Time-frequency domain feature extraction method based on voiceprint features of offshore wind turbine blade

Abstract

The invention discloses an offshore wind turbine blade voiceprint characteristic-based time-frequency domain feature extraction method, which relates to the technical field of sound signal processing and comprises the following steps of: acquiring offshore wind turbine blade voiceprint information; noise reduction is carried out on the voiceprint data through a self-adaptive wavelet transform and EMD combined algorithm and Kalman filtering, and a time domain feature is obtained by calculating short-time average energy and a zero-crossing rate; the method comprises the following steps: firstly, carrying out pre-emphasis, framing, windowing, fast Fourier transform, Mel filtering, logarithmic transformation and discrete cosine transform on a voiceprint to extract frequency domain characteristics such as a voiceprint spectrum centroid, a spectrum bandwidth and MFCC, and finally, carrying out time-frequency domain characteristic analysis and extraction by using wavelet transform and STFT to comprehensively describe the time-varying characteristics of a sound signal for voiceprint blade state recognition. Therefore, the accuracy and the stability of the signal are effectively improved, time-frequency domain analysis is carried out during feature extraction, the comprehensiveness and the universality of the voiceprint features of the leaf are effectively improved, and the extracted features improve the accuracy of identifying the state of the leaf.

Description

Technical Field

[0001] The present invention relates to the technical field of sound signal processing, and in particular to a time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades. Background Art

[0002] Nowadays, the global demand for renewable energy continues to grow. As a key part of clean energy, offshore wind power is expanding rapidly in scale and quantity. Offshore wind farms are often located in remote seas, with harsh environments and high maintenance costs. As a core component, the status of wind turbine blades is directly related to the performance and life of the wind turbine. It is crucial to develop a system to monitor the status of wind turbine blades in real time. However, existing monitoring technologies (such as telescope monitoring, vibration detection, and drone inspections) cannot accurately reflect the working status of wind turbine blades in real time.

[0003] In recent years, patent papers have proposed using neural network models (such as convolutional neural networks (CNN) or recurrent neural networks (RNN)) to extract the time-frequency domain features of sound signals. Convolutional neural networks (CNN) capture local features and hierarchical structures to extract the correlation between frequency and time. Recurrent neural networks (RNN) and their variants (such as LSTM and GRU) process sequence data and model long-term dependencies in the time dimension to capture the temporal characteristics of sound. In addition, the RegNet network composed of three parts, stem, body and head, is used to train the speed of voiceprint signal analysis. The optimized RegNet neural network is applied to the recognition of spectrograms to achieve feature learning and classification of spectrograms, thereby achieving voiceprint identification.

[0004] However, the prior art also has the following disadvantages: (1) Poor noise filtering effect: When processing the soundprint signal of offshore wind turbine blades, the existing technology has limited noise filtering ability for complex non-stationary signals and cannot effectively separate the noise components in the sound signal, resulting in low quality of the picked-up signal, affecting the accuracy of subsequent feature extraction and state recognition.

[0005] (2) Feature extraction is not comprehensive enough: Extracting soundprint features only from a single time domain or frequency domain cannot fully describe the time-varying characteristics of the sound signal, resulting in the extracted features being insufficiently accurate and universal in identifying blade status, making it difficult to meet the high-precision requirements of offshore wind turbine blade status monitoring.

[0006] (3) Insufficient real-time and reliability: Existing monitoring technologies such as telescope monitoring, vibration detection, and drone inspections cannot accurately reflect the working status of wind turbine blades in real time, making it difficult to achieve early fault warning and status monitoring, and cannot meet the needs of offshore wind farms for real-time monitoring of wind turbine blades. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, comprising the following steps: S1. Collect the sound signals of offshore wind turbine blade operation as soundprint data; S2, preprocessing the collected sound signal, using adaptive wavelet transform and empirical mode decomposition combined with noise reduction algorithm and Kalman filter to separate the noise component in the sound signal, and filter it out to obtain a noise-reduced sound signal; S3. Use the noise reduction sound signal to analyze the time domain characteristics of the offshore wind turbine blade vibration, calculate the short-time average energy of each frame, and use it to distinguish the blade vibration signal from the background noise; by setting a threshold, remove the background noise and irrelevant signals, and retain the blade vibration signal with short-time energy exceeding the threshold; further calculate the short-time zero-crossing rate to distinguish high-frequency vibration from low-frequency vibration, and distinguish the effective vibration segment from the silent segment; S4, pre-emphasize the processed data obtained in the previous step to enhance the energy of the high-frequency part of the voiceprint signal to obtain a pre-emphasized signal; S5, dividing the pre-emphasized signal obtained in the previous step into frames to obtain each frame of data; S6, applying a window function to each frame of data obtained in the previous step to smooth the signal boundary, thereby obtaining each frame of signal after windowing; S7, converting each frame of the windowed signal obtained in the previous step into the frequency domain by fast Fourier transform, obtaining the spectrum of each frame, and the corresponding spectrum centroid and spectrum bandwidth characteristics; S8, setting a Mel filter for feature extraction, decomposing the spectrum into an energy distribution adapted to the Mel scale, and obtaining the energy of each frequency band; S9, the energy of each frequency band obtained in the previous step is represented by logarithmic transformation to simulate the perceptual characteristics and obtain the logarithmic Mel frequency band energy; S10, the logarithmic Mel band energy obtained in the previous step is processed using discrete cosine transform, the data is converted from the frequency domain back to the cepstrum domain, the first 20 Mel cepstrum coefficients are extracted, and the 20-dimensional voiceprint MFCC features are obtained; S11, processing the noise-reduced sound signal obtained in step S2 by wavelet transform to obtain wavelet transform coefficients, providing frequency features for subsequent recognition; S12. For the noise-reduced sound signal obtained in step S2, segment the signal using a sliding window function through STFT, apply Fourier transform to each segment, obtain a time-frequency distribution diagram of the signal, retain the time evolution information, and reveal the characteristics of the spectrum characteristics changing with time.

[0008] The technical solution further defined in the present invention is: Furthermore, in step S2, the real state of the system is estimated by integrating the observed data through Kalman filtering, and its state prediction is as follows: Among them, k represents the current time step, k-1 represents the previous time step, is the prediction of the current state based on the previous state, F k-1 is the state transition matrix, is the optimal estimate of the state at the previous moment, B k-1 is the control input matrix, u k-1 is the control input; The observation update is as follows: Among them, K k is the Kalman gain, P k∣k-1 is the forecast error covariance matrix, H k is the observation matrix, R k is the observation noise covariance matrix, z k is the observed value at the current moment, is the optimal estimate of the current state, P k∣k is the estimation error covariance matrix at the current moment.

[0009] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S3, the calculation formula of the short-time average energy is: Among them, x(n) is the sound signal, w(m) is the window function, and N is the window length; The calculation formula of short-time zero-crossing rate is: where sgn[·] is the sign function; By calculating the short-time average energy of the sound signal and comparing it with the preset threshold, it is determined whether the sound signal exists. When the short-time average energy exceeds the threshold, the sound signal begins or continues to exist; when the short-time average energy is lower than the threshold, the sound signal is considered to have ended. The formula is: Among them, E n is the short-time average energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length.

[0010] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S4, pre-emphasis is performed through a pre-emphasis filter, and the pre-emphasis is implemented by performing a first-order difference operation on the signal, and the formula is: Among them, x[n] represents the current sampling point of the input signal, x[n-1] represents the previous sampling point, α It is the pre-emphasis coefficient, which is used to control the degree of enhancement of the high-frequency part.

[0011] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S5, the pre-emphasized signal is divided into short time frames of 16 ms, the frame shift is 10 ms, and each frame signal has 37.5% overlap.

[0012] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S6, the window function is set to a Hamming window, and the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame.

[0013] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S7, a fast Fourier transform is performed on the windowed signal of each frame to convert the time domain signal into a frequency domain signal. The formula is as follows: Where X[k] is the complex value of the kth frequency component of the frequency domain signal, x[n] is the nth sampling point of the time domain signal, and N is the number of sampling points of each frame of the signal. The spectrum amplitude and phase information of each frame of the signal are obtained by fast Fourier transform calculation. The spectrum centroid reflects the center of spectrum distribution: Where f(x) is the frequency of the kth band, is the amplitude of the kth frequency component of the frequency domain signal; The spectrum bandwidth measures the width of the spectrum distribution: .

[0014] As described above, in the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S8, a Mel filter with a center frequency in the frequency range of 2000~8000Hz is selected for subsequent Mel cepstral coefficient extraction, and finally the filter F is obtained. total=[f5,f4,f3,...,f1,F0,F1,F2,...,F7], after applying Mel filter to perform weighted processing on the spectrum, the output of each filter is obtained; the output of each filter represents the energy of the signal in the Mel band; the spectrum signal obtained by fast Fourier transform is convolved with the Mel filter to convert the spectrum signal from the linear frequency scale to the Mel scale, and the energy value of each Mel band is obtained.

[0015] As described above, a time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, in step S9, the energy value output by each Mel filter is logarithmically transformed, assuming that e m is the energy output of the mth Mel filter, and the energy after logarithmic transformation is expressed as: Among them, E m is the energy value after logarithmic transformation, A floating point number used to avoid zero logarithmic value problems.

[0016] As described above, a time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades is used. In step S10, a logarithmic Mel band energy sequence E=[E1, E2, E3, ..., E 20 ], the formula for discrete cosine transform is as follows: Among them, C k is the kth coefficient in the cepstrum domain, E m is the mth logarithmic Mel band energy, M is the number of Mel bands, and L is the MFCC dimension. Through the above calculation, the 20-dimensional Mel cepstral coefficient is obtained, which is the extracted voiceprint MFCC feature.

[0017] The beneficial effects of the present invention are: (1) In the present invention, an adaptive wavelet transform and EMD joint noise reduction algorithm and Kalman filter are used to integrate the localization characteristics of wavelet transform in the time-frequency domain and the ability of EMD decomposition to process complex non-stationary signals, thereby effectively filtering out background noise, improving the quality of the picked-up signal, and providing a clearer and more accurate signal basis for subsequent feature extraction; (2) In the present invention, not only the time domain features such as short-time average energy and short-time zero-crossing rate are calculated, but also the frequency domain features such as spectrum centroid, spectrum bandwidth and MFCC are extracted by means of fast Fourier transform, Mel filtering and the like, and the time-frequency domain feature analysis and extraction are performed using wavelet transform and STFT, which comprehensively describes the time-varying characteristics of the sound signal, making the extracted features richer and more comprehensive, thereby improving the accuracy and universality of identifying the blade status; (3) In the present invention, by performing time-frequency domain analysis on the sound signals generated by the wind turbine blades during operation, the status of the wind turbine blades can be monitored in real time, the characteristic patterns therein can be identified, early fault warning and status monitoring can be achieved, the operation and maintenance costs can be reduced, the demand for real-time monitoring of wind turbine blades in offshore wind farms can be met, and the operation efficiency and safety of offshore wind farms can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0019] This embodiment provides a time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, such as Figure 1 As shown, the following steps are included: S1. Collect the sound signals of offshore wind turbine blades as soundprint data. Use a high-precision acoustic sensor array to collect blade soundprint signals, including sound level meters, vibration sensors, and temperature sensors, to collect key parameters of soundprint, vibration, and temperature respectively.

[0020] The collection is divided into two stages: preliminary collection and refined collection: The primary collection stage records blade sound and key environmental parameters such as unit status, wind speed, and wind direction to compare the soundprint characteristics under different conditions and preliminarily reveal how blade sound changes with operating conditions.

[0021] The refined collection stage focuses on wind turbines with suspected blade cracks or faults. After preliminary analysis, the target units are screened, the microphone position and angle are adjusted, the collection frequency is increased, and the sound signals of blade operation are accurately captured; sound signal samples under different blade states (normal, holes, cracks, corrosion, deformation) are collected to construct voiceprint features.

[0022] S2. Preprocess the collected sound signal, use adaptive wavelet transform and empirical mode decomposition (EMD) combined with noise reduction algorithm and Kalman filter to effectively separate the noise component in the sound signal, and effectively filter it out to obtain a noise-reduced sound signal.

[0023] Noise reduction is the basic step of sound signal preprocessing, and its main goal is to filter out the interference of environmental noise and wind noise. Wavelet transform is a powerful signal analysis tool that can provide high resolution in both time domain and frequency domain, and is particularly suitable for the processing of non-stationary signals. By selecting appropriate wavelet basis functions and decomposition layers, the noise components in the sound signal can be effectively separated and filtered out.

[0024] On the basis of wavelet transform, an adaptive wavelet transform and EMD joint noise reduction algorithm is proposed, which integrates the localization characteristics of wavelet transform in the time and frequency domain and the ability of empirical mode decomposition (EMD) to decompose and process complex non-stationary signals, effectively filters out background noise, improves the quality of picked-up signals, decomposes and reconstructs signals, and extracts vibration frequency, amplitude and modal energy distribution characteristic parameters. At the same time, the algorithm parameters are optimized to achieve high-sensitivity acquisition of long-distance sound waves, and can capture clear blade soundprint signals even in the complex environment of offshore wind farms.

[0025] The specific steps of the adaptive wavelet transform and EMD joint denoising algorithm are as follows: S2.1, signal preprocessing: first input the signal and receive the original voiceprint signal, which usually contains background noise and target signal; then perform frame processing and divide the signal into multiple short time frames for subsequent processing.

[0026] S2.2, Empirical Mode Decomposition (EMD): First decompose the signal, perform EMD decomposition on the original voiceprint signal, and decompose it into several intrinsic mode functions (IMFs) and a residual term; IMF is a local feature representation of the signal and can reflect the non-stationary characteristics of the signal; the residual term usually represents the low-frequency trend or background noise of the signal; then filter the IMF, and according to the energy distribution or correlation of each IMF, filter out the IMF components related to the background noise, and retain the IMF components related to the target signal.

[0027] S2.3, Adaptive wavelet transform: First, select the wavelet basis. According to the signal characteristics, select the appropriate wavelet basis function (such as Morlet wavelet, Daubechies wavelet, etc.). The choice of wavelet basis will affect the noise reduction effect; then decompose the signal, perform wavelet decomposition on the retained IMF components, and decompose them into wavelet coefficients at different levels (including approximate coefficients and detail coefficients); then threshold processing, threshold processing is performed on the wavelet coefficients to remove noise components. The threshold selection can be fixed or adaptive (such as dynamic adjustment based on the statistical characteristics of the signal); then soft thresholding, set the coefficients less than the threshold to zero, and shrink the coefficients greater than the threshold; then hard thresholding, directly set the coefficients less than the threshold to zero, and retain the coefficients greater than the threshold; finally reconstruct the signal, perform wavelet reconstruction on the processed approximate coefficients and detail coefficients to obtain the denoised signal.

[0028] S2.4, signal fusion and optimization: First, signal fusion is performed to fuse the denoised IMF component with the residual term (if necessary) to obtain the final denoised voiceprint signal; then parameter optimization is performed to optimize the wavelet basis selection, EMD decomposition parameters and threshold processing strategy according to the denoising effect evaluation indicators (such as signal-to-noise ratio, mean square error, etc.) to further improve the denoising effect.

[0029] S2.5. Output noise reduction signal: Output the result, output the noise-reduced voiceprint signal for subsequent feature extraction and fault diagnosis.

[0030] Furthermore, Kalman filtering is a recursive linear minimum variance estimation algorithm that can integrate a series of observation data to estimate the most realistic state of the system under the premise of noise interference. Its state prediction is as follows: Among them, k represents the current time step, k-1 represents the previous time step, is the prediction of the current state based on the previous state, F k-1 is the state transition matrix, is the optimal estimate of the state at the previous moment, B k-1 is the control input matrix, u k-1 is the control input.

[0031] The observation update is as follows: Among them, K k is the Kalman gain, P k∣k-1 is the forecast error covariance matrix, H k is the observation matrix, R k is the observation noise covariance matrix, z k is the observed value at the current moment, is the optimal estimate of the current state, P k∣k is the estimation error covariance matrix at the current moment.

[0032] S3. Use the noise reduction sound signal to analyze the time domain characteristics of the offshore wind turbine blade vibration, calculate the short-time average energy of each frame, and use it to distinguish the representative blade vibration signal from the background noise; by setting the threshold, remove the background noise and non-correlated signals, and retain the blade vibration signal with higher short-time energy; further calculate the short-time zero-crossing rate to distinguish high-frequency vibration from low-frequency vibration, and identify the effective vibration segment from the silent segment.

[0033] According to GB / T 41973-2022 and ISO 10816-1 standards, when the vibration displacement of a small fan (power less than 15kW) exceeds 0.05mm, the vibration displacement of a medium-sized fan (power between 15kW and 300kW) exceeds 0.1mm, the vibration displacement of a large fan (power greater than 300kW) exceeds 0.2mm, or the vibration speed of a centrifugal fan exceeds 4.5mm / hour, the blade vibration signal is considered to have high short-term energy.

[0034] The collected blade operation sound signals are processed in frames, and the short-time energy and zero-crossing rate of each frame signal are calculated. By setting the thresholds of short-time energy and zero-crossing rate, the atypical audio signals of offshore wind turbine blades are identified and filtered out, and only the blade vibration signals with high short-time energy and low zero-crossing rate are retained. The retained high-energy signal part can more accurately reflect the vibration mode and harmonic characteristics of the blade, provide a more reliable basis for subsequent feature extraction, and further enhance the accuracy and stability of wind turbine blade soundprint recognition.

[0035] Short-Time Average Energy (STAE) measures the energy of a sound signal in a short period of time and reflects the change in sound intensity. It is crucial to distinguish low-energy signals, high-energy vibration segments, and background noise segments. The calculation formula is: Among them, x(n) is the sound signal, w(m) is the window function, and N is the window length.

[0036] The short-time zero-crossing rate refers to the number of times a sound signal crosses the zero level in a short period of time, reflecting the frequency characteristics of the signal and assisting in distinguishing high-frequency and low-frequency sounds. The calculation formula is: where sgn[·] is the sign function.

[0037] By calculating the short-time average energy of the sound signal and comparing it with the preset threshold, it is determined whether the sound signal exists. When the short-time average energy exceeds the threshold, the sound signal begins or continues to exist; when the short-time average energy is lower than the threshold, the sound signal is considered to have ended. The formula is: Among them, E n is the short-time average energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length.

[0038] S4. Pre-emphasize the processed data obtained in the previous step to enhance the energy of the high-frequency part of the voiceprint signal to obtain a pre-emphasized signal.

[0039] A pre-emphasis filter is applied to the retained high-energy vibration signal. Pre-emphasis is achieved by performing a first-order difference operation on the signal, as shown in the formula: Among them, x[n] represents the current sampling point of the input signal, x[n-1] represents the previous sampling point, αis the pre-emphasis coefficient, which is 0.97 and is used to control the degree of enhancement of the high-frequency part. The frequency range of random wind noise is concentrated below 500 Hz, which is a low-frequency signal. The sound signal of the faulty blade is a high-frequency signal compared to the wind noise frequency, and there is an obvious frequency difference. Therefore, through pre-emphasis processing, the high-frequency part of the signal can be made more prominent.

[0040] S5. Divide the pre-emphasized signal obtained in the previous step into frames to obtain each frame of data; divide the originally non-stationary signal into short-time frames of 16ms, and each frame signal is considered to be stable; the frame shift is 10ms, and each frame signal has 37.5% overlap to ensure continuity between frames and reduce information loss between frames.

[0041] S6. Apply a window function to each frame of data obtained in the previous step to smooth the signal boundary and obtain each frame of signal after windowing.

[0042] A window function is applied to the signal samples of each frame to reduce the distortion effect of the signal at the frame boundary. In this embodiment, a Hamming window is used, and the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame. By multiplying the window function point by point with each frame of signal, the boundary of the signal can be smoothed and the spectrum leakage effect can be reduced.

[0043] S7. Convert each frame of the windowed signal obtained in the previous step into the frequency domain through Fast Fourier Transform (FFT) to obtain the spectrum of each frame, as well as the corresponding spectrum centroid and spectrum bandwidth characteristics.

[0044] Perform fast Fourier transform on the windowed signal of each frame to convert the time domain signal into the frequency domain signal. The formula is as follows: Wherein, X[k] is the complex value of the kth frequency component of the frequency domain signal, x[n] is the nth sampling point of the time domain signal, N is the number of sampling points of each frame signal, and N=512 in this embodiment; the spectrum amplitude and phase information of each frame signal are obtained by fast Fourier transform calculation.

[0045] The spectrum centroid reflects the center of spectrum distribution: Where f(x) is the frequency of the kth band, It is the amplitude of the kth frequency component of the frequency domain signal; the centroid of the spectrum reflects the brightness or sharpness of the sound.

[0046] The spectrum bandwidth measures the width of the spectrum distribution: .

[0047] S8. Set a Mel filter for feature extraction, decompose the spectrum into an energy distribution adapted to the Mel scale, and obtain the energy of each frequency band.

[0048] A Mel filter with a center frequency in the range of 2000~8000Hz is selected for subsequent Mel cepstral coefficient extraction, and the filter F is finally obtained. total =[f5,f4,f3,...,f1,F0,F1,F2,...,F7], after applying Mel filter to perform weighted processing on the spectrum, the output of each filter is obtained; the output of each filter represents the energy of the signal in the Mel band; the spectrum signal obtained by fast Fourier transform is convolved with the Mel filter to convert the spectrum signal from the linear frequency scale to the Mel scale, and the energy value of each Mel band is obtained.

[0049] S9. The energy of each frequency band obtained in the previous step is represented by logarithmic transformation to simulate the perceptual characteristics and obtain the logarithmic Mel frequency band energy.

[0050] Perform logarithmic transformation on the energy value output by each Mel filter, assuming that e m is the energy output of the mth Mel filter, and the energy after logarithmic transformation is expressed as: Among them, E m is the energy value after logarithmic transformation, It is a very small floating point number used to avoid the logarithmic zero problem. By performing a logarithmic transformation on the energy value of each frequency band, a wide range of energy values ​​can be effectively compressed so that stronger frequency components do not overly dominate the feature representation.

[0051] S10. The logarithmic Mel band energy obtained in the previous step is processed using Discrete Cosine Transform (DCT), the data is converted from the frequency domain back to the cepstral domain, the first 20 Mel cepstral coefficients are extracted, and the 20-dimensional voiceprint Mel-Frequency Cepstral Coefficients (MFCC) features are obtained.

[0052] DCT maps the logarithmic Mel band energy from the frequency domain back to the cepstrum domain. The cepstrum domain is usually more suitable for sound processing and feature extraction. It can effectively compress information and highlight the main features of the signal. Given a logarithmic Mel band energy sequence of length 20, E=[E1,E2,E3,...,E 20], the formula for discrete cosine transform is as follows: Among them, C k is the kth coefficient in the cepstrum domain, E m is the mth logarithmic Mel band energy, M is the number of Mel bands, L is the MFCC dimension, in this embodiment, M=13, L=20; the cepstrum domain can compress the noise component in the spectrum information, highlight the core mode of the sound signal, and make the feature more stable and discriminative in the subsequent recognition process; through the above calculation, the 20-dimensional Mel cepstrum coefficient is obtained, which is the extracted voiceprint MFCC feature.

[0053] S11. The wavelet transform coefficients obtained by processing the denoised sound signal through wavelet transform reflect the subtle changes of the signal at different scales and positions; the wavelet energy reveals the energy distribution of the signal in different frequency bands, providing frequency characteristics for subsequent identification; the wavelet transform uses wavelet functions to perform multi-scale analysis of the signal, retaining the time information and revealing the details of the signal at different frequency scales; in voiceprint recognition, the wavelet transform coefficients are the key, which can accurately reflect the changes of the signal at different scales and positions, and provide an important basis for frequency feature extraction.

[0054] S12. For the noise-reduced sound signal obtained in step S2, the signal is segmented by using a sliding window function through a short-time Fourier transform (STFT), and Fourier transform is applied to each segment to obtain a time-frequency distribution diagram of the signal, retaining the time evolution information and revealing the characteristics of the spectrum characteristics changing with time.

[0055] STFT uses a sliding window function to segment the signal, applies Fourier transform to each segment, obtains the signal time-frequency distribution diagram, retains the time evolution information, and reveals that the spectrum characteristics change over time. In voiceprint recognition, the time-frequency peak reflects the time-varying spectrum characteristics of the signal and is an identification clue; the time-frequency energy reveals the energy distribution of the signal at different times and frequencies, providing comprehensive time-frequency characteristics.

[0056] According to the time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades provided in this embodiment, a large amount of sound data of wind turbine blades in normal and faulty states (such as cracks, holes, corrosion) is collected, and an adaptive wavelet transform and EMD joint noise reduction algorithm are used. Compared with a single noise reduction method, this method combines the localization characteristics of wavelet transform in the time-frequency domain and the ability of EMD decomposition to process complex non-stationary signals, effectively filtering out background noise and improving the quality of picked-up signals. The recognition accuracy and false alarm rate in the test training phase are shown in Table 1 below.

[0057] Table 1

[0058] The intelligent noise reduction algorithm automatically identifies and filters background noise, retaining the effective information of the blade soundprint. Advanced feature extraction technology, namely the frequency domain feature extraction method, is used to further extract the unique features of the blade soundprint. After the system turns on the noise reduction function, the recognition accuracy rate is significantly improved by about 10%, reaching more than 94%. Even in extreme cases with noise levels as high as 100dB, the system can still maintain stable recognition performance, and the false alarm rate is controlled below 5%, which fully proves its robustness and reliability in complex noise environments. The test data is shown in Table 2 below.

[0059] Table 2

[0060] In this embodiment, an adaptive wavelet transform and EMD joint noise reduction algorithm is used. Compared with a single noise reduction method, this method combines the localization characteristics of wavelet transform in the time and frequency domain and the ability of EMD decomposition to process complex non-stationary signals, thereby effectively filtering out background noise and improving the quality of picked-up signals.

[0061] The voiceprint data is analyzed and extracted in the time and frequency domains, and the short-time average energy and short-time zero-crossing rate time domain features are calculated; the sound signal is converted from the time domain to the frequency domain by Fourier transform, and the signal frequency features are extracted, including the spectral centroid, spectral bandwidth and MFCC; finally, the time and frequency domain analysis is performed to connect the time domain and frequency domain information, and the wavelet transform and STFT method are used. The STFT and wavelet transform (WT) are combined to simultaneously reflect the time and frequency changes of the sound signal. STFT divides the sound signal into short-time signal segments and performs Fourier transform to reveal the changes of frequency components over time. WT uses wavelet functions to decompose the signal to obtain details and approximate information of different frequency segments, which is suitable for non-stationary signal processing. Combining the advantages of both, the noise reduction data is analyzed in time and frequency, and the voiceprint feature parameters are extracted to more comprehensively describe the time-varying characteristics of the voiceprint signal. Finally, richer voiceprint features are obtained.

[0062] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades, characterized by: The following steps are involved: S1. Collect the sound signals of offshore wind turbine blade operation as soundprint data; S2, preprocessing the collected sound signal, using adaptive wavelet transform and empirical mode decomposition combined with noise reduction algorithm and Kalman filter to separate the noise component in the sound signal, and filter it out to obtain a noise-reduced sound signal; S3, using the noise reduction sound signal, analyzing the time domain characteristics of the offshore wind turbine blade vibration, and calculating the short-time average energy of each frame to distinguish the blade vibration signal from the background noise; By setting a threshold, background noise and irrelevant signals are removed, and blade vibration signals with short-term energy exceeding the threshold are retained; the short-term zero-crossing rate is further calculated to distinguish high-frequency vibration from low-frequency vibration, and to distinguish effective vibration segments from silent segments; S4, pre-emphasize the processed data obtained in the previous step to enhance the energy of the high-frequency part of the voiceprint signal to obtain a pre-emphasized signal; S5, dividing the pre-emphasized signal obtained in the previous step into frames to obtain each frame of data; S6, applying a window function to each frame of data obtained in the previous step to smooth the signal boundary, thereby obtaining each frame of signal after windowing; S7, converting each frame of the windowed signal obtained in the previous step into the frequency domain by fast Fourier transform, obtaining the spectrum of each frame, and the corresponding spectrum centroid and spectrum bandwidth characteristics; S8, setting a Mel filter for feature extraction, decomposing the spectrum into an energy distribution adapted to the Mel scale, and obtaining the energy of each frequency band; S9, the energy of each frequency band obtained in the previous step is represented by logarithmic transformation to simulate the perceptual characteristics and obtain the logarithmic Mel frequency band energy; S10, the logarithmic Mel band energy obtained in the previous step is processed using discrete cosine transform, the data is converted from the frequency domain back to the cepstrum domain, the first 20 Mel cepstrum coefficients are extracted, and the 20-dimensional voiceprint MFCC features are obtained; S11, processing the noise-reduced sound signal obtained in step S2 by wavelet transform to obtain wavelet transform coefficients, providing frequency features for subsequent recognition; S12. For the noise-reduced sound signal obtained in step S2, segment the signal using a sliding window function through STFT, apply Fourier transform to each segment, obtain a time-frequency distribution diagram of the signal, retain the time evolution information, and reveal the characteristics of the spectrum characteristics changing with time.

2. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S2, the real state of the system is estimated by integrating the observed data through Kalman filtering, and its state prediction is as follows: Among them, k represents the current time step, k-1 represents the previous time step, is the prediction of the current state based on the previous state, F k-1 is the state transition matrix, is the optimal estimate of the state at the previous moment, B k-1 is the control input matrix, u k-1 is the control input; The observation update is as follows: Among them, K k is the Kalman gain, P k∣k-1 is the forecast error covariance matrix, H k is the observation matrix, R k is the observation noise covariance matrix, z k is the observed value at the current moment, is the optimal estimate of the current state, P k∣k is the estimation error covariance matrix at the current moment.

3. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S3, the calculation formula of short-time average energy is: Among them, x(n) is the sound signal, w(m) is the window function, and N is the window length; The calculation formula of short-time zero-crossing rate is: where sgn[·] is the sign function; By calculating the short-time average energy of the sound signal and comparing it with the preset threshold, it is determined whether the sound signal exists. When the short-time average energy exceeds the threshold, the sound signal begins or continues to exist; when the short-time average energy is lower than the threshold, the sound signal is considered to have ended. The formula is: Among them, E n is the short-time average energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length.

4. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S4, pre-emphasis is performed through a pre-emphasis filter, and the pre-emphasis is implemented by performing a first-order difference operation on the signal, and the formula is: Among them, x[n] represents the current sampling point of the input signal, x[n-1] represents the previous sampling point, α It is the pre-emphasis coefficient, which is used to control the degree of enhancement of the high-frequency part.

5. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S5, the pre-emphasized signal is divided into short time frames of 16 ms, the frame shift is 10 ms, and each frame signal has 37.5% overlap.

6. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S6, the window function is set to a Hamming window, and the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame.

7. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S7, a fast Fourier transform is performed on the windowed signal of each frame to convert the time domain signal into a frequency domain signal, and the formula is as follows: Where X[k] is the complex value of the kth frequency component of the frequency domain signal, x[n] is the nth sampling point of the time domain signal, and N is the number of sampling points of each frame of the signal. The spectrum amplitude and phase information of each frame of the signal are obtained by fast Fourier transform calculation. The spectrum centroid reflects the center of spectrum distribution: Where f(x) is the frequency of the kth band, is the amplitude of the kth frequency component of the frequency domain signal; The spectrum bandwidth measures the width of the spectrum distribution: 。 8. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S8, a Mel filter with a center frequency in the frequency range of 2000-8000 Hz is selected for subsequent Mel cepstral coefficient extraction, and finally the filter F is obtained. total =[f5,f4,f3,...,f1,F0,F1,F2,...,F7], after applying Mel filter to perform weighted processing on the spectrum, the output of each filter is obtained; the output of each filter represents the energy of the signal in the Mel band; the spectrum signal obtained by fast Fourier transform is convolved with the Mel filter to convert the spectrum signal from the linear frequency scale to the Mel scale, and the energy value of each Mel band is obtained.

9. The method for extracting time-frequency domain features based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S9, the energy value output by each Mel filter is logarithmically transformed. Assume that e m is the energy output of the mth Mel filter, and the energy after logarithmic transformation is expressed as: Among them, E m is the energy value after logarithmic transformation, A floating point number used to avoid zero logarithmic value problems.

10. The time-frequency domain feature extraction method based on the soundprint characteristics of offshore wind turbine blades according to claim 1 is characterized by: In step S10, a logarithmic Mel band energy sequence E=[E1, E2, E3, ..., E 20 ], the formula for discrete cosine transform is as follows: Among them, C k is the kth coefficient in the cepstrum domain, E m is the mth logarithmic Mel band energy, M is the number of Mel bands, and L is the MFCC dimension. Through the above calculation, the 20-dimensional Mel cepstral coefficient is obtained, which is the extracted voiceprint MFCC feature.

Images