A method and system for predictive maintenance of a wind turbine

Abstract

The present invention discloses a method and system for predictive maintenance of a wind turbine, comprising the following steps: 1). Collect multi-source data of the wind turbine, including operation data, environmental data, and historical maintenance data; 2). Preprocess the collected data, extract high-frequency features and low-frequency trends of vibration signals using wavelet transform, reduce the dimension and redundant data by principal component analysis, and eliminate noise through Kalman filtering; 3). Construct a deep learning prediction model that fuses multi-source data, the model inputs include time series data, environmental parameters, and device image data, and outputs the health status score and potential fault category of the device; 4). Conduct fault precursor analysis based on the local outlier clustering algorithm; This method and system have significant improvements in terms of data comprehensiveness, model prediction accuracy, real-time performance, and maintenance automation.

Description

Technical Field

[0001] The present invention belongs to the field of maintenance and fault prediction of wind power generation equipment, and specifically relates to a predictive maintenance method and system for wind turbines. Background Art

[0002] The operating environment of wind turbines is complex, and they are long-term exposed to adverse conditions such as high wind speeds, temperature and humidity changes, and dust. Key components are prone to wear, fatigue, and faults. Once a wind turbine fails, it will not only lead to a decrease in power generation efficiency, but may also cause equipment damage and economic losses. Therefore, how to efficiently and reliably predict and maintain wind turbines has become an urgent technical problem in the industry.

[0003] In the prior art, the maintenance methods of wind turbines mainly include regular maintenance and condition-based maintenance. Traditional regular maintenance usually conducts inspections according to fixed cycles, without fully considering the real-time state of the equipment and environmental impacts, which is prone to cause waste of resources or untimely maintenance.

[0004] Moreover, the prior art has obvious deficiencies in data acquisition, fault prediction, maintenance strategies, and intelligence. Therefore, there is an urgent need for an intelligent predictive maintenance method that can integrate multi-source data, improve fault prediction accuracy, and dynamically optimize maintenance strategies to enhance the operating reliability and maintenance efficiency of wind turbines. Summary of the Invention

[0005] The purpose of the present invention is to provide a predictive maintenance method and system for wind turbines, which have significant improvements in terms of data comprehensiveness, model prediction accuracy, real-time performance, and maintenance automation.

[0006] The technical solution adopted by the present invention is as follows: A predictive maintenance method for wind turbines, comprising the following steps:

[0007] 1). Collect multi-source data of the wind turbine, including operating data, environmental data, and historical maintenance data;

[0008] 2). Preprocess the collected data, use wavelet transform to extract the high-frequency features and low-frequency trends of vibration signals, perform principal component analysis for dimensionality reduction and reduce redundant data, and at the same time eliminate noise through Kalman filtering;

[0009] 3). Construct a deep learning prediction model that integrates multi-source data, the model input includes time series data, environmental parameters, and equipment image data, and the output is the health status score and potential fault category of the equipment;

[0010] 4). Conduct fault precursor analysis based on the local outlier clustering algorithm, including:

[0011] Divide the real-time data into fixed time windows and extract the statistical features within the sliding window;

[0012] Perform clustering analysis on the outliers based on the K-Means or DBSCAN algorithm;

[0013] Match the abnormal clustering with the historical fault patterns through the dynamic time warping algorithm to judge the possibility of potential faults;

[0014] 5). Dynamically adjust the maintenance period, and the adjustment formula is: ; where is the adjusted maintenance period, is the basic maintenance period, and are adjustment coefficients, is the current operating load, is the health status score;

[0015] 6). Optimize the health status score by combining the dynamic environmental impact factor, and the optimization formula is: ; where is the adjusted health score, is the initial health score, is the dynamic environmental weight, and the dynamic environmental weight is determined by the principal component weights and eigenvalues of the environmental characteristics;

[0016] 7). Automatically generate a maintenance plan based on the prediction results and historical maintenance records. The maintenance plan includes the maintenance components, required tools, and operation steps, and tasks are sorted according to the priority;

[0017] 8). Send the maintenance plan to the operation and maintenance personnel or the automated maintenance system for execution;

[0018] Among them, in step 1), the operation data includes the real-time status information of vibration, current, and rotational speed components; the environmental data includes environmental parameters such as wind speed, wind direction, temperature and humidity, and dust concentration; the historical maintenance data includes the fault category, cause of occurrence, and maintenance records; the operation data in the multi-source data also obtains the temperature distribution of key components through a thermal imaging device, and combines the image processing algorithm to detect abnormal hot spots to optimize the fault prediction results;

[0019] Among them, the local outlier clustering algorithm in step 4) uses the sliding window technique to extract the mean, standard deviation, and peak value of the time series data as clustering features, and the clustering results of the outliers are matched with the historical fault data through the dynamic time warping algorithm;

[0020] Among them, the operating load in step 5) Calculated by real-time monitoring of power generation, rotational speed, and wind speed, the relationship between wind speed and power generation is represented by a power curve model;

[0021] Among them, the environmental dynamic weight in step 6) is calculated by the following formula: ; among them, is the environmental dynamic weight, is the principal component weight of environmental characteristics, is the environmental characteristic value, is the number of environmental parameters; j is the environmental parameter index, indicating the jth environmental parameter;

[0022] Among them, in step 7), the task priority of the maintenance plan is based on the health status score and environmental risk score, and the wind turbines with low scores or high risks are given priority for processing;

[0023] Among them, in step 3), the deep learning prediction model uses a multi-modal Transformer model to process time series data, environmental parameters, and image data respectively, and the fusion result is used to output the health status score;

[0024] Among them, edge computing nodes are deployed inside the control cabinet of the wind turbine, and the edge nodes process sensor data in real time and upload the anomaly detection results to the cloud;

[0025] Among them, in step 7), the automatically generated maintenance plan combines dynamic priorities to optimize the maintenance path and task scheduling, reducing the operation and maintenance costs;

[0026] Among them, in step 5), and parameters are dynamically updated through real-time environmental data to adapt to changes in environmental conditions.

[0027] A predictive maintenance system for a wind turbine, comprising:

[0028] A data acquisition module for collecting the operation data, environmental data, and historical maintenance data of the wind turbine;

[0029] An edge computing module for denoising the collected operation data and environmental data and extracting key feature values;

[0030] A multi-modal deep learning prediction module for fusing time series data, environmental parameters, and image data to generate the health status score and potential fault categories of the wind turbine;

[0031] A dynamic maintenance optimization module for adjusting the maintenance cycle in real time based on the health status score, environmental dynamic impact factor, and operating load, and automatically generating a maintenance plan including maintenance components, tools, and operation steps;

[0032] A central monitoring and management platform for visually monitoring the real-time operating status, health score, and maintenance plan of wind turbines, as well as the allocation and scheduling of maintenance tasks.

[0033] Among them, the multimodal deep learning prediction module includes:

[0034] A time series data processing unit for extracting time series features of vibration, current, and rotational speed;

[0035] An environmental data processing unit for analyzing environmental features such as wind speed, temperature, humidity, and dust concentration;

[0036] An image data processing unit that processes the thermal imaging images of key components through a convolutional neural network to detect temperature anomaly regions;

[0037] A feature fusion and prediction unit that fuses the features of the time series data, environmental data, and image data through a Transformer model and outputs a health status score and potential fault categories.

[0038] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0039] Multi-source data collection and integration: By integrating operation data, environmental data, and historical maintenance data, comprehensively understand the operating status of wind turbines and avoid information loss that may be caused by a single data source.

[0040] Edge computing and real-time analysis: Deploy an edge computing module to process data in real time, reduce data transmission latency, and improve the real-time performance and efficiency of fault prediction.

[0041] Multimodal deep learning prediction model: Fuse time series, environmental parameters, and image data, and use a multimodal Transformer model to achieve equipment health status scoring and fault category prediction, significantly improving prediction accuracy.

[0042] Local anomaly clustering analysis: Extract statistical features through a sliding window and combine with the K-Means or DBSCAN algorithm for anomaly clustering analysis to further improve the sensitivity of fault precursor detection.

[0043] Dynamic maintenance cycle adjustment: Combine operating load, health score, and environmental impact factors to dynamically adjust the maintenance cycle, improve the adaptability of the maintenance plan and resource utilization efficiency.

[0044] Automated maintenance plan generation: Based on prediction results and historical records, automatically generate a prioritized maintenance plan, optimize the allocation and scheduling of maintenance tasks, and reduce operation and maintenance costs.

[0045] Optimize the health scoring model: Incorporate environmental dynamic impact factors into the calculation of the health score to ensure the score is more applicable and accurate.

[0046] Omnidirectional data storage and visualization management: Use a time series database and a relational database to manage different types of data, and intuitively present the fan status through a visualization platform to facilitate decision-making by maintenance personnel.

[0047] The present invention has significantly improved in terms of data comprehensiveness, model prediction accuracy, real-time performance, and maintenance automation. Description of the Drawings

[0048] Figure 1 It is a schematic flowchart of the maintenance method of the present invention. Detailed Embodiments

[0049] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0050] See Figure 1 , a predictive maintenance method for a wind turbine, comprising the following steps:

[0051] 1). Collect multi-source data of the wind turbine, including operation data, environmental data, and historical maintenance data; the operation data includes real-time status information of components such as vibration, current, and rotational speed; the environmental data includes environmental parameters such as wind speed, wind direction, temperature, humidity, and dust concentration; the historical maintenance data includes fault categories, causes of occurrence, and maintenance records; the operation data in the multi-source data also obtains the temperature distribution of key components through a thermal imaging device, combines image processing algorithms to detect abnormal hot spots, and optimizes the fault prediction results.

[0052] Specifically,

[0053] ① Operation data collection

[0054] For the data collection device, the operation data mainly reflects the real-time status information of the key components of the wind turbine, and the following sensors need to be deployed:

[0055] Vibration sensor: Installed on the main shaft bearing, gearbox, and generator housing to collect vibration data of mechanical components. An accelerometer or a piezoelectric vibration sensor is used. Data collection frequency: 10 - 1000 Hz, and the specific frequency depends on the collected component;

[0056] Current sensor: Deployed in the output circuit of the generator to monitor current fluctuations and identify possible overloads or electrical faults. A Hall effect current sensor is used;

[0057] Rotational speed sensor: Installed at the end of the rotor shaft, it is used to measure the rotational speed of the blade in real time. An optical encoder or a magnetic induction rotational speed sensor is adopted;

[0058] Data transmission and storage

[0059] Edge computing module: Deploy edge computing devices inside the turbine to preliminarily process the operation data, denoise it and calculate key feature values (mean, standard deviation);

[0060] Data protocol: The MQTT (a lightweight Internet of Things protocol) is adopted to transmit the operation data to the central server to ensure low latency and high efficiency;

[0061] Storage method: Use a time series database to store the operation data to support subsequent time series analysis.

[0062] ② Environmental data acquisition

[0063] Data acquisition devices. The environmental data reflects the operating conditions of the area where the turbine is located. The following sensors need to be deployed:

[0064] Wind speed and direction sensor: Deployed at the top of the turbine tower, it is used to monitor the changes in wind speed and direction in real time. An ultrasonic anemometer and a wind vane are adopted;

[0065] Temperature and humidity sensor: Deployed in the middle and at the top of the tower, it is used to measure the environmental temperature and humidity to judge the cooling efficiency of the fan and the operating environment. A digital temperature and humidity sensor is adopted;

[0066] Dust concentration sensor: Deployed at the foundation of the fan or in the nearby area, it is used to monitor the concentration of particulate matter in the air and identify the possible impacts of dust on the equipment. A laser scattering particulate sensor is adopted;

[0067] Rainfall / snowfall monitoring: Deploy rain and snow sensors to detect the impacts of weather changes on the operation of the turbine;

[0068] Data transmission and storage

[0069] Wireless transmission module: The data of the environmental sensors is transmitted to the edge computing device through a LoRa or NB-IoT wireless module;

[0070] Real-time monitoring: Dynamically visualize the environmental data on the central server to provide an overall environmental monitoring view of the wind farm.

[0071] ③ Integration of historical maintenance data

[0072] Data sources. The historical maintenance data needs to integrate the information from the following sources:

[0073] Operation and maintenance records: The maintenance logs of the turbine, including regular maintenance and fault repair records;

[0074] Contains information: maintenance time, components, fault categories, repair measures, spare parts used, etc.;

[0075] Fault data: fault alarm records of the turbine and corresponding cause analysis;

[0076] Contains information: fault category, occurrence time, faulty component, downtime;

[0077] Spare parts usage record: includes the model, quantity of spare parts used in each maintenance and the reason for replacement;

[0078] Data storage and retrieval:

[0079] Maintenance database: design a maintenance data storage system based on a relational database, and the table structure includes:

[0080] Fault category table: records the classification of all known faults;

[0081] Component table: records information on key components of the wind turbine;

[0082] Maintenance record table: associated with the fault category and component table, records specific maintenance details;

[0083] Data tagging: add timestamp, component identification, and fault classification tags to maintenance records for quick retrieval and modeling;

[0084] Data integration:

[0085] Associate historical maintenance data with real-time operation data:

[0086] Match operation data with maintenance data according to the turbine number;

[0087] Conduct cluster analysis on historical data to generate typical fault patterns for training prediction models.

[0088] ④ Overall data integration architecture

[0089] Data acquisition architecture:

[0090] Perception layer: deploy various sensor devices to collect operation data and environmental data;

[0091] Transport layer: use edge computing nodes for data preprocessing and upload to the central server via wireless communication (LoRa, 4G / 5G);

[0092] Data processing layer: the central server integrates operation data, environmental data, and historical maintenance data to form a complete fan status database;

[0093] Data management platform:

[0094] Time series analysis platform: for trend analysis of operation data and environmental data;

[0095] Historical record retrieval system: provides a fast retrieval and analysis interface for operation and maintenance personnel;

[0096] Real-time monitoring platform: displays the real-time status of the turbine through a dashboard, including health score, environmental parameters, and historical fault information.

[0097] 2). Preprocess the collected data, extract the high-frequency features and low-frequency trends of the vibration signal using wavelet transform, reduce the dimension and redundant data by principal component analysis, and eliminate noise through Kalman filter;

[0098] Specifically,

[0099] ① Extract the high-frequency features and low-frequency trends of the vibration signal using wavelet transform

[0100] Basic principle of wavelet transform: Wavelet transform is a time-frequency analysis tool that can decompose a signal into multiple components, representing the high-frequency part (local mutations or abnormal features) and the low-frequency part (overall trend) respectively.

[0101] Implementation steps:

[0102] Input signal: The vibration signal is collected from key components of the turbine, such as the main shaft, gearbox, or generator. The input signal is time series data X(t);

[0103] Select wavelet basis function: Select Daubechies wavelet (db4 or db6) or Symlet wavelet, as it is suitable for feature extraction of mechanical vibration signals;

[0104] Set decomposition level: Set the number of decomposition layers according to the sampling frequency and the target feature frequency band , ; where Sampling frequency. , Frequency range obtained by decomposition;

[0105] Signal decomposition: Use discrete wavelet transform (DWT) to decompose the signal X(t) into multiple layers of high-frequency components and low-frequency components . ; where X(t): original signal (time series data), such as vibration signal, unit: (acceleration). L: number of layers of wavelet decomposition, determining the depth of decomposition, usually set according to the sampling frequency and the target frequency band. : the i-th layer high-frequency component, representing local mutation features, with the same unit as the original signal. : The low-frequency component of the L-th layer, representing the overall trend, with the unit being the same as that of the original signal;

[0106] High-frequency and low-frequency signal extraction: The high-frequency component is used to extract features (such as abnormal signals), and the low-frequency component is used to monitor the overall operating state.

[0107] ②Principal component analysis (PCA) for dimensionality reduction and redundancy reduction

[0108] Implementation steps:

[0109] Construct a data matrix: Organize multi-source data (vibration, current, rotational speed, etc.) into a matrix: ; where : Number of features, n: Length of the time series;

[0110] Standardization processing: Standardize the data matrix with a mean of 0 and a variance of 1: ; where : The i-th feature data after standardization. The original i-th feature data, unit: specific feature unit (such as vibration is , current is A, rotational speed is rpm. μ: Feature mean, with the unit being the same as Same. σ Feature standard deviation, with the unit being the same as Same;

[0111] Calculate the covariance matrix: Covariance matrix: ; where, X: Multi-source data matrix, with the dimension of , where is the number of samples, is the number of features. C: Covariance matrix, used to measure the correlation between features, with the dimension of ;

[0112] Feature decomposition: Find the eigenvalues and eigenvectors , The eigenvector of the covariance matrix, representing the direction of the i-th principal component, unit: dimensionless;

[0113] Principal component selection: Select the first K principal components according to the cumulative contribution rate: ; where : Feature matrix after dimensionality reduction, retaining the K principal components with higher information contribution rates, with the dimension of .

[0114] 3). Build a deep learning prediction model that integrates multi-source data. The inputs of the model include time series data, environmental parameters, and device image data, and the outputs are the health status score and potential failure categories of the device. The deep learning prediction model uses a multi-modal Transformer model to process time series data, environmental parameters, and image data respectively, and the fusion result is used to output the health status score.

[0115] Specifically,

[0116] ① Time series data processing module

[0117] Input format, input shape: ( , t, f), where: : Number of samples; t: Number of time steps; f: Number of time series features (vibration, current, rotational speed, etc.);

[0118] Embedding generation, formula: ; where Time series embedding representation; Input time series data; Embedding weight matrix; Embedding bias; Position encoding, used to retain time step information; Time series embedding dimension;

[0119] Feature extraction: Extract time series features through the Transformer Encoder .

[0120] ② Environmental parameter processing module

[0121] Input format, input shape: ( , e), where: : Number of samples; e: Number of environmental parameter features (wind speed, temperature, humidity, etc.);

[0122] Embedding generation, formula: ; where Environmental parameter embedding representation; : Input environmental parameters; Embedding weight matrix; Embedding bias; Environmental parameter embedding dimension;

[0123] Feature extraction: Extract environmental features through the Transformer Encoder .

[0124] ③ Image data processing module

[0125] Input format, input shape: ( , h, w, c), where: : number of samples; h: image height; w: image width; c: number of image channels;

[0126] Embedding generation, the image generates embedding features through a convolutional neural network: ; where Image embedding representation; Input image;

[0127] Feature extraction: Extract image features through the Transformer Encoder . Where Image embedding dimension.

[0128] ④ Feature fusion module

[0129] Fusion operation, concatenate and fuse features of each modality: Formula: ; where Fused feature vector;

[0130] Fully connected fusion, formula: ; where Fused feature; Fully connected weight matrix; Fully connected bias; Fused feature dimension.

[0131] ⑤ Output module

[0132] Health status score, output the health status score of the device through the regression layer: Formula: ; where : Health status score; : Weight matrix; Bias;

[0133] Fault classification, output the probability distribution of device fault categories through the classification layer: Formula: ; where : Probability distribution of fault categories; : Weight matrix; : Bias; C: Number of fault categories.

[0134] ⑥ Model training

[0135] Loss function, health status score loss (mean squared error):

[0136] ; where True health score; Predicted health score;

[0137] Fault classification loss (cross-entropy): ; where : True fault category label; Predicted fault category probability;

[0138] Total loss: ; where α, β: Weight factors.

[0139] ⑦ Model output

[0140] Health status score: A scalar in the range [0, 1], reflecting the overall health status of the device;

[0141] Fault classification: A vector representing the probability distribution of potential fault categories of the device.

[0142] 4). Conduct fault precursor analysis based on the local outlier clustering algorithm, including:

[0143] Divide the real-time data into fixed time windows and extract the statistical features within the sliding window;

[0144] Conduct clustering analysis on the outliers based on the K-Means or DBSCAN algorithm;

[0145] Match the abnormal clustering with the historical fault patterns through the dynamic time warping algorithm to judge the possibility of potential faults;

[0146] The local outlier clustering algorithm uses the sliding window technique to extract the mean, standard deviation, and peak value of the time series data as clustering features, and the clustering results of the outliers are matched with the historical fault data through the dynamic time warping algorithm.

[0147] The specific implementation steps are as follows:

[0148] ① Statistical feature extraction within the sliding window:

[0149] Data division, divide the real-time data according to the fixed time window and process it using the sliding window technique: Input data shape: Time series data where n: Number of data points; f: Number of features (vibration, current, rotational speed, etc.). Sliding window parameters: Window size w: Number of data points included in each window; Sliding step s: Number of points the window moves each time;

[0150] Feature extraction, within each window, calculate the following statistical features: Mean: ; where μ: Mean of the data within the window; : The i-th data point; w: Window size. Standard deviation: ; σ: Standard deviation of the data within the window. Peak value: ; p: Maximum value within the window;

[0151] Feature matrix generation: The features calculated for each window (mean, standard deviation, peak value) form a feature vector: F = [μ, σ, p]. The feature vectors of all windows are stacked into a feature matrix: ; m: number of windows; k = 3f: dimensionality of features extracted from each window (3 statistical features of each original feature).

[0152] ② Outlier clustering analysis

[0153] K-Means clustering: Use K-Means to cluster the feature matrix F, ; where, : feature vector of the i-th window; the K-th cluster; center of the cluster; K: number of clusters (including one outlier cluster);

[0154] DBSCAN clustering: Use DBSCAN to cluster the feature matrix F based on density: Define parameters: radius threshold ε: define the density range; minimum number of neighbors minPts: specify the minimum density of core points. Marking results: Points with sufficient density are classified as normal points; points in sparse regions are marked as outliers;

[0155] Output: The clustering result marks each window as: Normal points: belonging to dense clusters; Outliers: isolated points or sparse clusters.

[0156] ③ Dynamic Time Warping (DTW) matching

[0157] Dynamic Time Warping formula: Use Dynamic Time Warping to match the outlier sequence with historical fault patterns: ; where, D(i, j): cumulative matching distance; , : feature vectors of the outlier sequence and historical fault patterns; min: path selection of dynamic programming;

[0158] Threshold determination: Calculate the cumulative matching distance D(m, n) between the outlier sequence and each historical fault pattern: If D(m, n) < threshold, then it is determined that the outlier sequence matches a certain fault pattern, and there may be a potential fault.

[0159] 5). Dynamically adjust the maintenance period, and the adjustment formula is: ; where, is the adjusted maintenance period, is the basic maintenance period, and are adjustment coefficients, is the current operating load, Health status score; operating load Calculated by real-time monitoring of power generation, rotational speed, and wind speed. The relationship between wind speed and power generation is represented by a power curve model. Dynamically updated through real-time environmental data and parameters to adapt to changes in environmental conditions.

[0160] 6). Optimize the health status score by combining environmental dynamic impact factors. The optimization formula is: ; where is the adjusted health score, is the initial health score, is the environmental dynamic weight, and the environmental dynamic weight is determined by the principal component weight and eigenvalue of environmental characteristics;

[0161] Environmental dynamic weight Calculated by the following formula: ; where is the environmental dynamic weight, is the principal component weight of environmental characteristics, is the environmental eigenvalue, and m is the number of environmental parameters.

[0162] 7). Automatically generate a maintenance plan based on the prediction results and historical maintenance records. The maintenance plan includes maintenance components, required tools, and operation steps, and tasks are sorted according to priority;

[0163] The task priority of the maintenance plan is based on the health status score and environmental risk score. Wind turbines with low scores or high risks are given priority for processing;

[0164] The automatically generated maintenance plan combines dynamic priorities to optimize the maintenance path and task scheduling, reducing operation and maintenance costs.

[0165] 8). Send the maintenance plan to the operation and maintenance personnel or an automated maintenance system for execution;

[0166] To support this method, a wind turbine predictive maintenance system includes:

[0167] A data acquisition module for collecting the operation data, environmental data, and historical maintenance data of the wind turbine;

[0168] An edge computing module for denoising the collected operation data and environmental data and extracting key feature values;

[0169] A multi-modal deep learning prediction module for fusing time series data, environmental parameters, and image data to generate the health status score and potential fault categories of the wind turbine;

[0170] A dynamic maintenance optimization module, which is used to adjust the maintenance cycle in real time based on the health status score, environmental dynamic impact factors, and operating load, and automatically generate a maintenance plan including maintenance components, tools, and operation steps;

[0171] A central monitoring and management platform, which is used to visually monitor the real-time operating status, health score, and maintenance plan of the wind turbine, as well as the allocation and scheduling of maintenance tasks.

[0172] Furthermore, the multimodal deep learning prediction module includes:

[0173] A time series data processing unit, which is used to extract the time series features of vibration, current, and rotational speed;

[0174] An environmental data processing unit, which is used to analyze the environmental characteristics of wind speed, temperature, humidity, and dust concentration;

[0175] An image data processing unit, which processes the thermal imaging images of key components through a convolutional neural network to detect temperature anomaly regions;

[0176] A feature fusion and prediction unit, which fuses the features of the time series data, environmental data, and image data through a Transformer model and outputs the health status score and potential fault categories.

[0177] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predictive maintenance of a wind turbine, characterized in that, It includes the following steps: 1). Collect multi-source data of the wind turbine, including operation data, environmental data, and historical maintenance data; 2). Preprocess the collected data. Use wavelet transform to extract the high-frequency features and low-frequency trends of vibration signals, perform principal component analysis for dimensionality reduction and reduce redundant data, and at the same time eliminate noise through Kalman filtering; 3). Construct a deep learning prediction model that fuses multi-source data. The inputs of the model include time series data, environmental parameters, and device image data, and the outputs are the health status score and potential fault categories of the device; 4). Conduct pre-fault analysis based on the local outlier clustering algorithm, including: Divide the real-time data into fixed time windows and extract the statistical features within the sliding window; Perform clustering analysis on the outlier points based on the K-Means or DBSCAN algorithm; Match the outlier clustering with the historical fault patterns through the dynamic time warping algorithm to judge the possibility of potential faults; 5). Dynamically adjust the maintenance period, and the adjustment formula is: T maint-adaptive = T base - k1·L load - k2·S health ; where, T maint-adaptive is the adjusted maintenance period, T base is the basic maintenance period, k1 and k2 are adjustment coefficients, L load is the current operating load, and S health is the health status score; 6). Optimize the health status score by combining the environmental dynamic impact factor, and the optimization formula is: S adjusted = S raw - W env ; where S adjusted is the adjusted health score, S raw is the initial health score, and W env is the environmental dynamic weight, and the environmental dynamic weight is determined by the principal component weight and eigenvalue of the environmental characteristics; 7). Automatically generate a maintenance plan based on the prediction results and historical maintenance records. The maintenance plan includes maintenance components, required tools, and operation steps, and sorts the tasks according to priorities; 8). Send the maintenance plan to the operation and maintenance personnel or the automated maintenance system for execution; In step 1), the operation data includes the real-time status information of components such as vibration, current, and rotational speed; the environmental data includes environmental parameters such as wind speed, wind direction, temperature, humidity, and dust concentration; the historical maintenance data includes fault categories, causes of occurrence, and maintenance records; the operation data in the multi-source data also obtains the temperature distribution of key components through a thermal imaging device, combines image processing algorithms to detect abnormal hot spots, and optimizes the fault prediction results; In the local outlier clustering algorithm in step 4), the sliding window technology is used to extract the mean, standard deviation, and peak value of the time series data as clustering features, and the clustering results of the outlier points are matched with the historical fault data through the dynamic time warping algorithm; The operating load L in step 5) load Calculated by real-time monitoring of power generation, rotational speed, and wind speed. The relationship between wind speed and power generation is represented by a power curve model. The parameters k1 and k2 are dynamically updated based on real-time environmental data to adapt to changes in environmental conditions; The environmental dynamic weight W in step 6) env is calculated by the following formula: where W env is the environmental dynamic weight, λ j is the principal component weight of the environmental feature, F j is the environmental feature value, and m is the number of environmental parameters.

2. The predictive maintenance method of a wind turbine according to claim 1, characterized in that: In step 7), the task priorities of the maintenance plan are based on the health status score and environmental risk score, and the wind turbines with low scores or high risks are given priority for processing; the automatically generated maintenance plan combines dynamic priorities, optimizes the maintenance path and task scheduling, and reduces the operation and maintenance costs.

3. A predictive maintenance method for a wind turbine according to claim 1, characterized in that: In step 3), the deep learning prediction model uses a multi-modal Transformer model to process time series data, environmental parameters, and image data respectively, and the fusion result is used to output the health status score.

4. A predictive maintenance method for a wind turbine according to claim 1, characterized in that: Edge computing nodes are deployed inside the control cabinet of the wind turbine, and the edge nodes process sensor data in real time and upload the abnormal detection results to the cloud.

5. A wind turbine predictive maintenance system for supporting a predictive maintenance method of a wind turbine according to any one of claims 1 to 4, characterized in that, It includes: A data acquisition module for collecting the operation data, environmental data, and historical maintenance data of the wind turbine; An edge computing module for denoising the collected operation data and environmental data and extracting key feature values; A multi-modal deep learning prediction module for fusing time series data, environmental parameters, and image data to generate the health status score and potential fault categories of the wind turbine; A dynamic maintenance optimization module for adjusting the maintenance cycle in real time based on the health status score, environmental dynamic impact factor, and operation load, and automatically generating a maintenance plan including maintenance components, tools, and operation steps; A central monitoring and management platform for visualizing the real-time operating status, health score, and maintenance plan of wind turbines, as well as the allocation and scheduling of maintenance tasks.

6. The predictive maintenance system for a wind turbine according to claim 5, characterized in that, It includes: The multimodal deep learning prediction module includes: A time series data processing unit for extracting time series features of vibration, current, and rotational speed; An environmental data processing unit for analyzing environmental features such as wind speed, temperature, humidity, and dust concentration; An image data processing unit for processing thermal imaging images of key components through a convolutional neural network to detect temperature anomaly regions; A feature fusion and prediction unit for fusing the features of the time series data, environmental data, and image data through a Transformer model and outputting a health status score and potential fault categories.

Images