Wind power blade health diagnosis system based on AI

Abstract

The invention relates to the technical field of wind power equipment state monitoring, and discloses an AI-based wind power blade health diagnosis system comprising a sensing module used for collecting multi-source monitoring data in a wind power blade operation process, the multi-source monitoring data comprising vibration data, acoustic data, strain data, environmental data and visual image data; the data preprocessing module is in communication connection with the sensing module and is used for performing cleaning, standardization, data fusion and feature extraction on the acquired multi-source monitoring data and outputting a standardized feature data set; and the AI diagnosis model training module is used for constructing a deep learning model and is based on a training set composed of historical health state data and fault data. The method effectively solves the problems of low efficiency, poor accuracy, incapability of early warning and the like in existing wind power blade health diagnosis, provides reliable guarantee for safe and stable operation of a wind turbine generator, and has wide application prospect and practical value.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment condition monitoring technology, specifically to an AI-based wind turbine blade health diagnosis system. Background Technology

[0002] Wind turbine blades are one of the core components of wind turbine generators, and their operating status directly affects the power generation efficiency and safety stability of the entire wind turbine. Wind turbine blades are constantly exposed to a complex natural environment, facing erosion from harsh conditions such as strong winds, heavy rain, lightning strikes, and sandstorms. They also endure periodic aerodynamic loads, centrifugal loads, and vibration impacts, making them prone to cracks, wear, fatigue deformation, and other failures. Failure to detect and address these failures in a timely manner can lead to serious accidents such as blade breakage and turbine shutdown, resulting in significant economic losses.

[0003] Current methods for diagnosing wind turbine blade health mainly combine manual inspection, traditional sensor monitoring, and simple data analysis. Manual inspection is inefficient and costly due to environmental conditions and personnel experience, and it struggles to detect internal or hidden faults within the blade. Traditional sensor monitoring methods often analyze single-type data, such as relying solely on vibration signals to diagnose faults, ignoring the correlation between multi-source data. This results in low diagnostic accuracy and an inability to effectively capture the temporal patterns of fault development, hindering early warning systems. Furthermore, existing diagnostic methods have limited ability to identify complex faults, failing to accurately distinguish fault types and severity, thus complicating maintenance work.

[0004] With the rapid development of artificial intelligence technology, deep learning algorithms have demonstrated powerful feature extraction and pattern recognition capabilities in the field of fault diagnosis. Based on this, this invention proposes an AI-based wind turbine blade health diagnosis system. Through multi-source data fusion and a deep learning model, it achieves accurate, real-time diagnosis and early warning of the health status of wind turbine blades, addressing the shortcomings of existing technologies. Summary of the Invention

[0005] This invention provides an AI-based wind turbine blade health diagnosis system, which solves the problems mentioned in the background.

[0006] This invention provides the following technical solution: an AI-based wind turbine blade health diagnosis system, comprising: Sensing module: used to collect multi-source monitoring data during the operation of wind turbine blades, including vibration data, acoustic data, strain data, environmental data and visual image data; Data preprocessing module: Communicatively connected to the sensing module, used to clean, standardize, fuse, and extract features from the collected multi-source monitoring data, and output a standardized feature dataset; AI diagnostic model training module: used to build deep learning models. Based on a training set composed of historical health status data and fault data, the deep learning models are trained, validated and optimized to obtain pre-trained AI diagnostic models. Real-time diagnostic module: It is communicatively connected to the data preprocessing module and the AI ​​diagnostic model training module, respectively. It is used to input the standardized feature dataset into the AI ​​diagnostic model and obtain the leaf health status diagnosis result through model inference. The diagnosis result includes health level, fault type, fault location and fault severity. Early warning output module: It is connected to the real-time diagnosis module and is used to generate early warning information based on the diagnosis results and push it to the designated terminal.

[0007] The preferred sensing module includes a vibration sensor, an acoustic sensor, a strain gauge, a temperature and humidity sensor, and a high-definition camera. The vibration sensor is installed at the blade root and tip to collect three-dimensional vibration acceleration data during blade operation. The acoustic sensor is installed on the side of the nacelle near the blade to collect acoustic signals generated by the blade's operation. The strain gauge is attached to the critical stress area of ​​the blade to collect strain data. The temperature and humidity sensor is installed outside the nacelle to collect ambient temperature and humidity data. The high-definition camera is installed on the top of the nacelle to capture real-time images of the blade surface.

[0008] The selected feature extraction process of the data preprocessing module includes: performing Fourier transform and wavelet packet decomposition on vibration data to extract frequency domain features and time domain features; extracting Mel frequency cepstral coefficients from acoustic data; performing edge detection, texture feature extraction, and defect region segmentation on blade surface images; extracting statistical features from strain data and environmental data, including mean, variance, peak value, and trend features; and finally, using principal component analysis algorithm to perform dimensionality reduction and fusion of multi-dimensional features to obtain a standardized feature dataset.

[0009] The selected AI diagnostic model is a hybrid deep learning model based on CNN-LSTM, including a CNN feature extraction layer, an LSTM temporal analysis layer, and a fully connected classification layer. The CNN feature extraction layer is used to extract spatial features from the standardized feature dataset. The LSTM temporal analysis layer is used to capture the temporal dependencies of the feature data and analyze the changing trend of leaf health status. The fully connected classification layer is used to map the extracted features to health status categories and output diagnostic results.

[0010] The AI ​​diagnostic model training module further includes a model optimization unit, which uses an adaptive moment estimation optimization algorithm to adjust model parameters, avoids model overfitting through cross-validation, and sets a dynamic learning rate that adaptively adjusts the learning rate based on the number of training iterations to improve model training efficiency and generalization ability.

[0011] The selected health levels include normal state, minor fault state, moderate fault state, and severe fault state; the fault types include surface cracks, edge wear, coating peeling, lightning damage, fatigue deformation, and foreign object adhesion; the warning information includes text prompts, audible and visual alarms, and fault visualization reports, and the fault visualization reports include fault location markings, quantitative analysis of fault severity, and maintenance priority suggestions.

[0012] Specifically, the data preprocessing module and the real-time diagnosis module are communicatively connected to a data storage module. The data storage module is used to store raw monitoring data, standardized feature datasets, diagnostic results, and early warning records. The data storage module adopts a distributed database architecture, which supports real-time data writing and fast querying.

[0013] The present invention has the following beneficial effects: 1. Multi-source data fusion: By integrating multiple types of monitoring data such as vibration, acoustics, strain, and vision, it comprehensively captures blade operating status information, avoids the limitations of single data diagnosis, and improves diagnostic accuracy; 2. AI Intelligent Diagnosis: Employing a CNN-LSTM hybrid deep learning model, it can extract spatial features of data and capture temporal dependencies, achieving accurate fault identification and severity assessment with a diagnostic accuracy of over 95%, far exceeding traditional diagnostic methods. 3. Real-time performance and early warning: The system response time is less than 10 seconds, and it can output diagnostic results in real time. It can also provide early warning of potential serious faults up to 72 hours in advance, which saves time for operation and maintenance work and reduces the risk of fault escalation. 4. Automation and Intelligence: No manual intervention is required throughout the entire process, realizing full automation from data collection, processing, diagnosis to early warning, reducing the cost of manual inspection and improving operation and maintenance efficiency; 5. Dynamic update capability: Dynamic updates are achieved through incremental model training, continuously improving the ability to identify new types of faults and adapting to complex and ever-changing operating environments.

[0014] This invention effectively solves the problems of low efficiency, poor accuracy, and inability to provide early warning in existing wind turbine blade health diagnosis, providing a reliable guarantee for the safe and stable operation of wind turbine units, and has broad application prospects and practical value. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] An AI-based wind turbine blade health diagnosis system includes: Sensing module: Used to collect multi-source monitoring data during the operation of wind turbine blades. The multi-source monitoring data includes vibration data, acoustic data, strain data, environmental data, and visual image data. The sensing module includes vibration sensors, acoustic sensors, strain gauges, temperature and humidity sensors, and a high-definition camera. Vibration sensors are installed at the blade root and tip to collect three-dimensional vibration acceleration data during blade operation. Acoustic sensors are installed on the nacelle side near the blade to collect acoustic signals generated by the blade's movement. Strain gauges are attached to key stress areas of the blade to collect strain data. Temperature and humidity sensors are installed outside the nacelle to collect ambient temperature and humidity data. The high-definition camera is installed on the top of the nacelle to capture real-time images of the blade surface. Vibration sensors: Three-axis accelerometers with a measurement range of ±10g and a sampling frequency of 1000Hz are used, bolted to the upper and lower surfaces of the blade root and the blade tip; four are installed in total. Acoustic sensors: Wideband acoustic sensors with a frequency response range of 20Hz-20kHz and a sampling frequency of 44.1kHz are used, mounted on the front of the nacelle near the blade using a bracket, with the installation height corresponding to the blade's rotation trajectory. Strain gauges: Foil strain gauges with a sensitivity coefficient of 2.1 are used, and are attached to the upper and lower surfaces of the blade at the maximum chord length, as well as key stress areas such as the leading and trailing edges of the blade. A total of 8 gauges are attached, and waterproof sealing is used. Temperature and humidity sensor: A digital temperature and humidity sensor is used, with a measurement range of -40℃ to 85℃ and 0 to 100%RH, and an accuracy of ±0.5℃ and ±2%RH. It is installed in an unobstructed location outside the nacelle. High-definition camera: An industrial-grade high-definition camera with a resolution of 1920×1080 and a frame rate of 30fps is used. It is equipped with an infrared supplementary light module and is mounted on the top of the nacelle via a rotatable bracket. The lens faces the blade rotation area, and the shooting angle can be adjusted remotely. Data transmission adopts a combination of wireless and wired communication: the raw data collected by the sensor is transmitted to the edge gateway in the nacelle through the LoRa wireless communication module. The edge gateway is connected to the central server of the wind farm via Ethernet to realize real-time data upload. Data preprocessing module: Communicates with the sensing module and is used to clean, standardize, fuse, and extract features from the collected multi-source monitoring data, and output a standardized feature dataset; The feature extraction process of the data preprocessing module includes: performing Fourier transform and wavelet packet decomposition on vibration data to extract frequency domain and time domain features; extracting Mel frequency cepstral coefficients from acoustic data; performing edge detection, texture feature extraction, and defect region segmentation on blade surface images; extracting statistical features from strain data and environmental data, including mean, variance, peak value, and trend features; and finally, using principal component analysis to reduce and fuse multi-dimensional features to obtain a standardized feature dataset. The specific processing steps of the data preprocessing module are as follows: Data cleaning: Outliers in vibration and strain data are removed using the 3σ criterion; background noise in acoustic data is adaptively filtered; and blade surface images are preprocessed for denoising and defogging. Standardization: Normalize the data of each type, map the data to the [0,1] interval, and eliminate the dimensional differences between data from different sensors; Feature extraction: Vibration data: Fourier transform was performed to obtain the frequency spectrum, and frequency domain features such as peak frequency, spectral centroid, and harmonic components were extracted; wavelet packet decomposition was performed to 5 levels to extract time domain features such as energy proportion, kurtosis, and skewness of each frequency band, resulting in a total of 24-dimensional features; Acoustic data: Mel frequency cepstral coefficients were extracted to obtain 13-dimensional MFCC features and first-order and second-order difference features, for a total of 39 features; Blade surface images: edge features were extracted using the Canny edge detection algorithm, texture features were extracted using the gray-level co-occurrence matrix, and defect regions were segmented using the U-Net network to extract morphological features such as defect area, perimeter, and roundness, totaling 16-dimensional features; Strain data and environmental data: Extract statistical features such as mean, variance, peak value, trough value, and trend slope, totaling 12 dimensions; Feature fusion and dimensionality reduction: The features of the above dimensions are combined to form an initial feature set of 81 dimensions. Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the initial feature set, retaining the principal components with a cumulative contribution rate of 95%, and finally obtaining a 20-dimensional standardized feature dataset.

[0017] AI diagnostic model training module: used to build deep learning models. Based on a training set composed of historical health status data and fault data, the deep learning models are trained, validated and optimized to obtain pre-trained AI diagnostic models. The AI ​​diagnostic model is a hybrid deep learning model based on CNN-LSTM, including a CNN feature extraction layer, an LSTM temporal analysis layer, and a fully connected classification layer. The CNN feature extraction layer is used to extract spatial features from the standardized feature dataset; the LSTM temporal analysis layer is used to capture the temporal dependencies of the feature data and analyze the changing trends of leaf health status; the fully connected classification layer is used to map the extracted features to health status categories and output diagnostic results. The AI ​​diagnostic model training module also includes a model optimization unit, which uses an adaptive moment estimation optimization algorithm to adjust model parameters, avoids model overfitting through cross-validation, and sets a dynamic learning rate that adaptively adjusts the learning rate according to the number of training iterations to improve model training efficiency and generalization ability. The AI ​​diagnostic model uses a CNN-LSTM hybrid model, with the following specific structure: The CNN feature extraction layer consists of three convolutional layers and two pooling layers. The first convolutional layer uses 32 3×3 convolutional kernels with ReLU activation and a stride of 1. The first pooling layer is a 2×2 max pooling layer. The second convolutional layer uses 64 3×3 convolutional kernels with ReLU activation. The second pooling layer is a 2×2 max pooling layer. The third convolutional layer uses 128 3×3 convolutional kernels with ReLU activation. LSTM time series analysis layer: contains 2 LSTM units, each LSTM unit has a hidden layer dimension of 64, and uses the dropout mechanism to avoid overfitting; Fully connected classification layer: It contains two fully connected layers. The first layer has an output dimension of 32 and the activation function is ReLU. The second layer has an output dimension of 5, corresponding to five state categories: "normal state, surface crack, edge wear, fatigue deformation, and lightning damage", and the activation function is Softmax.

[0018] The model training process is as follows: Training set construction: Historical monitoring data of blades from 50 wind turbines in a wind farm were collected, including normal operation data and different types of fault data (obtained through artificially simulated faults and actual fault records), totaling 100,000 sample data, which were divided into training set, validation set and test set in a ratio of 7:2:1; Model training: The Adam optimization algorithm is used, with an initial learning rate of 0.001. The learning rate is halved when the validation set loss does not decrease for 5 consecutive iterations. The cross-entropy loss function is used, and the number of training iterations is set to 100. Model optimization: The model hyperparameters were adjusted by cross-validation (5-fold cross-validation). After optimization, the model achieved a diagnostic accuracy of 96.8% and a fault type identification accuracy of 95.3% on the test set, with an average response time of 8.2 seconds.

[0019] Real-time diagnosis module: It communicates with the data preprocessing module and the AI ​​diagnosis model training module respectively. It is used to input the standardized feature dataset into the AI ​​diagnosis model and obtain the leaf health status diagnosis result through model inference. The diagnosis result includes health level, fault type, fault location and fault severity. Health levels include normal, minor, moderate, and severe fault conditions; fault types include surface cracks, edge wear, coating peeling, lightning damage, fatigue deformation, and foreign object adhesion; warning information includes text prompts, audible and visual alarms, and fault visualization reports. The fault visualization reports include fault location markings, quantitative analysis of fault severity, and maintenance priority settings. Early warning output module: It communicates with the real-time diagnosis module and is used to generate early warning information based on the diagnosis results and push it to the designated terminal.

[0020] The data preprocessing module and the real-time diagnosis module are connected to a data storage module. The data storage module is used to store raw monitoring data, standardized feature datasets, diagnostic results and early warning records. The data storage module adopts a distributed database architecture, which supports real-time writing and fast querying of data.

[0021] The workflow of this system is as follows: Step 1: The sensing module collects vibration data, acoustic data, strain data, environmental data, and surface image data of the wind turbine blades in real time, with a sampling interval of 1 second; Step 2: The data preprocessing module cleans, standardizes, extracts features, and performs dimensionality reduction and fusion on the collected raw data to generate a standardized feature dataset; Step 3: The real-time diagnostic module inputs the standardized feature dataset into the pre-trained AI diagnostic model. The model then outputs the health level, fault type, fault location, and severity of the blade through inference. Step 4: The early warning output module generates early warning information based on the diagnostic results: if it is in a normal state, only data is recorded; if it is in a minor fault state, a yellow warning and routine maintenance suggestions are generated; if it is in a moderate fault state, an orange warning, a fault visualization report, and priority maintenance suggestions are generated; if it is in a severe fault state, a red warning, an audible and visual alarm signal, and an emergency shutdown suggestion are generated and pushed to the wind farm monitoring center terminal and the mobile APP of the operation and maintenance personnel. Step 5: The data storage module stores raw monitoring data, standardized feature datasets, diagnostic results, and early warning records in real time, and supports historical data query and analysis; Step 6: Every 3 months, the model update module calls the newly added data in the data storage module to perform incremental training on the AI ​​diagnostic model, realize the dynamic update of the model, and improve the model's ability to identify new types of faults.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based wind turbine blade health diagnosis system, characterized in that, The application relates to a wind turbine blade health state diagnosis system based on AI, which comprises the following modules: a perception module: used for collecting multi-source monitoring data in the operation process of a wind turbine blade, wherein the multi-source monitoring data comprises vibration data, acoustic data, strain data, environmental data and visual image data; a data preprocessing module: in communication connection with the perception module, used for cleaning, standardizing, data fusing and feature extracting of the collected multi-source monitoring data, and outputting a standardized feature data set; an AI diagnosis model training module: used for constructing a deep learning model, training, verifying and optimizing the deep learning model based on a training set composed of historical health state data and fault data, and obtaining a pre-trained AI diagnosis model; a real-time diagnosis module: in communication connection with the data preprocessing module and the AI diagnosis model training module, used for inputting the standardized feature data set into the AI diagnosis model, obtaining a blade health state diagnosis result through model reasoning, and the diagnosis result comprising a health grade, a fault type, a fault position and a fault severity; an early warning output module: in communication connection with the real-time diagnosis module, used for generating early warning information according to the diagnosis result and pushing the early warning information to a specified terminal.

2. The AI-based wind turbine blade health diagnosis system of claim 1, wherein: The perception module comprises a vibration sensor, an acoustic sensor, a strain gauge, a temperature and humidity sensor and a high-definition camera; the vibration sensor is installed at the root and the tip of the blade to collect three-dimensional vibration acceleration data when the blade is running; the acoustic sensor is installed on the side of the cabin close to the blade to collect acoustic signals generated when the blade is running; the strain gauge is pasted on the key stress area of the blade to collect strain data of the blade; the temperature and humidity sensor is installed outside the cabin to collect environmental temperature and humidity data; and the high-definition camera is installed on the top of the cabin to shoot real-time images of the blade surface.

3. The AI-based wind turbine blade health diagnosis system of claim 1, wherein: The feature extraction process of the data preprocessing module comprises the following steps: Fourier transform and wavelet packet decomposition are carried out on the vibration data to extract frequency domain features and time domain features; mel frequency cepstral coefficient extraction is carried out on the acoustic data; edge detection, texture feature extraction and defect area segmentation are carried out on the blade surface images; statistical feature extraction is carried out on the strain data and the environmental data, including mean value, variance, peak value and trend features; and finally, multi-dimensional features are reduced and fused through a principal component analysis algorithm to obtain a standardized feature data set.

4. The AI-based wind turbine blade health diagnosis system of claim 1, wherein: The AI diagnosis model is a hybrid deep learning model based on CNN-LSTM, which comprises a CNN feature extraction layer, an LSTM time sequence analysis layer and a full connection classification layer; the CNN feature extraction layer is used for extracting spatial features in the standardized feature data set; the LSTM time sequence analysis layer is used for capturing time sequence dependent relationships of the feature data and analyzing the change trend of the blade health state; and the full connection classification layer is used for mapping the extracted features to health state categories and outputting diagnosis results.

5. The AI-based wind turbine blade health diagnosis system of claim 4, wherein: The AI diagnosis model training module further comprises a model optimization unit, which adopts an adaptive moment estimation optimization algorithm to adjust model parameters, avoids model overfitting through a cross-validation method, sets a dynamic learning rate, adaptively adjusts the learning rate according to the number of training iterations, and improves the model training efficiency and generalization ability.

6. The AI-based wind turbine blade health diagnosis system of claim 1, wherein: The health level includes normal state, slight fault state, moderate fault state and serious fault state; and the fault type includes surface crack, edge wear, coating peeling, lightning damage, fatigue deformation and foreign matter adhesion. The early warning information includes text prompt, sound and light alarm and fault visual report, and the fault visual report includes fault position marking, fault severity quantitative analysis and maintenance priority suggestion.

7. The AI-based wind turbine blade health diagnosis system of claim 1, wherein: The data preprocessing module and the real-time diagnosis module are in communication connection with a data storage module, the data storage module is used for storing original monitoring data, standardized feature data set, diagnosis result and early warning record, the data storage module adopts a distributed database architecture, and supports real-time writing and quick query of data.