CN116153333A - Wind turbine blade fault diagnosis method based on aerodynamic noise - Google Patents

Abstract

The invention discloses an aerodynamic noise-based wind turbine blade fault diagnosis method in the field of wind turbine blade fault detection, which converts the historical data of aerodynamic noise samples of wind turbine blades into image data after filtering and noise reduction, and constructs convolution Neural network model, establish a training set, input image data into the convolutional neural network model for pre-training, the training set includes sample history data and expansion data; then obtain the aerodynamic noise of the wind turbine blade to be tested as the acoustic signal, and convert the acoustic signal The signals are aggregated to the server of the data acquisition and transmission center, and then the data is transmitted to the background server of the workstation through the network protocol; the data input to the background server is converted into image data after noise reduction, and then input into the convolutional neural network model, The fault type of the wind turbine blade is determined by the fault characteristics. The invention is easy to operate, can know the status of the blades of the wind turbine in time, and can improve the fault detection efficiency of the blades of the wind turbine.

Description

A Fault Diagnosis Method for Wind Turbine Blades Based on Aerodynamic Noise

technical field

The present invention relates to the technical field of wind turbine blade fault detection and machine learning, in particular to a wind turbine blade fault diagnosis method based on aerodynamic noise, which is used for judging blade tip damage, blade icing, trailing edge cracking, and leading edge Faults such as wear and surface dirt.

Background technique

As a clean and renewable energy, wind energy has been paid more and more attention by countries all over the world. However, my country's installed capacity and proportion are increasing year by year, and the development potential of the wind power industry is huge. However, due to the influence of wind turbine layout, environmental factors, and long-term operation, wind turbine blades are easily damaged, and problems such as blade tip damage, blade freezing, trailing edge cracking, and surface dirt will occur. If the problem cannot be found in time, it will cause further damage to the wind turbine blades.

At present, wind farms generally adopt the method of manual periodic inspection, but there are certain problems in this method. For example, with the continuous investment in wind farms, the scale of wind farms is getting bigger and bigger, and some wind turbines are installed in remote locations. If the method of manual inspection is adopted, it will take a lot of time and energy, and the inspection efficiency is low; at the same time, manual periodic inspection will cause problems to be discovered untimely. Although there are drone inspections at present, the use of drone inspections can only be used in a small area due to the inspection distance and flight time. How to improve the inspection efficiency and find problems in a more timely manner is an urgent technical problem to be solved in the management of wind turbines.

Contents of the invention

The purpose of the present invention is to aim at the technical defect in above-mentioned background, provide a kind of wind turbine blade fault diagnosis method based on aerodynamic noise, make it be possible to carry out fast detection to existing wind turbine blade state, discover the blade failure of wind turbine and its fault in time. Fault type.

In order to achieve the above object, the technical solution adopted in the present invention is as follows: a method for diagnosing faults of wind turbine blades based on aerodynamic noise, said method comprising:

(1) Convert the historical sample data and labels of the aerodynamic noise of the target wind turbine blade into image data after noise reduction processing, construct a convolutional neural network model, establish a training data set, and input the image data into the convolutional neural network model pre-training in

(2) Obtain the operating parameters, atmospheric parameters and aerodynamic noise of different wind turbine blades in the same wind farm at a fixed time scale as the collected acoustic signals; by summarizing the acoustic signals to the server of the data acquisition and transmission center, through the network protocol The data is transmitted to the background server of the workstation; the aerodynamic noise data of the wind turbine blade to be tested in the background server is obtained, and after noise reduction processing, it is converted into image data, and then input into the convolutional neural network model, and the corresponding eigenvalues are output. In the interactive operation interface, the fault type of the wind turbine blade judged is displayed according to the sample historical data of the aerodynamic noise of the wind turbine blade and the corresponding fault label;

The training data set of the convolutional neural network includes sample historical data and expansion data. The sample historical data is actually detected sample historical data, but because the sample historical data is often not complete and sufficient, more data is needed, so it is necessary to Expansion data, the expansion data includes simulation data and data set augmentation performed on historical sample data.

Further, the fault type corresponding to the fault tag includes at least one of icing, crack or defect, leading edge wear, and surface sand hole, which is obtained by means of locally constrained sparse autoencoder, which is seamlessly analyzed for existing faults. Supervised learning divides many categories, and then randomly selects 5 pieces of audio in each category to judge the fault type of each category and obtain the corresponding fault label.

Simulation data acquisition method among the present invention is as follows:

The airfoil data of the wind farm wind turbine in the sample historical data is adjusted according to the airfoil data of the wind turbine, the CST class-shape function and the Latin hypercube sampling, and the airfoil at 80% of the distance from the blade root is adjusted. Increase or decrease the shape to achieve new modeling, so as to obtain new airfoil data and corresponding labels, and then expand the data set under this airfoil; then input the adjusted new airfoil data into openfast and Bladed Or HAWC2 wind turbine simulation platform, set the corresponding parameters, and get the aerodynamic noise data at the observation point. Its beneficial effect is that on the basis of a limited data set, the data set is expanded, on the one hand, the distribution type of the sample becomes wider, and the over-fitting of the data is avoided; on the other hand, the increase of the data will make it possible to train The larger the number of samples, the more information the computer can learn from them, which improves the convergence.

Further, the acquired aerodynamic noise data is acquired through the acoustic signal acquisition and transmission device within a specified collection time period on sunny days; in rainy days or non-specified time periods, the rainproof and dustproof covers are opened to protect the safety of the equipment, and its power supply mode It can be powered by a solar panel on top of the rain shield.

Further, when the sample historical data is denoised and converted into image data, the following sub-steps are sequentially included:

(1-1) Filter the sample historical data through a Butterworth bandpass filter to remove background noise and noise generated during transmission and conversion;

(1-2) Use empirical wavelet transform for secondary filtering to filter out the mechanical noise generated inside the wind turbine, and obtain the aerodynamic noise data after noise reduction;

(1-3) Convert the denoised aerodynamic noise data into image data through the Mel spectrum.

In the present invention, the data added to the data set of the sample historical data is to flip the image data obtained after noise reduction and Mel spectrum, and adjust the image contrast, so as to obtain new data to expand the original training set , by improving the convergence of the acquired data.

Further, the above step (2) may include the following sub-steps:

(2-1) Collect the aerodynamic noise of the wind turbine blade to be tested, and transmit the acoustic signal to the background server through the network protocol through the data collection and transmission center. Convert to image data by Mel spectrum;

(2-2) Input the image data into the convolutional neural network model for classification, and the convolutional neural network model outputs the corresponding feature values according to the fault classification in the pre-training;

(2-3) Determine whether the feature value is within the preset threshold range. If the feature value is within the preset range, it will be classified according to the feature value. If the feature value exceeds the preset range, it will be regarded as undeterminable.

When the characteristic value exceeds the preset range and the judgment result is undeterminable, it shall be manually confirmed, and the audio signal and time-frequency diagram showing that the fault cannot be determined are retrieved through the computer operation interface of the workstation. Acoustic signals and time-frequency diagrams are used to judge the type of fault sound at the workstation. If the judgment still cannot be made, then on-site inspection is carried out. After the fault result is obtained, the result is stored and the original classification result is corrected. The combination of unsupervised and manual identification makes fault identification more accurate.

The image data converted from the aerodynamic noise and the classification results of the fault type of the wind turbine blades are all stored in the background server. Further, the stored picture data and the corrected classification results are input into the convolutional neural network as a training set, and then the convolutional neural network model is regularly updated.

Compared with the prior art, the beneficial effect of the present invention is: the present invention uses the aerodynamic noise of the wind turbine blade as the signal source, establishes the training set of the convolutional neural network through the sample historical data of the sample and the expanded sample data, and then obtains the required The noise data of the judged wind turbine blades are input into the convolutional neural network to determine the type of fault. For example, when there is clear ice in the icing state, the energy of the acoustic signal is mainly concentrated at 605Hz and 190Hz; when it is frosty, the energy of the acoustic signal is mainly concentrated at 108Hz and 830Hz; The probability is the highest, high-frequency signals will be generated, the center of gravity of the spectrum will become larger, and those with greater energy in the sound signal will shift to the high-frequency band, that is, the measured frequency will be greater than the frequency under normal working conditions. When the leading edge is worn, the energy of the acoustic signal is mainly distributed in 1-2kHz. In the case of trachoma defects, the energy of the acoustic signal is mainly distributed at 6-8kHz; different fault conditions correspond to different distribution frequencies, so it can be distinguished by combining the time domain, frequency domain and energy of the acoustic signal. It can reduce unnecessary losses caused by sudden events and reduce the workload of human judgment, and has the advantages of fast judgment speed, simple operation, low time cost and labor cost.

Description of drawings

Fig. 1 is a schematic diagram of a method for diagnosing a fault of a wind turbine blade based on aerodynamic noise provided by an embodiment of the present invention.

FIG. 2 is a top view schematic diagram of the location of the microphone for acoustic signal collection provided by the embodiment of the present invention.

Fig. 3 is a schematic diagram of a rainproof cover for acoustic signal collection provided by an embodiment of the present invention.

Fig. 4 is a structural diagram of a wind turbine and acoustic signal acquisition provided by an embodiment of the present invention.

Fig. 5 is a training flow chart of blade state detection provided by an embodiment of the present invention.

In the figure, 1-wind turbine blade; 2-wind turbine tower; 3-wind turbine tower base; 4-data concentration module and transmission component; 5-acquisition and transmission module: 101-microphone; 102-data transmission module; 103 -Empty module for spare capacity expansion; 6-wall surface; 7-solar panel.

Implementation

In order to make the object, technical solution and effect of the present invention more clear and definite, the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The aerodynamic noise-based wind turbine blade fault diagnosis method provided by the present invention can be applied to a terminal with computing power and interaction capability, and the terminal can execute the aerodynamic noise-based wind turbine blade fault diagnosis method provided by the present invention to check the state of the blade.

The wind turbine blade state detection system based on the aerodynamic noise convolutional neural network provided by the present invention, as shown in the figure, the wind turbine tower 2 is installed on the wind turbine tower base 3, and the wind turbine blade 1 and the generator body are arranged on the wind turbine On the top of the tower 2, the microphone 101 is used for acoustic signal collection, and the acoustic signal is transmitted to the collection and transmission module 5, and then connected to the background server through the data concentration module and the transmission component 4.

Figure 2 shows the noise signal acquisition device, the microphone 101 is used to collect sound signals, and is transmitted to the acquisition and transmission module 5 through the data transmission module 102; the spare capacity expansion empty module 103 is used for capacity expansion; the noise signal acquisition device is installed on the wind turbine tower On the wall 6 of 2, can supply power through solar panel 7.

In this embodiment, the microphone 101 is installed at a height of 850-950mm from the ground, in order to avoid the sound reflection caused by the microphone being too close to the ground, and the microphone being too high will cause the microphone to shake greatly and expand the data error; the distance from the tower is 4m, It is to avoid the sound wave reflection from the wind turbine tower too close to the tower; when the rainy day comes or the specified collection time is not met on a sunny day, the rainproof and dustproof cover will be raised from the data acquisition and transmission module, and the transmission will be interrupted at the same time. The power supply of the collection device ensures the safety of the equipment. The rainproof and dustproof cover first raises the baffle from the surroundings, and then rotates the top baffle to seal through the rotating structure at the top of the baffle farthest from the tower. The baffle is a solar panel. At the same time, the entire data acquisition and transmission module and the rain cover are all sloped structures. This not only effectively prevents rain and sand from damaging the data acquisition device, but also avoids rain, sand and dust on the top. For the problem of accumulation, during the designated collection time on a sunny day, the rainproof and dustproof covers will be put away, while the top baffle can convert solar energy into electrical energy and store it in the battery, which will be used by the device to realize self-production and self-use, and increase resource utilization. Utilization rate, as shown in Figure 3; using three microphones at the same time is to avoid the inaccurate experimental data obtained due to the fact that the yaw single microphone may be just at the minimum aerodynamic noise of the wind turbine, which will affect the judgment of the three microphones on the wind power. The tower of the wind turbine is distributed in a circle, and the angle between each two microphones is 120°, and they are aligned with the blades of the wind turbine, so that the data of the aerodynamic noise of the wind turbine can be obtained more comprehensively, and the sampling rate of the microphone is set to 2.5kHz .

The acquisition period of the three microphones is set to one week, and the acquisition time is set to 8:00, 14:00, and 20:00 of the day, respectively, and the acquisition is set for 3 minutes, and then the three microphones randomly intercept a period of 30 seconds of noise data during each acquisition time , while recording the operating parameters of the wind turbine blades and the atmospheric parameters within the collection time.

The collected data is unified and aggregated to the data collection and transmission center, and when the collection cycle reaches the specified time, it is uniformly transmitted to the background server through the network protocol.

A wind turbine blade fault diagnosis method based on aerodynamic noise carried out by the above-mentioned device, the method comprising:

(1) Convert the historical sample data and labels of the aerodynamic noise of the target wind turbine blade into image data after noise reduction processing, construct a convolutional neural network model, establish a training data set, and input the image data into the convolutional neural network model pre-training in

(2) Obtain the operating parameters, atmospheric parameters and aerodynamic noise of different wind turbine blades in the same wind farm at a fixed time scale as the collected acoustic signals; by summarizing the acoustic signals to the server of the data acquisition and transmission center, through the network protocol The data is transmitted to the background server of the workstation; the aerodynamic noise data of the wind turbine blade to be tested in the background server is obtained, and after noise reduction processing, it is converted into image data, and then input into the convolutional neural network model, and the corresponding eigenvalues are output. In the interactive operation interface, the fault type of the wind turbine blade judged is displayed according to the sample historical data of the aerodynamic noise of the wind turbine blade and the corresponding fault label;

The training data set of the convolutional neural network includes sample historical data and expansion data. The sample historical data is actually detected sample historical data, but because the sample historical data is often not complete and sufficient, more data is needed, so it is necessary to Expansion data, where the expansion data includes simulation data and data set augmentation performed on historical sample data.

Further, the fault type corresponding to the fault tag includes at least one of icing, crack or defect, leading edge wear, and surface sand hole, which is obtained by means of locally constrained sparse autoencoder, which is seamlessly analyzed for existing faults. Supervised learning divides many categories, and then randomly selects 5 pieces of audio in each category to judge the fault type of each category and obtain the corresponding fault label.

The method of obtaining the above simulation data is as follows:

The airfoil data of the wind farm wind turbine in the sample historical data is adjusted according to the airfoil data of the wind turbine, the CST class-shape function and the Latin hypercube sampling, and the airfoil at 80% of the distance from the blade root is adjusted. Increase or decrease the shape to achieve new modeling, so as to obtain new airfoil data and corresponding labels, and then expand the data set under this airfoil; then input the adjusted new airfoil data into openfast and Bladed Or HAWC2 wind turbine simulation platform, set the corresponding parameters, and get the aerodynamic noise data at the observation point.

The CST type function uses a category function and a shape function to describe the geometric shape of the airfoil, which can give a high-precision fitting curve with fewer parameters. Take x as the abscissa, y as the ordinate, and y _TE as the ordinate of the trailing edge to represent the CST type function, the formula is:

Among them, C(x) is a category function, and different values of N1 and N2 can be used to construct different geometric shapes, which are stipulated by the National Aviation Advisory Committee of the United States. Arc airfoil, N1=1.0, N2=1.0; S(x) is the shape function, Ai represents the weight factor introduced, called the shape function coefficient, i=0,1,..n; S(x) is the shape The function is the weighted sum of n-order Bernstein polynomials, and CST parameterized curves with different precisions can be obtained by adjusting the order of the polynomials.

Latin hypercube sampling is a method of stratified sampling. First, it splits the design space of variables into N equidistant intervals, and then selects a random data point in each interval, then each variable There are N data points, and then these data points are randomly combined. Its advantage is that no matter how much the variable dimension of the optimization problem is, the sample points can be evenly distributed in the design space.

Its beneficial effect is that on the basis of a limited data set, the data set is expanded, on the one hand, the distribution type of the sample becomes wider, and the over-fitting of the data is avoided; on the other hand, the increase of the data will make it possible to train The larger the number of samples, the more information the computer can learn from them, which improves the convergence.

Further, the acquired aerodynamic noise data is acquired through the acoustic signal acquisition and transmission device within a specified collection time period on sunny days; in rainy days or non-specified time periods, the rainproof and dustproof covers are opened to protect the safety of the equipment, and its power supply mode It can be powered by a solar panel on top of the rain shield.

Further, when the sample historical data is denoised and converted into image data, the following sub-steps are sequentially included:

(1-1) Filter the sample historical data through a Butterworth bandpass filter to remove background noise and noise generated during transmission and conversion;

(1-2) Use empirical wavelet transform for secondary filtering to filter out the mechanical noise generated inside the wind turbine, and obtain the aerodynamic noise data after noise reduction;

(1-3) Convert the denoised aerodynamic noise data into image data through the Mel spectrum.

Considering that the frequency response of the Butterworth filter is stable both inside and outside the passband, the Butterworth bandpass filter is used to filter out low-frequency wind noise and ensure that the signal does not attenuate within the fault frequency range as much as possible. , while taking into account the type of fault, set the lower limit cut-off frequency to 100Hz, and the upper limit cut-off frequency to 20kHz.

The empirical wavelet transform first divides the Fourier frequency bands, divides the spectrum into N frequency bands according to the segmentation algorithm, and then constructs a filter bank according to the band boundary coordinates, and then obtains component signals through wavelet basis function filtering, and then reconstructs the signals to obtain Filtered aerodynamic noise data.

Pass the data obtained after filtering and the simulation expansion data into the Mel language spectrogram, and convert the audio signal into a time-frequency diagram of the image. In the time-frequency diagram, the aerodynamic noise can be better observed in the time domain, frequency domain and energy distribution, which is conducive to manual judgment of faults. Mel Spectrogram is to pre-emphasize and frame the input signal, divide the original signal into several small blocks according to time, and each block is a frame; then add a window function to each frame to obtain better side lobe reduction Amplitude; then perform Fourier transform on each frame to obtain the energy spectrum; the last step is to apply the Mel filter to the energy spectrum obtained in the previous step, and then obtain the time-frequency map based on the Mel spectrogram.

In the present invention, the data added to the data set of the sample historical data is to flip the image data obtained after noise reduction and Mel spectrum, and adjust the image contrast, so as to obtain new data to expand the original training set , by improving the convergence of the acquired data.

Further, the above step (2) may include the following sub-steps:

(2-1) Collect the aerodynamic noise of the wind turbine blade to be tested, and transmit the acoustic signal to the background server through the network protocol through the data collection and transmission center. Convert to image data by Mel spectrum;

(2-2) Input the image data into the convolutional neural network model for classification, and the convolutional neural network model outputs the corresponding feature values according to the fault classification in the pre-training;

(2-3) Determine whether the feature value is within the preset threshold range. If the feature value is within the preset range, it will be classified according to the feature value. If the feature value exceeds the preset range, it will be regarded as undeterminable.

The acoustic signals obtained by different types of faults will be quite different. When there is clear ice in the icing state, the energy of the acoustic signal is mainly concentrated at 605Hz and 190Hz. When it is frosty, the energy of the acoustic signal is mainly concentrated at 108Hz and 830Hz; in crack or defect faults, the probability of damage to a single blade is the highest, high-frequency signals will be generated, the center of gravity of the spectrum will become larger, and those with greater energy in the sound signal will shift to the high-frequency band, that is, the measured frequency will be greater than that of the normal working frequency. Under normal conditions; when the leading edge is worn, the energy of the acoustic signal is mainly distributed at 1-2kHz; in the case of sand holes, the energy of the acoustic signal is mainly distributed at 6-8kHz; different fault conditions correspond to different distribution frequencies, so it can be passed Distinguish the way in which the time domain, frequency domain and energy of the acoustic signal are combined.

The frequency distribution of various faults is shown in Table 1:

The above-mentioned cracks or defects include various series of cracks on the blade, damage to the blade, and the like.

The obtained image data of the aerodynamic noise of the blade of the wind turbine to be measured is input into the fault type judgment module in the acoustic signal processing module, thereby obtaining the eigenvalue through the convolutional neural network model, and judging whether the eigenvalue is within the preset threshold range, If the eigenvalue is within the preset range, it will be classified according to the eigenvalue. If the eigenvalue exceeds the preset range, it will be regarded as undetermined, so it needs to be further manually confirmed. It can be called through the computer operation interface of the workstation. The audio signal and time-frequency diagram of the fault that cannot be determined are displayed. The operation and maintenance personnel judge the type of fault sound at the workstation according to the fault sound signal and time-frequency diagram. result and correct the original classification result;

Input the judgment result to the operation interface of human-computer interaction;

The generated feature value is a range value. If there are 5 types of faults, there are 6 numbers corresponding to 0-5, and 0 is normal. When the detected feature value is (0.6,1.4), it is number 1. Fault, when the detected feature value is (1.6,2.4), it is the fault number 2, and so on; if the detected feature value is [1.4,1.6], [2.4,2.6], etc., it is unable to judge, that is It exceeds the threshold range, so it needs to be manually confirmed on the operation interface and the classification result is corrected.

The aerodynamic noise audio and picture data and classification results are all stored.

The stored picture data and the corrected classification results are also input into the convolutional neural network as a training set, the database is supplemented and perfected, and the convolutional neural network model is updated. After one year of operation in the wind field, the database tends to be complete and it is possible to judge Most of the noise information, to achieve fully automated fault monitoring.

The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solutions disclosed in the present invention, those skilled in the art can make some replacements and modifications to some of the technical features according to the disclosed technical content without creative work. Deformation, these replacements and deformations are all within the protection scope of the present invention.

Claims (10)

1. A wind turbine blade fault diagnosis method based on aerodynamic noise is characterized by comprising the following steps:

(1) Sample historical data and labels of pneumatic noise of a target wind turbine blade are converted into image data after noise reduction treatment, a convolutional neural network model is built, a training data set is built, and the image data is input into the convolutional neural network model for pre-training;

(2) Acquiring operation parameters, atmospheric parameters and aerodynamic noise of different wind turbine blades of the same wind farm in a fixed time scale as acquired acoustic signals; the sound signals are summarized to a server of a data acquisition and transmission center, and data are transmitted to a background server of a workstation through a network protocol; acquiring aerodynamic noise data of a wind turbine blade to be detected in a background server, converting the aerodynamic noise data into image data after noise reduction treatment, inputting a convolutional neural network model, outputting corresponding characteristic values, and displaying and judging the fault type of the wind turbine blade according to sample historical data of aerodynamic noise of the wind turbine blade and corresponding fault labels in a man-machine interaction operation interface;

the training data set of the convolutional neural network comprises sample historical data and capacity expansion data, wherein the sample historical data are truly detected data, and the capacity expansion data comprise simulation data and data for increasing the data set of the sample historical data.

2. The wind turbine blade fault diagnosis method based on aerodynamic noise according to claim 1, wherein the fault type corresponding to the fault label at least comprises one of icing, cracking or defect, front edge abrasion and surface sand hole, the acquisition is to perform unsupervised learning on the existing fault by means of local constraint sparse self-encoder, a plurality of categories are separated, then 5 sections of audio in each category are randomly selected, and the fault type of each category is judged to obtain the corresponding fault label.

3. The aerodynamic noise-based wind turbine blade fault diagnosis method according to claim 1, characterized in that the simulation data acquisition method is as follows:

acquiring sample historical data of wing profile data of a wind turbine of a wind power plant, adjusting the wing profile according to the wing profile data of the wind turbine, a CST class-shape function and Latin hypercube sampling, and increasing or decreasing the wing profile at the position 80% away from a blade root to realize new modeling, thereby obtaining new wing profile data and corresponding labels, and further expanding a data set under the wing profile;

and inputting the new wing profile data obtained after adjustment into a openfast, bladed or HAWC2 wind turbine generator simulation platform, and setting corresponding parameters to obtain aerodynamic noise data at the observation point.

4. The aerodynamic noise-based wind turbine blade failure diagnosis method according to claim 1, characterized in that the obtained aerodynamic noise data is obtained by means of an acoustic signal acquisition and transmission device during a specified acquisition time period on a sunny day.

5. The aerodynamic noise-based wind turbine blade fault diagnosis method according to claim 1, wherein the noise reduction of the sample history data into image data comprises the following sub-steps in sequence:

(1-1) filtering the sample history data by means of a butterworth band-pass filter to remove background noise and noise generated during transmission and conversion;

(1-2) performing secondary filtering treatment by using empirical wavelet transformation, and filtering mechanical noise generated in the wind turbine to obtain noise-reduced pneumatic noise data;

(1-3) converting the noise-reduced aerodynamic noise data into image data by means of mel frequency spectrum.

6. The aerodynamic noise-based wind turbine blade fault diagnosis method according to claim 5, wherein the data added to the data set of the sample history data is data obtained by inverting the image data obtained by noise reduction and mel frequency spectrum, and adjusting the image contrast, thereby obtaining new data expansion original training set, and obtaining data by improving convergence.

7. The aerodynamic noise based wind turbine blade failure diagnosis method according to claim 1, characterized in that step (2) comprises the sub-steps of:

(2-1) acquiring aerodynamic noise of a wind turbine blade to be detected, transmitting an acoustic signal to a background server through a network protocol through a data acquisition and transmission center, reducing noise through a Butterworth band-pass filter and empirical wavelet variation, and converting the noise into image data through a Mel frequency spectrum;

(2-2) inputting the image data into a convolutional neural network model for classification, and outputting corresponding characteristic values by the convolutional neural network model according to fault classification in pre-training;

and (2-3) judging whether the characteristic value is within a preset threshold range, classifying according to the characteristic value if the characteristic value is within the preset threshold range, and judging that the characteristic value cannot be determined if the characteristic value exceeds the preset range.

8. The wind turbine blade fault diagnosis method based on aerodynamic noise according to claim 7, wherein the characteristic value exceeds a preset range, when the judging result is that the characteristic value cannot be confirmed, the characteristic value is manually confirmed, an audio signal and a time-frequency diagram which cannot confirm the fault are called and displayed through a computer operation interface of a workstation, an operation and maintenance person judges the type of fault sound in the workstation according to the fault sound signal and the time-frequency diagram, when the fault sound cannot be judged, the operation and maintenance person checks the operation and maintenance person on site again, and after the fault result is obtained, the result is stored and the original classification result is corrected.

9. The aerodynamic noise-based wind turbine blade fault diagnosis method according to claim 8, wherein the aerodynamic noise converted picture data and wind turbine blade fault type classification results are stored in a background server.

10. The wind turbine blade fault diagnosis method based on aerodynamic noise according to claim 9, wherein the stored picture data and the corrected classification result are input into a convolutional neural network as a training set, and the convolutional neural network model is updated periodically.

Images