A wind turbine blade fault detection method based on aeroacoustics and deep learning

Abstract

The application discloses a wind turbine blade fault detection method based on aeroacoustics and deep learning, and relates to the technical field of wind turbine blade damage detection. The detection method comprises the following steps: acquiring aeroacoustic data of a to-be-detected wind turbine blade under a healthy state and a damaged state; performing zero-mean and filtering processing on the aeroacoustic data based on aeroacoustic theory; calculating power spectral density and extracting power spectral density under a characteristic frequency index based on the aeroacoustic data subjected to the zero-mean and filtering processing, and establishing a BP neural network detection model based on power spectral density characteristics; inputting the to-be-detected data into the BP neural network detection model based on power spectral density characteristics after feature extraction, and calculating based on the BP neural network detection model; and establishing an evaluation system based on an output result to judge the state of the blade. The application avoids overfitting of the detection model, and improves the accuracy of blade damage identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine blade damage detection, and more particularly to a wind turbine blade fault detection method based on aeroacoustics and deep learning. BACKGROUND

[0002] As a key component of wind turbine operation, the blade often operates in complex natural environments such as strong wind, lightning, icing, and rain erosion, and is prone to damage. Due to the lack of effective real-time blade damage detection means, the traditional manual inspection method cannot timely detect blade damage, and such damage, if not timely detected and repaired, will lead to reduced power generation efficiency of the wind turbine, shutdown, and even accidents such as damage, greatly increasing the operation and maintenance cost of the wind turbine. Therefore, wind turbine blade damage detection is of great significance to ensure safe and efficient operation of the wind turbine and reduce the power generation cost of the wind turbine during its life cycle.

[0003] At present, the existing wind turbine damaged blade diagnosis methods mainly include:

[0004] (1) Strain monitoring: when using a resistance strain gauge to monitor wind turbine blade damage, the wiring is complex, the installation points are numerous, and it is difficult to monitor the damage in the internal structure of the wind turbine blade. When using an optical fiber strain gauge, the optical fiber sensor needs to be pre-embedded in the wind turbine blade, which increases the difficulty of manufacturing the wind turbine blade, and in addition, the special structure of the wind turbine blade also affects the detection accuracy of the optical fiber sensor;

[0005] (2) Vibration monitoring: the size of the wind turbine blade is large, and a large number of sensors are needed to realize global detection, which will increase the monitoring cost, and vibration monitoring is not sensitive to early minor defects of the wind turbine blade;

[0006] (3) Machine vision detection: the damage information of the wind turbine blade is easily disturbed by background information, and due to the lack of effective means to extract the blade damage features, the detection model input too many feature parameters, which leads to overfitting of the detection model, reduces the detection accuracy, and increases the detection time.

[0007] Therefore, it is necessary to develop a real-time and efficient wind turbine damaged blade diagnosis method to monitor the blade state in real time and reduce the operation and maintenance cost. SUMMARY

[0008] To solve the above problems, the present application provides a wind turbine blade fault detection method based on aeroacoustics and deep learning.

[0009] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0010] A wind turbine blade fault detection method based on aeroacoustics and deep learning, comprising the following steps:

[0011] Step 1: Obtain the aerodynamic acoustic data of the wind turbine blade under the healthy state and the damaged state; the aerodynamic acoustic data contains real-time sound pressure data X(j) of the wind turbine from start-up to rated operating conditions;

[0012] Step 2: Based on the aerodynamic acoustic theory, the aerodynamic acoustic data is zero-mean and filtered;

[0013] Step 3: Based on the aerodynamic acoustic data preprocessed in step 2, the power spectral density is calculated and the power spectral density under the characteristic frequency index is extracted, and a BP neural network detection model based on the power spectral density characteristics is established;

[0014] Step 4: After feature extraction, the test data is input into the BP neural network detection model based on the power spectral density characteristics for calculation, and an evaluation system is established based on the output results to judge the state of the blade.

[0015] Further, in step 1, the sampling time of the aerodynamic acoustic data t needs to meet:

[0016] ; wherein, F s is the sampling frequency of the acoustic sensor, is the frequency resolution requirement.

[0017] Further, step 2 specifically includes the following steps:

[0018] Step 2-1: Zero-mean processing:

[0019] ;

[0020] ; wherein, represents the i-th sound pressure value in the sound pressure data of any group of wind turbines to be tested, represents the i-th sound pressure value after mean processing; N is the number of current processing data;

[0021] Step 2-2: According to the aerodynamic noise frequency band [a, b] radiated by the wind turbine blade, band-pass filtering is performed,

[0022] ;

[0023] ;

[0024] ;

[0025] ; wherein, represents the frequency spectrum of the filtered data; represents the transfer function of the band-pass filter; X(f)a spectrum representing a time-domain acoustic pressure signal to be processed; filtered time-domain data; inverse Fourier transform.

[0026] Further, the step 3 specifically comprises the following steps:

[0027] Step 3-1: divide the aerodynamic acoustic data after zero-mean and filtering into a test set and a verification set;

[0028] Step 3-2: divide the acoustic pressure data X(j) in the test set into K segments, each with a length of L data points, and calculate the Fourier transform result of the data in the kth segment at the frequency index K n :

[0029] ;

[0030] wherein X(k, j) represents the jth acoustic pressure data in the kth segment, here k = 1, 2, …, K, j = 1, 2, …, L, and X(k, j) = X(j + (k-1)L), and K j j is the time-domain sample index, w(j) is a window function, and W(j) represents a base function for converting a time-domain acoustic pressure signal into a frequency-domain signal;

[0031] then the power spectral density (PSD) of the signal in the kth segment at the frequency K f n is:

[0032] ;

[0033] wherein U is the average energy of the window function, used for compensation in the power spectral density, and is calculated by the following formula:

[0034] ;

[0035] obtain the estimation of the power spectral density at the frequency f n :

[0036] ;

[0037] Step 3-3: calculate the power spectral density difference between the healthy blade and the damaged blade:

[0038] ;

[0039] wherein represents the power spectral density difference between the healthy blade and the damaged blade, ​​​​​​​Power spectral density value representing healthy blades, Power spectral density value representing damaged blades;

[0040] Step 3-4: Peak value retrieval is performed with the mean value of the power spectral density difference value as the threshold, and peak data greater than the mean value is retained to obtain the power spectral density characteristic peak value of each working condition Frequency index ;

[0041] Step 3-5: The number of occurrences of each working condition frequency index is counted, and the mean value of the number of occurrences is calculated. The frequency index greater than the mean value of the number of occurrences is retained and n 1 corresponding power spectral density characteristic value ;

[0042] Step 3-6: The n 1 power spectral density characteristic values are normalized, and the n 1 PSD characteristic values are taken as input, the output is set to 2, the activation function is selected as the sigmod function, the training method is selected as the adaptive momentum gradient descent method, the hidden layer neurons are selected as the average of the sum of the input and output neurons, and the target error is set to ε;

[0043] Step 3-7: The neural network output result [1, 1] represents a healthy blade, and the neural network output result [0, 1] represents a damaged blade. A BP neural network detection model is established and trained. The number of iterations is adjusted, and when the error of the model reaches the set error ε, the model iteration stops, and the trained model function is derived;

[0044] Step 3-8: The data in the test set is used to verify the BP neural network detection model. When the output result of the detection model is within the error allowable range, it is proved that the model is effective, the detection function is derived, otherwise the training model parameters are returned.

[0045] Further, the step 4 specifically comprises the following steps:

[0046] Step 4-1: The time domain data y(i) filtered by step 2 is subjected to power spectral density calculation according to step 3-2, and the power spectral density index characteristic value is extracted and normalized, and then input into the detection function obtained in step 3-8 to obtain a two-digit output result [U, V];

[0047] Step 4-2: The output result [U, V] is evaluated by setting a threshold α, and the state of the blade is judged.

[0048] Further, in the output result [U, V], the second bit value close to 1 indicates that the data conforms to the aerodynamic noise characteristics, and the closer the first bit value is to 0, the higher the probability of blade failure. The first bit output is taken as 1-U, and the two digits are reserved to output in percentage form, that is, the probability of failure. Finally, a reasonable threshold is set to determine whether to issue a warning message.

[0049] In the above scheme, aiming at the focus problem that the wind turbine blade failure is difficult to be accurately detected, based on the wind turbine aerodynamic acoustics and deep learning theory, the filter technology based on aerodynamic noise is used to reduce the data dimension and speed up the operation time of the model; the characteristic values of the damaged and healthy blades are obtained through the power spectral density calculation and peak extraction technology to train the BP neural network model, so as to avoid the overfitting of the detection model and improve the recognition accuracy, thereby providing a new technical approach for the real-time monitoring technology of the wind turbine blade. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 FIG. 1 is a general flowchart of a wind turbine blade failure detection method based on aerodynamic acoustics and deep learning in an embodiment of the present application;

[0051] Figure 2 FIG. 2 is a wind turbine blade aerodynamic sound power spectral density feature extraction flowchart in an embodiment of the present application;

[0052] Figure 3 FIG. 3 is a BP neural network detection model construction flowchart based on the power spectral density characteristics in an embodiment of the present application. DETAILED DESCRIPTION

[0053] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0054] In this embodiment, the 450W wind turbine of SD2030 airfoil is taken as a typical analysis object, the sound pressure data of the wind turbine under the blade damage and health state is collected, the blade failure diagnosis model is established, and a wind turbine blade failure detection method based on aerodynamic acoustic signal characteristics is specifically described. As shown in FIG. 1, it includes the following steps: Figure 1

[0055] Step 1, artificially set the common damage form of the wind turbine blade, arrange the acoustic sensors according to the far-field test standard, collect the aerodynamic sound signals of the healthy blade and the damaged blade of the same type of wind turbine under different working conditions for multiple times, and the collected data contains the real-time sound pressure data x(j) of the wind turbine from start to rated working condition;​

[0056] Step 2, the frequency band of the turbulent boundary layer trailing edge aerodynamic noise of the 450W small wind turbine with SD2030 airfoil is about 1000-2000Hz, according to which the collected acoustic signal is processed by zero mean value and 1000-2000Hz band-pass filtering; Specifically:

[0057] Step 2-1: zero mean value processing:

[0058] ;

[0059] In the formula, represents the i-th sound pressure value in the sound pressure data of any group of wind turbines to be measured, represents the i-th sound pressure value after mean value processing.

[0060] Step 2-2: 1000-2000Hz band-pass filtering processing;

[0061] ;

[0062] ;

[0063] ;

[0064] In the formula, represents the frequency spectrum of the filtered data; represents the transfer function of the band-pass filter; X(f) represents the frequency spectrum of the time-domain sound pressure signal to be processed; represents the filtered time-domain data; represents the inverse Fourier transform.

[0065] Step 3: see Figure 3 , based on the acoustic data of the blade health and damage, calculate the power spectral density and extract the power spectral density under the characteristic frequency index, and establish a BP neural network detection model based on the power spectral density characteristics, that is, the blade detection model; Specifically:

[0066] Step 3-1 divides the collected sound pressure data x(j) into training set and test set two parts, wherein the test set contains a group of acoustic data of healthy blades and fault blades under different working conditions, and the remaining data is composed of the training set;

[0067] Step 3-2: divide the sound pressure data X(j) in the test set into K segments, each segment has L data points, calculate the Fourier transform and the power spectral density (PSD) at the frequency index n; Wherein:

[0068] ;

[0069] wherein, represents the K th sound pressure data in the j th segment, wherein j is the time domain sample index, W(j) is the window function, represents the base function for converting the time domain sound pressure signal into the frequency domain signal; the power spectral density (PSD) of the signal in the K th segment at the frequency f n is:

[0070]

[0071] wherein U is the average energy of the window function, used for compensation in the power spectral density; which is calculated by:

[0072]

[0073] obtain the estimation of the power spectral density at the frequency f n :

[0074]

[0075] Step 3-3: Based on the obtained power spectral density result, calculate the power spectral density difference between the healthy blade and the damaged blade in each working condition training set:

[0076]

[0077] wherein, represents the power spectral density difference between the healthy blade and the damaged blade, represents the power spectral density value of the healthy blade, represents the power spectral density value of the damaged blade;

[0078] Step 3-4: Calculate the mean value of the peak value of the power spectral density difference, and use the mean value of the power spectral density difference as the threshold for peak value retrieval, retain the peak value data greater than the average value, and obtain the power spectral density characteristic peak value and the frequency index

[0079] Step 3-5: Calculate the frequency index corresponding to the power spectral density difference peak value in the training set, and calculate the mean value of the frequency index. Use the mean value as the threshold, and retain the frequency index greater than the mean value of the frequency index , according to the retained frequency index obtain the main power spectral density difference peak value, and obtain n 1 ​​​​​​a corresponding power spectral density characteristic value ;

[0080] Step 3-6: normalize the power spectral density characteristic value n 1 Step 3-6: normalize the power spectral density characteristic value n 1 The number of input neurons is set to 2, the training function is selected as traingdx, the training method is selected as adaptive momentum gradient descent method, the number of hidden layer neurons is selected as the average of the sum of input and output neurons, the target error is set to 0.001, the number of hidden layer neurons is calculated according to the number of input and output neurons, and the BP neural network training model is obtained;

[0081] Step 3-7: the neural network output result [1,1] represents a healthy blade, and the neural network output result [0,1] represents a fault blade. The peak data set of the main power spectral density difference is normalized and input into the BP neural network for iterative calculation. The number of iterations is determined according to the fitting degree of the training error and the target error, until the error of the training model is equal to the set target error, and the detection model for judging the health state of the blade is obtained;

[0082] Step 3-8: verify the detection model using the data in the test set. When the output result of the detection model is within the error allowable range, it proves that the model is effective, and the detection function is derived. Otherwise, return to adjust the training model parameters; Specifically: extract the power spectral density characteristic value under the frequency index of the test set data and normalize the processing. Input the power spectral density characteristic value of each verification data into the blade health state detection model, and the output result is within the set error allowable range, which means that the detection model is effective, and the detection function is derived. Otherwise, return to adjust the training model parameters;

[0083] Step 4: input the data to be measured after feature extraction into the blade detection model based on BP neural network for calculation, and establish an evaluation system based on the output result to judge the state of the blade; Specifically:

[0084] Step 4-1: calculate the power spectral density of the time domain data y(i) after pre-processing and filtering according to step 2 according to step 3-2, and obtain the power spectral density index extract the power spectral density characteristic value, normalize it and input it into the detection function obtained in step 3-8 to obtain a two-digit output result [U, V];

[0085] Step 4-2, the state of the blade is judged by setting a threshold value a and evaluating the output result [U, V] according to the results in the following table. In the two-digit number of the detection model output, the second digit value close to 1 indicates that the data conforms to the aerodynamic noise characteristics, and the closer the first digit value is to 0, the higher the probability of blade failure. The first output is taken as 1-U, which is reserved as two digits and output in percentage form, that is, the probability of failure. Finally, a reasonable threshold value is set to determine whether to issue a warning message.

[0086] The wind turbine blade fault detection method based on aerodynamic acoustic signal characteristics provided by the present application filters the acoustic data of the wind turbine to be measured based on aerodynamic acoustic theory, compares and analyzes the aerodynamic sound power spectrum density characteristics of the damaged and healthy wind turbine to establish a blade state feature extraction method, inputs the blade state features into the BP neural network for training with the help of deep learning theory, obtains a detection model, verifies the detection model using a verification data set, inputs the acoustic characteristics of the wind turbine to be measured into the detection model, and finally judges the health state of the wind turbine blade by combining the evaluation system with the output results of the model. The method can be applied to fault identification of various types of wind turbine blades, has fast detection timeliness and high accuracy, so as to early detect the damage of the wind turbine blade, reduce the safety accidents of the wind turbine, and reduce the operation and maintenance cost of the wind farm.

[0087] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various modifications or changes within the scope of the claims, which does not affect the essential content of the present application.

Claims

1. A wind turbine blade fault detection method based on aeroacoustics and deep learning, characterized in that, Comprising the following steps: Step 1: obtaining aerodynamic acoustic data of the wind turbine blade to be tested under healthy state and damaged state; Step 2: zero-mean and filtering processing of the aerodynamic acoustic data based on aerodynamic acoustic theory; Step 3: calculating power spectral density and extracting power spectral density under the index of characteristic frequency based on the aerodynamic acoustic data preprocessed in step 2, and establishing a BP neural network detection model based on power spectral density characteristics; Step 3-1: dividing the aerodynamic acoustic data after zero-mean and filtering processing into a test set and a verification set; Step 3-2: Divide the sound pressure data X(j) in the test set into K segments, each of length L data points, and calculate the Fourier transform result of the segment data at the frequency index K n ; ;​​ In the formula, Indicates the first K The first in the paragraph j Each sound pressure level data point, here j For time-domain sample index, W(j) For window functions, Let the basis functions be used to convert the time-domain sound pressure signal into a frequency-domain signal; then the first... K Segment signal at frequency f n Power spectral density at for: ； where U is the average energy of the window function used to compensate in the power spectral density; obtaining the power spectral density at the frequency f n n ; Step 3-3: Calculate the power spectral density difference between healthy and broken blades: ; wherein represents the difference in power spectral density between healthy and damaged blades, represents the power spectral density value of healthy blades, represents the power spectral density value of damaged blades; Step 3-4: Peak searching with the mean of the power spectral density difference as the threshold, and the power spectral density characteristic peak value of each working condition is obtained with frequency index ; Step 3-5: Count the number of times each frequency index appears and take the average of the number of times, use the average as the threshold, keep the frequency index greater than the average number of times and n 1 corresponding power spectral density characteristic value ; Step 3-6: the n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized, so that n 1 characteristic values of the power spectral density are normalized Step 3-7: representing a healthy blade by neural network output result [1, 1] and a damaged blade by neural network output result [0, 1], establishing and training the BP neural network detection model, adjusting the number of iterations, stopping the model iteration when the error of the model reaches the set error epsilon, and deriving the trained model function; Step 3-8: verifying the BP neural network detection model using data in the verification set, proving the model effective when the output result of the detection model is within the error allowable range, deriving the detection function, otherwise returning to adjust the training model parameters; Step 4: inputting the data to be tested after feature extraction into the BP neural network detection model based on power spectral density characteristics for calculation, and establishing an evaluation system based on the output result to judge the state of the blade.

2. The wind turbine blade fault detection method based on aeroacoustics and deep learning according to claim 1, wherein, The sampling time of the aeroacoustic data in step 1 t The following needs to be satisfied: ; where is the frequency resolution requirement.

3. The method of wind turbine blade fault detection based on aeroacoustics and deep learning as claimed in claim 1, wherein, The step 2 specifically comprises the following steps: Step 2-1: zero-mean processing: ； In the formula, represents the i-th sound pressure value in the sound pressure data of any one group of wind turbines to be measured, represents the i-th sound pressure value after mean value processing; N is the number of current processing data; Step 2-2: band-pass filtering according to the aerodynamic noise frequency band [a, b] radiated by the wind turbine blade, ； ； ； wherein represents the filtered data spectrum; represents the transfer function of the band-pass filter; X(f) represents the spectrum of the time-domain sound pressure signal to be processed; represents the filtered time-domain data; represents the inverse Fourier transform.

4. The method of wind turbine blade fault detection based on aeroacoustics and deep learning as claimed in claim 1, wherein, The step 4 specifically comprises the following steps: Step 4-1: The time domain data y(i) filtered in step 2 is calculated according to step 3-2 to obtain the power spectrum density index The power spectrum density characteristic value is extracted, normalized, and introduced into the detection function obtained in step 3-8 to obtain a two-digit output result [U, V]. Step 4-2: evaluating the output result [U, V] by setting a threshold alpha, and judging the state of the blade.

5. The method of wind turbine blade fault detection based on aeroacoustics and deep learning as claimed in claim 4, wherein, In the output result [U, V], the second value close to 1 indicates that the data conforms to the aerodynamic noise characteristics, and the first value closer to 0 indicates a higher probability of blade failure. Taking 1-U of the first output and retaining two digits, the output is in the form of percentage, which is the probability of failure. Finally, a reasonable threshold is set to determine whether to issue a warning message.

Images