Wind turbine blade voiceprint recognition detection method

Abstract

The invention relates to a wind turbine blade voiceprint recognition detection method, which comprises the following steps of: acquiring an initial sound signal of a wind turbine blade to be detected, preprocessing the initial sound signal, acquiring a sound signal, performing time-frequency domain decomposition on the sound signal, and acquiring a time-frequency domain decomposition result; carrying out feature vector extraction and fusion on a time-frequency domain decomposition result to obtain a comprehensive feature vector; the comprehensive feature vector is input into a blade state diagnosis model, the blade state of the wind turbine to be detected is output, the blade state diagnosis model is obtained by training an existing blade state diagnosis model through a target training set, and the existing blade state diagnosis model is obtained by training a large-scale training set; the target training set is historical sound signals of the same model as the to-be-detected wind turbine and corresponding state labels, and the large-scale training set is historical sound signals of blades of different models and corresponding state labels. According to the invention, the blade state monitoring accuracy and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine blade detection, and in particular to a wind turbine blade soundprint recognition and detection method. Background Art

[0002] Wind turbine blades generate complex aerodynamic noise and structural vibrations during operation. These sound signals contain rich information about the blade status. However, the outdoor environment where wind turbines are located is complex and changeable, and the on-site noise interference is serious, resulting in a low signal-to-noise ratio of the blade soundprint signal, which increases the difficulty of soundprint recognition. In addition, different types of wind turbine blades have different structures and materials, and their soundprint characteristics are also different, making it challenging to establish a unified soundprint recognition model. At the same time, there are many types of blade deviations, including aerodynamic shape deviations, installation angle deviations, structural damage, etc. Different types of deviations have different effects on soundprint characteristics, which increases the complexity of deviation identification. Wind turbine blades operate in harsh outdoor environments for a long time, and their soundprint signals are easily affected by environmental factors such as temperature, humidity, and wind speed, resulting in the non-stationarity and uncertainty of soundprint characteristics, which brings difficulties to soundprint recognition. In summary, wind turbine blade soundprint recognition and detection faces technical challenges such as large environmental noise interference, strong blade differences, diverse deviation types, and non-stationary soundprint signals. It is urgent to study targeted soundprint feature extraction, noise elimination, and adaptive recognition methods to improve the accuracy and reliability of blade deviation detection. Summary of the invention

[0003] The purpose of the present invention is to provide a method for wind turbine blade soundprint recognition and detection to improve the accuracy and efficiency of blade status monitoring.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A wind turbine blade soundprint recognition and detection method, comprising:

[0006] Collecting the initial sound signal of the wind turbine blade to be detected and preprocessing it to obtain the sound signal, performing time-frequency domain decomposition on the sound signal, and obtaining the time-frequency domain decomposition result;

[0007] Extracting and fusing feature vectors from the time-frequency domain decomposition results to obtain a comprehensive feature vector;

[0008] The comprehensive feature vector is input into a blade state diagnosis model, and the state of the wind turbine blade to be detected is output, wherein the blade state diagnosis model is obtained by training an existing blade state diagnosis model with a target training set, and the existing blade state diagnosis model is constructed by an attention mechanism, a convolutional neural network, and a support vector machine classifier and is obtained by training with a large-scale training set, the target training set is a historical sound signal of the same model as the wind turbine to be detected and the corresponding state label, and the large-scale training set is a historical sound signal of blades of different models and the corresponding state label.

[0009] Optionally, preprocessing the initial sound signal includes:

[0010] Performing filtering on the initial sound signal to obtain a sound signal after filtering;

[0011] Adopting an adaptive filtering algorithm to estimate the current environmental noise spectrum characteristics in real time, adjusting the parameters of the spectrum subtraction algorithm according to the estimated current environmental noise spectrum characteristics, and obtaining an optimized spectrum subtraction algorithm;

[0012] The sound signal after the filtering process is subjected to noise reduction according to the optimized spectral subtraction algorithm to obtain the sound signal.

[0013] Optionally, performing time-frequency domain decomposition on the sound signal includes: performing time-frequency domain decomposition on the sound signal using a wavelet analysis method to obtain decomposition results in the time domain, frequency domain and time-frequency joint domain.

[0014] Optionally, extracting and fusing the time-frequency domain decomposition results to obtain a comprehensive feature vector includes:

[0015] According to the time domain decomposition result, the waveform characteristics of the sound signal in the time domain are extracted to obtain a time domain feature vector, wherein the waveform characteristics include signal amplitude, energy and zero-crossing rate;

[0016] According to the frequency domain decomposition result, extract the spectrum characteristics of the sound signal in the frequency domain to obtain a frequency domain feature vector, wherein the spectrum characteristics include the frequency distribution, spectrum peak position, and spectrum peak amplitude of the sound signal;

[0017] According to the time-frequency joint domain decomposition result, extract the characteristics of the sound signal in the time-frequency joint domain, and obtain the time-frequency joint domain feature vector, wherein the characteristics of the time-frequency joint domain are the distribution of the sound signal on the time-frequency plane;

[0018] The time domain feature vector, the frequency domain feature vector and the time-frequency joint domain feature vector are fused to obtain the comprehensive feature vector.

[0019] Optionally, training the existing blade status diagnosis model through a large-scale training set includes:

[0020] Obtain historical comprehensive feature vectors of historical sound signals of blades of different models;

[0021] Input the historical comprehensive feature vector into the attention mechanism, calculate the attention weight coefficients in the time domain, frequency domain and time-frequency domain through weight learning, and obtain the adjusted feature vector;

[0022] Inputting the adjusted feature vector into the convolutional neural network for feature extraction to obtain high-discriminative features;

[0023] The high-discrimination feature is input into the support vector machine classifier to output the blade status, wherein the blade status includes a normal state, slight wear, and severe wear.

[0024] Optionally, obtaining the adjusted feature vector includes: multiplying the attention weight coefficient by the corresponding time-frequency area value, frequency domain area value and time-frequency area value through weighted processing to obtain the adjusted feature vector.

[0025] Optionally, acquiring the blade state diagnosis model includes: training a fully connected layer of the existing blade state diagnosis model according to a target training set to acquire the blade state diagnosis model.

[0026] Optionally, collecting the initial sound signal of the wind turbine blade to be detected includes: deploying a sound sensor array on the wind turbine blade to be detected, collecting the initial sound signal and transmitting the collected initial sound signal to a cloud server in real time via a wireless communication module.

[0027] The beneficial effects of the present invention are as follows: the present invention uses wavelet analysis to perform time-frequency domain decomposition, and extracts the denoised blade sound signal according to an adaptive noise elimination algorithm. Through the attention mechanism, the key features that are sensitive to the blade state are highlighted. In order to solve the problem of differences in soundprint features of blades of different models, a transfer learning method is introduced to achieve model generalization across blade models. The present invention realizes the full life cycle management and predictive maintenance of wind turbine blades, and improves the accuracy and efficiency of blade state monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 The present invention is a flowchart of a method for wind turbine blade soundprint recognition and detection according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] In wind turbines, blades, as a medium for wind power energy conversion, play a vital role in the power generation process. Common methods for detecting blade damage are: 1) Visual inspection method: The visual inspection method mainly observes defects on the outer surface of wind turbine blades and the surface of the accessible area of ​​the inner cavity with the naked eye or with the help of tools such as magnifying glasses. The steps are as follows: After the blade resin is infused and cured and before mold closing, visually inspect the more obvious surface defects such as dry fibers, bulges, cracks, bubbles, scratches, etc. During the blade manufacturing process, such as after mold closing and surface painting, the visual method is limited to the detection of paint defects in the accessible area of ​​the inner cavity and the blade surface. 2) Infrared thermal wave method: The infrared thermal wave method receives the infrared radiation of the blade through an infrared thermal imager, converts it into an image, and then judges the damage of the blade based on the infrared imaging image. The steps are as follows: Apply a uniform heat flow to the wind turbine blade to make the temperature of the material surface consistent. When the blade is damaged to varying degrees, the temperature and infrared radiation intensity of the surface of the damaged part will react abnormally. These abnormal reactions are converted into visual thermal distribution images, and the damage of the blade is judged by the image. 3) X-Ray method: X-ray detection method has been widely used in the field of composite material detection through real-time imaging detection technology. The steps are as follows: Use X-ray detection technology to detect common defects of wind turbine blades such as glue-deficient voids, inclusions, cracks perpendicular to the surface of FRP, fat-enriched, and partial wrinkles. Determine volume defects such as glue-deficient voids and inclusions in blades. 4) Tapping method: The tapping method uses hard objects such as a small hammer and a steel ruler to tap the bonding area of ​​the blade to determine whether there is a bonding void. The steps are as follows: During the inspection of the wind turbine blade manufacturing process, tap the blade bonding area. Determine whether there is a bonding void by sound. When there is no obvious difference in the sound, it will cause the defect to be unable to be determined or missed. 5) AE technology (acoustic emission technology): Acoustic emission technology determines the damage of the blade by detecting the acoustic signal generated when the blade is damaged. The steps are as follows: Collect noise signals, which are easily affected by factors such as wind speed, temperature, humidity, terrain, and sensor installation location. Noise signals must undergo complex filtering and denoising to obtain effective information. The algorithm is used for further processing, which increases the difficulty of noise monitoring. 6) Ultrasonic method: Ultrasonic detection technology is a method of judging blade damage by detecting the reflection and scattering of ultrasonic waves inside the blade. The steps are as follows: When the ultrasonic wave encounters defects or cracks inside the blade, it will produce reflection and scattering, thus forming some echoes. The damage of the blade is judged by analyzing these echoes.

[0033] However, infrared heat wave method, X-Ray method, tapping method and ultrasonic method are mostly used for static inspection before blade assembly, and have little effect on the detection of damage during the service of blades; visual inspection method and AE technology can detect damage caused by blades during operation, but the former relies on manpower and is too inefficient; although the latter can monitor the health of blades in real time, it requires the installation of sensors inside the blades, which can easily damage the internal structure of the blades; the comprehensive damage detection method is just a combination and optimization of single detection methods, and the defects in actual engineering applications still exist, and it is not possible to detect wind turbine blades well.

[0034] The detection of large equipment is developing in the direction of high speed, high precision and high efficiency. Establishing a reliable health monitoring system is a necessary measure to ensure the safe operation of these equipment. Traditional mechanical health monitoring systems use simple time domain and frequency domain indicators to describe the operating status of the equipment. The selection of these specific indicators often comes from professional technicians and expert experience. They only have good diagnostic effects on specific equipment and their faults and are not universally applicable. In complex industrial testing sites, fault information is complex and changeable, often mixed with background noise and multi-sensor measurement noise, and may also include internal and external excitations and the coupling of multiple faults. This makes the diagnostic methods based on these characteristic indicators have poor stability and low diagnostic accuracy.

[0035] like Figure 1 As shown, this embodiment provides a method for wind turbine blade soundprint recognition and detection, including:

[0036] Collecting the initial sound signal of the wind turbine blade to be detected and preprocessing it to obtain the sound signal, decomposing the sound signal in the time-frequency domain, and obtaining the decomposition result in the time-frequency domain;

[0037] Extract and fuse the feature vectors of the time-frequency domain decomposition results to obtain a comprehensive feature vector;

[0038] The comprehensive feature vector is input into the blade state diagnosis model, and the state of the wind turbine blade to be detected is output, wherein the blade state diagnosis model is obtained by training the existing blade state diagnosis model with the target training set, and the existing blade state diagnosis model is constructed by the attention mechanism, convolutional neural network and support vector machine classifier and obtained by training with a large-scale training set, the target training set is the historical sound signals and corresponding state labels of the same model as the wind turbine to be detected, and the large-scale training set is the historical sound signals and corresponding state labels of blades of different models.

[0039] Furthermore, preprocessing the initial sound signal includes:

[0040] Performing filtering on the initial sound signal to obtain the sound signal after filtering;

[0041] Adopting an adaptive filtering algorithm to estimate the current environmental noise spectrum characteristics in real time, adjusting the parameters of the spectrum subtraction algorithm according to the estimated current environmental noise spectrum characteristics, and obtaining an optimized spectrum subtraction algorithm;

[0042] The sound signal after filtering is subjected to noise reduction according to the optimized spectral subtraction algorithm to obtain the sound signal.

[0043] Specifically, low-frequency noise interference is removed by a high-pass filter, or the main frequency band of blade vibration (such as 50Hz to 2kHz) is retained by a band-pass filter. Adaptive noise elimination technology (such as adaptive least mean square algorithm, LMS) is used to dynamically estimate the spectrum distribution of ambient noise by comparing the blade sound signal and the reference noise signal (such as background noise far away from the blade). Assume that at a certain moment, the estimated ambient noise spectrum shows an obvious peak at 500Hz, which may be due to the operation of nearby mechanical equipment. According to the estimated ambient noise spectrum characteristics, the parameters of the spectral subtraction algorithm are dynamically adjusted to eliminate the background noise in a targeted manner. For example, spectral subtraction reduces the impact of noise by subtracting the estimated noise spectrum from the spectrum of the original signal. Assuming that the amplitude of the ambient noise at 500Hz is 10dB, the spectral subtraction method will reduce the signal amplitude of this frequency band by 10dB accordingly, thereby effectively weakening the background noise.

[0044] Furthermore, performing time-frequency domain decomposition on the sound signal includes: performing time-frequency domain decomposition on the sound signal using a wavelet analysis method to obtain decomposition results in the time domain, the frequency domain, and the time-frequency joint domain.

[0045] Furthermore, feature vectors are extracted and fused from the time-frequency domain decomposition results to obtain a comprehensive feature vector including:

[0046] According to the time domain decomposition result, the waveform characteristics of the sound signal in the time domain are extracted to obtain the time domain feature vector, wherein the waveform characteristics include signal amplitude, energy and zero-crossing rate;

[0047] According to the frequency domain decomposition result, the spectrum characteristics of the sound signal in the frequency domain are extracted to obtain the frequency domain characteristic vector, wherein the spectrum characteristics include the frequency distribution, spectrum peak position and spectrum peak amplitude of the sound signal;

[0048] According to the decomposition result of the joint time-frequency domain, the characteristics of the sound signal in the joint time-frequency domain are extracted to obtain the feature vector of the joint time-frequency domain, wherein the characteristics of the joint time-frequency domain are the distribution of the sound signal on the time-frequency plane;

[0049] The time domain feature vector, frequency domain feature vector and time-frequency joint domain feature vector are fused to obtain a comprehensive feature vector.

[0050] Specifically, the preprocessed sound signal is decomposed in the time-frequency domain by wavelet analysis method to obtain the representation of the signal in the time domain, frequency domain and time-frequency joint domain. Wavelet analysis is a multi-scale time-frequency analysis method that can decompose the signal into wavelet components of different scales and frequencies, thereby revealing the characteristics of the signal at different times and frequencies. For example, the sound signal can be decomposed using discrete wavelet transform (DWT) or continuous wavelet transform (CWT). Through wavelet decomposition, various representations of the signal such as time domain waveform, spectrum diagram and time-frequency diagram can be obtained. The denoised sound signal is subjected to discrete wavelet transform, and db4 wavelet basis function is used for 4-layer decomposition. In this way, the time-frequency information of the signal at different scales and frequencies is obtained. According to the results of the time-frequency domain decomposition, the characteristics of the sound signal in the time domain, frequency domain and time-frequency joint domain are extracted to form a feature vector. The purpose of feature extraction is to extract effective information that can reflect the state of the blade from the results of the time-frequency domain decomposition. The time domain features can reflect the time-varying characteristics of the signal, such as amplitude, energy, and zero-crossing rate; the frequency domain features can reflect the frequency distribution, peak position, peak amplitude, and other frequency characteristics of the signal; the time-frequency joint domain features can reflect the distribution of the signal energy on the time-frequency plane. For example, the time domain features such as the root mean square value, kurtosis, and skewness of the signal can be calculated; the frequency domain features such as the power spectrum density, spectrum center of gravity, and harmonic amplitude of the signal can be calculated; and the time-frequency joint domain features such as the energy distribution of the signal in different time-frequency regions can be calculated. The extracted time domain feature vector, frequency domain feature vector, and time-frequency joint domain feature vector are fused to construct a comprehensive feature vector as a feature parameter reflecting the blade state information. The time domain features can provide basic statistical features about the signal, while the frequency domain features can help identify the frequency components of the signal, thereby extracting key information. The time-frequency joint domain feature vector can provide the frequency characteristics of the signal at different time points, which helps to capture the dynamic changes of the signal. The time domain, frequency domain, and time-frequency joint domain feature vectors provide comprehensive, accurate, and anti-interference feature parameters for blade state recognition, which helps to achieve effective state recognition.

[0051] The extracted time domain feature vector, frequency domain feature vector and time-frequency joint domain feature vector are fused to construct a comprehensive feature vector as a feature parameter reflecting the blade status information. The purpose of feature fusion is to integrate different types of features to obtain more comprehensive and accurate blade status information. For example, the time domain feature vector, frequency domain feature vector and time-frequency joint domain feature vector can be directly spliced ​​together to form a high-dimensional comprehensive feature vector; different types of features can also be fused using weighted averaging, principal component analysis and other methods. The calculated root mean square value, spectral center of gravity and the first layer of wavelet decomposition energy ratio are spliced ​​to obtain a comprehensive feature vector.

[0052] Furthermore, the existing blade status diagnosis model is trained through a large-scale training set, including:

[0053] Obtain historical comprehensive feature vectors of historical sound signals of blades of different models;

[0054] The historical comprehensive feature vector is input into the attention mechanism, and the attention weight coefficients in the time domain, frequency domain, and time-frequency domain are calculated through weight learning to obtain the adjusted feature vector;

[0055] The adjusted feature vector is input into the convolutional neural network for feature extraction to obtain high discriminative features;

[0056] The high discriminative features are input into the support vector machine classifier to output the blade status, where the blade status includes normal status, slight wear, and severe wear.

[0057] Specifically, for blades of different models, such as model A, model B, and model C, their voiceprint data in normal state and various fault states are collected respectively. The voiceprint data in each state can be collected by a high-precision microphone under different working conditions to ensure the diversity and representativeness of the data. The voiceprint dataset of each model contains thousands of voiceprint samples, and each sample is labeled with its corresponding blade model and state, such as "Model A-Normal", "Model A-Crack", etc., to obtain a large-scale training set.

[0058] This embodiment adaptively adjusts the importance of each time-frequency region according to the attention weight coefficient, and highlights the key features that are sensitive to the state of the leaves. Through weighted processing, the attention weight coefficient is multiplied by the corresponding time-frequency region value, so that the features of the important regions are more significant. For example, if a frequency band shows a higher energy change when the blade is abnormal, its corresponding attention weight coefficient will be larger. After weighting, the features of the frequency band will be amplified, so that it is easier to be identified in subsequent analysis. According to the attention weight coefficients of the time domain, frequency domain and time-frequency domain, the adjusted feature vector is obtained, and the adjusted feature vector is feature extracted by a convolutional neural network (CNN) to obtain a high-discriminative sound feature. CNN is good at processing high-dimensional data and can automatically extract deep features in the data. For example, the adjusted feature vector is input into a CNN model containing multiple convolutional layers and pooling layers, and the model extracts more discriminative sound features through layer-by-layer abstraction. These features can more accurately reflect the state of the blade. The extracted sound features are input into a support vector machine (SVM) classifier to judge the state of the blade. SVM is an efficient classification algorithm that can map data to a high-dimensional space.

[0059] For the above comprehensive feature vector, the self-attention mechanism calculates the correlation between each time-frequency region and all other time-frequency regions, thereby obtaining the attention weight coefficient of each time-frequency region. Assuming that the attention weight coefficient of a time-frequency region is 0.8, while the weight coefficient of another region is 0.2, this indicates that the former region is more important in describing the blade state. According to the attention weight coefficient, the importance of each time-frequency region is adaptively adjusted to highlight the key features that are sensitive to the blade state. Through weighted processing, the attention weight coefficient is multiplied by the corresponding time-frequency region value to make the features of the important region more significant. After weighting, the features of the frequency band will be amplified, making it easier to identify in subsequent analysis. CNN extracts spatial features, while support vector machine (SVM) handles classification tasks, combined with the attention mechanism to enhance the attention of important spatial regions, and finally integrates spatial and temporal features to improve the expressiveness of the model. This multi-level and multi-technical combination enables the adaptive adjustment of feature importance and improves the reliability of wind turbine blade sound analysis.

[0060] Furthermore, obtaining the adjusted feature vector includes: multiplying the attention weight coefficient by the corresponding time-frequency area value, the frequency domain area value and the time-frequency area value through weighted processing to obtain the adjusted feature vector.

[0061] Furthermore, obtaining the blade state diagnosis model includes: training a fully connected layer of an existing blade state diagnosis model according to a target training set to obtain the blade state diagnosis model.

[0062] Specifically, for the target blade model of the wind turbine to be tested, a small number of voiceprint samples of this model are selected, and the pre-trained basic model is used to extract features to obtain the voiceprint feature representation of the target blade model. Some network layer parameters of the basic model are fixed, and the voiceprint samples of the target blade model are used to train the fully connected layer. The model parameters are optimized through the back propagation algorithm to quickly adapt to the sound feature distribution of the target blade model. During the fine-tuning training process, the number of training rounds is controlled through strategies such as early stopping to avoid overfitting and improve the generalization ability of the model. The fine-tuned model is applied to the sound recognition task of the target blade model to achieve voiceprint recognition across blade models and improve the adaptability and efficiency of wind turbine fault diagnosis.

[0063] Furthermore, collecting the initial sound signal of the wind turbine blade to be detected includes: deploying a sound sensor array on the wind turbine blade to be detected, collecting the initial sound signal and transmitting the collected initial sound signal to a cloud server in real time through a wireless communication module.

[0064] Specifically, multiple high-precision microphones can be installed at different parts of the blade, and these microphones can capture various sound signals generated by the blade during rotation. Assume that four microphones are installed on a wind turbine blade, located at the leading edge, trailing edge, middle and root of the blade, so that the sound characteristics of the blade at different positions can be fully collected. The collected sound signal data is transmitted to the cloud server in real time through the wireless communication module. In this process, data compression and encryption are performed. For example, using the JPEG2000 algorithm to compress the sound data can greatly reduce the amount of data transmission while ensuring data quality. At the same time, the AES encryption algorithm is used to encrypt the data to ensure that the data is not stolen or tampered with during transmission. Assuming that the size of the original sound data is 100MB, it can be reduced to 20MB after compression, and the transmission efficiency is significantly improved. On the cloud server, the received sound signal data is preprocessed.

[0065] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A wind turbine blade soundprint recognition and detection method, characterized in that: include: Collecting the initial sound signal of the wind turbine blade to be detected and preprocessing it to obtain the sound signal, performing time-frequency domain decomposition on the sound signal, and obtaining the time-frequency domain decomposition result; Extracting and fusing feature vectors from the time-frequency domain decomposition results to obtain a comprehensive feature vector; The comprehensive feature vector is input into a blade state diagnosis model, and the state of the wind turbine blade to be detected is output, wherein the blade state diagnosis model is obtained by training an existing blade state diagnosis model with a target training set, and the existing blade state diagnosis model is constructed by an attention mechanism, a convolutional neural network, and a support vector machine classifier and is obtained by training with a large-scale training set, the target training set is a historical sound signal of the same model as the wind turbine to be detected and the corresponding state label, and the large-scale training set is a historical sound signal of blades of different models and the corresponding state label.

2. The wind turbine blade soundprint recognition and detection method according to claim 1, characterized in that: Preprocessing the initial sound signal includes: Performing filtering on the initial sound signal to obtain a sound signal after filtering; Adopting an adaptive filtering algorithm to estimate the current environmental noise spectrum characteristics in real time, adjusting the parameters of the spectrum subtraction algorithm according to the estimated current environmental noise spectrum characteristics, and obtaining an optimized spectrum subtraction algorithm; The sound signal after the filtering process is subjected to noise reduction according to the optimized spectral subtraction algorithm to obtain the sound signal.

3. The wind turbine blade soundprint recognition and detection method according to claim 1, characterized in that: Decomposing the sound signal in the time-frequency domain includes: decomposing the sound signal in the time-frequency domain by using a wavelet analysis method to obtain decomposition results in the time domain, the frequency domain and the time-frequency joint domain.

4. The wind turbine blade soundprint recognition and detection method according to claim 2, characterized in that: Extracting and fusing the time-frequency domain decomposition results to obtain a comprehensive feature vector includes: According to the time domain decomposition result, the waveform characteristics of the sound signal in the time domain are extracted to obtain a time domain feature vector, wherein the waveform characteristics include signal amplitude, energy and zero-crossing rate; According to the frequency domain decomposition result, extract the spectrum characteristics of the sound signal in the frequency domain to obtain a frequency domain feature vector, wherein the spectrum characteristics include the frequency distribution, spectrum peak position, and spectrum peak amplitude of the sound signal; According to the time-frequency joint domain decomposition result, extract the characteristics of the sound signal in the time-frequency joint domain, and obtain the time-frequency joint domain feature vector, wherein the characteristics of the time-frequency joint domain are the distribution of the sound signal on the time-frequency plane; The time domain feature vector, the frequency domain feature vector and the time-frequency joint domain feature vector are fused to obtain the comprehensive feature vector.

5. The wind turbine blade soundprint recognition and detection method according to claim 1, characterized in that: Training the existing blade status diagnosis model through a large-scale training set includes: Obtain historical comprehensive feature vectors of historical sound signals of blades of different models; Input the historical comprehensive feature vector into the attention mechanism, calculate the attention weight coefficients in the time domain, frequency domain and time-frequency domain through weight learning, and obtain the adjusted feature vector; Inputting the adjusted feature vector into the convolutional neural network for feature extraction to obtain high-discriminative features; The high-discrimination feature is input into the support vector machine classifier to output the blade status, wherein the blade status includes a normal state, slight wear, and severe wear.

6. The wind turbine blade soundprint recognition and detection method according to claim 5, characterized in that: Obtaining the adjusted feature vector includes: multiplying the attention weight coefficient by the corresponding time-frequency area value, frequency domain area value and time-frequency area value through weighted processing to obtain the adjusted feature vector.

7. The wind turbine blade soundprint recognition and detection method according to claim 1, characterized in that: Acquiring the blade state diagnosis model includes: training the fully connected layer of the existing blade state diagnosis model according to a target training set to acquire the blade state diagnosis model.

8. The wind turbine blade soundprint recognition and detection method according to claim 1, characterized in that: The collecting of the initial sound signal of the wind turbine blade to be detected includes: arranging a sound sensor array on the wind turbine blade to be detected, collecting the initial sound signal and transmitting the collected initial sound signal to a cloud server in real time through a wireless communication module.

Images