database according to the target entity nodes and the target connection relationships.

[0008] Optionally, the step of inputting the target fault correlation subgraph into a preset fault diagnosis model, so as to use the preset fault diagnosis model to infer an initial diagnosis scheme for the initial fault problem object based on the target fault correlation subgraph, includes: The target fault-related subgraph is input into a preset fault diagnosis model. Through the attention mechanism of the preset fault diagnosis model, the connection relationships and preset constraint rules defined in the target fault-related subgraph are focused, and paths that conflict with the structure of the target fault-related subgraph are filtered to obtain the initial inference path. A preset causal chain reasoning operation is performed on the initial reasoning path to determine each fault root cause and the probability distribution of the fault root causes in the target fault-related subgraph. Based on the probability distribution of the root cause of the fault, a corresponding investigation target sequence is generated according to a preset arrangement order; Each root cause of the fault in the target sequence is diagnosed to generate an initial diagnostic plan.

[0009] Optionally, the step of converting the initial diagnostic scheme into a structured target fault work order through preset entity extraction and slot filling operations includes: Obtain a preset set of fault work order templates; the preset set of fault work order templates includes fault work order templates for different types of work orders; The initial diagnostic scheme is analyzed to determine the target filling content and the corresponding target fault type from the initial diagnostic scheme; The target fault work order template is determined from the preset fault work order template set according to the target fault type; The target content is filled into the target fault work order template through preset entity extraction and slot filling operations to generate the target fault work order.

[0010] Optionally, generating a corresponding target fault troubleshooting tree based on the target fault work order includes: Analyze the target diagnosis report in the target fault work order; Based on the obtained parsing results and a preset prompt word template, the target diagnostic report is transformed into an initial fault diagnosis tree; Using search enhancement generation technology, retrieve historical fault work orders that are the same as the target fault type, and obtain historical troubleshooting path data and equipment manuals for the historical fault work orders; Based on the historical troubleshooting path data and equipment manual, the initial fault troubleshooting tree is optimized to generate a target fault troubleshooting tree. This allows the user to obtain troubleshooting instructions for each fault troubleshooting step through the target fault troubleshooting tree in a question-and-answer interactive manner, and to troubleshoot the wind farm.

[0011] Secondly, this application provides a wind farm fault handling device, comprising: The object determination module is used to acquire multi-source heterogeneous data of the wind farm and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology. The subgraph acquisition module is used to determine the target knowledge graph library, and retrieve the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault-related subgraph; The solution reasoning module is used to input the target fault correlation subgraph into a preset fault investigation model, so as to use the preset fault investigation model to reason about the initial diagnosis solution of the initial fault problem object based on the target fault correlation subgraph. The troubleshooting tree generation module is used to transform the initial diagnostic scheme into a structured target fault work order through preset entity extraction and slot filling operations, and generate a corresponding target fault troubleshooting tree based on the target fault work order, so that the user terminal can use the target fault troubleshooting tree to troubleshoot the wind farm.

[0012] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the wind farm fault handling method as described above.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the wind farm fault handling method as described above.

[0014] In summary, this application first acquires multi-source heterogeneous data of a wind farm, and then determines the initial fault problem object of the wind farm based on the multi-source heterogeneous data using a preset multimodal alignment technology; it then determines a target knowledge graph library, retrieves the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault-related subgraph; it inputs the target fault-related subgraph into a preset fault investigation model, and uses the preset fault investigation model to infer the initial diagnosis scheme of the initial fault problem object based on the target fault-related subgraph; it then transforms the initial diagnosis scheme into a structured target fault work order through preset entity extraction and slot filling operations, and generates a corresponding target fault investigation tree based on the target fault work order, so that the user terminal can perform fault investigation on the wind farm through the target fault investigation tree. As described above, this application acquires multi-source heterogeneous data from wind farms, uses a pre-defined multimodal alignment technique to determine the initial fault problem object, retrieves the object from the target knowledge graph library to obtain the corresponding target fault-related subgraph, inputs the subgraph into a pre-defined fault investigation model, infers and generates an initial diagnostic solution, then transforms the solution into a structured target fault work order through pre-defined entity extraction and slot filling operations, and generates a target fault investigation tree based on the work order, ultimately providing the user end with the means to conduct fault investigation work on the wind farm. In this way, by introducing the equipment topology knowledge graph as a hard constraint during the inference process, and using attention masks or cue word engineering to force the large language model to follow physical connection logic, the problem of "illusions" easily generated by general models in industrial diagnosis is effectively solved, ensuring the physical consistency of the diagnostic solution. Secondly, the static diagnostic text is transformed into a dynamic interactive fault investigation tree containing If-Then-Else logic, and incremental inference based on on-site feedback is supported, realizing dynamic navigation of the fault investigation path, significantly reducing the reliance on the personal experience of maintenance personnel, and improving on-site investigation efficiency. Furthermore, a knowledge closed-loop mechanism based on the professional confidence weight of personnel was constructed, which can automatically filter high-quality expert experience for model fine-tuning and eliminate low-quality data, realizing the automated accumulation of wind farm operation and maintenance knowledge and the continuous adaptive evolution of the system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This application discloses a flowchart of a wind farm fault handling method. Figure 2This is a schematic diagram of the knowledge update process of a closed-loop acceptance and adaptive knowledge evolution subsystem disclosed in this application; Figure 3 This application discloses a specific flowchart of a wind farm fault handling method. Figure 4 This is a schematic diagram of the structure of a wind farm fault handling device disclosed in this application; Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0018] Currently, fault handling relies heavily on manual experience, and maintenance plans lack standardized and intelligent handling suggestions, resulting in inconsistent solutions for the same type of fault handled by different personnel. Static and fixed on-site troubleshooting paths are unable to cope with complex operating conditions, requiring repeated telephone consultations or consulting paper manuals, increasing downtime and causing power generation losses. Work order entry involves a large amount of repetitive work, and historical high-value experience is difficult to effectively accumulate and reuse. Traditional diagnostic systems are mostly based on fixed rules or static fault trees, unable to understand complex on-site phenomena described in natural language, and struggling to provide dynamic and in-depth reasoning suggestions. To address the above technical problems, this application discloses a wind farm fault handling method, device, equipment, and storage medium, which can improve the efficiency of wind farm fault handling.

[0019] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a wind farm fault handling method, including: Step S11: Obtain multi-source heterogeneous data of the wind farm, and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology.

[0020] In this embodiment, during operation, the multi-source data sensing and preprocessing subsystem first collects the wind farm's operational data in real time. This multi-source heterogeneous data includes time-series waveform data, alarm logs, and natural language descriptions from the wind farm monitoring system. Specifically, the collected data includes time-series waveform data output by the wind turbine SCADA (Supervisory Control And Data Acquisition) system, such as turbine speed, bearing temperature, and vibration signals. It also includes wind turbine alarm logs, alarm information generated by the intelligent early warning system, and natural language problem descriptions entered by maintenance personnel into the production management system.

[0021] Next, when the system detects abnormal operating parameters, alarms, or warnings, it automatically triggers the problem identification logic and instantiates an initial fault problem object. The process involves acquiring fault alarm notifications from wind farms; constructing a time window based on the time-series waveform data; and using a preset sliding window technique to extract a preset length of time-series waveform data before and after the fault alarm notification time for normalization processing, resulting in normalized time-series waveform data; inputting the normalized time-series waveform data into a preset time-series encoder to obtain a time-series feature vector; inputting the alarm log and the natural language description into a pre-trained language model for semantic extraction to generate a text feature vector; determining the association weight matrix between the time-series feature vector and the text feature vector through a cross-modal attention mechanism; weighting and aggregating each time-series feature vector based on the weight matrix to obtain an aggregated time-series feature vector; mapping the aggregated time-series feature vector and the text feature vector to a preset dimension through a multi-layer perception mechanism to obtain an aligned feature representation; and performing a preset problem identification operation on the aligned feature representation to determine the initial fault problem object; the initial fault problem object includes any one or more of the following: the turbine stand number, fault code, and ambient wind speed at the time of the fault occurrence. Specifically, firstly, a time window is constructed for the time-series waveform data output by the wind turbine SCADA system. A sliding window technique is used to extract a pre-defined length of operational data segment before and after the fault trigger moment, and this segment is normalized to eliminate the influence of different physical dimensions. Subsequently, the processed time-series data is input into a Transformer-based time encoder. This encoder utilizes a multi-head self-attention mechanism to capture the dynamic evolution characteristics of wind turbine speed, bearing temperature, and vibration signals over time, outputting a high-dimensional time-series feature vector containing equipment operating status information. Simultaneously, for the natural language text and alarm logs entered by maintenance personnel, a pre-trained language model based on the BERT architecture is used as the text embedding layer. This model performs word segmentation, positional encoding, and semantic extraction on the unstructured text, generating text feature vectors containing contextual semantic information. To achieve deep fusion of time-series and text data in the feature space, a cross-modal attention mechanism is further introduced. This mechanism uses the text feature vector as the query vector and the time-series feature vector as the key and value vectors. It calculates the association weight matrix between the two in the semantic space through dot product scaling, thereby identifying the time-series segment most relevant to a specific fault description. After weighting and aggregating the temporal features based on this weight matrix, a multilayer perceptron projects the aggregated temporal features and the original text features onto a latent space of the same dimension. By minimizing the distribution difference between the two in the latent space, the alignment of multimodal data is completed, ultimately forming a unified joint feature representation that combines physical state information and semantic description information, i.e., the aligned feature representation. The aligned feature representation automatically triggers problem identification logic to determine the initial fault problem object.It is important to know that the initial fault problem object encapsulates data such as the machine position number, fault code, and ambient wind speed corresponding to the occurrence of the anomaly, and stores the collected multimodal data as the problem context.

[0022] Step S12: Determine the target knowledge graph library, and retrieve the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault related subgraph.

[0023] In this embodiment, to obtain a target knowledge graph library for wind farm equipment topology, the various components of the wind turbine and the connection relationships between them can be determined. Each component of the wind turbine is mapped to a physical node, and connection edges between the physical nodes are determined based on the connection relationships between the components. These connection edges include any one or more of electrical connections, mechanical connections, signal control, and hydraulic pipeline relationships. The target knowledge graph library is determined based on the physical nodes and the connection edges. Specifically, the wind farm equipment topology knowledge graph library contains physical nodes and physical connection edges for each component of the wind turbine. The physical connection edges include at least electrical connections, mechanical connections, signal control, and hydraulic pipeline relationships. Physical nodes cover: blades, pitch system, drive train, generator, converter, main control system, etc.; the connection edge relationships in the graph cover: electrical connections, mechanical connections, signal control, etc. For example, the graph explicitly defines an electrical connection between the "pitch slip ring" and the "pitch driver," and a failure of the "pitch slip ring" will result in "communication loss."

[0024] Then, after acquiring the target knowledge graph, the initial fault problem object is queried in the target knowledge graph to determine the target entity node and target connection relationship corresponding to the initial fault problem object. Based on graph retrieval enhancement generation technology, the target fault-related subgraph is extracted from the target knowledge graph according to the target entity node and the target connection relationship. Specifically, a parallel retrieval mechanism is triggered, and a two-layer precise retrieval is carried out on the wind farm equipment topology knowledge graph based on the technical architecture of graph retrieval enhancement generation. The first layer is entity location retrieval: the core fault entity is extracted from the initial fault problem object through named entity recognition technology, and the target entity node is quickly located in the knowledge graph based on the unique identifier of the entity, while the attribute information of the node is extracted. The second layer is subgraph structure retrieval: with the target entity node as the core, according to the predefined physical connection edge relationship in the knowledge graph, a breadth-first traversal algorithm is used to recursively retrieve all component nodes and connection relationships directly or indirectly associated with the core entity, forming a subgraph structure knowledge slice centered on the fault entity and covering the "fault propagation path". Understandably, the subgraph structure knowledge slices not only contain entity information of related components but also explicitly label the connection types, signal transmission directions, and fault impact logic between components, ensuring that the extracted knowledge slices completely match the actual physical topology of the wind power equipment and possess strong constraints. Subsequently, the subgraph structure knowledge slices are structurally formatted and transformed into a constraint context prompt that can be parsed by the large language model. The prompt adopts a three-part structure of "entity-relationship-constraint": the first part clearly defines the core fault entities and the list of related components; the second part lists the physical connection relationships and constraint rules between components in a structured table format; and the third part clearly defines the inference forbidden zone constraints. At the same time, to enhance the constraint effect, the prompt is fused with the multimodal feature vector of the initial fault problem object and mapped to a vector space consistent with the input dimension of the large language model through a text embedding layer, ensuring that the model can simultaneously understand data features and physical constraints.

[0025] Step S13: Input the target fault correlation subgraph into the preset fault investigation model, so as to use the preset fault investigation model to deduce the initial diagnosis scheme of the initial fault problem object based on the target fault correlation subgraph.

[0026] In this embodiment, the target fault-related subgraph is input into a preset fault diagnosis model. Through the attention mechanism of the preset fault diagnosis model, the model focuses on the connection relationships and preset constraint rules defined in the target fault-related subgraph, filtering out paths that conflict with the structure of the target fault-related subgraph to obtain an initial inference path. A preset causal chain inference operation is performed on the initial inference path to determine each fault root cause in the target fault-related subgraph and its probability distribution. Based on the probability distribution of the fault root causes, a corresponding investigation target sequence is generated according to a preset arrangement order. Each fault root cause in the investigation target sequence is diagnosed to generate an initial diagnostic plan. Specifically, after receiving the constraint context prompt, the large language model initiates a reasoning process based on physical topology constraints. The large language model has built-in reasoning logic fine-tuned from wind power operation and maintenance corpus. When generating diagnostic suggestions, it uses an attention mechanism to focus on the physical connection relationships and constraint rules defined in the prompt, automatically filtering out inference paths that conflict with the knowledge graph subgraph structure. Simultaneously, the model combines anomaly indicators from time-series data snapshots with fault propagation logic in subgraphs to perform causal chain reasoning, calculate the probability distribution of each potential root cause of the fault, and generate a structured sequence of investigation targets according to the physical logic order of "from easy to troubleshoot to difficult to handle, from core components to related components." During the reasoning process, a physical logic consistency verification mechanism is simultaneously activated to perform real-time reverse verification on each diagnostic conclusion and each investigation step generated by the model: by searching the knowledge graph to confirm whether there is physical connectivity between the components corresponding to the investigation path, and to verify whether the root cause of the fault and the fault phenomenon conform to the physical propagation law. If any logical inconsistency is detected, a dynamic correction instruction is immediately triggered, guiding the model to readjust the reasoning direction based on the subgraph structure knowledge slices through prompt word engineering until the generated diagnostic content fully conforms to the physical topology constraints of the wind power equipment, resulting in an initial diagnostic plan. The initial diagnostic plan includes five modules: fault root cause probability distribution, investigation target sequence, physical connection basis explanation, key indicator anomaly analysis, and fault exclusion scope. The physical connection specification clearly marks the knowledge graph subgraph connection relationship corresponding to each diagnostic conclusion, ensuring that the diagnostic suggestions are coherent and reachable in terms of equipment structure, connection relationship and fault propagation logic. From a technical point of view, it completely avoids generating invalid diagnostic conclusions that do not conform to the actual structure of the equipment, and solves the core pain point that general large language models are prone to physical illusion in industrial scenarios.

[0027] Step S14: The initial diagnostic scheme is transformed into a structured target fault work order through preset entity extraction and slot filling operations. A corresponding target fault investigation tree is generated based on the target fault work order so that the user terminal can troubleshoot the wind farm through the target fault investigation tree.

[0028] In this embodiment, after obtaining the initial diagnostic report, the work order lifecycle management subsystem is responsible for converting the unstructured diagnostic results into business documents that conform to power production standards. This involves obtaining a preset set of fault work order templates; this set contains fault work order templates for different types of work orders; parsing the initial diagnostic scheme to determine the target content and corresponding target fault type from the initial diagnostic scheme; determining the target fault work order template from the preset set of fault work order templates based on the target fault type; and filling the target content into the target fault work order template through preset entity extraction and slot filling operations to generate the target fault work order. Specifically, the system automatically identifies the problem type and fills in fields such as "work order planning," "required spare parts," "safety measures," and "investigation content" for the problem work order. Based on the nature of the work order, it automatically extracts key information to generate a work order or operation ticket draft and establishes a unique index association. It includes built-in standard templates for different types of work orders; the fault work order template focuses on emergency repairs, with fields including fault symptoms, downtime, and defect elimination deadline. Defect Work Order Template: Focuses on planned defect elimination, with fields including defect level and suggested processing time. The initial diagnostic report is parsed, and entity extraction and slot filling technologies are used to automatically map natural language suggestions to work order fields. For example, "Suggest replacing blade 3 pitch encoder" is automatically filled into the "Planned Content" field of the work order, and "No. 3 engine compartment" is filled into the "Work Location" field. Entity extraction and slot filling technologies are used to automatically map unstructured natural language text in the intelligent diagnostic report into structured target fault work orders that conform to the power production management system specifications.

[0029] Furthermore, the target diagnostic report in the target fault work order is parsed; based on the parsing results and a preset prompt word template, the target diagnostic report is transformed into an initial fault troubleshooting tree; historical fault work orders of the same type as the target fault are retrieved using retrieval enhancement generation technology, and historical troubleshooting path data and equipment manuals of the historical fault work orders are obtained; the initial fault troubleshooting tree is optimized based on the historical troubleshooting path data and equipment manuals to generate a target fault troubleshooting tree, so that the user terminal can obtain troubleshooting instructions for each fault troubleshooting step through the target fault troubleshooting tree in a question-and-answer interaction format, and perform fault troubleshooting on the wind farm. Specifically, the target diagnostic report is semantically parsed, and key information such as faulty equipment, phenomena, and locations are extracted through entity extraction to construct a query vector; secondly, historical fault work orders are retrieved based on the wind power knowledge graph, and the most relevant historical troubleshooting path data is recalled using fault type index and semantic similarity matching; at the same time, standard parameters and maintenance specifications of corresponding components are retrieved from the equipment manual library. Then, the retrieved historical paths are compared and analyzed with the initial fault diagnosis tree. Through logical consistency checks and step efficiency evaluations, the branches of the initial tree are adjusted: redundant steps are merged, the inspection order is optimized, missing judgment criteria are added, and standard parameters from the equipment manual are filled into the "JudgmentStandard" field to ensure the accuracy of the troubleshooting basis. The optimized troubleshooting tree is output in JSON format, including step ID, inspection item, judgment criteria, positive / negative result jumps, safety measures, and spare tools. The generated JSON structure is as follows: { "StepID": "S1", "CheckItem": "Inspect the carbon powder buildup in the slip ring cavity of wind turbine No. 3 (Work location: inside the hub; Operation method: open the slip ring cavity protective cover and visually observe the thickness of the carbon powder buildup)". "JudgmentStandard": "Toner buildup thickness ≥ 2mm is abnormal, < 2mm is normal". "Outcome_Positive": "S2 (Toner buildup abnormal, perform cleanup operation)", "Outcome_Negative": "S3 (Toner buildup is normal; check communication pin contact pressure)", "SafetyMeasure": "Execute hub locking and disconnect pitch control power", "ToolsSpareParts": "Protective gloves, brushes, vacuum cleaners" }, { "StepID": "S2", "CheckItem": "Use a specialized cleaning agent to clean the carbon powder from the pitch changer slip ring cavity and wipe the slipway surface (Procedure: After spraying the cleaning agent, gently wipe with a lint-free cloth to avoid residue)". "JudgmentStandard": "The surface of the slide is free of visible carbon powder and oil stains, indicating that the cleaning is qualified." "Outcome_Positive": "S4 (Cleanup successful, communication status tested)", "Outcome_Negative": "S2 (Cleanup failed, repeat cleanup steps)", "SafetyMeasure": "Maintain ventilation inside the wheel hub and prevent cleaning agent residue from corroding parts", "ToolsSpareParts": "Pitch ring cleaning agent, lint-free cloth" } Understandably, the JSON structure is then parsed by the mobile app's built-in inference engine and transformed into a step-by-step question-and-answer interface. Each step only displays the current operation instruction and a binary option (such as "Yes / No" or "Normal / Abnormal"). After the maintenance personnel make a selection based on the situation, they are automatically redirected to the next logical step, thus achieving dynamic guidance.

[0030] In addition, such as Figure 2 As shown, after a work order is completed, a closed-loop process and experience learning are implemented. During the work order acceptance phase, maintenance personnel are required to fill in the "actual cause of the fault" and "actual handling measures," and score the accuracy of the generated recommendations. Upon successful acceptance, the production management system record is automatically changed to "Completed," and based on the existence of related work orders, a command is sent to the centralized control system to revert the alarm status to "Work Order Processing Closed" or "Problem Directly Closed." Subsequently, the maintenance personnel's "Work Summary" is analyzed. If the actual handling solution deviates from the model's suggestion and is verified as effective, the case is structured and stored in the vector database, enabling automatic updates to the knowledge base. Simultaneously, the confidence-weighted learning module maintains an "expert profile" for each maintenance personnel, including dimensions such as their seniority, historical work order pass rate, and job level, calculating the personnel's "professional confidence weight." The update strategy uses reinforcement learning based on human feedback as its core framework, combining the knowledge characteristics of the wind farm operation and maintenance field with the training rules of large language models, to construct a full-process technical solution of "high-quality sample screening - reward model construction - PPO incremental fine-tuning", ensuring that the model can accurately absorb the expert experience of high-confidence personnel: 1. High-quality positive sample screening and structuring Screening Mechanism: The system automatically selects work orders that meet the following conditions as positive sample candidate sets: the professional confidence weight of the handler is ≥0.7 (high confidence personnel); the work order acceptance result is "passed" and the maintenance personnel's rating of the model recommendation accuracy is ≥4 stars (1-5 star system); when there is a difference between the actual handling measures and the model's initial diagnostic suggestions, the difference is verified to be valid on-site, such as the model suggesting "check the pitch slip ring", the actual handling is "replace the pitch drive power module" and the fault is eliminated; the work order contains complete closed-loop information of "fault phenomenon - actual root cause - handling measures - verification results", with no missing key fields.

[0031] RAG (Retrieval-augmented Generation) assisted validity verification: For work orders in the candidate set, the wind farm equipment topology knowledge graph and historical work order database are retrieved through retrieval-augmented generation technology to verify whether the actual handling measures conform to the physical topology constraints of the equipment, such as whether the replaced parts have a direct physical connection with the faulty entity. At the same time, the optimal handling solutions for similar faults in history are compared to confirm the uniqueness and applicability of the work order experience and eliminate duplicate or invalid experience data.

[0032] Structured Transformation: A dedicated entity recognition and relation extraction model for the wind power industry is used to extract core semantic units from the selected high-quality work orders. These units include entities of fault phenomena, handling actions, related components, and safety measures. A triple relationship of "fault phenomenon-root cause-handling measures" is constructed to transform the unstructured work order text into structured data that meets the model training requirements. At the same time, the data is mapped to the unified semantic space of the wind farm equipment topology knowledge graph to ensure that the data and model input formats are compatible.

[0033] 2. Reward Model Training

[0034] A reward model specifically designed for wind power operation and maintenance is constructed to quantitatively evaluate the on-site applicability and accuracy of the diagnostic solutions generated by the model, providing feedback signals for reinforcement learning. Dataset Construction: High-quality structured work order data (label 1.0) was selected as positive samples, and the following data were used as negative samples (label 0.0): invalid work orders submitted by low-confidence personnel (weight < 0.5) (work orders that failed acceptance or whose handling measures were ineffective); diagnostic solutions generated by the initial model that did not conform to physical logic; handling experiences from historical work orders that have been verified as erroneous; and non-standard operation cases selected from publicly available wind power operation and maintenance literature. A training dataset with a positive-to-negative sample ratio of 3:1 was ultimately constructed, containing no fewer than 1000 entries.

[0035] Model Architecture and Training: The reward model is fine-tuned based on the BERT-base model. The input is the diagnostic solution text generated by the model, and the output is a quantitative score from 0 to 10. During training, the cross-entropy loss function is used, the learning rate is set to 2e-5, the batch size is 32, and the training epochs are 10. After each epoch, the model performance is evaluated using a validation set. Training stops when the validation set accuracy reaches above 92%, ensuring that the reward model can accurately distinguish between high-quality and low-quality diagnostic solutions.

[0036] 3. Incremental fine-tuning of large models based on the PPO (Proximal Policy Optimization) algorithm

[0037] Based on a large language model pre-trained on wind power corpus, an incremental training algorithm using a proximal policy optimization method is employed to solidify high-confidence experience. Training data input format: Structured positive sample data is fused with relevant subgraph information from the wind farm equipment topology knowledge graph to construct a triplet input prompt of "fault context - physical constraints - high-quality handling solution". For example: "Fault context: No. 3 wind turbine pitch shaft 1 communication failure, SCADA alarm code EG_8092, nacelle temperature 42℃, shaft 1 drive temperature 65℃; Physical constraints: pitch shaft 1 - electrical connection - pitch drive - signal control - main control system; High-quality handling solution: replace the pitch drive power module, implement hub locking and high-voltage power failure safety measures."

[0038] PPO training parameter configuration: Learning rate is set to 1e-5, batch size to 16, KL divergence constraint coefficient to 0.05, dominance estimation window size to 128, and discount factor to 0.95. During training, after the base model generates diagnostic solutions, the reward model scores them, and the scores are fed back to the model as reward signals. By adjusting the model parameters to maximize the reward value, the model gradually learns the fault handling logic and operating procedures of high-confidence personnel.

[0039] Mini-batch iterative training mechanism: Considering the incremental nature of new high-quality work orders in wind farms, a mini-batch iterative training method is adopted. Each training iteration only inputs 100-200 new positive samples, with 3-5 training rounds. After each batch of training, a temporary model version is generated and its performance is compared with the current online model. When the overall performance of the temporary model improves by more than 5%, it is replaced by the online model, ensuring that the model can continuously absorb the latest high-value experience while avoiding system resource consumption and service interruptions caused by large-scale training. If the feedback submitted by low-weight junior personnel differs significantly from the model prediction, it is marked as "knowledge pending review" and is not directly updated into the database; it requires manual review by experts.

[0040] As described above, this application's embodiments acquire multi-source heterogeneous data from wind farms, determine the initial fault problem object using a preset multimodal alignment technology, retrieve the object from the target knowledge graph library to obtain the corresponding target fault-related subgraph, input the subgraph into a preset fault investigation model, infer and generate an initial diagnostic solution, then transform the solution into a structured target fault work order through preset entity extraction and slot filling operations, and generate a target fault investigation tree based on the work order, ultimately providing the user terminal with the means to conduct fault investigation work on the wind farm. In this way, by introducing the equipment topology knowledge graph as a hard constraint during the inference process, and using attention masks or cue word engineering to force the large language model to follow physical connection logic, the problem of "illusions" easily generated by general models in industrial diagnosis is effectively solved, ensuring the physical consistency of the diagnostic solution. Secondly, the static diagnostic text is transformed into a dynamic interactive fault investigation tree containing If-Then-Else logic, and incremental inference based on on-site feedback is supported, realizing dynamic navigation of the fault investigation path, significantly reducing the reliance on the personal experience of maintenance personnel, and improving on-site investigation efficiency. Furthermore, a knowledge closed-loop mechanism based on the professional confidence weight of personnel was constructed, which can automatically filter high-quality expert experience for model fine-tuning and eliminate low-quality data, realizing the automated accumulation of wind farm operation and maintenance knowledge and the continuous adaptive evolution of the system.

[0041] As can be seen from the previous embodiment, this application discloses a wind farm fault handling method, which can improve the fault handling efficiency of wind farms. Next, we will address... Figure 3 A detailed explanation of a wind farm fault handling method is provided.

[0042] (1) Multi-source alarm capture and problem instantiation: Real-time monitoring of SCADA and early warning system data streams. When a threshold is triggered or a manually entered natural language description is received, a "problem record object" is instantiated. This object contains the problem source, device location number, occurrence time, and original alarm snapshot. An example of a "problem record object" is as follows: #Problem Name: Communication Failure on Pitch Shaft 1 of Wind Turbine No. 3

[0043] #Source of the problem: SCADA system

[0044] #Time of occurrence: 2023-10-15 14:30:05

[0045] #Equipment Position Number: 3

[0046] #Original alarm: [Code: EG_8092][Nacelle temperature: 42℃][Axis 1 drive temperature: 65℃]

[0047] (2) Physical Consistency Diagnostic Reasoning: When the user clicks "Processing Suggestions", the large language model reasoning module is invoked. By combining real-time alarm data, multimodal context information, and physical connection relationships defined in the knowledge graph, possible causes of failure are reasoned and analyzed to generate preliminary diagnostic suggestions. The physical connectivity of the diagnostic path is verified, and troubleshooting steps that do not have physical consistency are eliminated. The intelligent diagnostic solution after physical logic correction is output. "The root cause of the fault is suspected to be dirt or wear on the pitch slip ring raceway or carbon brushes, which interferes with the communication signal and affects the drive." The device temperature is high, which could be due to dust accumulation on the cooling fan. It is recommended to first check the slip ring housing, clean off any carbon powder, and also check the communication pin connector. Touch pressure. (3) Structured generation of work orders and linkage with two tickets (work ticket / operation ticket): Based on the intelligent diagnostic scheme, problems are classified and mapped to determine the work order type. The natural language content in the diagnostic scheme is parsed, and information such as fault level and troubleshooting content is automatically filled into the standard fields of the work order. If the work order involves high voltage or special operations, the associated "two tickets" information is automatically generated for maintenance personnel to review. After the specialist quickly reviews and confirms that there are no errors, he / she clicks "issue", and the work order is transferred to the mobile terminal of the maintenance team leader.

[0048] Category mapping: Determine the work order type (fault / defect / energy efficiency / reliability).

[0049] Field population: Through entity extraction and slot filling technology, the system automatically completes the detailed description of the work order, recommended processing method, and required spare parts, transforming the above reasoning suggestions into standard work order fields. [Fault Level]: High (causing shutdown)

[0050] [Work Order Planning]: Inspect and clean the pitch slip ring; measure the slip ring communication channel resistance; check the shaft 1 driver cooling fan. fan.

[0051]

Required spare parts

[0052] [Safety Measures]: Hub locking must be implemented; the pitch system control power must be disconnected.

[0053] [Investigation Content]: {"Step1":"Check_A","Outcome_Positive":"Go_Step2","Outcome_ Negative:"Go_Step3"}

[0054] (4) Dynamic Interactive Maintenance Execution: After arriving at the site, maintenance personnel enter the "Intelligent Assistance Mode" via mobile terminal. Instead of displaying lengthy instructions, the system pushes troubleshooting commands step by step according to the decision tree logic generated in step S3. Based on real-time feedback from personnel, such as "measured voltage is normal," the system dynamically adjusts the subsequent troubleshooting path in real time until the fault point is located.

[0055] (5) Cascade Status Reset and Acceptance: After maintenance is completed, the maintenance personnel submit a work order. The on-duty personnel see in the centralized control system that the fan status has been restored to "Running" and the fault code has disappeared. They then click "Acceptance Passed" to conduct the operation acceptance. After acceptance, the production management system updates the problem status to "Completed" and executes the cascade shutdown logic. It sends an "Alarm Clearing Command" to the centralized control system through the interface, realizing cross-system cascade status synchronization and fault handling closed loop.

[0056] (6) Knowledge Accumulation Based on Confidence Weights: Obtain the "Work Summary" and "Acceptance Results" of this work order, extract the fault phenomena and actual handling of this work order. Assign confidence weights to this experience based on the historical qualifications of the handling personnel (e.g., senior engineer vs. intern), and save standard cases. When encountering a similar fault again, the model will prioritize recommending solutions validated by high-weighted experts.

[0057] By maintaining an "expert profile" database and constructing a multi-dimensional quantitative model, we can achieve confidence weight allocation and knowledge updates for structured cases. The specific implementation steps are as follows: Step 1: Constructing a Multi-Dimensional Quantitative Model and Calculating Confidence Weights. The system first retrieves the qualification data of the personnel involved and performs standardized quantitative processing through job level, historical work order performance, and professional qualification dimensions. The job level dimension (weight coefficient 0.35) presets quantitative values ​​for different positions, such as 1.0 for senior specialists and 0.2 for interns; the historical work order performance dimension (weight coefficient 0.45) is calculated by statistically analyzing work order samples from the past two years, according to the formula: Historical work order pass rate × 0.6 + success rate of handling difficult faults × 0.4 The calculation is mapped to the interval [0, 1.0]; the professional qualification dimension (weight coefficient 0.2) is calculated by accumulating the holding of advanced technical certificates, special equipment operation qualifications and special training sessions.

[0058] Final confidence weight The formula is derived by weighted summation of the quantified values ​​from the three dimensions mentioned above: W =Job Level Quantitative Value 0.35+ Historical work order performance quantification value 0.45+ professional qualification quantification value 0.2.

[0059] according to W The value classifies cases into high confidence levels ( ≥0.7), medium confidence level (0.5≤ W ≤0.7) and low confidence level ( WThree levels (<0.5). Furthermore, a dynamic calibration mechanism is implemented: when the case acceptance score is ≥4 stars and the fault is completely resolved, the weight increases by 0.1; if the personnel handling the case are cross-functional support staff, the weight decreases by 0.1, to ensure that the weight allocation aligns with actual experience value.

[0060] The second step involves extracting structured data from high-confidence cases into a large language model fusion optimization system based on high-confidence experience. This data is encapsulated as an input prompt, consisting of "fault context + physical constraints + core logic chain," and a command dataset is constructed using "high-quality processing solutions" as supervisory signals. During the incremental pre-training phase, a low learning rate (1e-5) and weight decay (0.01) are used to fine-tune the pre-trained large model. The accuracy of "fault root cause prediction" and other metrics are evaluated using a validation set, and intermediate parameters are saved when the metrics improve by ≥3%.

[0061] During the reinforcement learning fine-tuning phase, a reward model is used to quantitatively score the generated solutions, triggering positive rewards for results with scores ≥8. A proximate policy optimization algorithm is employed to adjust model parameters, and a KL (Kullback-Leibler Divergence) divergence constraint coefficient of 0.05 is used to suppress the model's forgetting of existing domain knowledge. The fine-tuned model is validated on a test set containing 50 typical faults, ensuring an improvement of ≥8% in the diagnostic accuracy for faults related to high-confidence cases.

[0062] Step 3: Weighted Vector Update and Graph Completion of the External Knowledge Base. For updating the vector database, a weighted average fusion algorithm is used. Let the original vector be... The new high-confidence case vector is Then the updated vector : ; This algorithm ensures that high-confidence experiences have a higher retrieval priority in the vector space.

[0063] Simultaneously, dynamic completion of the knowledge graph is performed: if a high-confidence case involves a new "root cause-treatment" relationship, an edge relationship is automatically added to the graph and its source and weight are labeled; if the device attribute data in the case is inconsistent with the existing attributes in the graph, the node attributes are updated based on the high-confidence case data and logged. Finally, the updated content is pushed in real time to the reasoning and decision-making subsystem based on physical topology constraints through a synchronization interface, realizing closed-loop knowledge reuse under the RAG architecture.

[0064] It's important to know that the generated investigation content is not static text, but rather a conditional probability decision tree structure. During the generation phase, the large language model outputs a nested structure containing If-Then-Else logical relationships, for example: {"Step1":"Check_A","Outcome_Positive":"Go_Step2","Outcome_Negative":"Go_Step3"}.

[0065] During the execution phase, the mobile application uses a lightweight inference engine to parse the structure. When the on-site situation reported by maintenance personnel does not belong to a branch already defined in the current decision tree, an incremental inference mechanism is triggered. This mechanism calls the cloud model in real time to perform "incremental inference," generating a new decision branch and sending it to the front end, thereby achieving dynamic expansion of the troubleshooting path.

[0066] For example, in a practical application test at a wind farm with an installed capacity of 200MW, the average time to handle wind turbine failures was reduced from 8.5 hours to 5.2 hours, the average time to fill out work orders was reduced from 15 minutes using the traditional method to 3 minutes, the first-time repair rate of difficult failures increased from 75% to 92%, and the frequency of reuse of historical experience was significantly improved. With the assistance of wind farm failure handling methods, newly hired maintenance personnel have failure handling capabilities close to those of senior engineers.

[0067] See Figure 4 As shown in the figure, an embodiment of the present invention discloses a wind farm fault handling device, comprising: The object determination module 11 is used to acquire multi-source heterogeneous data of the wind farm and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology. Subgraph acquisition module 12 is used to determine the target knowledge graph library, and retrieve the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault related subgraph; Solution reasoning module 13 is used to input the target fault correlation subgraph into a preset fault investigation model, so as to use the preset fault investigation model to reason about the initial diagnosis solution of the initial fault problem object based on the target fault correlation subgraph. The troubleshooting tree generation module 14 is used to transform the initial diagnosis scheme into a structured target fault work order through preset entity extraction and slot filling operations, and generate a corresponding target fault troubleshooting tree based on the target fault work order, so that the user terminal can troubleshoot the wind farm through the target fault troubleshooting tree.

[0068] As described above, this application acquires multi-source heterogeneous data from wind farms, uses a pre-defined multimodal alignment technique to determine the initial fault problem object, retrieves the object from the target knowledge graph library to obtain the corresponding target fault-related subgraph, inputs the subgraph into a pre-defined fault investigation model, infers and generates an initial diagnostic solution, then transforms the solution into a structured target fault work order through pre-defined entity extraction and slot filling operations, and generates a target fault investigation tree based on the work order, ultimately providing the user end with the means to conduct fault investigation work on the wind farm. In this way, by introducing the equipment topology knowledge graph as a hard constraint during the inference process, and using attention masks or cue word engineering to force the large language model to follow physical connection logic, the problem of "illusions" easily generated by general models in industrial diagnosis is effectively solved, ensuring the physical consistency of the diagnostic solution. Secondly, the static diagnostic text is transformed into a dynamic interactive fault investigation tree containing If-Then-Else logic, and incremental inference based on on-site feedback is supported, realizing dynamic navigation of the fault investigation path, significantly reducing the reliance on the personal experience of maintenance personnel, and improving on-site investigation efficiency. Furthermore, a knowledge closed-loop mechanism based on the professional confidence weight of personnel was constructed, which can automatically filter high-quality expert experience for model fine-tuning and eliminate low-quality data, realizing the automated accumulation of wind farm operation and maintenance knowledge and the continuous adaptive evolution of the system.

[0069] In some specific implementations, the multi-source heterogeneous data includes time-series waveform data, alarm logs, and natural language descriptions from the wind farm monitoring system; Accordingly, the object determination module 11 may specifically include: The notification acquisition unit is used to acquire fault alarm notifications from the wind farm. The data acquisition unit is used to construct a time window based on the time-series waveform data, and use a preset sliding window technique to extract the time-series waveform data of a preset length before and after the time of the fault alarm notification according to the time window, and perform normalization processing to obtain normalized time-series waveform data. The first vector acquisition unit is used to input the normalized timing waveform data into a preset timing encoder to obtain a timing feature vector. The vector generation unit is used to input the alarm log and the natural language description into the pre-trained language model to perform semantic extraction operations to generate text feature vectors. The evidence determination unit is used to determine the association weight matrix between the temporal feature vector and the text feature vector through a cross-modal attention mechanism; The second vector acquisition unit is used to perform weighted aggregation on each of the time-series feature vectors based on the weight matrix to obtain the aggregated time-series feature vector; The feature representation acquisition unit is used to map the aggregated temporal feature vector and the text feature vector to a preset dimension through a multi-layer perceptron mechanism to obtain the aligned feature representation; The object determination unit is used to perform a preset problem identification operation on the aligned feature representation to determine the initial fault problem object; the initial fault problem object includes any one or more of the following: the station number, the fault code, and the ambient wind speed at the time of the fault.

[0070] In some specific implementations, the subgraph acquisition module 12 may specifically include: The component and relationship determination unit is used to determine the various components of the wind turbine and the connection relationships between the various components; A connection edge determination unit is used to map each component of the wind turbine into a physical node, and determine the connection edges between the physical nodes based on the connection relationships between the components; wherein, the connection edges include any one or more of electrical connections, mechanical connections, signal control, and hydraulic pipeline relationships; The knowledge graph library determination unit is used to determine the target knowledge graph library based on the entity nodes and the connecting edges.

[0071] In some specific implementations, the subgraph acquisition module 12 may specifically include: The node and relationship determination unit is used to query the initial fault problem object in the target knowledge graph database to determine the target entity node and target connection relationship corresponding to the initial fault problem object. The subgraph extraction unit is used to extract target fault-related subgraphs from the target knowledge graph database based on graph retrieval enhancement generation technology, according to the target entity nodes and the target connection relationships.

[0072] In some specific implementations, the scheme reasoning module 13 may specifically include: The path acquisition unit is used to input the target fault-related subgraph into a preset fault investigation model, and through the attention mechanism of the preset fault investigation model, focus on the connection relationships and preset constraint rules defined in the target fault-related subgraph, filter out paths that conflict with the structure of the target fault-related subgraph, and obtain the initial inference path. The probability distribution determination unit is used to perform a preset causal chain reasoning operation on the initial reasoning path to determine each fault root cause and the probability distribution of the fault root causes in the target fault-related subgraph. A sequence generation unit is used to generate a corresponding investigation target sequence based on the probability distribution of the root cause of the fault according to a preset arrangement order; The scheme generation unit is used to diagnose each root cause of the fault in the investigation target sequence in order to generate an initial diagnostic scheme.

[0073] In some specific implementations, the investigation tree generation module 14 may specifically include: The set acquisition unit is used to acquire a preset set of fault work order templates; the preset set of fault work order templates includes fault work order templates of different types of work orders; The content and type determination unit is used to parse the initial diagnostic scheme to determine the target filling content and the corresponding target fault type from the initial diagnostic scheme; The template determination unit is used to determine the target fault work order template from the preset fault work order template set according to the target fault type; The work order generation unit is used to fill the target content into the target fault work order template through preset entity extraction and slot filling operations to generate a target fault work order.

[0074] In some specific implementations, the investigation tree generation module 14 may specifically include: The report parsing unit is used to parse the target diagnosis report in the target fault work order; The troubleshooting tree conversion unit is used to convert the target diagnostic report into an initial troubleshooting tree based on the obtained parsing results and a preset prompt word template. The path data and manual acquisition unit is used to retrieve historical fault work orders with the same type as the target fault using retrieval enhancement generation technology, and to acquire the historical troubleshooting path data and equipment manuals of the historical fault work orders. The troubleshooting tree generation unit is used to optimize the initial fault troubleshooting tree based on the historical troubleshooting path data and equipment manual, and generate a target fault troubleshooting tree so that the user terminal can obtain the troubleshooting instructions for each fault troubleshooting step through the target fault troubleshooting tree in the form of question and answer interaction, and perform fault troubleshooting on the wind farm.

[0075] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0076] Figure 5This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the wind farm fault handling method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0077] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0078] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0079] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the wind farm fault handling method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0080] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned wind farm fault handling method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0082] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0083] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0084] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0085] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for handling faults in a wind farm, characterized in that, include: Acquire multi-source heterogeneous data of the wind farm, and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology; Determine the target knowledge graph base, and retrieve the initial fault problem object in the target knowledge graph base to obtain the corresponding target fault-related subgraph; The target fault correlation subgraph is input into a preset fault diagnosis model so that the preset fault diagnosis model can be used to deduce the initial diagnosis scheme for the initial fault problem object based on the target fault correlation subgraph. The initial diagnostic scheme is transformed into a structured target fault work order through preset entity extraction and slot filling operations. Based on the target fault work order, a corresponding target fault investigation tree is generated so that the user terminal can use the target fault investigation tree to investigate the faults of the wind farm.

2. The wind farm fault handling method according to claim 1, characterized in that, The multi-source heterogeneous data includes time-series waveform data, alarm logs, and natural language descriptions from the wind farm monitoring system. Accordingly, the step of determining the initial fault problem object of the wind farm based on the multi-source heterogeneous data using a preset multi-modal alignment technique includes: Receive fault alarm notifications from the wind farm; A time window is constructed based on the time-series waveform data, and a preset sliding window technique is used to extract a preset length of the time-series waveform data before and after the time of the fault alarm notification according to the time window and perform normalization processing to obtain normalized time-series waveform data. The normalized timing waveform data is input into a preset timing encoder to obtain a timing feature vector; The alarm log and the natural language description are input into a pre-trained language model to perform semantic extraction to generate text feature vectors; The association weight matrix between the temporal feature vector and the text feature vector is determined by a cross-modal attention mechanism. Based on the weight matrix, the time series feature vectors are weighted and aggregated to obtain the aggregated time series feature vectors. The aggregated temporal feature vector and the text feature vector are mapped to a preset dimension through a multi-layer perceptron mechanism to obtain an aligned feature representation; A preset problem identification operation is performed on the aligned feature representation to determine the initial fault problem object; the initial fault problem object includes any one or more of the following: the station number, the fault code, and the ambient wind speed at the time of the fault.

3. The wind farm fault handling method according to claim 1, characterized in that, The determination of the target knowledge graph base includes: Identify the individual components of the wind turbine and the connections between them; Each component of the wind turbine is mapped as a physical node, and the connection edges between the physical nodes are determined based on the connection relationships between the components; wherein, the connection edges include any one or more of electrical connections, mechanical connections, signal control, and hydraulic pipeline relationships; The target knowledge graph base is determined based on the entity nodes and the connecting edges.

4. The wind farm fault handling method according to claim 3, characterized in that, The step of retrieving the initial fault problem object in the target knowledge graph database to obtain the corresponding target fault-related subgraph includes: The initial fault problem object is queried in the target knowledge graph to determine the target entity node and target connection relationship corresponding to the initial fault problem object. Based on graph retrieval enhancement generation technology, target fault-related subgraphs are extracted from the target knowledge graph database according to the target entity nodes and the target connection relationships.

5. The wind farm fault handling method according to claim 1, characterized in that, The step of inputting the target fault correlation subgraph into a preset fault diagnosis model, and using the preset fault diagnosis model to infer an initial diagnosis plan for the initial fault problem object based on the target fault correlation subgraph, includes: The target fault-related subgraph is input into a preset fault diagnosis model. Through the attention mechanism of the preset fault diagnosis model, the connection relationships and preset constraint rules defined in the target fault-related subgraph are focused, and paths that conflict with the structure of the target fault-related subgraph are filtered to obtain the initial inference path. A preset causal chain reasoning operation is performed on the initial reasoning path to determine each fault root cause and the probability distribution of the fault root causes in the target fault-related subgraph. Based on the probability distribution of the root cause of the fault, a corresponding investigation target sequence is generated according to a preset arrangement order; Each root cause of the fault in the target sequence is diagnosed to generate an initial diagnostic plan.

6. The wind farm fault handling method according to any one of claims 1 to 5, characterized in that, The process of converting the initial diagnostic plan into a structured target fault work order through preset entity extraction and slot filling operations includes: Obtain a preset set of fault work order templates; the preset set of fault work order templates includes fault work order templates for different types of work orders; The initial diagnostic scheme is analyzed to determine the target filling content and the corresponding target fault type from the initial diagnostic scheme; The target fault work order template is determined from the preset fault work order template set according to the target fault type; The target content is filled into the target fault work order template through preset entity extraction and slot filling operations to generate the target fault work order.

7. The wind farm fault handling method according to claim 6, characterized in that, The step of generating a corresponding target fault troubleshooting tree based on the target fault work order includes: Analyze the target diagnosis report in the target fault work order; Based on the obtained parsing results and a preset prompt word template, the target diagnostic report is transformed into an initial fault diagnosis tree; Using search enhancement generation technology, retrieve historical fault work orders that are the same as the target fault type, and obtain historical troubleshooting path data and equipment manuals for the historical fault work orders; Based on the historical troubleshooting path data and equipment manual, the initial fault troubleshooting tree is optimized to generate a target fault troubleshooting tree. This allows the user to obtain troubleshooting instructions for each fault troubleshooting step through the target fault troubleshooting tree in a question-and-answer interactive manner, and to troubleshoot the wind farm.

8. A wind farm fault handling device, characterized in that, include: The object determination module is used to acquire multi-source heterogeneous data of the wind farm and determine the initial fault problem object of the wind farm based on the multi-source heterogeneous data through a preset multi-modal alignment technology. The subgraph acquisition module is used to determine the target knowledge graph library, and retrieve the initial fault problem object in the target knowledge graph library to obtain the corresponding target fault-related subgraph; The solution reasoning module is used to input the target fault correlation subgraph into a preset fault investigation model, so as to use the preset fault investigation model to reason about the initial diagnosis solution of the initial fault problem object based on the target fault correlation subgraph. The troubleshooting tree generation module is used to transform the initial diagnostic scheme into a structured target fault work order through preset entity extraction and slot filling operations, and generate a corresponding target fault troubleshooting tree based on the target fault work order, so that the user terminal can use the target fault troubleshooting tree to troubleshoot the wind farm.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the wind farm fault handling method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the wind farm fault handling method as described in any one of claims 1 to 7.

Images