Health monitoring system and monitoring method for wind turbine blades

Abstract

Disclosed in the present invention are a health monitoring system and monitoring method for wind turbine blades. In the present invention, fault signals are extracted and collected on the basis of multi-sensor monitoring, and the collected signals are amplified and denoised. Preprocessed signal data is then further fused with spectral features of an acoustic emission signal and a vibration signal, self-supervised health representation learning is performed to generate severely scarce negative samples, and classification of measured data collected from healthy and unhealthy blades is performed by means of auxiliary tasks, thereby implementing pseudo-label-based supervised learning. In the present invention, data fusion, EMD denoising, self-supervised learning and LSTM model training technology are used, such that blades of a wind turbine generator can be monitored in real time online, high-precision diagnosis can be performed on faults that have occurred, and inspection and fault type determination for evaluating the state of health of the current blade are performed, thereby meeting the requirements of a wind power market for online monitoring of wind turbine blades.

Description

Wind turbine blade health monitoring system and monitoring method Technical Field

[0001] The present invention relates to the field of wind turbine blade fault diagnosis, and in particular to a wind turbine blade health monitoring system and a monitoring method thereof based on the fusion of acoustic emission signals and vibration signals. Background Art

[0002] Wind energy is currently the most widely used renewable energy source, but wind turbines operate in harsh environments and are prone to frequent system failures. Wind turbine blades, as the core component of the turbine, account for 15%-20% of the total cost of the entire unit and are susceptible to damage from external shocks such as thermal stress, moisture absorption, and lightning strikes, resulting in high operating and maintenance costs. Currently, wind turbine generators, gearboxes, pitch control, and yaw mechanisms are equipped with remote online monitoring systems that provide real-time status monitoring. However, wind turbine blades are particularly susceptible to damage from lightning strikes and wind erosion, resulting in cracking, delamination, and other failures. Currently, remote, real-time status monitoring methods are incomplete, and operating the turbine without promptly identifying any issues can have disastrous consequences.

[0003] Currently, research on wind turbine blade fault diagnosis has yet to be fully developed. Existing early warning models primarily focus on modeling and analyzing a specific component using a specific monitoring method. These models lack high accuracy and generalization capabilities, and cannot achieve self-diagnosis and self-warning for the entire wind turbine blade, let alone a comprehensive assessment of wind turbine health. For example, a wind turbine blade fault diagnosis system based on acoustic emission (AE) technology detects damaged areas on the blades and then uses wavelet analysis to extract features from the blade AE ​​signals, enabling identification of different types of damage. However, relying solely on a single technology can lead to technical flaws that can result in insufficient detection accuracy and timeliness.

[0004] Summary of the Invention

[0005] In view of this situation, the present invention proposes a wind turbine blade health monitoring system and a monitoring method thereof to solve the above-mentioned problems existing in the existing technology and applications.

[0006] The present invention is achieved through the following technical solutions:

[0007] The wind turbine blade health monitoring system includes a vibration sensor: a mounting section is set up approximately 2-3 meters from the root of the wind turbine blade. Within this section, a vibration sensor is installed according to the positioning requirements of a dual-channel vibration sensor, approximately 20-30 cm from the trailing edge tip. The vibration sensor is installed perpendicular to the axis to ensure that axial and radial vibrations are collected. An acoustic emission sensor: a mounting section is set up at the center of the blade, and a leading edge and trailing edge acoustic emission sensor are installed within this mounting section. Two sensors should be installed on each blade, with the pressure-side sensor approximately 20-30 cm from the trailing edge tip, and the tension-side sensor approximately 20-30 cm from the leading edge tip. The two AE sensors are installed on the same cross-section. A data acquisition system is responsible for converting the vibration and acoustic emission signals collected by the sensors placed on various parts of the wind turbine blade into digital signals for subsequent analysis and processing. The outputs of the multiple vibration and acoustic emission signal sensors are connected to the inputs of the signal conditioning and amplification circuit. Then, through the multi-channel ADC sampling module, they are connected to the data acquisition hardware for calculation, control, and transmission of the digital signals.

[0008] The present invention also proposes a health monitoring method implemented by the above-mentioned wind turbine blade health monitoring system based on the fusion of acoustic emission signals and vibration signals, including extracting and collecting fault signals under multi-sensor monitoring, amplifying and denoising the collected signals; then fusing the spectral characteristics of the acoustic emission signal and the vibration signal with the pre-processed signal data, performing self-supervised health representation learning to generate severely scarce negative samples, and classifying the measurement data collected from healthy and unhealthy blades through auxiliary tasks to achieve supervised learning based on pseudo-labels. Beneficial effects

[0009] 1) Time domain and wavelet feature extraction technology based on empirical mode decomposition filtering and noise reduction improves the quality of blade acoustic emission signals. Machine learning-based acoustic emission blade defect recognition technology based on multi-dimensional feature fusion is a domestic first;

[0010] 2) The real-time blade defect monitoring system based on the fusion of vibration signals and acoustic emission signals is the first of its kind in China. It can monitor blade defects in real time, quantify the extent of faults, and detect early-stage faults.

[0011] 3) Comprehensive application and optimization comparison of multi-source data fusion technology at the data, feature, and decision ends;

[0012] 4) The spectrum characteristics and other characteristic statistics of the acoustic emission signal and the spectrum characteristics and other characteristic statistics of the vibration signal are combined into a characteristic vector, thereby efficiently completing the signal fusion processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Fig. 1 is a flow chart of the present invention;

[0014] FIG2 is a schematic diagram of the deployment positions of the acoustic emission signal sensor and the vibration signal sensor in the present invention;

[0015] FIG3 is a schematic cross-sectional view of the deployment positions of the acoustic emission signal sensor and the vibration signal sensor in the present invention;

[0016] FIG4 is a flow chart of a data acquisition system according to the present invention;

[0017] Figure 5 Flowchart of the hybrid multi-scale convolution combined with double-layer LSTM model in the present invention DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification.

[0019] The process of the wind turbine blade defect health monitoring method proposed in the present invention is shown in Figure 1: Figure 2 is a plan view schematic diagram of the deployment positions of the acoustic emission signal sensor and the vibration signal sensor, and Figure 3 is a cross-sectional schematic diagram of the deployment positions of the acoustic emission signal sensor and the vibration signal sensor.

[0020] Step 1: Installation of vibration sensor and acoustic emission sensor

[0021] Vibration Sensor 2 Installation Location: Create a mounting section approximately 2-3 meters from the 1 / 3 point of the wind turbine blade root. Within this section, install Vibration Sensor 2 according to the positioning requirements for dual-channel vibration sensors, approximately 20-30 cm from the trailing edge. Vibration Sensor 2 should be installed perpendicular to the axis to ensure that both axial and radial vibrations are detected. The connection method for the vibration sensor should be determined based on the sensor model and connection requirements to ensure stable signal transmission.

[0022] Acoustic Emission Sensor 1 Installation Location: A mounting section is located at the center of the blade (approximately one-third of the total length from the blade root). Within this section, two acoustic emission sensors (AE sensors) are installed, one for the leading edge and one for the trailing edge. Two sensors should be installed on each blade. The pressure-side sensor is approximately 20-30 cm from the trailing edge tip, and the tension-side sensor is approximately 20-30 cm from the leading edge tip. Special attention should be paid to the installation of both AE sensors on the same cross-section.

[0023] Step 2: Signal acquisition based on acoustic emission sensor and vibration sensor, as shown in Figure 4

[0024] In this embodiment, the data acquisition system integrates acoustic emission detection and vibration monitoring technologies. Using two different probes, integrated into the same channel number conditioning stage, the same data acquisition and processing card detects, collects, and processes the two different signals, significantly expanding the bandwidth of the monitored signals. The data acquisition system converts vibration and acoustic emission signals collected by sensors located on various parts of the wind turbine blades into digital signals for subsequent analysis and processing. The outputs of multiple vibration and acoustic emission sensors are connected to the inputs of a signal conditioning and amplification circuit. These signals are then connected to the data acquisition hardware via a multi-channel ADC sampling module for calculation, control, and transmission.

[0025] Step 3: noise reduction processing of acoustic emission signals and vibration signals;

[0026] Acoustic emission signal noise reduction: First, perform empirical mode decomposition (EMD). Then, calculate the correlation coefficient between each intrinsic mode function (IMF) and the original signal, remove the components with low correlation, reconstruct the remaining IMF components, and complete EMD filtering.

[0027] Specifically, you can follow the steps below:

[0028] 1) The original signal is adaptively decomposed into N IMF components through EMD.

[0029] 2) Calculate the correlation coefficient between each IMF and the original signal. corr = IMF i (txt)

[0030] 3) Eliminate the components whose correlation coefficients are less than the set threshold and reconstruct the remaining components.

[0031] Where N is the total number of IMFs; M is the number of IMF components that need to be reconstructed.

[0032] Subsequently, the time domain features such as acoustic emission duration, peak amplitude, center frequency, and energy are extracted. Through wavelet transform, the peak characteristic frequency of the sub-band after wavelet packet classification is extracted, and the energy distribution of the frequency band after wavelet decomposition is calculated. Finally, the above features are combined to obtain the wavelet packet feature combination into a feature matrix for the next step of blade fault discrimination model training.

[0033] For noise reduction of vibration signals, discrete wavelet transform (DWT) is used. DWT is a wavelet transform method that quickly processes discrete signals. It decomposes the signal into multiple wavelet subbands, each representing a different frequency range. We can select an appropriate wavelet basis based on the characteristics of the signal to achieve signal noise reduction. Specifically, we can follow the following steps:

[0034] 1. Decompose the collected vibration signal into multiple wavelet sub-bands. Use the tool function wavedec to build a DWT, thereby decomposing the signal into multiple wavelet sub-bands.

[0035] 2. Use the wavelet coefficients and their corresponding subbands to select an appropriate threshold for denoising. Use the wdencmp function for threshold denoising. The parameter "t" in the wdencmp function controls the threshold value.

[0036] 3. Finally, use the waverec function to reconstruct the processed signal to obtain the noise-reduced vibration signal.

[0037] Step 4: Preprocess the denoised data

[0038] The present invention provides a method for preprocessing a vibration signal and an acoustic emission signal, comprising the following steps:

[0039] 1. Collect vibration signals and acoustic emission signals after noise reduction.

[0040] 2. Perform short-time Fourier transform (STFT) on the vibration signal and acoustic emission signal respectively to obtain their time-frequency domain characteristics.

[0041] 3. Collect characteristic statistics of vibration signals, including power spectrum density, maximum amplitude, etc.

[0042] 4. Collect characteristic statistics of acoustic emission signals, including power spectrum density, peak value, etc.

[0043] 5. The spectrum characteristics and other characteristic statistics of the acoustic emission signal and the spectrum characteristics and other characteristic statistics of the vibration signal are combined into a characteristic vector.

[0044] Step 5: Perform self-supervised health representation learning to generate negative samples

[0045] In this step, we use two data augmentation strategies, signal component fusion and noise addition, from self-supervised health representation learning (SHRL) to generate negative samples. Specifically, we treat the pre-processed data (i.e., original training data) as positive samples and then generate negative samples of wind turbine blades with different health conditions through data augmentation to compensate for the scarcity of negative samples. Through data augmentation, unhealthy samples are generated for auxiliary tasks. This case employed a data augmentation strategy called component fusion, supervised by pseudo-labels. Specifically, the authors randomly replaced some components in the preprocessed training data to achieve data augmentation with a fusion ratio of 0 < η < 1. They also added Gaussian white noise with a standard deviation of E to generate falsely labeled negative samples of varying health states, assigning pseudo-labels to all samples based on their health state. This data augmentation scheme considers both differences and similarities between components, ensuring the rationality and representativeness of the augmentation of unhealthy samples, improving data diversity, and increasing robustness to adversarial attacks.

[0046] Step 6: Auxiliary tasks and pseudo-label supervision

[0047] The auxiliary task in the SHRL method is to classify the measurement data collected from healthy and unhealthy leaves, which is supervised learning based on pseudo labels. The training objective of this auxiliary task can be expressed as Eq. (5):

[0048] To express it, where y1 = [1,0] and y2 = [0,1] are one-hot vectors. In the SHRL method, supervised learning is performed by using the supervisory signal generated by the original data, and the high layer of the network is used to obtain an effective representation of the measurement to understand the health status of the blade. In the SHRL method, two sub-networks θ are connected in series. map and θ pro The mapping and projection tasks are implemented separately. The mapping network consists of multiple convolutional layers and activation functions. It obtains the representation of the original data in the latent space through nonlinear transformation and learns the transformation parameters θ based on pseudo-label supervised learning. map The projection head is a multi-layer perceptron with a depth of H, which is used to separate the effects of auxiliary and health monitoring tasks. Through the supervised learning of auxiliary tasks and pseudo labels of the SHRL method, we can effectively optimize Eq. (5) and obtain a representation suitable for conditional monitoring in a semi-supervised manner.

[0049] Step 7: Online monitoring and diagnosis of blade health

[0050] For online health monitoring of leaves, we use the representation patterns in the learned high-level mapping network and use kernel density estimation (KDE) to extract the distribution of health data representations. In this case, the kernel density estimator corresponding to the representation of healthy leaves is:

[0051] Where K is the kernel function. In the developed method, the Gaussian kernel function is adopted because of its convenient mathematical properties. The function is expressed as:

[0052] The bandwidth matrix N is determined based on the covariance matrix of the data representation, thereby obtaining the internal model of the healthy blade data representation. N represents the difference between the representation distributions of the healthy sample and the monitoring sample, revealing the health status of the blade based on the learned representation pattern. Then, the running measurement values ​​are collected and the negative log likelihood is calculated to evaluate the difference with the healthy data model, so as to understand the health status of the blade. The negative likelihood metric is usually normalized with the help of the fault index (FI) to improve the interpretability of the monitoring results and judge the health status of the blade:

[0053] The value of FI is set in the range of [0,1]. The larger the FI is, the higher the possibility of wind blade failure is.

[0054] Step 8: Identify the fault type of the blade defect

[0055] We use the trained model to diagnose blade defects. Specifically, we determine whether a blade fault has occurred based on the blade health assessment results. If a blade fault is detected, we further analyze the blade's operating data and, combined with prior knowledge and experience, determine the fault type.

[0056] In this embodiment, the fault type is preliminarily determined based on the spectral characteristics of the vibration signal and the acoustic emission signal in combination with prior knowledge and experience.

[0057] First, data preparation is performed to train the model, dividing the original data into a training set and a test set. The training set is derived from the fused signal data after noise reduction and preprocessing in step 4, ensuring real-world application. The test set consists of signal data of different fault types collected in real-world scenarios and public datasets. Normalization is performed on each data point to avoid the vanishing gradient problem. Finally, the corresponding signal fault type features are obtained.

[0058] This example proposes a hybrid multi-scale convolution combined with a two-layer LSTM model to classify fault types. The hybrid multi-scale convolution combined with a two-layer long short-term memory network model is shown in Figure 5. It mainly includes a hybrid multi-scale composed of three multi-scale convolutions, two long short-term memory neural networks, and one fully connected layer.

[0059] The fused data from step 4 is used as the dataset for training a hybrid multi-scale convolutional model combined with a two-layer LSTM. Once the dataset is prepared, the model hyperparameters are optimized, including experimenting with different learning rates to find the optimal one. Early stopping is used during model training, stopping training when the validation set loss function begins to increase to avoid overfitting and find the optimal number of iterations.

[0060] This design uses cross entropy loss as the loss function to quantify the gap between the predicted value and the true value of the target. The loss function for a single word or a single segment in time series data is as follows:

[0061] Model output is the probability value, and y (t) is a series of fixed values ​​(labels), so y (t) =1.

[0062] The average cross entropy loss is:

[0063] The model is trained using a prepared dataset. Because the fault types in this paper are numerous and mutually exclusive, representing a multi-classification task, the softmax activation function of the multi-classification logistic regression model is used to convert the model output into class probabilities. The softmax function accepts a vector as input and returns a vector with the same dimension as the input vector, where each element is a probability value between 0 and 1. The function expression is as follows:

[0064] in is the model output for a certain fault type, the denominator The model output set for all fault types is used to obtain the probability P of the current fault type. The purpose of using the data set to train the neural network model is to approximate the fault type with the highest probability θ T Finally, after the training is completed, the test set is used to evaluate the performance of the model.

[0065] Ultimately, we can combine prior knowledge and experience to make a preliminary judgment, use the trained hybrid multi-scale convolution combined with a two-layer LSTM model to classify the new input data, and determine the type of fault caused by the wind blade based on the final probability of various types of faults.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wind turbine blade health monitoring system, characterized in that: include Vibration sensor: Set up an installation section at a position 2-4m away from the root of the wind turbine blade, 1 / 3 of the way to the tip. In this section, install the vibration sensor according to the position requirements of the dual-channel vibration sensor, about 20-30cm away from the trailing edge tip. The vibration sensor is installed in a direction perpendicular to the axis to ensure that the axial and radial vibrations are collected; Acoustic emission sensor: an installation section is set at the center of the blade, and an acoustic emission sensor for the leading edge and the trailing edge is installed in the installation section. Two sensors are installed on each blade, of which the pressure side sensor is 20-30cm away from the trailing edge tip, and the tension side sensor is 20-30cm away from the leading edge tip. The two acoustic emission sensors are installed on the same cross section; Data acquisition system: responsible for converting the vibration signals and acoustic emission signals collected by sensors placed on various parts of the wind turbine blades into digital signals for subsequent analysis and processing; The outputs of multiple vibration signals and acoustic emission signal sensors are connected to the input of the signal conditioning and amplification circuit, and are connected to the data acquisition hardware via the multi-channel ADC sampling module to calculate, control and transmit the digital signals.

2. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 1 is characterized in that: It includes extracting and collecting fault signals under multi-sensor monitoring, amplifying and denoising the collected signals; then fusing the spectral characteristics of acoustic emission signals and vibration signals with the pre-processed signal data, performing self-supervised health representation learning to generate scarce negative samples, and classifying the measurement data collected from healthy and unhealthy blades through auxiliary tasks to achieve supervised learning based on pseudo-labels.

3. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 2 is characterized in that: The noise reduction processing includes performing noise reduction processing on the acoustic emission signal and the vibration signal, The AE signal denoising includes first performing empirical mode decomposition, then calculating the correlation coefficient between each intrinsic mode component and the original signal, removing the components with low correlation, reconstructing the remaining IMF components, and completing EMD filtering. Specifically, follow these steps: 1) The original signal is adaptively decomposed into N IMF components through EMD; 2) Calculate the correlation coefficient between each IMF and the original signal corr=IMF i (t)*x(t) 3) Eliminate the components whose correlation coefficient is less than the set threshold and reconstruct the remaining components Where N is the total number of IMFs; M is the number of IMF components that need to be reconstructed; Subsequently, the acoustic emission duration, peak amplitude, center frequency, and energy time domain characteristics are extracted. Through wavelet transform, the peak characteristic frequency of the sub-band after wavelet packet classification is extracted, and the energy distribution of the frequency band after wavelet decomposition is calculated. Finally, the above features are combined to obtain the wavelet packet features into a feature matrix for the next step of blade fault identification model training.

4. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 3 is characterized in that: The vibration signal denoising process uses discrete wavelet transform, specifically, the following steps are performed: The collected vibration signal is decomposed into multiple wavelet sub-bands, and the DWT is established using the tool function "wavedec" to decompose the signal into multiple wavelet sub-bands; a suitable threshold is selected for denoising using the wavelet coefficients and their corresponding sub-bands, and the "wdencmp" function is used for threshold denoising; finally, the processed signal is reconstructed using the "waverec" function to obtain the denoised vibration signal.

5. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 4 is characterized in that: The denoised data is preprocessed, including the following steps: 1) Collect the vibration signal and acoustic emission signal after noise reduction; 2) Perform short-time Fourier transform on the vibration signal and acoustic emission signal respectively to obtain their time-frequency domain characteristics; 3) Collect characteristic statistics of vibration signals, including power spectrum density and maximum amplitude; 4) Collect characteristic statistics of acoustic emission signals, including power spectrum density and peak value; 5) The frequency spectrum characteristics and other characteristic statistics of the acoustic emission signal and the frequency spectrum characteristics and other characteristic statistics of the vibration signal are combined into a characteristic vector.

6. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 2 is characterized in that: The self-supervised health representation learning to generate negative samples includes: using two data enhancement strategies of signal component fusion and noise addition in the self-supervised health representation learning to generate negative samples, treating the pre-processed original training data as positive samples, and then generating negative samples of wind turbine blades with different health conditions by component fusion and noise addition, so as to compensate for the scarcity of negative samples; Generate unhealthy samples for auxiliary tasks through data augmentation Component fusion is performed and pseudo labels are used for supervision. Specifically, some components are randomly replaced in the preprocessed training data to achieve data enhancement with a fusion ratio of 0<η<1. At the same time, Gaussian white noise with a standard deviation of E is added to generate false label negative samples of different health states, and pseudo labels are assigned to all samples according to the health status.

7. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 2 is characterized in that: The auxiliary task and the supervision of pseudo labels, wherein the auxiliary task is to classify the measurement data collected from healthy and unhealthy leaves, thereby realizing supervised learning based on pseudo labels. The training objective of this auxiliary task can be expressed by formula (5): To express, where y1 = [1,0], y2 = [0,1] are one-hot vectors.

8. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 2, characterized in that: It also includes online monitoring and diagnosis of blade health status: For the online health monitoring of the blade, the representation mode in the learned high-level mapping network is used, and the distribution of health data representation is extracted using kernel density estimation. The kernel density estimator corresponding to the representation of healthy blades is: Where K is the kernel function; in the developed method, the Gaussian kernel function is adopted because of its convenient mathematical properties, which is expressed as: The bandwidth matrix N is determined according to the covariance matrix of the data representation, so as to obtain the internal model of the healthy blade data representation. N represents the difference between the representation distributions of the healthy sample and the monitoring sample. The health status of the blade is revealed based on the learned representation mode. Then, the running measurements are collected and the negative log likelihood is calculated to evaluate the difference with the healthy data model. In this way, the health status of the blade can be understood. The negative likelihood metric is usually normalized with the help of the fault index (FI) to improve the interpretability of the monitoring results and judge the health status of the blade: The value of FI is set in the range of [0,1]. The larger the FI is, the higher the possibility of fan blade failure is.

9. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 2, characterized in that: It also includes fault type identification of blade defects: based on the spectral characteristics of vibration signals and acoustic emission signals, combined with prior knowledge and experience, a preliminary judgment of the fault type is made. Specifically, the first is the data preparation process, in which the model is trained to divide the original data into a training set and a test set. The training set comes from the fused signal data after noise reduction and preprocessing to achieve the effect of being applied to real scenarios. The test set is signal data of different fault types collected in actual scenarios. Each data point is normalized to avoid the gradient vanishing problem, and finally the corresponding signal fault type characteristics are obtained.

10. The health monitoring method implemented by the wind turbine blade health monitoring system according to claim 9, characterized in that: The fault classification adopts a hybrid multi-scale convolution combined with a double-layer LSTM model to classify fault types. The hybrid multi-scale convolution combined with a double-layer long short-term memory network model mainly includes a hybrid multi-scale composed of 3 multi-scale convolutions and 2 long short-term memory neural networks and 1 fully connected layer.

Images