Draught fan abnormity monitoring and operation inspection optimization method based on knowledge graph and deep learning
By constructing a TCN-BiGRU-Attention network and a two-layer knowledge graph, combined with a large language model, the problem of wind turbine data fusion was solved, enabling efficient wind turbine anomaly monitoring and operation and maintenance optimization, and improving operation and maintenance efficiency and interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN UNIV
- Filing Date
- 2025-12-04
- Publication Date
- 2026-05-05
AI Technical Summary
The existing condition monitoring systems for wind turbines suffer from inconsistent data distribution and a lack of effective integration. Traditional anomaly detection methods have high false alarm rates and poor scalability. Deep learning models lack interpretability, and the knowledge graph construction has limited coverage, making it difficult to adapt to complex fault requirements, resulting in low operation and maintenance efficiency.
Data is collected through SCADA and CMS systems, and a TCN-BiGRU-Attention network is constructed for anomaly monitoring. A two-layer knowledge graph and a large language model are combined to perform fault-operation and maintenance strategy reasoning, and the optimal operation and maintenance strategy under multiple constraints is output.
It enables real-time equipment status monitoring and multimodal operation and maintenance strategy generation for wind turbine units, improving operation and maintenance efficiency and interpretability, and supporting intelligent decision-making and real-time fault handling.
Smart Images
Figure CN121976918A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine equipment operation and maintenance technology, and in particular to a wind turbine anomaly monitoring and operation and maintenance optimization method based on knowledge graphs and deep learning. Background Technology
[0002] With the rapid development of wind power technology and the centralized grid connection of large-scale wind farms, the operational stability and reliability of wind turbine units have become crucial to ensuring the safe and efficient operation of wind farms. Wind turbine units have complex structures, including blades, main shafts, and gearboxes, among other highly coupled components. These components operate for extended periods in complex environments such as high altitudes, high wind speeds, and strong corrosion, making them highly susceptible to wear and cracks. Failure to detect and address these issues promptly can lead to equipment damage, economic losses, and even grid disturbances.
[0003] To enhance the operational status awareness and fault early warning capabilities of wind turbines, wind power systems widely deploy Condition Monitoring Systems (CMS) and Supervisory Control and Data Acquisition (SCADA) systems. CMS primarily focuses on acquiring and analyzing the vibration, temperature, and other signals of key rotating components to identify early mechanical faults. SCADA systems, on the other hand, enable real-time acquisition and remote monitoring of overall turbine operating parameters (such as wind speed, rotational speed, current, voltage, and power). However, current CMS and SCADA data are typically distributed across different platforms, exhibiting inconsistent sampling frequencies, complex dimensions, and heterogeneous data semantics, lacking effective data fusion and correlation interpretation mechanisms. Traditional anomaly detection and fault diagnosis rely mainly on threshold settings, expert rules, or empirical models, which suffer from high false alarm rates, poor scalability, and difficulty in adapting to complex nonlinear faults.
[0004] In recent years, deep learning methods have demonstrated powerful modeling capabilities for processing multi-source temporal data. In particular, the combined structure of temporal convolutional networks, bidirectional gated recurrent units, and attention mechanisms (TCN-BiGRU-Attention) can automatically extract key features from complex multi-source temporal data and achieve high-precision anomaly identification. However, deep models often lack good interpretability, making it difficult to support subsequent engineering decisions and operational processes.
[0005] Knowledge graph technology provides a structured representation for equipment knowledge modeling, enabling the representation of the relationships between wind power equipment's composition, failure modes, operational experience, and maintenance strategies through entity-relationship-entity triples, supporting logical reasoning based on semantic paths and embeddings. However, current knowledge graph construction often relies on manually extracted rules, resulting in limited coverage and insufficient integration with real-time monitoring data. Large Language Models (LLMs), a deep learning-based natural language processing technique, have achieved significant results in various general-purpose natural language processing tasks. However, in the field of wind turbine operation and maintenance, technical manuals and procedures contain numerous technical terms and concepts, limiting their applicability and making it difficult to meet complex needs. Traditional LLMs can lead to fabricated answers, incorrect responses, and machine illusions, causing wind turbine maintenance personnel to be unable to promptly address maintenance challenges and reducing maintenance efficiency. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for wind turbine anomaly monitoring and operation and maintenance optimization based on knowledge graphs and deep learning.
[0007] The objective of this invention is achieved through the following technical solution: a wind turbine anomaly monitoring and operation and maintenance optimization method based on knowledge graphs and deep learning, the method comprising,
[0008] S1. Collect wind turbine operation data through the SCADA system and wind turbine status data through the CMS system; obtain wind turbine operation and maintenance data, and perform data preprocessing on the operation data, status data and operation and maintenance data to obtain data-aligned multidimensional data.
[0009] S2. By constructing a TCN-BiGRU-Attention network model to analyze multidimensional data, the model can perform anomaly monitoring of wind turbines and output the anomaly type and confidence level predicted by the model.
[0010] S3. Construct a two-layer knowledge graph, including an anomaly-fault knowledge graph and a fault-operation and maintenance strategy knowledge graph;
[0011] S4. Based on a two-layer knowledge graph, take the anomaly type as input, extract feasible solutions according to the graph reasoning algorithm, and output the operation and maintenance strategy triplet.
[0012] S5. Input the operation and maintenance strategy triplet and multi-dimensional data into the large language model, perform multi-round semantic reasoning and strategy optimization, and output the optimal operation and maintenance strategy under multiple constraints.
[0013] Specifically, in S2, the temporal convolutional network (TCN) of the deep neural network first performs a convolutional transformation on the original time series to extract local dynamic features; let the input time series be:
[0014] ;
[0015] In the formula, T is the length of the input sequence; d is the feature dimension at each time step;
[0016] The output of the Temporal Convolutional Network (TCN) is:
[0017] ;
[0018] In the formula, For the (l-1)th TCN convolutional layer at time step eigenvectors on; The parameters are those of the i-th convolutional kernel in the l-th layer; is the bias term; k is the kernel size; d is the dilation factor; by stacking multiple convolutional layers, progressive temporal features are extracted.
[0019] The extracted temporal features are input into a Bi-Gated Recurrent Unit (BiGRU) to obtain deeper forward and backward time dependencies. The output of the BiGRU is: ;
[0020] Attention mechanisms are used to calculate the similarity score between features and the global context at each time step, and attention weights are then assigned.
[0021] ;
[0022] In the formula, The importance score of the t-th time step to the final decision; Here is the weight matrix of the attention layer. This is a learnable attention query vector; This is the bias vector for the attention layer; This represents the attention weight at time t; The weighted context representation; For each sub-term to be summed, sum from k=1 to k=T;
[0023] The predicted probability is:
[0024] ;
[0025] In the formula, y represents the classification target; The weight matrix of the output layer. This is the bias vector for the output layer.
[0026] Specifically, the hidden state update process for a single GRU unit is as follows:
[0027] ;
[0028] ;
[0029] ;
[0030] ;
[0031] In the formula, This is the hidden state vector from the previous time step; These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the hidden layer weight matrices for the update gate, reset gate, and candidate state, respectively. These are the bias terms for the update gate, reset gate, and candidate state, respectively. To update the door; To reset the door; This is a candidate new state; ⊙ represents the current state; ⊙ represents the Hadamard product.
[0032] Specifically, the specific steps of S3 include:
[0033] A two-layer knowledge graph is constructed using the Relation Extraction Model (RERE). For any sequence of wind turbine operation and maintenance data c, the RERE model outputs each relation. The probability of:
[0034] ;
[0035] In the formula, This refers to the BERT (Bidirectional Encoder Representations from Transformers) encoding representation of the first character [CLS] of a sequence; The sigmoid activation function is used; BERT is a language representation model. Let r be the classification weight matrix of relation r; For the bias term of relation r;
[0036] Subsequently, the relation extraction model RERE extracts the corresponding subject and object (s,o) for each identified relation r, with the input as: And encoded into a matrix using BERT. The relation extraction model RERE locates sequence boundaries using four pointers:
[0037] ;
[0038] In the formula, This represents the predicted probability distribution of belonging to boundary type k at each position in the sequence; This is the sigmoid activation function; Let k be the weight matrix of the pointer k; For the bias term of pointer k, These represent four types of pointers: subject start point, subject end point, object start point, and object end point.
[0039] The overall objective function of the relation extraction model RERE is to maximize the likelihood function of all training samples:
[0040] ;
[0041] In the formula, c is the number of samples in the training dataset. i T represents the wind turbine operation and maintenance data sequence of the i-th training sample; i To be with c i The corresponding target relation and entity boundary triple set;
[0042] Decomposed using the chain rule of conditional probability:
[0043] .
[0044] Specifically, step S4 includes,
[0045] Let a knowledge graph be a set of entities. and edge set Directed graph:
[0046] ;
[0047] The initial input set is derived from the knowledge graph by inferring the possible fault types:
[0048] ;
[0049] in, f is the set of fault type nodes in the graph; i Let be the i-th candidate fault type node, and m be the number of candidate fault types;
[0050] set up As candidate policy nodes, and For its initial embedding, the semantic matching degree between faults and policies is normalized and scored using the following formula:
[0051] ;
[0052] In the formula, To share the linear transformation matrix; This is a trainable attention weight vector; Faulty node The initial embedding vector is the candidate policy node. The initial embedding vector, || is the feature concatenation operation; For a strategy among multiple strategy candidate paths Compared to The semantic importance of faults in graphs;
[0053] Define the set of all paths from fault f to policy s as:
[0054] ;
[0055] In the formula, Let s be the currently considered faulty node, and s be a candidate operation and maintenance strategy node corresponding to fault f. In a knowledge graph, starting from the fault node p is the set of all feasible paths from a starting point to the policy node s. i For set The i-th specific path in the path, where n is the number of paths;
[0056] Each path P is represented as a sequence of nodes. The path scoring function is defined as:
[0057] ;
[0058] In the formula, For the i-th edge on the path semantic weight; Score the semantic credibility or contextual consistency of intermediate entity nodes in the path;
[0059] The path relevance score of each policy node to the current failure is the maximum path score:
[0060] ;
[0061] The comprehensive candidate strategy score is calculated using a dual fusion mechanism:
[0062] ;
[0063] In the formula, To adjust the parameters; For strategy The overall relevance score relative to fault f; These are the semantic matching weights obtained through the graph attention mechanism;
[0064] Introducing a strategy resource matching index Combining strategy execution modeling with resource costs and feasibility constraints:
[0065] ;
[0066] In the formula, This is the weighting coefficient between semantic-driven and resource-driven approaches; This is the comprehensive evaluation score of strategy s after simultaneously considering semantic relevance and resource constraints.
[0067] The present invention has the following advantages:
[0068] This invention enables real-time equipment status monitoring and multimodal operation and maintenance strategy generation for wind turbines. It generates a two-layer knowledge graph based on operational data and maintenance information, and automatically identifies potential problems and candidate operation and maintenance strategies for wind turbines using graph reasoning technology. Finally, leveraging the modeling capabilities and multimodal processing capabilities of the large-scale model, it generates real-time wind turbine operation and maintenance strategies, improving the efficiency of wind turbine operation and maintenance. Attached Figure Description
[0069] Figure 1 This is a schematic diagram of the wind turbine abnormality monitoring and operation and maintenance strategy optimization method of the present invention;
[0070] Figure 2 This is a schematic diagram of two types of triples. Detailed Implementation
[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0072] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0073] It should be noted that relational terms such as "first" and "second" are used merely 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 limitation, 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.
[0074] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0075] like Figures 1 to 2 As shown, a wind turbine anomaly monitoring and operation and maintenance optimization method based on knowledge graphs and deep learning is presented. This method includes:
[0076] S1. Collect wind turbine operation data through the SCADA system and wind turbine status data through the CMS system; obtain wind turbine operation and maintenance data, and perform data preprocessing on the operation data, status data and operation and maintenance data to obtain data-aligned multidimensional data.
[0077] Key operating parameters of the wind turbines, such as wind speed, wind direction, and power generation, are collected at low frequency through the SCADA system. Minor changes in the turbine's operating status are collected at high frequency through the CMS system, with particular attention paid to the status parameters of key components such as the gearbox, main shaft, and generator. Maintenance data includes unstructured text data such as historical maintenance logs, fault work orders, and operation and maintenance records. Data preprocessing is performed on these multi-source data sources to standardize input and output formats and achieve multi-dimensional data alignment. After data access, a series of standardization processes are required, including timestamp unification, data completion, and data standardization. For high-frequency monitoring signals from the CMS, a multi-level preprocessing workflow is used for bandpass filtering and wavelet denoising to extract typical diagnostic features reflecting the equipment's health status. For maintenance text data, based on a pre-trained language model and rule engine, combined with named entity recognition and relation extraction technologies, key semantic tags such as equipment name, component location, fault type, and maintenance measures are extracted. Standardized text records are constructed using a unified field structuring method to ensure usability and comparability in subsequent knowledge fusion processes. Ultimately, all accessed data will be encapsulated and output in a unified format. Structured time-series data will be stored in a tabular format, high-frequency signals will be represented by a combination of feature matrices and original waveforms, and text data will be converted into a JSON structure with semantic tags for use by the knowledge graph construction module.
[0078] S2. By constructing a TCN-BiGRU-Attention network model to analyze multidimensional data, the model can perform anomaly monitoring of wind turbines and output the anomaly type and confidence level predicted by the model.
[0079] The Temporal Convolutional Network (TCN) of a deep neural network first performs a convolutional transformation on the original time series to extract local dynamic features; let the input time series be:
[0080] ;
[0081] In the formula, T is the length of the input sequence; d is the feature dimension at each time step; TCN implements causal convolution and dilated convolution operations through a multi-layer one-dimensional convolutional network, which increases the receptive field to capture dependencies over a longer time range while maintaining temporal causality. Let the output of the l-th layer be... If the kernel size is k and the inflation factor is d, then:
[0082] ;
[0083] In the formula, For the (l-1)th TCN convolutional layer at time step eigenvectors on; The parameters are those of the i-th convolutional kernel in the l-th layer; is the bias term; k is the kernel size; d is the dilation factor; by stacking multiple convolutional layers, progressive temporal features are extracted while preserving the sequence structure; BiGRU runs two GRU units in opposite directions simultaneously, which can better capture the sequential relationships in the sequence data; the hidden state update process of a single GRU unit is as follows:
[0084] ;
[0085] ;
[0086] ;
[0087] ;
[0088] In the formula, This is the hidden state vector from the previous time step; These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the hidden layer weight matrices for the update gate, reset gate, and candidate state, respectively. These are the bias terms for the update gate, reset gate, and candidate state, respectively. To update the door; To reset the door; This is a candidate new state; ⊙ represents the current state; ⊙ represents the Hadamard product.
[0089] The extracted temporal features are input into a Bi-Gated Recurrent Unit (BiGRU) to obtain deeper forward and backward time dependencies. The output of the BiGRU is: It can simultaneously utilize information from the temporal context to provide semantically rich contextual representations for subsequent attention mechanisms.
[0090] Attention mechanisms are used to calculate the similarity score between features and the global context at each time step, and attention weights are then assigned.
[0091] ;
[0092] In the formula, The importance score of the t-th time step to the final decision; This is the weight matrix of the attention layer; This is a learnable attention query vector; This is the bias vector for the attention layer; This represents the attention weight at time t; The weighted context representation serves as the final semantic expression vector used for classification or reasoning. For each sub-item in the summation, sum from k=1 to k=T. If the final classification target is y, then the predicted probability is:
[0093] ;
[0094] In the formula, y represents the classification target; This is the weight matrix of the output layer; This serves as the bias vector for the output layer. Through this structure, TCN captures short-term and medium-to-long-term sequence changes, BiGRU encodes bidirectional semantics, and the Attention mechanism highlights key moment information, thereby effectively enabling the identification of abnormal operating conditions and the inference of fault types in wind power data. This enhances the system's ability to perceive anomalies in complex operating states, laying a data foundation for the subsequent construction of anomaly knowledge triples and the recommendation of operation and maintenance strategies.
[0095] S3. Construct a two-layer knowledge graph, including an anomaly-fault knowledge graph and a fault-operation and maintenance strategy knowledge graph; achieve semantic connectivity between the two-layer graphs through intermediate entity fault types. Traditional relation extraction methods based on sentence-level context are difficult to accurately capture the deep semantic relationships between entities in multimodal and multi-granular data such as wind turbine operation logs, fault records, and maintenance texts. The two-layer knowledge graph module of this invention uses a RERE (Relation Extraction via Relation-first Entity Recognition) relation extraction model that introduces reasoning ability and evidence information enhancement to construct the two-layer knowledge graph.
[0096] A two-layer knowledge graph is constructed using the Relation Extraction Model (RERE). The RERE model employs a strategy of extracting triples from the data first, then entities. For any wind turbine operation and maintenance data sequence c, the RERE model outputs each relation. The probability of:
[0097] ;
[0098] In the formula, This refers to the BERT (Bidirectional Encoder Representations from Transformers) encoding representation of the first character [CLS] of a sequence; This is the sigmoid activation function; Let r be the classification weight matrix of relation r; For the bias term of relation r;
[0099] Subsequently, the relation extraction model RERE extracts the corresponding subject and object (s,o) for each identified relation r, with the input as: And encoded into a matrix using BERT. The relation extraction model RERE locates sequence boundaries using four pointers:
[0100] ;
[0101] In the formula, This represents the predicted probability distribution of belonging to boundary type k at each position in the sequence; This is the sigmoid activation function; Let k be the weight matrix of the pointer k; For the bias term of pointer k, These represent four types of pointers: subject start point, subject end point, object start point, and object end point.
[0102] Given an instance The overall objective function of the relation extraction model RERE is to maximize the likelihood function of all training samples:
[0103] ;
[0104] In the formula, c is the number of samples in the training dataset. i T represents the wind turbine operation and maintenance data sequence of the i-th training sample; i To be with c i The corresponding target relation and entity boundary triple set;
[0105] Decomposed using the chain rule of conditional probability:
[0106] ;
[0107] The final results yield an anomaly-fault knowledge graph and a fault-operation strategy two-layer knowledge graph, providing data support for the subsequent generation of operation and maintenance strategies.
[0108] S4. Based on a two-layer knowledge graph, using the anomaly types inferred from the knowledge graph as input, feasible solutions are extracted according to the graph reasoning algorithm. Relevance modeling and optimization ranking of operation and maintenance strategies are performed, outputting operation and maintenance strategy triples; achieving efficient and interpretable intelligent decision support; assuming the knowledge graph is a graph containing entity sets... and edge set Directed graph:
[0109] ;
[0110] Among them, entity nodes include fault types, components, operation and maintenance measures, strategy solutions, etc., and edges represent semantic relationships between entities;
[0111] The initial input set is derived from the knowledge graph by inferring the possible fault types:
[0112] ;
[0113] In the formula, Let f be the set of fault type nodes in the graph. i Let be the i-th candidate fault type node; m is the number of candidate fault types.
[0114] To model the relevance of policy entities, a Graph Attention Network (GAT) is introduced to perform weighted modeling of the relationships between nodes; let... As candidate policy nodes, and For its initial embedding, the semantic matching degree between faults and policies is normalized and scored using the following formula:
[0115] ;
[0116] In the formula, To share the linear transformation matrix; is a trainable attention weight vector; || is the feature concatenation operation; For a strategy among multiple strategy candidate paths Compared to The semantic importance of faults in graphs; Faulty node The initial embedding vector, Candidate policy nodes The initial embedding vector;
[0117] Operation and maintenance strategy knowledge graphs often express complex fault-cause-response chains through multi-dimensional semantic paths. Therefore, a semantic path retrieval mechanism is introduced to achieve cross-dimensional semantic reasoning of the graph. The set of all paths from fault f to strategy s is defined as:
[0118] ;
[0119] In the formula, s represents the currently considered faulty node; s is a candidate operation and maintenance strategy node corresponding to fault f. In a knowledge graph, starting from the fault node p is the set of all feasible paths from a starting point to the policy node s. i For set The i-th specific path in the path, where n is the number of paths;
[0120] Each path P is represented as a sequence of nodes. The path scoring function is defined as:
[0121] ;
[0122] In the formula, For the i-th edge on the path semantic weight; Score the semantic credibility or contextual consistency of intermediate entity nodes in the path;
[0123] The path relevance score of each policy node to the current failure is the maximum path score:
[0124] ;
[0125] This scoring system combines semantic structural features with the logic between complex multi-hop entities, helping to improve the semantic adaptability of policy selection and the interpretability of reasoning. It also yields a graph attention score. Path reasoning score Then, the comprehensive candidate strategy score is calculated through a dual fusion mechanism:
[0126] ;
[0127] In the formula, To adjust the parameters; These are the semantic matching weights obtained through the graph attention mechanism; For strategy The overall relevance score is relative to fault f; this parameter can be adaptively adjusted based on the semantic complexity of the fault context and the policy network topology. To make the policy recommendation results more closely match actual operation and maintenance resources and field conditions, a policy resource matching degree index is introduced. Model resource costs and feasibility constraints by combining information such as personnel, equipment, and time required for strategy execution:
[0128] ;
[0129] In the formula, This is the weighting coefficient between semantic-driven and resource-driven approaches; To obtain a comprehensive evaluation score for strategy s that simultaneously considers semantic relevance and resource constraints, the system sorts the set of candidate strategies and outputs a set of strategies labeled with priority and resource matching degree, which is then transmitted to the large language model. The large language model is based on existing general large models (such as GPT and Deepseek) and is implemented by calling the APIs of various large models.
[0130] S5. Input the operation and maintenance strategy triples and multidimensional data into the large language model, perform multi-round semantic reasoning and strategy optimization, and output the optimal operation and maintenance strategy under multiple constraints. The semantic scores and resource constraint labels contained in the triples are embedded as structured inputs into a standardized prompt template and fused with multimodal information collected in the current operation and maintenance scenario. Multimodal inputs include, but are not limited to, on-site images, sensor monitoring data, operation and maintenance logs, and historical strategy text records. After unified encoding, the multimodal data is injected into the prompt words of the large language model along with the strategy template, thereby constructing a semantic understanding and generation process with multimodal context awareness capabilities.
[0131] Building upon multimodal fusion, the large language model performs multiple rounds of internal semantic reasoning to assess the logical consistency between policy semantics and the current operational condition, safety boundary constraints, and the rationality of resource consumption. In this process, the model can not only identify potential policy conflicts and redundancies based on structured prior information, but also make linguistic-level causal inferences and predictions about risk factors implicit in unstructured modalities. The large language model can then provide reasons for refusal to execute or offer alternative policy suggestions through natural language generation.
[0132] Finally, the semantic consistency scores of each candidate strategy, obtained by integrating multimodal information, linguistic context, and operational condition knowledge, are ranked, and a set of multimodal optimal strategies under multiple constraints, including safety, executability, and resource adaptability, is output. Simultaneously, interactive dialogue verification with frontline personnel is supported, enabling real-time strategy adjustments, explanations, and tracing based on user feedback, ensuring the entire operation and maintenance recommendation system possesses excellent interpretability and real-time interactivity.
[0133] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.
Claims
1. A method for wind turbine anomaly monitoring and operation and maintenance optimization based on knowledge graphs and deep learning, characterized by: The method includes, S1. Collect wind turbine operation data through the SCADA system and wind turbine status data through the CMS system; Acquire wind turbine operation and maintenance data, perform data preprocessing on operation data, status data, and operation and maintenance data to obtain data-aligned multidimensional data; S2. By constructing a TCN-BiGRU-Attention network model to analyze multidimensional data, the model can perform anomaly monitoring of wind turbines and output the anomaly type and confidence level predicted by the model. S3. Construct a two-layer knowledge graph, including an anomaly-fault knowledge graph and a fault-operation and maintenance strategy knowledge graph; S4. Based on a two-layer knowledge graph, take the anomaly type as input, extract feasible solutions according to the graph reasoning algorithm, and output the operation and maintenance strategy triplet. S5. Input the operation and maintenance strategy triplet and multi-dimensional data into the large language model, perform multi-round semantic reasoning and strategy optimization, and output the optimal operation and maintenance strategy under multiple constraints.
2. The wind turbine anomaly monitoring and operation and maintenance optimization method based on knowledge graphs and deep learning according to claim 1, characterized in that: The temporal convolutional network (TCN) of the deep neural network in S2 first performs a convolutional transformation on the original time series to extract local dynamic features; let the input time series be: ； In the formula, T is the length of the input sequence; d is the feature dimension at each time step; The output of the Temporal Convolutional Network (TCN) is: ； In the formula, The parameters are those of the i-th convolutional kernel in the l-th layer; For the (l-1)th TCN convolutional layer at time step eigenvectors on; is the bias term; k is the kernel size; d is the dilation factor; by stacking multiple convolutional layers, progressive temporal features are extracted. The extracted temporal features are input into a Bi-Gated Recurrent Unit (BiGRU) to obtain deeper forward and backward time dependencies. The output of the BiGRU is: ; Attention mechanisms are used to calculate the similarity score between features and the global context at each time step, and attention weights are then assigned. ； In the formula, The importance score of the t-th time step to the final decision; Here is the weight matrix of the attention layer. This is a learnable attention query vector; This is the bias vector for the attention layer; This represents the attention weight at time t; The weighted context representation; The predicted probability is: ； In the formula, y represents the classification target; This is the weight matrix of the output layer. This is the bias vector for the output layer.
3. The wind turbine anomaly monitoring and operation and maintenance optimization method based on knowledge graphs and deep learning according to claim 1, characterized in that: The hidden state update process for a single GRU unit is as follows: ； ； ； ； In the formula, To update the door; To reset the door; This is a candidate new state; This is the current state; Let be the hidden state vector from the previous time step. These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the hidden layer weight matrices for the update gate, reset gate, and candidate state, respectively. These are the bias terms for the update gate, reset gate, and candidate state, respectively; ⊙ is the Hadamard product.
4. The wind turbine anomaly monitoring and operation optimization method based on knowledge graphs and deep learning according to claim 1, characterized in that: The specific steps of S3 include: A two-layer knowledge graph is constructed using the Relation Extraction Model (RERE). For any sequence of wind turbine operation and maintenance data c, the RERE model outputs each relation. The probability of: ； In the formula, This refers to the BERT (Bidirectional Encoder Representations from Transformers) encoding representation of the first character [CLS] of a sequence; Let r be the classification weight matrix of relation r; For the bias term of relation r; This is the sigmoid activation function; Subsequently, the relation extraction model RERE extracts the corresponding subject and object (s,o) for each identified relation r, with the input as: And encoded into a matrix using BERT. The relation extraction model RERE locates sequence boundaries using four pointers: ； in, This is the predicted probability distribution of belonging to boundary type k at each position in the sequence. For sigmoid activation function, Let k be the weight matrix of the pointer. For the bias term of pointer k, These represent four types of pointers: subject start point, subject end point, object start point, and object end point. The overall objective function of the relation extraction model RERE is to maximize the likelihood function of all training samples: ； in, c is the number of samples in the training dataset. i Let T be the wind turbine operation and maintenance data sequence of the i-th training sample. i To be with c i The corresponding target relation and entity boundary triple set; Decomposed using the chain rule of conditional probability: 。 5. The wind turbine anomaly monitoring and operation optimization method based on knowledge graphs and deep learning according to claim 1, characterized in that: Step S4 specifically includes: Let a knowledge graph be a set of entities. and edge set Directed graph: ； The initial input set is derived from the knowledge graph by inferring the possible fault types: ； in, Let f be the set of fault type nodes in the graph. i Let be the i-th candidate fault type node, and m be the number of candidate fault types; set up As candidate policy nodes, and For its initial embedding, the semantic matching degree between faults and policies is normalized and scored using the following formula: ； In the formula, To share the linear transformation matrix; This is a trainable attention weight vector; Faulty node The initial embedding vector is the candidate policy node. The initial embedding vector, || is the feature concatenation operation; For a strategy among multiple strategy candidate paths Compared to The semantic importance of faults in graphs; Define the set of all paths from fault f to policy s as: ； In the formula, Let s be the currently considered faulty node, and s be a candidate operation and maintenance strategy node corresponding to fault f. In a knowledge graph, starting from the fault node p is the set of all feasible paths from a starting point to the policy node s. i For set The i-th specific path in the path, where n is the number of paths; Each path P is represented as a sequence of nodes. The path scoring function is defined as: ； In the formula, For the i-th edge on the path semantic weight; Score the semantic credibility or contextual consistency of intermediate entity nodes in the path; The path relevance score of each policy node to the current failure is the maximum path score: ； The comprehensive candidate strategy score is calculated using a dual fusion mechanism: ； In the formula, To adjust the parameters; These are the semantic matching weights obtained through the graph attention mechanism; For strategy The overall relevance score relative to fault f; Introducing a strategy-resource matching index Combining strategy execution modeling with resource costs and feasibility constraints: ； In the formula, This is the weighting coefficient between semantic-driven and resource-driven approaches; This is the comprehensive evaluation score of strategy s after simultaneously considering semantic relevance and resource constraints.