Wind turbine blade defect detection method based on acoustic emission

Abstract

The application discloses a kind of wind power blade defect detection methods based on acoustic emission, and the application relates to wind power blade detection technical field, the application is equipped with multiple laser, multiple spectral camera, thermal imaging sensor, several acoustic emission sensors on unmanned aerial vehicle, obtain multiple spectral image and thermal imaging image in fixed time, obtain total discriminant sequence of multiple spectral image and total discriminant sequence of thermal imaging image by processing, suspicious pixel points in two sequences are all mapped as actual three-dimensional coordinates of wind power blade, light reflection defect area, thermal distribution defect area are obtained by fitting, the specific area of wind power blade defect is obtained by the intersection of two areas, the area is positioned to the area to be detected acoustic signal using acoustic emission sensor, and after LMD is combined with MFSE denoising, feature vector is extracted, and substituted into GRU wind power blade defect judgment model, to judge whether there is internal fault.The application realizes the comprehensive perception and correlation judgment of internal and external defects of wind power blade.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade inspection technology, specifically to a method for detecting defects in wind turbine blades based on acoustic emission. Background Technology

[0002] As the core load-bearing component of wind power generation systems, wind turbine blades operate in complex and ever-changing environments for a long time. Their structural integrity is directly linked to the overall operating efficiency of the unit. Existing detection technologies mainly focus on internal and external defects of the blades, but due to the limited detection methods and insufficient information integration, the problem of missed detection occurs frequently.

[0003] For internal defects, traditional methods mainly rely on acoustic emission detection technology. However, wind turbine blades are mostly made of fiber-reinforced composite materials, and their anisotropy makes the acoustic signal prone to scattering and attenuation during propagation. In addition, the complex environmental noise around the acoustic emission sensor can seriously contaminate the original signal, causing weak defect signals to be masked. For external defects, existing technologies generally use image recognition methods, but such methods are often affected by factors such as lighting conditions and surface stains. Furthermore, external defects often occur together with internal damage. For example, surface cracks may extend into the interior and cause structural layer failure. However, existing technologies only detect in a single dimension and cannot establish a correlation analysis between internal and external defects, which increases the risk of missing composite defects.

[0004] Therefore, there is an urgent need for a comprehensive detection method that can integrate multi-source detection data and take into account both internal and external defect characteristics, in order to overcome the limitations of single detection methods in existing technologies and improve the comprehensiveness and accuracy of defect detection. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a wind turbine blade defect detection method based on acoustic emission, which solves the problems of easy omissions and lack of multi-modal fusion in existing technologies for detecting internal and external defects of wind turbine blades when detected separately.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a wind turbine blade defect detection method based on acoustic emission, comprising:

[0007] S101. Based on the acquired multispectral images, calculate the four-dimensional spectral vector corresponding to each pixel, calculate the angle between it and the reference spectral vector of the normal area, and if it is greater than the angle threshold, mark the pixel as a suspicious defect pixel to obtain a single multispectral image discrimination sequence. Process multiple discrimination sequences to obtain the total discrimination sequence, and map the suspicious pixels to the actual three-dimensional coordinates of the wind turbine blade and mark them with 0.

[0008] S102. Calculate the temperature value of each pixel in each thermal imaging image to obtain the temperature value sequence within the time period. Calculate the total rate of change, mean offset rate, standard deviation coefficient, periodicity, and coefficient of determination. If the total rate of change is not within the specified range, or the mean offset rate is greater than the first threshold, or the periodicity is less than the second threshold, initiate a first-level warning and proceed to the second-level confirmation. If the standard deviation coefficient is greater than the third threshold and the coefficient of determination is less than the fourth threshold, determine that the pixel is a suspicious pixel. Obtain the total discrimination sequence of the thermal imaging image. Map the suspicious pixel to the actual three-dimensional coordinates and mark it with 1.

[0009] S103. Perform directional bounding box fitting on the three-dimensional coordinates marked 0 and 1 on the wind turbine blade to obtain the light reflection defect and heat distribution defect areas respectively. Take the intersection to determine the specific area of ​​the surface defect of the wind turbine blade. Use the UAV acoustic emission sensor to locate the area and obtain the acoustic signal to be detected. After LMD combined with MFSE denoising, extract the feature vector and substitute it into the GRU wind turbine blade defect judgment model to further determine whether there is an internal fault in the area.

[0010] As a further aspect of the present invention, it also includes S100, wherein S100 is achieved by mounting a multi-laser, a multispectral camera, a thermal imaging sensor and several acoustic emission sensors on a drone, and by illuminating the wind turbine blades with the multi-laser on the drone, and capturing several multispectral images and thermal images within a fixed time period using the multispectral camera and the thermal imaging sensor respectively.

[0011] As a further aspect of the present invention, the wavelength bands selected by the multiple lasers include the visible band, the near-infrared band, and the short-wave infrared band.

[0012] As a further aspect of the present invention, when collecting information from wind turbine blades, the specific requirements for the surrounding environment include: the ambient temperature must be stable between 10-30℃, the humidity must be <= 80%RH, the wind speed must be <= 5m / s, multispectral imaging requires uniform natural light, thermal imaging must avoid direct sunlight, stay away from strong electromagnetic sources, and the temperature difference between the background temperature of the shooting area and the blade temperature must be < 5℃.

[0013] As a further aspect of the present invention, the discrimination sequences of multiple multispectral images are processed, and the proportions of the number of suspicious and normal pixels in each pixel, PerDou and PerNor, are calculated. The max{PerDou,PerNor} is selected as the overall judgment of the pixel to obtain the overall discrimination sequence of the multispectral image.

[0014] As a further aspect of the present invention, if the number of multiple multispectral images is even, the percentage of suspicious pixels PerNei within a 3×3 neighborhood around the pixel is calculated. If PerNei>=0.5, the current pixel is determined to be suspicious; otherwise, it is determined to be normal. If the pixel is a corner pixel, the effective neighborhood is three pixels; if the pixel is an edge pixel, the effective neighborhood is five pixels. The corner pixels are the top left, bottom left, top right, and bottom right, and the edge pixels are the middle left, middle right, top middle, and bottom middle.

[0015] As a further aspect of the present invention, the specific method for constructing the dynamic threshold for each group of TVRs is as follows:

[0016] M sets of data were collected from the same area of ​​the defect-free blade. The TVR of each pixel was calculated, and a normal distribution TVRnormal~N(μ,σ) was constructed. 2 );

[0017] Let Crmin = μ - k × σ and Crmax = μ + k × σ, where k is the quality and reliability coefficient, which can be 2 or 3.

[0018] As a further aspect of the present invention, the specific steps for denoising acoustic signals using LMD combined with MFSE are as follows:

[0019] The original acoustic signal is decomposed into multiple product function components PF and residual terms using LMD;

[0020] Calculate the mean MFSE Argmfse of each PF. If Argmfse>=Argth, then classify the PF component as a noise component; otherwise, classify it as a defect signal component.

[0021] The denoised acoustic signal is obtained by adding the defect signal component and the residual term.

[0022] As a further aspect of the present invention, time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features are extracted from the denoised acoustic signal. The time-domain features include peak value, root mean square value, rise time, pulse width, and zero-crossing rate. The frequency-domain features include the dominant frequency, band energy, skewness, and kurtosis of the power spectral density (PSD) of the Fast Fourier Transform (FFT). The time-frequency-domain features include wavelet energy entropy and variance of the Short Time Fourier Transform (STFT). The nonlinear features include permutation entropy (PE) and fractal dimension (FD).

[0023] As a further aspect of the present invention, the initial parameters of the GRU model specifically include an input dimension of 13, a GRU layer of 2, a number of neurons in each layer of 64, an activation function of tanh, a dropout rate of 0.2, an output dimension of q+1, a learning rate of 0.001, a training epoch of 50, and a batch size of 32, where q is the number of types of internal defects in the wind turbine blade.

[0024] This invention provides a method for detecting defects in wind turbine blades based on acoustic emission, which has the following advantages compared with existing technologies:

[0025] (1) By integrating multiple lasers, multispectral cameras, thermal imaging sensors and acoustic emission sensors, this invention constructs a collaborative detection system for surface features and internal signals, effectively solving the problem of missed detection of composite defects by a single detection method, and realizing comprehensive perception and correlation judgment of internal and external defects of blades.

[0026] (2) This invention reduces the risk of misjudgment caused by fixed thresholds by using a dynamic threshold method in thermal imaging analysis and constructing a normal distribution based on defect-free blade data to determine a reasonable threshold range; in acoustic signal processing, LMD decomposition combined with MFSE is used to denoise the original signal, and GRU model is used to extract multi-dimensional features for intelligent judgment, which can accurately capture the defect features in complex nonlinear acoustic signals and enhance the robustness of the detection system under complex working conditions. Attached Figure Description

[0027] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] like Figure 1 This invention provides a method for detecting defects in wind turbine blades based on acoustic emission, comprising:

[0030] S100: The drone is equipped with multiple lasers, a multispectral camera, a thermal imaging sensor, and several acoustic emission sensors. The multilaser on the drone illuminates the wind turbine blades, and the multispectral camera and thermal imaging sensor capture several multispectral images and thermal images within a fixed time period.

[0031] The drone can be a six-rotor drone, which can hover stably in complex airflow environments and achieve high-precision scanning at close range. At the same time, since the drone needs to carry multiple acquisition devices, its multi-rotor payload redundancy can meet the actual needs.

[0032] The multi-laser can employ a four-band laser with wavelengths of 650nm, 532nm, 850nm, and 1064nm. 650nm / 532nm corresponds to the visible band, which is within the sensitive area of ​​the human eye and is used for visual identification of surface defects, such as cracks and coating peeling. 850nm corresponds to the near-infrared band, which can penetrate shallow coatings and detect damage to the composite matrix, such as fiber breakage. 1064nm corresponds to the short-wave infrared band, which is sensitive to the difference in dielectric constant between resin and fiber and can identify delamination defects.

[0033] A four-channel multispectral camera was chosen because RGB cameras only have three channels, which cannot separate near-infrared and short-wave infrared signals, such as the 850nm / 1064nm mentioned above, and are prone to missing key defect features. While hyperspectral cameras cover the full spectrum, the amount of data is enormous, which is difficult for the drone's computing power to process in real time, and the cost is also high. The four-channel multispectral camera has four custom bands that can match the laser wavelength, while taking into account both resolution and data efficiency, and supports real-time spectral vector calculation.

[0034] High-precision thermal imagers are selected for thermal imaging sensors because internal defects in wind turbine blades only cause temperature differences of 0.5-2℃, such as layering. High-precision equipment can capture minute temperature changes, avoiding missed detections. At the same time, it is also to achieve millimeter-level defect location and avoid positional deviations caused by pixel aliasing.

[0035] The acoustic emission sensor can be a piezoelectric ceramic acoustic emission sensor, which can cover the main frequency band of defect signals, capture signals from the initiation to the propagation of microcracks, avoid frequency blind spots, adapt to the rapid attenuation of acoustic emission signals of composite materials, effectively amplify weak signals, improve the signal-to-noise ratio, and adapt to blade curved surfaces to ensure acoustic coupling efficiency.

[0036] When using the above-mentioned equipment to collect information from wind turbine blades, the environmental requirements around the wind turbine blades are as follows:

[0037] The ambient temperature needs to be kept stable between 10-30℃ to prevent the thermal imager from causing measurement errors due to temperature drift.

[0038] Humidity should be <= 80%RH to prevent condensation on the lens or water vapor from absorbing infrared radiation, and to avoid moisture adsorbing on the surface of composite materials, which would affect temperature conduction.

[0039] The wind speed should be less than or equal to 5 m / s to prevent drone shaking from causing image misalignment and to reduce the interference of airflow on the temperature distribution of the blade surface;

[0040] Multispectral imaging requires uniform natural light and avoids shadows or overexposure; thermal imaging must avoid direct sunlight to prevent abnormal increases in surface temperature.

[0041] Keep away from strong electromagnetic sources, such as substations and frequency converters, to prevent interference with the sensor signal chain and ensure the accuracy of temperature data acquisition;

[0042] The temperature difference between the background temperature of the shooting area and the temperature of the leaves should be less than 5℃ to avoid background radiation interfering with the contrast of thermal imaging.

[0043] S101. Based on several acquired multispectral images, calculate the four-dimensional spectral vector corresponding to each pixel in each multispectral image, calculate the angle θ between the spectral vector of each pixel and the reference spectral vector of the normal region, and if θ is greater than η, mark the pixel as a suspected defect pixel; otherwise, mark the pixel as a normal pixel.

[0044] The dimension of the spectral vector corresponding to each pixel should be consistent with the number of laser emitters, and each component represents the reflection intensity at that wavelength;

[0045] Defects in wind turbine blades can cause changes in surface material, roughness, and reflectivity, which in turn can lead to specific changes in spectral characteristics, such as cracks, coating peeling, and fiber breakage. By normalizing the vector angle, the influence of light intensity fluctuations on the absolute value is eliminated, and only the differences in spectral shape are considered. At the same time, the multispectral information provided by multiple lasers forms a high-dimensional spectral vector, and defects will produce abnormal responses at specific wavelengths.

[0046] η needs to be dynamically adjusted based on the blade material, coating type, and laser wavelength combination, and is usually taken as 15°. The spectral angle range under normal operating conditions is:

[0047] For defect-free areas, the pixel spectrum is highly consistent with the reference spectrum, and the included angle is usually less than or equal to 10°, ideally close to 0°. However, due to the slight roughness of the blade surface and the uniformity of illumination, there are some fluctuations.

[0048] For areas with minor abnormalities, such as surface stains or slight wear, the included angle may be between 10° and 15°.

[0049] For defective areas, when the included angle is greater than 15°, the spectral differences are significant, and the more severe the defect, the larger the included angle, such as deep cracks and large-area peeling.

[0050] Each multispectral image is classified as either suspicious or normal to obtain a multispectral image discrimination sequence. Then, several multispectral image discrimination sequences within the same time period are processed to calculate the proportions of suspicious and normal numbers in each pixel to the total number of pixels, PerDou and PerNor. The maximum value of max{PerDou,PerNor} is selected as the total judgment value of that pixel to obtain the total multispectral image discrimination sequence.

[0051] If the number of several multi - spectral images within this period of time is even, the situation where PerDou equals PerNer may occur for pixel points, and it is impossible to make a judgment simply relying on max{PerDou, PerNor}. At this time, a priority determination rule needs to be added based on the original method. The specific processing steps are as follows:

[0052] Calculate the proportion PerNei of suspicious pixel points in the 3×3 neighborhood around this pixel point. If PerNei >= 0.5, the current pixel point is determined to be suspicious; otherwise, it is determined to be normal;

[0053] Defects usually have spatial aggregation in images. For example, the area around a crack may have abnormal spectral reflection due to changes in surface morphology. Neighborhood consistency checking can use spatial correlation to reduce the influence of isolated noise;

[0054] For corner pixels, such as the upper - left, lower - left, upper - right, and lower - right, the effective neighborhood is three pixels;

[0055] For edge pixels, such as the left - middle, right - middle, upper - middle, and lower - middle, the effective neighborhood is five pixels;

[0056] For example, there are 3 multi - spectral images, and each image has 5 pixel points. The discrimination sequence of the first multi - spectral image is {suspicious, normal, normal, suspicious, normal}, the discrimination sequence of the second multi - spectral image is {normal, suspicious, normal, normal, suspicious}, and the discrimination sequence of the third multi - spectral image is {suspicious, suspicious, normal, suspicious, normal}. For the first pixel point, the number of suspicious points is 2, the number of normal points is 1, PerDou = 2 / 3, PerNor = 1 / 3, max{PerDou, PerNor} = PerDou, so the first pixel point is judged to be suspicious. By analogy, the total discrimination sequence of the multi - spectral image is {suspicious, suspicious, normal, suspicious, normal};

[0057] Take out the pixel points marked as suspicious in the total discrimination sequence of the multi - spectral image, locate them on the actual three - dimensional coordinates of the wind turbine blade, and mark this three - dimensional coordinate with 0.

[0058] S102. Calculate the temperature value of each pixel point in each thermal imaging picture. For each pixel point, form a temperature value sequence within this period of time, and calculate the total change rate TVR, mean offset rate MOR, standard deviation coefficient CoV, period coincidence degree PC, and determination coefficient R2 of each pixel point temperature value sequence. If or MOR > miu1 or PC < miu2, a first - level warning is initiated and a second - level confirmation is entered. If CoV > miu3 and R2 < miu4, then this pixel point is determined to be a suspicious pixel point, where miu1, miu2, miu3, and miu4 are all set by the user themselves;

[0059] When determining the TVR for each group, the upper limit Crmin and the lower limit Crmax of the threshold interval are dynamically changing. Since the TVR of each group of pixels is affected by independent random factors such as environmental noise and material uniformity fluctuations, according to the central limit theorem, its distribution approaches a normal distribution, providing a theoretical basis for threshold statistical modeling. The specific method for constructing dynamic thresholds is as follows:

[0060] Collect M groups of data for the same area of the defect-free blade, where M≥50. Calculate the TVR of each group of pixels and construct a normal distribution TVRnormal~N(μ,σ2);

[0061] Let Crmin = μ - k×σ and Crmax = μ + k×σ, where k is the confidence coefficient, usually taking 2 or 3;

[0062] The specific reasons for selecting the total change rate TVR, the mean offset rate MOR, and the period coincidence degree PC as the first-level warning conditions include:

[0063] The total change rate TVR. TVR is the sum of the positive / negative change rates at adjacent moments in the temperature sequence, reflecting the severity of temperature change and the frequency of direction switching. If TVR > Crmax, it may indicate a sudden rise or fall in temperature, such as abnormal local heat dissipation caused by surface cracks; if TVR < Crmin, it may indicate that the temperature change is too slow, such as internal delamination hindering heat conduction and resulting in a stagnant temperature rise;

[0064] TVR is the most direct indicator of thermal response anomalies. Defects will significantly change the local thermal resistance, such as cracks and debonding, resulting in the temperature change rate deviating from the normal range and having strong sensitivity. For example, when irradiated with a laser, the temperature in the normal area should rise steadily at a certain rate, while in the debonding area, due to heat accumulation, the heating rate increases initially, the positive change rate increases, and later the change rate may drop suddenly due to thermal saturation, and the overall TVR exceeds the threshold;

[0065] The mean offset rate MOR. MOR measures the deviation degree of the mean value of the temperature sequence from the defect-free reference area, reflecting the abnormality of the long-term thermal equilibrium state. If MOR > miu1, it may mean that the defect area continuously absorbs or dissipates heat abnormally. For example, coating damage will cause a decrease in local reflectivity and absorb more laser energy, resulting in a mean value higher than the reference area, and damage to the internal honeycomb core may cause the mean value to be lower than the normal area due to the interruption of the heat conduction path;

[0066] The mean offset is the cumulative thermal effect caused by defects and forms a complement to the dynamic change rate of TVR. For example, for some progressive defects, the TVR change is not significant in the initial stage, but the long-term mean value will gradually deviate, such as minor debonding, and MOR can capture such chronic anomalies;

[0067] The period coincidence degree PC, which reflects the synchronization between the temperature change period and the laser irradiation period. The normal area should strictly follow the laser heating frequency. PC < miu2 indicates period offset or disorder, which may be caused by defects leading to heat conduction delay. For example, the formation of a debonding layer creates a thermal resistance, resulting in a temperature rise lagging behind the laser irradiation. Or there may be abnormal heat diffusion, such as fiber breakage changing the heat flow direction, leading to non-periodic fluctuations.

[0068] Laser irradiation has a clear periodicity, and a normal thermal response will show fluctuations of the same frequency. Period disorder is a strong characteristic of the existence of defects. For example, when there is delamination inside the blade, the laser energy needs to pass through the air layer, and the heat conduction delay causes the temperature rise peak to lag behind the laser pulse, resulting in a significant decrease in PC.

[0069] These three indicators cover the core characteristics of the thermal response from the three dimensions of TVR, MOR, and PC respectively, and are directly related to the physical mechanisms of defects, namely, thermal resistance change, heat accumulation, and abnormal heat conduction path. As the core of the primary warning, its role is to quickly filter out significantly abnormal pixel points, narrow the range of suspicious areas, and avoid missing serious defects.

[0070] The specific reasons for choosing the coefficient of standard deviation CoV and the coefficient of determination R2 for secondary confirmation include:

[0071] The coefficient of standard deviation CoV, which reflects the relative amplitude of temperature fluctuations and measures thermal stability. If CoV > miu3, it indicates that the temperature fluctuations are剧烈, which may be caused by defects leading to local thermal instability. For example, at the edge of a surface crack, due to increased air convection, the temperature fluctuates greatly during laser irradiation. For internal inclusion defects, due to differences in thermal expansion coefficients, periodic thermal stress release is triggered, resulting in an increase in the fluctuation amplitude.

[0072] Some environmental noises may cause instantaneous abnormalities in TVR, such as a sudden change in short-term wind speed, but CoV remains stable, that is, the fluctuation amplitude does not continue to increase. While real defects will破坏 the thermal equilibrium mechanism, leading to continuous increased fluctuations, and CoV can distinguish accidental noises from structural abnormalities.

[0073] The coefficient of determination R2, which reflects the goodness of the linear fitting of the temperature sequence and measures the trend consistency. If R2 < miu4, it indicates that the temperature change trend is disordered and there is no clear regularity. In the normal area, under stable heat input, it should show an approximate linear temperature rise or fall, such as a linear increase in temperature during continuous laser irradiation. In the defect area, due to non-linear changes in thermal resistance, such as the debonding area expanding with temperature, the trend curve may be bent.

[0074] CoV and R2 focus on the stability of fluctuations and the rationality of trends, solving the false alarm problem of the primary warning. The primary conditions may trigger a warning due to instantaneous interference, such as a short-term occlusion of the light spot. The secondary conditions require that the fluctuation amplitude is continuously abnormal and the trend is disordered, ensuring that only abnormalities with physical consistency are judged as suspicious.

[0075] All pixels in each thermal image are combined to determine whether they are suspicious or normal, resulting in a total discrimination sequence for the thermal image. Pixels marked as suspicious in the total discrimination sequence are extracted and located on the actual three-dimensional coordinates of the wind turbine blade, and these three-dimensional coordinates are marked with 1.

[0076] S103. Fit all three-dimensional coordinates marked as 0 on the wind turbine blade with an oriented bounding box to obtain the light reflection defect region, fit all three-dimensional coordinates marked as 1 with an oriented bounding box to obtain the heat distribution defect region, and find the intersection of the light reflection region and the heat distribution fault region to obtain the specific region of the wind turbine blade defect.

[0077] The acoustic emission sensor on the UAV is used to collect acoustic signals in the specific area of ​​the wind turbine blade defect, and multiple sets of acoustic signal data to be detected are obtained. After the acoustic signal data to be detected is denoised by Local Mean Decomposition (LMD) combined with Multi-Scale Fuzzy Slope Entropy (MFSE), the feature vector is extracted. The feature vector is substituted into the trained GRU wind turbine blade defect judgment model to obtain the specific defect type inside the wind turbine blade.

[0078] The specific steps for denoising acoustic signals using LMD combined with MFSE are as follows:

[0079] (1) The original acoustic signal is decomposed into multiple product function components PF and residual terms by LMD, and each PF corresponds to a signal component at a different frequency scale.

[0080] The acquired raw acoustic signals are mostly nonlinear and non-stationary signals. LMD can adaptively decompose complex signals into several PFs with clear physical meanings. Each PF corresponds to a specific frequency modulation feature, which facilitates the subsequent separation of noise and effective signals.

[0081] By decomposing, signals of different time scales and frequency ranges can be separated into different power factors (PFs), such as low-frequency signals generated by defects and high-frequency interference from environmental noise, providing a basis for noise identification.

[0082] (2) Calculate the mean MFSE Argmfse of each PF to measure its complexity at different time scales. The complexity of the noise component is higher. If Argmfse>=Argth, then the PF component is classified as a noise component; otherwise, it is classified as a defect signal component.

[0083] MFSE measures the complexity of a signal at different time scales. Calculating the mean can integrate information from multiple scales, avoiding the one-sidedness of single-scale analysis and more comprehensively reflecting the signal characteristics of the PF component.

[0084] Defect signals exhibit regularity at a specific scale and have low complexity; noise, on the other hand, exhibits random fluctuations across all scales and has high complexity. Calculating the mean value can highlight this overall difference and facilitate threshold setting.

[0085] The MFSE at a single scale may be affected by local noise. Mean calculation can smooth out random fluctuations, improve feature stability, and ensure reliable distinction between noise and signal components.

[0086] (3) Add the defect signal component and the residual term to obtain the denoised acoustic signal;

[0087] The residual term usually represents the low-frequency trend of the signal and does not contain periodic or high-frequency oscillation characteristics. Its complexity is extremely low, which is fundamentally different from the high complexity of noise. It directly reflects the overall trend of the signal and should be retained directly to avoid information loss.

[0088] The specific method for constructing the GRU wind turbine blade defect judgment model is as follows:

[0089] Several sets of normal and defective acoustic signals were obtained from historical data collection. After LMD combined with MFSE denoising of these acoustic signals, clean normal signals and defective signals were obtained.

[0090] Extract time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features from normal signals and various defective signals, and combine the feature vectors of each acoustic signal;

[0091] Time-domain characteristics include peak value, root mean square value, rise time, pulse width, and zero-crossing rate;

[0092] Peak value represents the maximum absolute value of a signal in a time series, reflecting the degree of stress concentration caused by defects;

[0093] The root mean square value characterizes the signal energy intensity and is sensitive to material damage, such as fiber breakage.

[0094] Rise time represents the time it takes for a signal to rise from 10% to 90% of its peak value. Defective areas may experience prolonged rise times due to damping variations.

[0095] The pulse width represents the duration for which the signal is above a threshold, such as when the threshold is half the peak value.

[0096] Zero-crossing rate represents the number of times a signal crosses zero per unit time, reflecting the oscillation frequency of the signal. Defects can change the stiffness of a material, leading to changes in the zero-crossing rate.

[0097] Frequency domain characteristics include the dominant frequency, band energy, and skewness and kurtosis of the power spectral density (PSD) in the Fast Fourier Transform (FFT).

[0098] The FFT frequency represents the frequency component with the highest energy, reflecting the defect size. Small defects correspond to high frequencies, and large defects correspond to low frequencies.

[0099] Frequency band energy; calculate the energy percentage of each frequency band. Different defect types will lead to differences in characteristic frequency band energy.

[0100] Skewness and kurtosis of power spectral density (PSD): Skewness reflects spectral symmetry, and kurtosis reflects spectral sharpness. Defects can lead to abnormal spectral distribution.

[0101] Time-frequency domain features include wavelet energy entropy and the variance of short-time Fourier transform (STFT).

[0102] Wavelet energy entropy reflects the uniformity of energy distribution across different frequency bands. Defects can lead to energy concentration in specific frequency bands, resulting in a decrease in entropy value.

[0103] The variance of STFT reflects the distribution characteristics of the signal in the time-frequency plane. The time-frequency distribution in the defect area is more disordered, and the variance increases.

[0104] Nonlinear characteristics include permutation entropy (PE) and fractal dimension (FD).

[0105] Permutation entropy calculates the probability distribution entropy value of different permutation patterns, but defects increase the complexity of the signal;

[0106] Fractal dimension reflects the self-similarity of a signal. Material damage can lead to changes in surface or internal structure, resulting in changes in the FD value.

[0107] The acoustic signals were divided into training and testing sets in a 7:3 ratio to ensure that the samples were balanced.

[0108] The initial parameters of the GRU model are as follows:

[0109] The input dimension is 13, the number of GRU layers is 2, the number of neurons in each layer is 64, the activation function is tanh, the dropout rate is 0.2, the output dimension is q+1, the learning rate is 0.001, the number of training rounds is 50, and the batch size is 32, where q is the number of types of internal defects in the wind turbine blade.

[0110] The training set is substituted into the GRU model for training. The classification cross-entropy is selected as the loss function. The internal parameters are adjusted by substituting the test set to obtain the final GRU wind turbine blade defect judgment model.

[0111] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0112] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for detecting defects in a wind turbine blade based on acoustic emission, characterized in that, Comprise: S101, based on the collected multiple multispectral images, calculate the four-dimensional spectral vector corresponding to each pixel in each multispectral image, calculate the angle θ between each pixel spectral vector and the normal area reference spectral vector, if θ is greater than η, mark the pixel point as a suspicious defect pixel point, obtain a single multispectral image discrimination sequence, process the multispectral image discrimination sequence in a period of time, obtain a multispectral image total discrimination sequence, map the suspicious pixel points in the multispectral image total discrimination sequence to the actual three-dimensional coordinates of the wind turbine blade, and mark them with 0, wherein η is the angle threshold, the method for obtaining the multispectral image total discrimination sequence is: calculate the proportion of the number of suspicious and normal pixels to the total number PerDou, PerNor, select max{PerDou, PerNor} as the total judgment of the pixel, if PerDou=PerNor, calculate the proportion of suspicious pixels in the 3*3 neighborhood of the pixel PerNei, if PerNei>=0.5, the current pixel is determined as suspicious, otherwise it is determined as normal; if the pixel is a corner pixel, the effective neighborhood takes three pixels, if the pixel is an edge pixel, the effective neighborhood takes five pixels, the corner pixel is the upper left, lower left, upper right, lower right, and the edge pixel is the left middle, right middle, upper middle and lower middle; S102, calculate the temperature value of each pixel point in each thermal imaging picture, obtain the temperature value sequence of each pixel point in the time period, calculate the total variation rate TVR, the mean offset rate MOR, the standard deviation coefficient CoV, the period coincidence degree PC, and the determination coefficient R2 of each pixel point temperature value sequence, if or MOR>miu1 or PC<miu2, then start the first level warning, enter the second level confirmation, if CoV>miu3 and R2<miu4, then determine that the pixel point is a suspicious pixel point, judge all pixel points as suspicious and normal, obtain the total judgment sequence of the thermal imaging image, map the suspicious pixel point in the total judgment sequence of the thermal imaging image to the actual three-dimensional coordinates of the wind power blade, and mark it with 1, wherein Crmin and Crmax are the lower limit and upper limit of the TVR threshold interval, miu1, miu2, miu3, and miu4 represent the first threshold value, the second threshold value, the third threshold value, and the fourth threshold value. S103, by fitting the three-dimensional coordinates marked as 0 and 1 on the fan blade through the oriented bounding box to obtain the light reflection defect area and the thermal distribution defect area, the intersection of the two is obtained to obtain the specific area of the surface defect of the wind turbine blade, the area is located by the acoustic emission sensor on the unmanned aerial vehicle to obtain the to-be-detected acoustic signal, after the to-be-detected acoustic signal is denoised by local mean decomposition LMD combined with multi-scale fuzzy slope entropy MFSE, the feature vector is extracted, and the feature vector is substituted into the GRU wind turbine blade defect judgment model to judge whether the area exists internal fault.

2. The acoustic emission based method of wind turbine blade defect detection according to claim 1, characterized in that, Further comprising S100, by carrying multiple lasers, multispectral cameras, thermal imaging sensors and several acoustic emission sensors on the unmanned aerial vehicle, illuminating the wind turbine blade by the multiple lasers on the unmanned aerial vehicle, and using the multispectral camera and the thermal imaging sensor to shoot several multispectral images and thermal imaging images in a fixed time period respectively.

3. The acoustic emission based method of wind turbine blade defect detection according to claim 2, characterized in that, The selected wave band of the multiple lasers includes visible wave band, near-infrared wave band and short-wave infrared wave band.

4. The acoustic emission based method of wind turbine blade defect detection as claimed in claim 2, wherein, When collecting information of the wind turbine blade, the specific requirements of the surrounding environment include that the environmental temperature needs to be stable at 10-30℃, the humidity needs to be <=80%RH, the wind speed needs to be <=5m / s, the multispectral imaging needs to be uniform natural light, the thermal imaging needs to avoid direct sunlight, be away from strong electromagnetic source, and the background temperature of the shooting area needs to be within 5℃ difference from the blade temperature.

5. The acoustic emission based method of wind turbine blade defect detection as claimed in claim 1, wherein, The specific dynamic threshold construction method of each group of TVR is: M groups of data are collected from the same area of the defect-free blade, and the TVR of each group of pixels is calculated to construct a normal distribution TVRnormal~N(μ,σ 2 ). According to μ and σ, the upper limit Crmin and the lower limit Crmax of the threshold interval of each group of TVR are obtained, the Crmin = μ-k×σ, the Crmax = μ+k×σ, wherein μ is the mean value of the TVR data of the same region of the defect-free blade, σ is the standard deviation of the TVR data of the same region of the defect-free blade, and k is a confidence coefficient, which can be 2 or 3.

6. The acoustic emission based method of wind turbine blade defect detection as claimed in claim 2, wherein, The specific steps of using LMD combined with MFSE to denoise the acoustic signal are as follows: The original acoustic signal is decomposed into a plurality of product function components PF and a residual term by LMD; The MFSE mean value Argmfse of each PF is calculated, if Argmfse >= Argth, the product function component is classified as a noise component, otherwise, it is classified as a defect signal component, wherein Argth is the threshold value of the MFSE mean value; The defect signal component and the residual term are added to obtain the denoised acoustic signal.

7. The acoustic emission based method of wind turbine blade defect detection as claimed in claim 2, wherein, The time domain features, the frequency domain features, the time-frequency domain features and the nonlinear features are extracted from the denoised acoustic signal, the time domain features include the peak value, the root mean square value, the rise time, the pulse width and the zero-crossing rate, the frequency domain features include the main frequency of the fast Fourier transform (FFT), the frequency band energy, the skewness and the kurtosis of the power spectral density (PSD), the time-frequency domain features include the wavelet energy entropy and the variance of the short-time Fourier transform (STFT), and the nonlinear features include the permutation entropy (PE) and the fractal dimension (FD).

8. The acoustic emission based method of wind turbine blade defect detection as claimed in claim 2, wherein, The initial parameters of the GRU model specifically include that the input dimension is 13, the number of GRU layers is 2, the number of neurons in each layer is 64, the activation function is tanh, the dropout rate is 0.2, the output dimension is q+1, the learning rate is 0.001, the number of training rounds is 50, and the batch size is 32, wherein q is the number of categories of the internal defects of the wind turbine blade.

Images