Wind turbine equipment fault diagnosis method and system based on large model

Abstract

A large-model-based wind turbine equipment fault diagnosis method and system includes the following steps: initializing a feature set based on normal data; selecting multiple candidate features based on the mixed score of each candidate feature and saving them into the feature set; using a time series segmentation algorithm to perform sample division on the normal data and construct a multi-dimensional spatiotemporal feature representation; constructing a fault diagnosis model, including: loading a large language model as the basic architecture, injecting the multi-dimensional spatiotemporal feature representation into the embedding input layer of the large language model; connecting an adaptive pooling layer after the output of the large language model to construct a two-layer MLP classifier for realizing the recognition and classification of different fault types; the first layer of the two-layer MLP classifier compresses the input features to half of the original feature dimension and applies GELU activation, and the second layer of the two-layer MLP classifier maps to the corresponding fault category space; constructing a knowledge mechanism library to provide the model with prior knowledge of wind turbines and design a loss function driven by wind turbine knowledge for model training.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine equipment fault diagnosis, and in particular to a wind turbine equipment fault diagnosis method and system based on a large model. Background Art

[0002] In recent years, as wind power technology has matured, operating costs have continued to decline. As the core component of wind power generation, the stable operation of wind turbines is of great significance to the entire wind energy industry. Wind turbines operate in harsh conditions and undergo drastic operating conditions for long periods of time, inevitably leading to degradation and damage to various components. Monitoring and fault diagnosis of wind turbine components during operation are crucial for improving equipment stability and reducing maintenance costs. Promptly identifying and addressing faults can prevent economic losses caused by these failures.

[0003] Existing wind turbine fault diagnosis methods include those driven by physical models. By constructing a physical model and combining it with state estimators such as Kalman filtering to generate residual signals, early fault detection is achieved using threshold judgment or statistical tests. In terms of data-driven methods, state identification is achieved by constructing a high-dimensional feature space, such as principal component analysis and support vector machines, to solve nonlinear classification problems to capture faults. With the iteration of machine learning technology, random forests and K-nearest neighbor algorithms have shown advantages in local feature matching. The current technological frontier has extended to the field of deep learning, and fault detection is performed through convolutional neural networks, long short-term memory networks, deep autoencoders, etc.

[0004] While existing wind turbine fault diagnosis algorithms are effective for general fault diagnosis, the complex and diverse nature of actual wind turbine operating data, along with imbalanced data types, makes it difficult to extract fault features and generalize poorly to small sample sizes, making it difficult to effectively diagnose faults. Therefore, there is an urgent need to develop an effective wind turbine fault diagnosis method to meet the needs of practical applications. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the present invention provides a wind turbine equipment fault diagnosis method and system based on a large model, which solves the problems existing in wind turbine equipment fault diagnosis technology, such as complex and diverse data, unbalanced data categories, weak feature extraction and complex fault identification capabilities, and poor generalization of small samples.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] In a first aspect, a large-scale model-based wind turbine equipment fault diagnosis method includes the following steps:

[0008] S1. Collect and clean wind turbine data to obtain cleaned data;

[0009] S2. Establishing a data anomaly determination rule; using the data anomaly determination rule to determine the cleaned data to obtain normal data;

[0010] S3, initializing a feature set based on the normal data; calculating a mixed score for each candidate feature, and selecting a plurality of the candidate features based on the mixed score and saving them into the feature set;

[0011] S4. Based on the feature set, using a time series segmentation algorithm to perform sample division on the normal data and construct a multi-dimensional spatiotemporal feature representation;

[0012] S5. Build a fault diagnosis model, including:

[0013] Loading a large language model as a basic architecture, and injecting the multi-dimensional spatiotemporal feature representation into the embedding input layer of the large language model;

[0014] An adaptive pooling layer is added to the output of the large language model to construct a two-layer MLP classifier for identifying and classifying different fault types. The first layer of the two-layer MLP classifier compresses the input features to half the original feature dimension and applies GELU activation. The second layer of the two-layer MLP classifier maps the input features to the corresponding fault category space.

[0015] S6. Construct a knowledge mechanism library, wherein the knowledge mechanism library is used to provide prior knowledge of wind turbines; design a wind turbine knowledge-driven loss function based on the prior knowledge of wind turbines; and train the fault diagnosis model using the loss function.

[0016] Preferably, the step S1 includes:

[0017] A wind speed and power scatter plot is constructed according to the wind turbine data; and a DBSCAN density clustering algorithm is used to delete abnormal data in the wind turbine data according to the wind speed and power scatter plot.

[0018] Preferably, the step S2 includes:

[0019] The cleaned data is fitted to generate a wind speed power curve using an adaptive segmentation strategy; and a data anomaly determination rule is established based on the wind speed power curve using a dynamic threshold detection mechanism;

[0020] The data anomaly determination rules include: taking the ideal wind speed corresponding to each power value as the average value, counting all wind speed data corresponding to the power value into an interval, calculating the standard deviation of the interval, and defining the normal data range according to the three sigma principle; if the wind turbine data exceeds the normal data range, it is considered abnormal data.

[0021] Preferably, the step S3 includes:

[0022] S31. Extract each feature column data based on the normal data, and calculate the mean and standard deviation of each feature column data;

[0023] S32, normalizing the feature column data according to the mean and the standard deviation, and returning the normalized feature column data;

[0024] S33, initializing a feature set, and calculating the mixed score of each candidate feature;

[0025] The calculation formula of the mixed score is:

[0026] ;

[0027] Wherein, MI is the normalized mutual information of each candidate feature data; Accuracy represents the diagnostic performance index based on the current candidate feature using the XGBoost (based on gradient boosting decision tree) model; Redundancy is the absolute value of the average Pearson correlation coefficient between the candidate feature and the selected feature; α and γ are weight parameters.

[0028] Preferably, the weight parameters α and γ are dynamically adjusted with the iteration rounds.

[0029] Preferably, in step S5, the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model is loaded as the infrastructure; the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model is stacked by 28 layers of Transformer layers, each layer containing multi-head self-attention and feedforward network.

[0030] In the second aspect, a large-model-based wind turbine equipment fault diagnosis system includes a memory, a processor, and a display; the memory is used to store relevant programs and data of the wind turbine fault diagnosis method based on the large language model; the processor is used to execute the program stored in the memory, drive the fault diagnosis model to run, and perform real-time diagnosis of wind turbine equipment; the display is used to visualize the model diagnosis results; the system executes the steps of the method provided in the first aspect or any possible implementation of the first aspect.

[0031] In a third aspect, a computer-readable storage medium stores a computer program thereon, wherein the computer-readable storage medium stores instructions, which, when the instructions are executed on a computer or a processor, cause the computer or processor to execute the method provided in the first aspect or any possible implementation of the first aspect.

[0032] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0033] 1. Unlike existing wind turbine fault diagnosis methods, which suffer from insufficient model complexity, difficulty capturing the complex characteristics of wind turbine data, and lack the ability to correlate heterogeneous data such as vibration and temperature, or to construct a unified fault representation space. Furthermore, they struggle to capture the spatiotemporal relationships of wind turbine data, and suffer from low learning efficiency for small samples. This invention utilizes a large language model to process wind turbine data. Leveraging the large-scale parameters and self-attention mechanism of the large language model, it captures nonlinear relationships in high-dimensional wind turbine data and explicitly models the spatiotemporal relationships between turbine components. Furthermore, the large language model's pre-trained knowledge transfer capabilities, combined with embedded wind turbine mechanism knowledge, create a knowledge-driven loss function. This allows for rapid adaptation of fault modes from a small number of labeled samples, improving the accuracy, precision, and generalization capabilities of wind turbine fault diagnosis.

[0034] 2. Unlike traditional greedy algorithms, which often rely on a fixed evaluation criterion, ignore interactions between features, and are prone to falling into local optimal solutions, this invention adds comprehensive evaluation indicators to the feature selection process using greedy algorithms. These indicators include mutual information, accuracy, and redundancy. The weight parameters α and γ are dynamically adjusted with each iteration. This allows the algorithm to simultaneously consider the correlation between features and the target variable, the diagnostic performance of the model, and the independence between features. This allows for a more comprehensive assessment of feature importance and adapts to the feature selection requirements of different stages. In the early stages, more emphasis is placed on the impact of mutual information on feature selection, while in the later stages, redundancy penalties are considered to reduce redundancy between features, helping the algorithm escape local optimality and thus optimizing the final feature set.

[0035] 3. Publicly disclose the fine-tuning and input-output design methods for model transfer learning. By freezing the underlying parameters of a pre-trained large language model and fine-tuning the top-level network, this approach achieves lightweight deployment and efficient computing resource utilization while retaining the general representation capabilities of the large language model. Furthermore, a multi-dimensional temporal feature input layer and multi-category classification output are constructed to enhance the model's ability to capture spatiotemporal correlation features under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the method of embodiment 1 of the present invention;

[0037] Figure 2 This is a schematic diagram of the method framework of Example 1 of the present invention;

[0038] Figure 3 This is a schematic diagram of the device system structure of Example 2 of the present invention;

[0039] Figure 4 1 is a schematic diagram of the results of the wind turbine equipment fault diagnosis method according to Example 1 of the present invention;

[0040] Figure 5 1 is a schematic diagram of the results of the wind turbine equipment fault diagnosis method according to Example 1 of the present invention. DETAILED DESCRIPTION

[0041] In order to make the technical means, creative features, objectives and effects of the invention easier to understand, the present invention is further described with reference to specific figures. However, the present invention is not limited to the following implementation cases.

[0042] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0043] Example 1:

[0044] like Figure 1 The wind turbine equipment fault diagnosis method based on the large model shown includes the following steps:

[0045] S1. Collect and clean wind turbine data to obtain cleaned data;

[0046] Step S1 includes:

[0047] A scatter plot of wind speed and power is constructed based on wind turbine data. Based on the scatter plot of wind speed and power, the DBSCAN density clustering algorithm is used to delete abnormal data.

[0048] S11. Collect wind turbine data, including SCADA (Supervisory Control and Data Acquisition) data and CMS (Condition Monitoring System) data, and collect key statistical information of the measured values ​​of each sensor within 10 minutes. The key statistical information includes the average, minimum, maximum, and standard deviation of the data of each sensor within 10 minutes;

[0049] S12, based on the key statistical information of the values ​​measured by each sensor within 10 minutes, that is, the wind turbine data, extract the characteristic columns of wind speed and power (x wind ,x power ), set wind speed as the horizontal axis and power as the vertical axis, and draw a scatter plot of wind speed and power;

[0050] S13. Based on the scatter plot of wind speed and power, the grid search DBSCAN density clustering algorithm (which is an existing technology) is used to set the optimal neighborhood distance threshold eps=0.1 and the minimum number of points in the neighborhood threshold min_samples=5, mark the abnormal data, and delete the abnormal data and the abnormal data with power lower than 0;

[0051] S14. Set the neighborhood distance threshold eps to a value range of [0.1, 0.3] and the minimum number of points in the neighborhood threshold min_samples to a value range of [2, 5] for grid search; perform DBSCAN density clustering for each (eps, min_samples) parameter combination and calculate the clustering quality index. The clustering quality index calculation formula is as follows:

[0052] ;

[0053] Where n is the number of clusters; σ i is the internal distance of cluster i, σ j is the internal distance of cluster j; d(c i ,c j ) is the Euclidean distance between the centers of clusters i and j; the internal distance refers to the average distance from all samples in the cluster to the cluster center;

[0054] Combined with the above clustering quality indicators, the data with low indicators are deleted to finally obtain the cleaned data.

[0055] This step achieves preliminary cleaning of wind turbine data and removes isolated outliers.

[0056] S2. Establish data anomaly determination rules; use the data anomaly determination rules to determine the cleaned data and obtain normal data;

[0057] Step S2 includes:

[0058] An adaptive segmentation strategy is used to fit the cleaned data to generate a wind speed power curve. Based on the wind speed power curve, a data anomaly determination rule is established using a dynamic threshold detection mechanism.

[0059] The data anomaly judgment rules include: taking the ideal wind speed corresponding to each power value as the average value, counting all wind speed data corresponding to the power value into an interval, calculating the standard deviation of the interval, and defining the normal data range according to the three sigma principle; if the wind turbine data exceeds the normal data range, it is abnormal data.

[0060] S21. Based on the cleaned data, the power range is divided into several equal-width intervals according to the adaptive segmentation strategy (existing technology), and divided into 5 intervals with a step size of 0.2MW. Each interval is represented by [P i ,Pi+1 ),in , calculate the density midpoint of each interval;

[0061] S22. Based on the midpoint of each interval density, a third-order polynomial is used to fit the wind speed power curve of the wind turbine using the least squares method. The calculation formula of the third-order polynomial is as follows:

[0062] ;

[0063] S23, ideal wind speed power curve, based on the dynamic threshold detection mechanism (existing technology), takes the ideal wind speed corresponding to each power value as the average value , all wind speed data corresponding to the power value are counted as an interval, and the standard deviation of the interval is calculated ,pass Principle, define the normal wind turbine equipment data range Therefore, the rule for determining data anomaly is to define the normal data range of wind turbine equipment by fitting the ideal wind speed power curve, and determine that data outside the range is abnormal data.

[0064] This step achieves in-depth cleaning of wind turbine data, extracts the dual constraints of dynamic local statistics and physical laws, and achieves more refined data cleaning than traditional methods, providing high-quality input for wind turbine data analysis.

[0065] S3. Initialize the feature set based on normal data; calculate the mixed score of each candidate feature, and select multiple candidate features based on the mixed score and save them into the feature set.

[0066] Step S3 includes,

[0067] S31. Extract each feature column data based on normal data , calculate the mean of each feature column data and standard deviation ;

[0068] S32. Standardize the feature column data according to the mean and standard deviation, and return the standardized feature column data;

[0069] S33, set initialization feature set By parallel calculation of the mixed scores of each candidate feature, the top 20 features are iteratively selected as the optimal feature subset and saved in the feature set. The calculation formula of the mixed score is:

[0070] ;

[0071] Where MI is the normalized mutual information of each candidate feature data, which is obtained by calculating the joint probability distribution of the candidate feature and the target variable. Here, the target variable is the fault type label. It can significantly improve the modeling ability of complex data features for high-dimensional wind turbine data with strong nonlinear coupling relationships; Accuracy represents the diagnostic performance indicator of the current candidate feature using the XGBoost (based on gradient boosting decision tree) model; Redundancy is the absolute value of the average Pearson correlation coefficient between the candidate feature and the selected feature. For high-dimensional multi-source heterogeneous wind turbine data, it can reduce redundant noise interference, enhance robustness, and improve the generalization ability of the model; the weight parameters α and γ are dynamically adjusted with the iteration rounds.

[0072] α decreases quadratically, and γ increases quadratically, which can be expressed as follows:

[0073] ;

[0074] ;

[0075] Where α0 is the initial mutual information weight coefficient, which is set to 0.7, γ0 is the initial redundancy penalty weight coefficient, which is set to 0.1, and t and T are the current iteration round and the maximum iteration round, respectively.

[0076] This step flexibly balances feature correlation, model performance and redundancy suppression to avoid overfitting, while reducing the impact of noise in high-dimensional multi-source heterogeneous data, enhancing robustness, and improving the modeling capabilities of strongly coupled wind turbine data, significantly improving the characterization efficiency of complex data features.

[0077] S4. Based on the feature set, the time series segmentation algorithm is used to divide the normal data samples and construct a multi-dimensional spatiotemporal feature representation.

[0078] Step S4 includes the following steps:

[0079] S41. Based on the event information in the wind turbine equipment data set, select the normal data at the beginning and end of each fault and the 7 days before and after the fault. In addition, randomly select some normal data and merge them in the order of events to generate time series training samples.

[0080] S42. Based on the generated time series training samples, a sliding window algorithm is used to extract continuous segments from the time series, and the main window size W=36 and the sliding step size S=6 are defined;

[0081] S43. Extract the data of the 20 selected feature columns within each window segment, and perform coding mapping on the device fault label at the end of each window;

[0082] S44. Design a dual-channel feature projection network to map the data of the selected 20 feature columns into a high-dimensional space through two fully connected layers, a GELU layer, and layer normalization. The specific formula is as follows:

[0083] Assume that the optimal input feature is , through 2 layers of stacking transformation:

[0084] ;

[0085] ;

[0086] ;

[0087] Where B represents the batch size, which is set to 16, W represents the sliding window size, which is set to 36, and d represents the feature dimension, which is set to 20.

[0088] ;

[0089] S45, injecting a learnable position encoding vector, using random initialization parameters to capture temporal position information, and forming the final multi-dimensional spatiotemporal feature representation;

[0090] This step captures local temporal patterns through a sliding window, fuses feature space information using a dual-channel feature projection network, and injects global temporal context in combination with position encoding to form a multi-dimensional spatiotemporal feature representation, providing robust input for downstream classification tasks. It retains the continuity of wind turbine data time series while capturing local dynamic changes to avoid missing key fault signals. It also enables the model to perceive the absolute position and relative order of the sliding window in the overall sequence, enhancing the ability to model the temporal dependency before and after the fault occurs.

[0091] S5. Build a fault diagnosis model, including:

[0092] Load the large language model as the infrastructure and inject the multi-dimensional spatiotemporal feature representation into the embedding input layer of the large language model;

[0093] An adaptive pooling layer is connected after the output of the large language model to construct a two-layer MLP classifier. The first layer of the two-layer MLP classifier compresses the input features to half of the original feature dimension and applies GELU activation. The second layer of the two-layer MLP classifier maps them to the corresponding fault category space.

[0094] The specific steps are as follows:

[0095] S51. Load the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model as the infrastructure and inject the multi-dimensional spatiotemporal feature representation into the embedding input layer of the model.

[0096] S52, the pre-trained model consists of 28 layers of Transformer stacking, each layer contains multi-head self-attention and feedforward network, which is specifically expressed as follows:

[0097] Output of layer L:

[0098] ;

[0099] ;

[0100] ;

[0101] Where Q = W q h,K=W k h,V=W v h, where W q 、W k 、W v 、W o They are query projection matrix, key projection matrix, value projection matrix, and output projection matrix respectively;

[0102] S53. After the model output, an adaptive pooling layer is connected to build a two-layer MLP classifier. The first layer compresses the input features to half of the original feature dimension and applies GELU activation. The second layer maps them to the corresponding fault category space. The specific representation is as follows:

[0103] ;

[0104] ;

[0105] Where, is the hidden state output by the model, and D is the hidden dimension;

[0106] This step builds an efficient and robust fault diagnosis framework through knowledge transfer, adaptive pooling and classifier optimization of the pre-trained model. It solves core problems such as the difficulty in modeling long-term time series dependencies in wind turbine data, poor generalization of small samples, and strong noise interference, and significantly improves the accuracy and precision of fault detection under complex working conditions.

[0107] Fault diagnosis models include:

[0108] a. Time series data input layer: Based on the characteristics of time series data, we design and build a time series input layer that is suitable for large language models to achieve effective encoding and input representation of time series features;

[0109] b. Large language model infrastructure: After the input layer, a pre-trained large language model is introduced as the core architecture. This model leverages contextual modeling and generalization capabilities to analyze the input temporal feature sequences.

[0110] c. Classification output layer: An adaptive pooling layer is connected to the output of the large language model to perform dimensionality compression and feature aggregation on the high-dimensional time series features output by the model. Based on the aggregated feature representation, a classification output module is constructed to identify and classify different fault types.

[0111] S6: Build a knowledge mechanism library based on the physical mechanism knowledge of wind turbines to provide prior knowledge of wind turbine data, convert it into soft constraints and embed it into the model training process; freeze the main parameters of the large language model and only fine-tune the top layer for joint training to enhance the model's generalization ability for specific fault modes.

[0112] S61. Build a knowledge mechanism database containing wind turbine fault physical rules, historical cases, and expert experience;

[0113] S62. Freeze the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model parameters and retain the top-level network for parameter update.

[0114] S63. Embed the knowledge attention mechanism and rule constraint layer in the top-level network, design a knowledge-driven loss function, and combine the classification task with physical rule constraints for model training. The physical rule constraints are divided into gearbox fault characteristic frequency capacity constraints, generator three-phase current imbalance rate constraints, and bearing fault characteristic frequency energy distribution constraints. The specific calculation formulas are:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, [f min ,f max ] is the fault characteristic frequency range defined in the knowledge base, which is set to [100 Hz, 300 Hz], and γ is the ideal energy threshold defined in the knowledge base, which is set to 0.8;

[0120] The calculation formula of the knowledge-driven loss function is:

[0121] ;

[0122] ;

[0123] Where, L task For the cross entropy loss of fault classification task, L konwledeg is the knowledge constraint loss, is the knowledge constraint weight, both are 1 / 3, y n.i is the true label of the nth sample belonging to the i-th category, is the predicted probability that the nth sample belongs to the i-th class;

[0124] S64, in the fine-tuning stage, uses Autocast and GradScaler for mixed precision training, uses AdamW optimizer, sets the initial learning rate to 2e-5, and the weight decay to 0.01. A total of 10 batches are trained based on the knowledge-driven loss function. After each batch, the model performance is evaluated on the test set. The wind turbine equipment fault diagnosis model with the best F1 score is selected and saved. For specific fault diagnosis results, refer to Figure 4 and Figure 5 ;

[0125] This step builds a knowledge base, provides physical rules, historical cases, and expert experience as prior knowledge for model training, introduces knowledge guidance to improve the model's ability to identify key fault characteristics, enhance interpretability, and jointly optimize classification and rules to improve model accuracy. Finally, mixed precision training is used to improve training efficiency and accelerate model convergence.

[0126] S7. Deploy the wind turbine equipment fault diagnosis method based on the large model to the system.

[0127] Step S7 includes,

[0128] S71, according to Figure 2 The large model pre-training framework shown in the figure is used to diagnose wind turbine equipment faults. This process involves storing the processed data in storage. This process uses advanced large model technology to improve the accuracy and efficiency of fault diagnosis.

[0129] S72, press Figure 3 The configuration shown here uses a data acquisition interface to acquire wind turbine operating status data, including but not limited to key parameters such as vibration signals, temperature, speed, voltage, and current. After preprocessing, the collected data is uploaded to the edge node to provide data support for subsequent local inference.

[0130] S73. Receive pre-processed data from multiple wind turbines via the input / output interface, deploy a trained wind turbine fault diagnosis model on the edge, and execute local inference tasks on the edge device's processor. Key intermediate data and final diagnostic results generated during the inference process are temporarily stored in edge storage and, based on policy, selectively uploaded to the cloud for further analysis and model optimization.

[0131] S74, the cloud receives diagnostic results and historical operating data uploaded from multiple edge nodes through the input / output interface, and uses its powerful computing power to combine historical data across units and regions to train and optimize the global model. The updated model is distributed to each edge node and end-side device through the input / output interface, achieving continuous model iteration and performance improvement.

[0132] S75, the cloud is equipped with a display terminal, providing a visual monitoring interface, supporting remote viewing of the health status of each wind turbine, real-time diagnostic results, and system-generated maintenance recommendations;

[0133] This step deploys the large-model-based wind turbine equipment fault diagnosis method into the system. Through the cloud-edge-end collaborative architecture, it not only improves the overall performance of the system, but also enhances its ability to cope with complex environments.

[0134] Example 2:

[0135] A large-model-based wind turbine equipment fault diagnosis system includes a memory, a processor, and a display. The memory is used to store programs and data related to the large language model-based wind turbine fault diagnosis method. The processor is responsible for executing the programs stored in the memory, driving the wind turbine equipment fault diagnosis model to run and perform real-time diagnosis of the wind turbine equipment. The display is used to visualize the model diagnosis results, displaying the diagnosis results in real time in an intuitive and clear interface, allowing operation and maintenance personnel to promptly understand the equipment's operating status.

Claims

1. A wind turbine equipment fault diagnosis method based on a large model, characterized in that: The following steps are involved: S1. Collect and clean wind turbine data to obtain cleaned data; S2. Establishing a data anomaly determination rule; using the data anomaly determination rule to determine the cleaned data to obtain normal data; S3. Initialize a feature set based on the normal data; Calculating a mixed score for each candidate feature, and selecting a plurality of the candidate features according to the mixed score and saving them into the feature set, including: S31. Extract each feature column data based on the normal data, and calculate the mean and standard deviation of each feature column data; S32, normalizing the feature column data according to the mean and the standard deviation, and returning the normalized feature column data; S33, initializing a feature set, and calculating the mixed score of each candidate feature; The calculation formula of the mixed score is: ； Where, MI is the normalized mutual information of each candidate feature data; Accuracy Indicates the diagnostic performance index based on the current candidate features using the gradient boosting decision tree model; Redundancy is the absolute value of the average Pearson correlation coefficient between the candidate feature and the selected feature; α 、 γ is Weight parameter; S4. Based on the feature set, using a time series segmentation algorithm to perform sample division on the normal data and construct a multi-dimensional spatiotemporal feature representation; S5. Build a fault diagnosis model, including: Loading a large language model as a basic architecture, and injecting the multi-dimensional spatiotemporal feature representation into the embedding input layer of the large language model; An adaptive pooling layer is added to the output of the large language model to construct a two-layer MLP classifier for identifying and classifying different fault types. The first layer of the two-layer MLP classifier compresses the input features to half the original feature dimension and applies GELU activation. The second layer of the two-layer MLP classifier maps the input features to the corresponding fault category space. S6. Construct a knowledge mechanism library; the knowledge mechanism library is used to provide prior knowledge of wind turbines; design a wind turbine knowledge-driven loss function based on the prior knowledge of wind turbines; and use the loss function to train the fault diagnosis model.

2. The large model-based wind turbine equipment fault diagnosis method according to claim 1, characterized in that: The step S1 comprises: A wind speed and power scatter plot is constructed according to the wind turbine data; and a DBSCAN density clustering algorithm is used to delete abnormal data in the wind turbine data according to the wind speed and power scatter plot.

3. The wind turbine equipment fault diagnosis method based on a large model according to claim 1, characterized in that: The step S2 comprises: The cleaned data is fitted to generate a wind speed power curve using an adaptive segmentation strategy; and a data anomaly determination rule is established based on the wind speed power curve using a dynamic threshold detection mechanism; The data anomaly determination rules include: taking the ideal wind speed corresponding to each power value as the average value, counting all wind speed data corresponding to the power value into an interval, calculating the standard deviation of the interval, and defining the normal data range according to the three sigma principle; if the wind turbine data exceeds the normal data range, it is considered abnormal data.

4. The wind turbine equipment fault diagnosis method based on a large model according to claim 1, characterized in that: The weight parameter α 、 γ Dynamically adjusted with iteration rounds.

5. The large model-based wind turbine equipment fault diagnosis method according to claim 1, characterized in that: In step S5, the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model is loaded as the infrastructure; the DeepSeek-R1-Distill-Qwen-1.5B pre-trained model is stacked by 28 layers of Transformer layers, each layer containing multi-head self-attention and feedforward networks.

6. A large-scale model-based wind turbine equipment fault diagnosis system is characterized by: It includes a memory, a processor, and a display; the memory is used to store relevant programs and data of a large-model-based wind turbine equipment fault diagnosis method; The processor is used to execute the program stored in the memory, drive the fault diagnosis model to run, and perform real-time diagnosis on the wind turbine equipment; the display is used to visualize the model diagnosis results; The system performs the steps of the method provided in claim 1.

7. A computer-readable storage medium, characterized in that A computer program is stored thereon, and instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer or a processor, the computer or the processor executes the steps of the method according to any one of claims 1 to 5.

Images