CN110792563B

CN110792563B - Wind turbine generator blade fault audio monitoring method based on convolution generation countermeasure network

Info

Publication number: CN110792563B
Application number: CN201911067884.1A
Authority: CN
Inventors: 王旻轩; 鲍亭文; 金超; 晋文静
Original assignee: Beijing Cyberinsight Technology Co ltd
Current assignee: Beijing Cyberinsight Technology Co ltd
Priority date: 2019-11-04
Filing date: 2019-11-04
Publication date: 2020-09-15
Anticipated expiration: 2039-11-04
Also published as: CN110792563A

Abstract

The application relates to a wind turbine generator blade fault audio monitoring method based on convolution generation countermeasure network, which comprises the steps of dividing blade rotating audio signals collected by sound collection equipment, using principal component analysis to each group of blade power spectrum matrixes to construct and generate the countermeasure network, using deep convolution to generate the countermeasure network structure, and using anoGAN to train the characteristic square matrixes of the blades in turn; after a group of blade data is trained to generate a confrontation model, respectively inputting the feature matrixes of the remaining group of blades into an anomaly detector, taking the input data to be tested as a first part of model output, carrying out iterative training again, calculating the fitting degree of the input data to the original training data, and taking the last model loss as a fitting error; and traversing and calculating errors to obtain the difference size of each group of blades relative to the rest groups of blades. The method and the device are high in efficiency, low in cost, free of manual operation and suitable for fault monitoring of the wind turbine generator.

Description

Wind turbine generator blade fault audio monitoring method based on convolution generation countermeasure network

Technical Field

The application relates to a wind turbine generator blade fault audio monitoring method based on a convolution generation countermeasure network, which is applicable to the technical field of wind turbine generator fault monitoring.

Background

When the wind turbine generator runs in an alternating load environment for a long time, the blades are easy to crack, corrode and other faults, the service life of the blades is shortened, and the capture of wind energy by the fan is influenced, so that early diagnosis and early warning of the blade faults are of great significance to safe and healthy operation of the fan. The current methods for monitoring the faults of the fan blade comprise image recognition, thermal imaging, laser detection, vibration signal characteristic recognition and the like.

The chinese patent with application number 201910603546.9 proposes an audio signal detection method based on the effect of profile coefficient optimization K-means clustering, which distinguishes a fault frame and a non-fault frame of a section of audio signal, plots a bar graph on a time domain with a sample label of a feature frame, and comprehensively analyzes the periodic characteristics of similar feature frames to judge whether a blade of a certain fan is faulty. The method judges the fault by judging whether the bar graph of the clustered class labels on the time domain has periodic characteristics. The two-clustering method is difficult to accurately detect faults with weak performance characteristics in audio, and is difficult to realize automatic high-efficiency detection because whether periodic characteristics exist or not needs manual judgment.

Chinese patent application No. CN201710641430.5 also proposes a method for diagnosing faults by audio frequency using a sound collection device, which mainly intercepts characteristic frequencies and compares the characteristic frequencies to diagnose faults, and this method is similar to fault diagnosis based on vibration signals, but the difference of analysis modes is determined by the characteristics that audio frequency is easily interfered by noise due to its different propagation modes from vibration signals. If the audio signal is simply subjected to frequency domain analysis, the detection accuracy of the traditional vibration signal by using the same method is difficult to achieve, and the periodic characteristics accompanying the blade wind sweeping are lost.

In the prior art, a method for monitoring the fault of the blade of the wind turbine generator, which has high efficiency and low cost and does not need manual operation, is urgently needed.

Disclosure of Invention

The invention aims to provide a wind turbine generator blade fault monitoring method which is high in efficiency, low in cost and free of manual operation.

The wind turbine generator blade fault audio monitoring method based on the convolution generation countermeasure network comprises the following steps:

(1) dividing the audio signals of the blade rotation collected by the sound collection equipment to obtain the starting and stopping time points of the wind sweeping sound of each blade of the fan;

(2) performing principal component analysis on each group of blade power spectrum matrixes, converting a single group of blade signal matrixes into characteristics using principal components to display data, and reducing the dimension of each original section of wind sweeping signal to a square matrix with the same row number and column number;

(3) establishing and generating a confrontation network, generating a confrontation network structure by using deep convolution, training the characteristic square matrix of the blade by using anoGAN in turn, fixedly generating parameter structures of a model and a discrimination model for the characteristic square matrix of each group of blades through repeated iteration confrontation training, and extracting the last layer of convolution layer of the discriminator to be used as a characteristic extractor;

(4) reconstructing a model, taking noise distribution of an implicit space as model input, and combining the output of a generator and the output of a feature extractor into a model output to be used as an abnormal detector;

(5) after a group of blade data is trained to generate a confrontation model, respectively inputting the feature matrixes of the remaining group of blades into an anomaly detector, taking the input data to be tested as a first part of model output, carrying out iterative training again, calculating the fitting degree of the input data to the original training data, and taking the last model loss as a fitting error;

(6) and traversing and calculating errors to obtain the difference size of each group of blades relative to the rest groups of blades, and taking the maximum value of the obtained difference as a final abnormal risk value.

Preferably, step (2) may further include the following steps:

(2.1) carrying out short-time Fourier transform on the audio signal to obtain a converted time-frequency spectrum matrix;

(2.2) calculating the corresponding relation with the power spectrum matrix width according to the sampling frequency and duration obtained when the audio is read in, converting time division points into division indexes of matrix columns to obtain the power spectrum of each section of wind sweeping signals of each blade, and splicing the power spectrum into the total power spectrum of each blade in the time section;

(2.3) unifying the time-frequency domain matrix of each group of blades in the time dimension, and extracting the principal component of the power matrix of each blade in the time domain;

and (2.4) converting a single-group blade signal matrix representing time-frequency-energy information into a characteristic using principal components to display data by using principal component analysis on each group of blade power spectrum matrix, and reducing the dimension of each original section of wind sweeping signal to a square matrix with the same row number and column number.

Preferably, the difference in step (6) is calculated by performing a relative anomaly score calculation to obtain a difference representing the difference between one group of blades and the remaining group of blades, and the maximum value of the obtained relative anomaly scores of the groups is used as a final anomaly risk value.

Preferably, the wind turbine generator has three blades, namely a blade a, a blade b and a blade c, and in the step (6), the method for performing error calculation in a traversing manner comprises the following steps: training by using the data of the blade a, and respectively inputting the blade b and the blade c to obtain a fitting error L_a，b，L_a，c(ii) a Training by using the data of the blade b, and respectively inputting the blade a and the blade c to obtain a fitting error L_b，a，L_b，c(ii) a Training by using blade c data, and respectively inputting blade a and blade b to obtain fitting error L_c，a，L_c，b。

According to the wind turbine generator blade fault audio monitoring method based on the convolution generation countermeasure network, follow-up detection is based on blade wind sweeping sound segmentation, other signal acquisition devices such as lasers and cameras do not need to be additionally arranged, a fan does not need to be stopped for monitoring, and data do not need to be manually marked; the method fully considers the frequency domain, time domain and amplitude characteristics of the audio signal, realizes unsupervised real-time anomaly monitoring and full-automatic detection, does not need manual participation, can perform anomaly detection on weak faults which are difficult to identify in traditional signal processing and vibration analysis, and does not need expert knowledge and priori knowledge; most common vanes with a faulty apparent aerodynamic performance can be diagnosed without specifying the type of fault.

Drawings

Fig. 1 is a schematic flow diagram of a wind turbine blade fault audio monitoring method according to the present application.

Fig. 2 shows an audio diagram of a full wind sweep in embodiment mode 1.

Fig. 3 shows an audio diagram of the blade 1 in embodiment mode 1.

Fig. 4 shows an audio diagram of the blade 2 in embodiment mode 1.

Fig. 5 shows an audio map of the blade 3 in embodiment mode 1.

Fig. 6 shows an audio diagram of a full sweep in embodiment mode 2.

Fig. 7 shows an audio diagram of a full sweep in embodiment mode 3.

Fig. 8 shows an audio map of the blade 1 in embodiment mode 3.

Fig. 9 shows an audio map of the blade 2 in embodiment mode 3.

Fig. 10 shows an audio map of the blade 3 in embodiment mode 3.

Detailed Description

To make the objects, technical solutions and advantages of the present application more apparent, embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments and features of the embodiments in the present application may be arbitrarily combined with each other without conflict.

The technical terms appearing in the present application are explained and illustrated below.

Short time Fourier transform

STFT (short time fourier transform) is a mathematical transform related to the fourier transform that determines the frequency and phase of the local area sinusoid of a time-varying signal. It defines a very useful class of time and frequency distributions that specify the complex amplitude of any signal over time and frequency. In practice, the process of computing the short-time fourier transform is to divide a longer time signal into shorter segments of the same length, and compute the fourier transform, i.e. the fourier spectrum, on each shorter segment. The mathematical expression of STFT is as follows:

where ω (n) is a sliding window emphasizing frequency information over a certain time domain. The STFT calculation procedure was as follows:

1. dividing an input signal into N parts, taking N points from each part, and intercepting a signal sequence with corresponding Length, wherein N depends on a Window width parameter (Window Length);

2. the selected Window function (Window Type) is multiplied point by point with the corresponding point of the intercepted signal sequence, and the obtained number is the value after windowing interception;

3. if the window width is smaller than the FFT length, filling operation is carried out on two sides of the window by using 0;

4. performing FFT calculation;

5. and obtaining the overlapping Size (Overlap Size) of the sliding window according to the selected step Size, and repeating the steps 1-4 until the sliding window reaches the end of the signal.

The STFT-processed audio signal thus has a localized nature in the time and frequency domain, by means of which the time-frequency characteristics of the signal can be analyzed.

Principal component analysis

Principal Component Analysis (PCA) is a method of statistical analysis, simplifying data sets. It uses orthogonal transformation to linearly transform the observed values of a series of possibly correlated variables, thus projecting as values of a series of linearly uncorrelated variables, which are called principal components. In particular, the principal component can be viewed as a linear equation that contains a series of linear coefficients to indicate the projection direction. PCA is sensitive to regularization or preprocessing of the raw data (relative scaling).

The basic idea of PCA is to orthogonalize the linear transformation:

moving the center of the coordinate axis to the center of the data, and then rotating the coordinate axis to make the variance of the data on the C1 axis maximum, that is, the projection of all n data individuals in the direction is most dispersed, which means that more information is retained, and C1 becomes the first principal component;

c2 second principal component: finding a C2 such that the covariance (correlation coefficient) of C2 and C1 is 0 to avoid overlapping with the C1 information and maximize the variance of the data in that direction;

by analogy, the third principal component, the p-th principal component of the fourth principal component … …, and p random variables may have p principal components, is found.

Principal component analysis is often used to reduce the dimensionality of a data set while maintaining features in the data set that contribute most to variance. This is done by keeping the lower order principal components and ignoring the higher order principal components. Such low order components tend to preserve the most important aspects of the data. Since principal component analysis relies on given data, the accuracy of the data has a large impact on the analysis results. The PCA is used for scaling and dimensionality reduction of the time domain characteristics of the time-frequency domain matrix, so that the data has the inherent advantages that: different columns represent the same unit, all column-wise time domain values represent the same frequency domain characteristics, and the order of magnitude is consistent. Therefore, PCA is particularly useful in analyzing complex data such as time-frequency domain matrices obtained through STFT.

Leaf anomaly detection method for generating antagonistic network anomaly detection model (anoGAN) based on deep convolution

AnoGAN, which generates normal behavior of countermeasure network (DCGAN) learning data based on deep convolution, and performs detection of abnormal data by calculating an abnormal value by mapping original data to a hidden space (late space). As a means of online monitoring, the model will score new data by the degree of match between incoming data and previously learned distributions for the purpose of real-time anomaly detection.

Model structure

For the purpose of anomaly detection, it is necessary to let the model learn normal behavior data by generating countermeasure networks (GANs). The method trains a generative model and a discriminant model for distinguishing real data from generative data. Different from the optimization mode of a single loss function of a common neural network, the structure aims to seek the 'Nash equilibrium' between losses, on one hand, the expression capability of a generated model to specific distribution is enhanced, and on the other hand, the discrimination of generated data and real data by a discrimination model is more accurate.

Unsupervised manifold learning for normal behavior

Manifold learning considers that the data we can observe is actually mapped to a high-dimensional space by a low-dimensional prevalence. Because of the limitation of the internal features of the data, the data in some high dimensions can generate dimensional redundancy, and in fact, the data can be uniquely represented only by a relatively low dimension, and the input data is expressed in the network structure through the mapping of the features in different spaces. Taking the picture as an input, the anomaly detection is taken as an example: m groups of normal pictures I are obtained before the model is constructed_mM is 1, 2, wherein

Is a matrix (two-dimensional intensity map) with a dimension a × b, for each matrix<I_m>Extracting K sub-matrices

(image block). Input I during training_mTraining GAN to learn fluids

(distribution of normal data), the variability of the training data is also represented in an unsupervised manner. Obtained during the test process<y_n，l_n>Wherein y is_nIs an unknown picture of dimension c × c extracted from test data J, l_n∈ {0, 1}, N1, 2.. the N is a binary label vector, and these labels of whether an anomaly is present are used by the test process to evaluate anomaly detection performance.

GAN coding for normal behavior

A GAN comprises two competing modules: a generator G and a discriminator D. The generator G maps a uniform distribution G (z) consisting of one-dimensional noise samples z randomly sampled from an implicit space z to a two-dimensional original sample space, i.e. a fluid constituted by normal behavioural data

Distribution p of learning-free training samples x_gThe network structure of the generator G is equivalent to a decoder using a convolutional structure. Discriminator D is a standard CNN structure that maps two-dimensional pictures into a single scalar D (-) whose output represents that the given input is sampled self-training data

Instead of the probability of the generated sample G (z) generated from the generator G. The model has no loss function, the optimization process is a 'binary minimum maximum game' problem, and D and G can simultaneously optimize themselves according to the value function V (G, D).

Wherein the arbiter self-distributes p the real training samples and samples by training maximization_gThe generator simultaneously trains to fool the arbiter by minimizing the loss of v (g) ═ log (1-D (g (z)), i.e., maximizing the loss of the arbiter, v (g) ═ D (g (z)). And fixing one party in the training process, updating the parameters of the other network, and alternately iterating to maximize the error of the other party, and finally generating a model G to estimate the distribution of the sample data. The generative model G implicitly defines a probability distribution P_gIdeally P_gCan converge to the data true distribution P_data. This minimalist grand game if and only if P_g＝P_dataAn optimal solution exists, namely Nash equilibrium is achieved, the generated model G recovers the distribution of training data, and the accuracy rate of the discrimination model D is equal to 50%.

Mapping to hidden spaces

By countertraining, the generator has learned how to map the noise samples z of the hidden space to the real samples x,

GANs do not directly derive the inverse transform from real samples to implicit space,

given an input x, an attempt is made to find a point z in the hidden space corresponding to the distribution G (z) in the fluid space

Closest to the input data x. The similarity between x and G (z) depends on the input data obeying the data distribution p used to train the generator_gTo the extent of (c). To find the optimal z, a random sampling z from the distribution z of the hidden space is initialized₁Inputting the data into a trained generator to obtain generated data G (z)₁). Data definition based on each generationA loss function representing the loss of mapping from the hidden space to the real data, resulting in z being mapped in the hidden space₁Updating the parameter lifting gradient to the next point, and obtaining z by analogy₂....z_γThat is, the position of z in hidden space is iteratively boosted by back propagation according to the gradient to obtain the most similar data G (z)_r). Instead of the generator or discriminator weights being updated in this step, the input noise z is updated_γSince both models are already trained well using normal behavior data. Since the transition of the samples in the hidden space is continuous, i.e. the target data generated by two consecutive points are also similar, the position where the optimal g (z) can be generated can be found by updating the noise.

The loss function from the original data to the implicit spatial mapping consists of two parts, residual loss and discriminant loss. Residual loss primary metric generating data G (z)_γ) Similarity or difference with input data x:

under ideal conditions, a perfect generator is matched with a hidden spatial mapping relationship, and the residual loss on ideal input data is 0. Discrimination loss the output of the discriminator is subjected to loss calculation, and data G (z) is generated_γ) Input to a discriminator to obtain discrimination loss

Where σ is sigmoid cross entropy, D (G (z)_γ) Taking the sigmoid activation function as an example, the cross entropy is calculated as L α ln (P) + (1- α) ln (1-P), where P is the sigmoid activation function, i.e. the logit transformation function.

Discriminant loss can also be improved by feature mapping. This process and the samples z that would have previously been from the hidden space_γUpdating to trick discriminator D differently, z may be stepped through another form of discrimination loss_γUpdated to form a distribution consistent with the learned distribution from normal dataDistribution G (z)_γ) I.e. feature mapping. The purpose of this processing is to preserve as much of the feature information as possible, with as little change in the distribution of the data after each iteration as possible. The objective function optimized for the generator may also vary, and unlike maximizing the discriminant probability output of the discriminant for generating samples to optimize the parameters of the generator, the generator may force the generation of data consistent with the statistical distribution of the training data, i.e. the expression of the intermediate features is similar to the true data, which limitation is particularly significant for the classification problem. Because no label data is used in the countertraining, the process does not learn features that accurately distinguish the classes, but rather good feature expressions. Therefore, the scalar output (classification probability) of the discriminator is not used for calculating the discrimination loss, and the intermediate feature expression (convolution layer output of the neural network) with richer feature information in the discriminator is used for defining new discrimination loss to measure x and generate data G (z)_γ) The difference of (a):

wherein a layer of intermediate f (-) of the discriminator is used to represent the statistical distribution of the input data. For new discriminant losses, the position adjustment of the implicit spatial samples z does not rely solely on the decision of the trained discriminant, i.e., the generation of data G (z)_γ) Whether to comply with the distribution of normal data and allow for richer expression of characteristic information. To some extent, the use of discriminators is no longer just classifiers, but rather functions as feature extraction.

The overall loss function consists of the weighted values of residual loss and discriminant loss:

where λ represents the loss weight, only the parameter z representing the position of the implicit spatial sample will be adjusted with the back propagation, and the parameters of the trained generator and discriminator are fixed.

Anomaly detectionMeasuring

In the anomaly detection phase, a decision is made as to whether normal or anomalous for the new input data. The defined loss function is a measure of the data G (z) generated in each update iteration gamma_γ) And the degree of matching of input data at the time of the countermeasure training. Similarly, an anomaly score scalar may be derived from the penalty function to represent the deviation of the input data from the normal data:

A(x)＝(1-λ)·R(x)+λ·D(x)

residual score R (x) and discriminative score D (x) from the last time (x)^th) Residual loss after iterative update of implicit spatial mapping

And discriminating the loss

After the abnormal input data is trained, the loss of the two parts is relatively large, the final abnormal score is also large, and the input data which is very similar to the data of the training anti-network can obtain smaller loss and abnormal score. When the training data is normal data, the data with a larger abnormal score is abnormal data.

Converting semi-supervised anomaly detection into unsupervised anomaly detection

Since no tag data is present, it is assumed that the input data x belongs to a finite set

Wherein when k is 3, there are three sets of sample data { x in the finite set₁，x₂，x₃And assuming that there is at most one set of abnormal data, the abnormal detection can be performed in a recursive manner:

1. from m 1 → 3, x is used_mTraining generation confrontation network to obtain generator G with fixed model parameters_mAnd a discriminator D_m；

2. For arbitrary data x₁The corresponding generation countermeasure network model can fully represent the mapping relation of the group of data and the hidden space through trainingIs a step of;

3. using the set of trained discriminators to perform feature mapping, extracting the output of the last convolutional layer as the feature extracted from the original data to obtain x₁Corresponding feature mapping model fe₁；

4. For the other two sets of data { x₂，x₃Respectively as input data, and obtaining respective judgment output x through a trained feature mapping model_dWhen the completely normal (consistent with the input data) data should reach the consistency of the output distribution of the generator and x, the output of the discriminator and x_dA consistent effect;

5. establishing a model from a hidden space random sample z to a generation output and a judgment output, and converting x, x_dTraining as corresponding labels to obtain residual loss and discriminant loss and corresponding abnormal score A₂，A₃；

6. Calculating { x ] according to the set weight₂，x₃Establishing a relative abnormal score of the abnormal score:

when using x₁During training, x is obtained₂，x₃Relative anomaly score of

Represents the deviation of the two sets of data from the training data, if x₂，x₃If a certain group of abnormal data exists, the score of the abnormal data is larger, the score of the other group of normal data is smaller, and the relative abnormal score is larger;

7. similarly, relative anomaly scores of two other sets of data are obtained using each set of data input, and anomaly diagnosis is performed based on the magnitude of the three relative anomaly scores.

The wind turbine generator set comprises a base, a stand column, a cabin positioned on the stand column and three fans arranged on the cabin, wherein the fans are connected with blades; the wind sweeping device also comprises a sound collecting device which is used for collecting wind sweeping sound generated when the blades rotate. Referring to fig. 1, the wind turbine blade fault audio monitoring method based on convolution generation countermeasure network according to the application includes the following steps:

(1) and (4) dividing the audio signal of the blade rotation to obtain the starting and stopping time points of the wind sweeping sound of each blade of the fan.

The sound collecting device records an audio signal generated by the wind swept by the blades, and the sound of the wind swept is in a waveform characteristic with wave crests and wave troughs along with the rotation of each blade and represents the change trend of sound energy when each blade rotates. By analyzing the audio signal, a feature that characterizes the trend of energy variation in the entire sound can be selected, and when the feature is at the lowest point, it can be used as a division point of each blade-turning sound signal.

(2) And carrying out short-time Fourier transform on the audio signal.

The length of the window determines the time resolution and the frequency resolution of the spectrogram, a reasonable window length range capable of obviously observing abnormal sound distribution is obtained through experiments, a window width with the length of M is selected (the value of M is usually between 128-plus-1024 and depends on the expected frequency resolution, the signal length and the step length with the length of N (N is less than M, and N is usually M/2), the number of the obtained frequency segments is M/2, namely the dimensionality of the time-frequency spectrum matrix after the whole sound conversion is M/2

(n is the time domain length, depending on the input data).

(3) According to the sampling frequency and duration obtained when the audio is read in, the corresponding relation with the power spectrum matrix width n is calculated, the time division points are converted into division indexes of matrix columns, the power spectrum of each section of wind sweeping signals of each blade is obtained, and the power spectrum [ a, b, c ] of each blade in the time is spliced into the total power spectrum [ a, b, c ] of each blade in the time. In this case, the frequency dimension of the total power spectrum is the same, and the time dimension is different.

(4) And unifying the time-frequency domain matrix of each group of blades in the time dimension, and extracting the principal component of the power matrix of each blade in the time domain.

Because different input signal lengths are different, the number of the respective wind sweeping segments of the three divided blades is also different, so that the time-frequency domain matrixes of each group of blades need to be unified in the time dimension, and the frequency domain dimensions are consistent. And extracting the principal component of the power matrix of each blade in the time domain because the wind sweeping sounds of each blade are relatively similar and a part with strong noise exists outside the main wind sweeping sound.

(5) The method comprises the steps of converting a single-group blade signal matrix representing time-frequency-energy information into characteristics using principal components to display data by using principal component analysis on each group of blade power spectrum matrix, and reducing the dimension of each original section of wind sweeping signal to the same row number and column number by considering the input structure and the operation efficiency of a generation countermeasure network of short-time Fourier transform

Form a new

An order square matrix;

(6) entering an on-line monitoring stage, and marking three groups of square matrixes (called characteristic square matrixes below) representing time-frequency domain and energy information of blade wind sweeping as M₁，M₂，M₃；

(7) And constructing and generating a countermeasure network, training the feature matrixes of the blades a, b and c by using a DCGAN (deep convolution generation countermeasure network) structure in turn using anoGAN, fixing the parameter structures of a generation model and a discrimination model by carrying out repeated iteration countermeasure training on the feature matrixes of each group of blades, and extracting the last layer of convolution layer of the discriminator to be used as a feature extractor IM. Wherein, the number of iterations of the actual training process may be 100-500.

(8) Reconstructing the model M, taking the noise distribution of the hidden space as the model Input, combining the Output of the generator and the Output of the feature extractor as the model Output to be used as an abnormal detector, and the structure is that Input is z (noise), Output is G_output，IM_output]。

(9) After one group of blade data is trained to generate a confrontation model, the characteristic matrixes of the other two groups of blades are respectively input into an anomaly detector, input data to be detected are used as a first part of M output, iterative training is carried out again, the fitting degree of the input data to original training data is calculated, and the last model loss is used as a fitting error L.

(10) And (3) traversing to calculate errors: training by using the data of the blade a, and respectively inputting the blade b and the blade c to obtain a fitting error L_a，b，L_a，c(ii) a Training by using the data of the blade b, and respectively inputting the blade a and the blade c to obtain a fitting error L_b，a，L_b，c(ii) a Training data of a blade c, and respectively inputting the blade a and the blade b to obtain a fitting error L_c，a，L_c，b. As shown in fig. 1, the three sets of feature matrices at the start of the dotted line are simultaneously input for training the anomaly detector and are simultaneously input for other blade models.

(11) If one group of blades is abnormal (taking the blade a as an example), when training is carried out by using the blades b and c and the blade a is used as an input to be detected, the abnormal value (fitting error) of the blade a is obviously larger than that of the other group of blades without the abnormality, so that the relative abnormal score calculation is carried out for quantitative representation to obtain the difference representing the two groups of blades relative to the other group of blades.

(12) And mapping the maximum value of the three groups of relative abnormal scores into a value range of [0, 100] as a final abnormal risk value.

According to the wind turbine generator blade fault audio monitoring method based on the convolution generation countermeasure network, follow-up detection is based on blade wind sweeping sound segmentation, other signal acquisition devices such as lasers and cameras do not need to be additionally arranged, a fan does not need to be stopped for monitoring, and data do not need to be manually marked; the frequency domain, time domain and amplitude characteristics of the audio signal are fully considered, unsupervised real-time anomaly monitoring and full-automatic detection are realized, and manual participation is not needed; diagnosing based on the difference between the blades, obtaining the difference value between the blades through the fitting error between the blades, wherein the fitting error between the normal blades is small, and the fitting error between the normal blades and the abnormal blades is large; establishing relative abnormal scores to represent abnormal indexes of the two leaves to be detected; the method can be used for carrying out abnormal detection on weak faults which are difficult to identify in traditional signal processing and vibration analysis without expert knowledge and priori knowledge; most common vanes with a faulty apparent aerodynamic performance can be diagnosed without specifying the type of fault.

Examples of embodiment

The blade wind sweeping sound 22 groups collected from a certain wind field comprise 11 groups with faults, and the specific fault types are different. A partial sample of a typical signal and its time-frequency domain power spectrum is as follows.

In a typical fault mode 1, a single blade has fault characteristics similar to high-frequency whistle in wind sweeping sound, and the fault characteristic is determined to be weak when the blade has fault.

FIG. 2 shows that the full sweep in mode 1 shows a whistle accompaniment above the frequency 4096 Hz; FIG. 3 shows that the blade 1 split in mode 1 shows no failure, wherein no whistle morphology is observed; FIG. 4 shows that the blade 2 divided in mode 1 shows a fault, in which a whistle accompaniment occurs in a high frequency band; fig. 5 shows that the blade 3 divided in mode 1 shows no failure, in which no whistle morphology is observed.

According to the experiment, on the basis of proving that the segmentation is accurate, the characteristic performance of a fault mode with an unknown reason, which is accompanied by high-frequency low-energy whistling generated by a single blade, is observed, the fault is difficult to capture by using the traditional time-frequency domain signal analysis technology, and the blade abnormity detection method based on the deep convolution generation antagonistic network abnormity detection model (anoGAN) mentioned in the method is used for carrying out 19 times of tests, so that the following results are obtained:

using a training model of (no-fault blade 1), the relative anomaly scores of (fault blade 2) and (no-fault blade 3) are respectively:

[0.30006403326824954,0.3135906943696607,0.31786652817275246,0.32219042514656976,0.32564399524316984,0.333992480541115,0.33549351567806796,0.34048784601899024,0.34368124116210425,0.34556006894799907,0.35121853042685286,0.36283907834167833,0.3633355842285909,0.3661980735345505,0.3758643455500911,0.37833823728925753,0.3786332505505233,0.38752331304714704,0.4024106816863637]；

using a fault blade 2 training model, the relative anomaly scores of a fault-free blade 1 and a fault-free blade 3 are respectively as follows:

[0.03666751988069426,0.04021439082850867,0.04815121464891185,0.033704536991766045,0.044724577426534226,0.04340430800432746,0.04179608689338652,0.043259298676139646,0.04080363486538868,0.03739510732141588,0.03761166997812392,0.04451871169607397,0.0472102077733969,0.034087485795451,0.040372378962769136,0.040489094303295906,0.04129196223229141,0.04796606108859879]；

using a training model of (faultless blade 3), the relative anomaly scores of (faultless blade 1) and (faulted blade 2) are respectively:

[0.2171931210538521,0.2547810230119036,0.2848908451435764,0.26526940833762547,0.2856949102946677,0.30759248799882155,0.2911501500443904,0.30924862137293063,0.30202136804071045,0.24530856503563625,0.3747275518287287,0.3110276689944409,0.28638430683268473,0.28554103791495494,0.26409322600380786,0.3017711428488341,0.31157105752775405,0.3084963026079462]

it can be seen that if a faulty blade appears in the test data, its anomaly score is significantly greater than that of the other non-faulty blade by more than 30%, and if the test blades are all non-faulty blades, its relative anomaly score is less than 5%. Similarly, multiple experiments were conducted on samples of the remaining 10 groups each with a slight failure with one blade, with the relative anomaly score for the failed blade being between 15% and 40% more, and the relative anomaly scores for both non-failed blades being below 5%.

In a typical failure mode 2, sound energy generated among blades of a wind sweeping signal is remarkably different, surface defects caused by lightning strike are preliminarily judged, and no obvious characteristic exists in frequency.

Figure 6 shows a full sweep in mode 2 showing a significant difference in energy for a single blade from the other two.

Using a training model of (a faultless blade 1), the relative anomaly scores of (a faultless blade 2) and (a faulted blade 3) are respectively:

[0.2643214881138685,0.25434605766365576,0.19175394568013682,0.1941144908126007,0.21416577900526637,0.29576353779705106,0.17541294768583307,0.2857264742959556,0.2551333527041035,0.18495716031207468,0.26722694185758755,0.1916968257905939,0.2771246000509297,0.17749755172147988,0.24723866327656907,0.2138355228037121,0.1995809937706425,0.22733406294041508]；

[0.11381207466082735,0.12244367183007258,0.1936356372403418,0.18165615348216388,0.19075304791793765,0.14789119529957845,0.16766844535542413,0.17983567105124929,0.19018668562173857,0.17737207993045298,0.15390898140264098,0.15557894842446618,0.10818440343641324,0.13281993282004048,0.16020192931022284,0.19701702531343576,0.19510802645185704,0.18110813122660507]；

using a fault blade 3 training model, the relative anomaly scores of a fault-free blade 1 and a fault-free blade 1 are respectively as follows:

[0.01303153312454802,0.02550213751096476,0.04582234394812545,0.03605304541591951,0.03365925226811296,0.01946339784958535,0.01158230453093735,0.02211947247124243,0.01144804406446984,0.02066313029011534,0.02299363713143194,0.02634601177549228,0.01086644500138187,0.01586928004666257,0.04833154229412311,0.0079349565295521,0.0193016288560986,0.01469769596747397]。

it can be seen that, for the abnormal mode characterized by the sound energy distinction, the method can still effectively diagnose whether the blade is abnormal or not through the relative abnormal score.

Typical failure mode 3 — no failure wind sweep sound.

Figure 7 shows that the full sweep in mode 3 shows no anomalies with the blades. Fig. 8 shows that the blade 1 in pattern 3 does not have any abnormality accompanying the blade. Fig. 9 shows that the blade 2 in pattern 3 does not have any abnormality accompanying the blade. Fig. 10 shows that the blade 3 in pattern 3 does not have any abnormality accompanying the blade.

Using a training model of (faultless blade 1), the relative anomaly scores of (faultless blade 2) and (faultless blade 3) are respectively:

[0.031231898848487514,0.03176514239016959,0.03338121197353578,0.03354453471126238,0.03375653140226851,0.033774400293838006,0.033971181964419894,0.03460118699383662,0.03473389553633074,0.03499823572206108,0.03555610834158918,0.03570138669907169,0.035905374077732574,0.03593912570306457,0.03643971660924821,0.03702828405686487,0.03718828565953739,0.037540912231631626]；

using a training model of (faultless blade 2), the relative anomaly scores of (faultless blade 1) and (faultless blade 3) are respectively:

[0.00989041563282651,0.009876817464110716,0.010153965189406552,0.010120103696780498,0.009368211487119475,0.009497615732370703,0.010190210695270254,0.010484019610907272,0.009388298538989722,0.008272910612883539,0.010829056636120785,0.010310911840172734,0.010009342844448392,0.00933976680359754,0.011164575791941024,0.00966043406264873,0.009980604766997548,0.011849989037034556]；

[0.023090083406850478,0.023280385834052817,0.02399305387696031,0.023910252620847883,0.022405459413050482,0.023979551733012496,0.024703723469317894,0.022827966353246416,0.02538051867791085,0.02683136586641309,0.02666688299697833,0.02278846724234273,0.026514541149034718,0.024137129037619033,0.024047339635227767,0.02521898253898177,0.024900972283053747,0.024373952764634332]。

it can be seen that when no defect occurs in the fan blades, the relative anomaly scores of the wind sweeping signals of the three blades are all 0-3%, and the relative anomaly scores of the other 10 groups of blades without faults are all within 5% when the test is performed. In conclusion, through the relative abnormal values among the blades, the appearance defects with different failure modes, different expression characteristics and different strengths can be accurately detected.

Although the embodiments disclosed in the present application are described above, the descriptions are only for the convenience of understanding the present application, and are not intended to limit the present application. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the disclosure as defined by the appended claims.

Claims

1. A wind turbine generator blade fault audio monitoring method based on a convolution generation countermeasure network is characterized by comprising the following steps:

2. The wind turbine blade fault audio monitoring method according to claim 1, wherein the step (2) comprises the following steps:

3. The wind turbine blade fault audio monitoring method according to claim 1 or 2, wherein the difference in step (6) is calculated by performing relative anomaly score calculation to obtain a difference representing one group of blades relative to the remaining groups of blades, and the maximum value of the obtained relative anomaly scores of the groups is used as a final anomaly risk value.

4. The wind turbine blade fault audio monitoring method according to claim 1 or 2, wherein there are three blades, namely a blade a, a blade b and a blade c, and in the step (6), the method for performing the error calculation in a traversal mode comprises the following steps: to be provided withTraining data of the blade a, and respectively inputting the blade b and the blade c to obtain a fitting error L_a,b,L_a,c(ii) a Training by using the data of the blade b, and respectively inputting the blade a and the blade c to obtain a fitting error L_b,a,L_b,c(ii) a Training by using blade c data, and respectively inputting blade a and blade b to obtain fitting error L_c,a,L_c,b。