Intelligent wind power blade defect identification method based on phased array detection

Abstract

The invention relates to an intelligent wind power blade defect identification method based on phased array detection, which aims at the anisotropic characteristic of a wind power blade glass fiber reinforced composite material, and can flexibly adjust the sound beam angle, effectively penetrate through the material and inhibit the scattering of an ultrasonic signal through a low-frequency linear array probe with characteristic parameters and an electronic focusing technology. Defects in different directions such as impurities perpendicular to layers, cracks in the fiber direction, wrinkles and the like can be accurately detected, meanwhile, by combining wavelet packet transformation noise reduction and a feature extraction algorithm, the distinguishing capacity of defect echoes and noise in complex materials is remarkably improved, the method is adaptive to complex structures and material characteristics of wind power blades, the application range of the detection technology is expanded, and the detection efficiency is improved. An ultrasonic phased array detection technology and an intelligent algorithm (a K-means clustering algorithm and a YOLOv8 target detection model) are deeply fused, the limitation that traditional detection depends on manual analysis is broken through, and a full-process automatic detection system of data acquisition, feature extraction and intelligent recognition is constructed.

Description

Technical Field

[0001] The present invention relates to an intelligent detection method for wind turbine blade defect detection. Specifically, it is used to construct a defect recognition model, locate defects and conduct quantitative analysis on detection signals when phased array detection of wind turbine blades, thereby realizing quantitative and classified identification of inclusion, delamination and wrinkle defects in wind turbine blades. Background Art

[0002] In recent years, the wind power industry has grown its share of installed power generation capacity nationwide due to its environmental performance and high energy conversion efficiency. As a key component of wind turbines, wind turbine blades capture wind energy, which drives the rotor and converts it into electricity. The production process for wind turbine blades is complex, with the primary material, GFRP, undergoing multiple steps, including layup, vacuuming, and resin infusion. Inadvertent introduction of impurities during the layup process can lead to inclusion defects; uneven layup can cause wrinkles, which can lead to compression failure and delamination; incomplete removal of air during vacuuming can cause air bubble defects; and insufficient resin infusion can cause debonding. In addition to the two aforementioned defect mechanisms, blades can be subject to collision and twisting during transportation and installation, further amplifying minor defects already present during production. Therefore, the loads exerted on blades during operation, including aerodynamics, gravity, and centrifugal forces, as well as defects during the manufacturing process, can seriously impact blade quality, leading to wind turbine shutdowns and tower collapses.

[0003] To ensure the safe use of wind turbine blades, it is essential to control the quality of composite material manufacturing through necessary inspection methods. Currently, existing methods for detecting wind turbine blade root defects include visual inspection and conventional ultrasonic testing. Visual inspection can only detect surface defects but cannot detect internal defects. Conventional ultrasonic testing uses the amplitude and propagation time of the defect wave to determine the defect location and size. However, its sensitivity is low, and various reflected signals are difficult to distinguish from the A-shaped display, which can easily lead to missed defects.

[0004] With the continuous development of ultrasonic phased array testing technology, it has become a research hotspot in the field of wind turbine blade root inspection. Ultrasonic phased arrays can generate two-dimensional and three-dimensional images such as B-scan and C-scan through electronic focusing and multi-angle scanning, clearly showing the location, shape, and size of defects. The accuracy of identifying tiny defects (such as millimeter-level cracks and delamination) is higher than that of traditional ultrasonic testing. By controlling the excitation sequence of the probe array elements, the acoustic beam angle can be flexibly adjusted (e.g., 0° to 70°), adapting to the anisotropic properties of the blade composite laminate structure and effectively detecting defects in different directions, such as debonding perpendicular to the interlayer and cracks along the fiber direction. However, wind turbine blades are made of glass fiber reinforced composite materials. The anisotropy of the material causes severe ultrasonic signal scattering, making it difficult to distinguish defect echoes from material noise. Currently, blade inspection still relies on manual analysis of A-scan, B-scan, and C-scan images. Different inspectors' judgments on defect size, location, and type can vary due to differences in experience, technical skills, and cognitive skills. Furthermore, manual frame-by-frame analysis of inspection images and signals is time-consuming, making it difficult to meet the requirements for efficient inspection, especially for comprehensive inspections of large wind turbine blades. Inspectors can easily experience visual and mental fatigue from concentrating on observing complex inspection data for extended periods, potentially overlooking smaller or hidden defects. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and thus provide an intelligent wind turbine blade defect identification method based on phased array detection, which can effectively solve the problem of rapid and accurate detection and identification of wind turbine blade defects.

[0006] The technical solution provided by the present invention is:

[0007] A method for intelligently identifying defects in wind turbine blades based on phased array detection includes the following steps:

[0008] Step 1: Use a low-frequency linear array probe to test a simulated wind turbine blade sample and obtain A-scan, B-scan, and C-scan views. Specifically:

[0009] (1) Determine the probe model

[0010] The probe model is determined to be 0.5L64-1.5×22-MW5-F2.5-D3;

[0011] (2) Preparation of simulated test blocks

[0012] 9 simulated wind turbine blade specimens with inclusions, delaminations, and wrinkles were processed and numbered;

[0013] (3) Step test block production

[0014] To establish the corresponding relationship between the defect depth and size of wind turbine blade simulation specimens and the detection signal characteristics and provide a quantitative identification basis, a stepped test block with dimensions of 100 mm in length × width × 50 mm and thicknesses of 10 mm, 20 mm, 30 mm, and 40 mm was designed and manufactured. The sides of the four stepped test blocks were connected end to end in ascending order along the length direction, with the bottom surfaces flush, forming a stepped shape.

[0015] (4) Basic parameter settings

[0016] After connecting the probe to the instrument, calibrate the device's zero point and sound velocity using the direct contact method on a stepped test block using chemical paste as a coupling agent. Then, place the probe on a defect-free area of ​​the simulated specimen and adjust the bottom wave to 80% of the screen's full scale, which serves as the test sensitivity.

[0017] (5) Scanning process

[0018] Nine simulated test blocks were tested using ultrasonic phased array technology, with data collected using a C-scan method. The length of the simulated test block was along the X-axis, and the width was along the Y-axis. The intersection of the X-axis and Y-axis at the lower left corner of the simulated test block was at point (0,0). The probe first scanned the simulated test block from point (0,0) along the positive X-axis until it reached the edge of the simulated test block. The probe then scanned from the edge of the simulated test block along the negative X-axis, with a 15% overlap with the previous scan path, until it reached the edge of the simulated test block. This process was repeated until the entire simulated test block was scanned, resulting in A-scan and B-scan data.

[0019] Step 2: Feature extraction and clustering training based on A-scan signals

[0020] (1) Normalized signal: Use the direct traversal method to scan all sampling points of the A-scan data signal x(t) one by one, and record the current maximum value x max and the minimum value x min , let the original signal be x(t)=[x1,x2,…,x N ], initialize: x max =x1,x min =x1, then i=2 to N:x max =max(x max , x i ), x min =min(x min ,x i ); for each sampling point value x i The normalized value x is calculated by the following formula / i , the signal interval range is [0,1], and all normalized values ​​are x(n):

[0021]

[0022] (2) Wavelet packet decomposition: The normalized numerical signal is recursively decomposed to the jth layer using wavelet packet transform to generate 2 j low-frequency and high-frequency sub-bands;

[0023] (3) Feature vector construction: The energy proportion E(n), signal-to-noise ratio SNR and comprehensive index F are integrated to construct a feature vector. The feature vector is expressed as

[0024] a. Calculate the target sub-band energy ratio E(n):

[0025] a.1 For the coefficient X(j,n) of the nth subband in the jth layer, calculate the subband energy:

[0026]

[0027] Where M is the subband coefficient length;

[0028] a.2 Calculate the total energy of all sub-bands:

[0029]

[0030] a.3 Energy proportion E(n) is:

[0031]

[0032] b. Calculate the signal-to-noise ratio (SNR):

[0033]

[0034] in, X is the original noisy signal; It is the result of normalized signal decomposition by wavelet packet, sub-band denoising and inverse transformation reconstruction;

[0035] c. Calculate comprehensive indicators:

[0036]

[0037] Where: α is the weight coefficient, which is 0.6, SNR max is the maximum SNR value of all filtered signals;

[0038] (4) Determine the number of clusters K and the standardization of eigenvectors: According to the detection properties of wind turbine blades, the data are divided into defective and non-defective data, that is, K = 2, and the eigenvector [E(n) i , SNR i 、F i ] to ensure that features of different dimensions have the same weight in distance calculation:

[0039] (5) Using K-means algorithm to set cluster centers

[0040] From the normalized set of eigenvectors {E / (n), SNR / 、F / } randomly select k vectors as the initial cluster centers

[0041] a. Sample allocation

[0042] For each eigenvector [E(n) i , SNR i 、F i ] Calculate its difference with all initial centers distance, [E(n) i , SNR i 、F i ]Assigned to the cluster with the nearest center In, S j represents the jth cluster, which contains all feature vector samples assigned to the jth cluster center; j corresponds to cluster centers 1, 2, ..., k;

[0043] b. Calculate the mean within the cluster and update the center

[0044] For each Calculate the mean vector of all its samples as the new center after the first iteration at this time It is not a random sample, but the centroid of the cluster sample, which is more representative than the initial center. The formula is:

[0045]

[0046] c. Repeat allocation and update until convergence

[0047] Based on the new center Reassign all samples to the nearest cluster , calculate the average value of each cluster again to get Similarly, when the center of two adjacent iterations no longer changes, the iteration stops, and the final C t It is the cluster center;

[0048] (6) Label assignment and C-scan image reconstruction based on K-means algorithm: Set label L i =1 means defective, L i = 0 means no defect, each eigenvector [E(n) i , SNR i 、F i ] After clustering, it belongs to a certain category j and its corresponding label Li Set to 1 or 0 and retain its spatial coordinates (x i ,y i );

[0049] The A-scan signal of each element of the phased array probe constitutes three-dimensional data [x, y, A(x, y, t)], where (x, y) is the spatial position and A(x, y, t) is the time domain signal of the position; the label L of each feature vector is i Corresponding to its spatial coordinates (x i ,y i ), that is, each position (x, y) gets a category label, defect or non-defect;

[0050] Set the color mapping rule: if the label of the position (x, y) is a defect, the pixel color is set to red; if the label of the position (x, y) is not a defect, the pixel color is set to a light color; arrange all pixels in order of spatial coordinates to generate a C-scan reconstructed image. The C-scan reconstructed image is a two-dimensional plane image, and the color can reflect the defect distribution.

[0051] (7) Defect location and quantification:

[0052] Extract the pixel coordinates (x, y) of the cluster label "defect" from the C-scan reconstructed image to determine the plane position of the defect. Combined with the time t of the A-scan signal and the material sound velocity c, calculate the defect depth d = (c × t) / 2;

[0053] The width of the image is calculated by the number of sound beams and the chip spacing. The total width = (number of sound beams - 1) × chip spacing. The center distance between adjacent sound beams in the width direction is the pixel width, and the actual scanning step length in the length direction is the pixel length. The actual physical size of each pixel in the C-scan reconstructed image is pixel width × pixel length. The number of pixels N marked as defect areas in the C-scan reconstructed image is counted. The actual area of ​​the defect size = number of pixels N × area of ​​a single pixel; the final defect size is expressed as area × depth, that is, the final K-means model is obtained; an ultrasonic phased array device is used to scan the wind turbine blades with a low-frequency linear array probe to obtain A-scan and B-scan signals. The A-scan signal is input into the K-means model, and the image recognition result is output. The result will clearly identify the possible defects in the blade root and the location and size of the defects.

[0054] The working principle of the present invention is:

[0055] 1. Data Collection: Based on the structural characteristics of wind turbine blades and failure analysis, defects such as inclusions, delamination, and wrinkles at the blade root are clearly the primary cause of failure. A dedicated low-frequency linear array probe (such as the 0.5L64-1.5×22-MW5-F2.5-D3 probe with a center frequency of 0.5 MHz) is used to inspect simulated defective specimens. A-scan, B-scan, and C-scan data are acquired using the C-scan method. Chemical paste is used as a coupling agent during scanning, and the equipment zero point and sound velocity are calibrated using the direct contact method. The bottom wave is adjusted to 80% of the screen's full scale to determine detection sensitivity. The defect length is measured using the half-wave height method.

[0056] 2. Defect quantification based on A-scan signals: The A-scan signals are normalized and decomposed to the jth layer through wavelet packet transform to generate multiple sub-bands. The energy proportion E(n), signal-to-noise ratio SNR and comprehensive index F of each sub-band are calculated and fused to construct a feature vector. The feature vectors are standardized and clustered using the K-means algorithm. A defect or no-defect label is assigned to each feature vector based on the clustering result. The C-scan image is then reconstructed based on the spatial coordinates. The defect pixel coordinates are extracted from the reconstructed image to determine the plane position. The defect depth is calculated based on the A-scan signal time and the material sound velocity. The actual area is calculated by counting the number of pixels in the defect area to achieve defect location and quantification.

[0057] 3. Defect characterization based on B-scan signals: Labelmg annotation software was used to select defect areas in ultrasound B-scan images and annotate them into four categories: normal, wrinkles, inclusions, and delamination. The annotated samples were divided into training, validation, and test sets at a ratio of 7:2:1. The YOLOv8 model was used for training, and the mean average approximation (mAP) was calculated on the validation set every 10 rounds. If the mAP was less than 0.7, the neck network was optimized to add a shallow feature detection layer. If the validation set mAP did not improve after 15 consecutive rounds of training, the optimal model was saved. The Network Slimming algorithm was used to prune and optimize the model. Finally, the test set was used to evaluate the model effect and achieve defect characterization.

[0058] 4. Scanning Implementation: Use ultrasonic phased array equipment and dedicated probes to scan wind turbine blades to obtain A-scan and B-scan signals. The A-scan signal is input into the K-means model to output the defect location and size, and the B-scan signal is input into the improved YOLOv8 model to output the defect category.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. Technological integration and innovation: Deeply integrating ultrasonic phased array detection technology with intelligent algorithms (K-means clustering algorithm, YOLOv8 target detection model), we break through the limitations of traditional detection that relies on manual analysis and build a full-process automated detection system of "data collection-feature extraction-intelligent recognition".

[0061] 2. Multi-dimensional defect analysis: Defect quantification is achieved based on A-scan signals, and the defect position, size, and depth are accurately located through C-scan image reconstruction. Defect qualitative analysis is achieved using the YOLOv8 model based on B-scan signals, accurately distinguishing categories such as wrinkles, inclusions, and delamination. This provides a "quantitative + qualitative" dual detection capability, comprehensively improving the accuracy and reliability of defect identification.

[0062] 3. Design of special detection scheme: Based on the structural characteristics of wind turbine blade roots and the properties of composite materials, a low-frequency linear array probe with specific parameters is used to optimize the detection parameters (such as center frequency, array spacing, etc.), and simulated specimens and stepped test blocks containing different types of defects are produced. The corresponding relationship between defect depth, size and detection signal characteristics is established, providing a quantitative identification basis for detection, thereby enhancing the pertinence and adaptability of the detection scheme.

[0063] 4. Model optimization and efficiency improvement: During YOLOv8 model training, the neck network was optimized by adding a shallow feature detection layer to enhance the model's defect recognition capabilities. The network slimming algorithm was used to prune the model, compressing the model volume by 30%-50% while maintaining accuracy. This made the model more suitable for the real-time and high-efficiency requirements of on-site inspections, achieving a balanced optimization of inspection efficiency and model performance.

[0064] 5. Strong detection adaptability: In view of the anisotropic characteristics of glass fiber reinforced composite materials of wind turbine blades, the low-frequency linear array probe with special parameters and electronic focusing technology can flexibly adjust the sound beam angle, effectively penetrate the material and suppress ultrasonic signal scattering. It can accurately detect defects in different directions such as inclusions perpendicular to the layers, cracks along the fiber direction, and wrinkles. At the same time, combined with wavelet packet transform noise reduction and feature extraction algorithms, it significantly improves the ability to distinguish defect echoes from noise in complex materials, adapts to the complex structure and material characteristics of wind turbine blades, and expands the scope of application of detection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a physical picture of the probe of the present invention.

[0066] Figure 2 This is a physical picture of the instrument and simulation test block of the present invention.

[0067] Figure 3 This is a physical picture of the stepped test block of the present invention.

[0068] Figure 4 This is a scanning path diagram of the simulated test block of the present invention.

[0069] Figure 5 Schematic diagram of typical A-scan view, B-scan view and C-scan view obtained after scanning according to the present invention.

[0070] Figure 6The original image of the typical defect simulation sample of the present invention and the C-scan image reconstructed by clustering algorithm, including: (a) macro image; (b) K-means reconstructed image; (c) 20mm depth defect; (d) 30mm depth defect.

[0071] Figure 7 This is the defect image annotation file of the present invention and the LabelImg software page display diagram.

[0072] Figure 8 This is a graph showing how the loss and evaluation accuracy of the present invention change with the number of training rounds.

[0073] Figure 9 This is a diagram showing the prediction results of B-scan images at different positions of a typical defective sample of the present invention.

[0074] Figure 10 This is a diagram of raw data collected by the ultrasonic phased array of the present invention.

[0075] Figure 11 This is an image diagram of quantitative defect identification in an application example of the invention.

[0076] Figure 12 This is an image diagram of defect qualitative identification in an application example of the present invention. Figure 13 This is a real picture of the root defect of a fan blade in an application example of the present invention. DETAILED DESCRIPTION

[0077] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0078] The present invention provides an intelligent wind turbine blade defect identification method based on phased array detection, comprising the following steps:

[0079] Step 1: Use a low-frequency linear array probe to test a simulated wind turbine blade sample and obtain A-scan, B-scan, and C-scan views. Specifically:

[0080] (1) Determine the probe model

[0081] The probe model is determined to be 0.5L64-1.5×22-MW5-F2.5-D3;

[0082] Specific technical parameters are shown in Table 1:

[0083] Table 1 Probe technical parameters

[0084]

[0085] (2) Preparation of simulated test blocks

[0086] Nine simulated wind turbine blade specimens containing inclusions, delaminations, and wrinkles were processed and numbered. The nine simulated specimens were: two specimens with inclusions measuring 300 mm × 300 mm × 50 mm in length, width, and height. Each specimen had two inclusion defects. One specimen had two inclusion defects measuring 33 mm × 27 mm × 20 mm (length × width × depth) and 35 mm × 33 mm × 40 mm, respectively. Another specimen had two inclusion defects measuring 47 mm × 41 mm × 20 mm (length × width × depth) and 56 mm × 33 mm × 30 mm, respectively.

[0087] Three test blocks with a length, width, and height of 300 mm × 300 mm × 55 mm and containing one delamination defect. The delamination defect dimensions of the three test blocks are 49 mm × 47 mm × 30 mm (length × width × depth), 138 mm × 130 mm × 30 mm, and 113 mm × 107 mm × 30 mm, respectively.

[0088] The four test blocks containing one wrinkle defect have a length, width and height of 300mm×300mm×55mm, 300mm×300mm×60mm, 300mm×300mm×40mm and 300mm×300mm×65mm respectively; the corresponding wrinkle defect aspect-to-height ratios are 0.03, 0.06, 0.1 and 0.14 respectively.

[0089] The specific sample numbers and corresponding sample sizes, sample defect types and sizes are shown in Table 2:

[0090] Table 2 Size and defect type of defective specimens

[0091]

[0092]

[0093] (3) Step test block production

[0094] To establish the corresponding relationship between the defect depth and size of wind turbine blade simulation specimens and the detection signal characteristics and provide a quantitative identification basis, a stepped test block with dimensions of 100 mm in length × width × 50 mm and thicknesses of 10 mm, 20 mm, 30 mm, and 40 mm was designed and manufactured. The sides of the four stepped test blocks were connected end to end in ascending order along the length direction, with the bottom surfaces flush, forming a stepped shape.

[0095] (4) Basic parameter settings

[0096] After connecting the probe to the instrument, the device zero point and sound velocity calibration was completed on a stepped test block using a direct contact method using chemical paste as a coupling agent. The zero point calibration is as follows: First, place the probe vertically at the center of the 10mm thick step test surface of the stepped test block. Adjust the instrument range so that the reflected echo B1 from the bottom of the test block is clearly displayed on the screen. Record the time t1 of this reflected echo on the time base axis. Then, place the probe vertically at the center of the 20mm thick step test surface of the stepped test block. Adjust the instrument range so that the reflected echo B1 from the bottom of the test block is clearly displayed on the screen. Record the time t2 of this reflected echo on the time base axis. According to the formula zero point t0 = test time - (thickness × 2) / sound velocity reference value, using the simultaneous equations of the two step thicknesses (10mm, 20mm) and the corresponding times (t1, t2), eliminate the unknown sound velocity and calculate the delay time of the sound wave inside the probe (zero point t0). The specific sound velocity calibration is as follows: After the zero point calibration is completed, first place the probe vertically at the center of the 30mm thick step detection surface of the step test block, adjust the instrument range so that the reflected echo B1 from the bottom surface of the test block is clearly displayed on the screen, record the time t of the reflected echo on the time base axis (t = bottom wave time - zero point t0), and according to the formula sound velocity c = 2×thickness d / t, substitute the thickness of the step 30mm and the corresponding bottom wave propagation time t to obtain the longitudinal wave sound velocity c1 of the material. Then place the probe vertically at the center of the 40mm thick step detection surface of the step test block, adjust the instrument range so that the reflected echo B1 from the bottom surface of the test block is clearly displayed on the screen, record the time t of the reflected echo on the time base axis (t = bottom wave time - zero point t0), and according to the formula sound velocity c = 2×thickness d / t, substitute the thickness of the step 40mm and the corresponding bottom wave propagation time t to obtain the longitudinal wave sound velocity c2 of the material. Finally, c = (c1 + c2) / 2 is taken to obtain the actual longitudinal wave velocity c; the probe is then placed on a defect-free portion of the simulated specimen, and the bottom wave is adjusted to 80% of the full scale of the screen as the detection sensitivity; the defect length is measured using the half-wave height method (-6dB method); the instrument is a phased array ultrasonic instrument;

[0097] (5) Scanning process

[0098] The ultrasonic phased array test was performed on 9 simulated test blocks, and the data acquisition adopted the C-scan method. The length direction of the simulated test block was the X axis, the width direction was the Y axis, and the intersection of the X axis and the Y axis at the lower left corner of the simulated test block was the (0,0) point. Figure 4 As shown in the figure, the probe first starts from point (0,0) and scans the simulated test block in the positive direction of the X axis until it reaches the edge of the simulated test block. Then, the probe scans from the edge of the simulated test block in the negative direction of the X axis with a 15% overlap with the previous scanning path in the Y axis until it reaches the edge of the simulated test block. The above steps are repeated until the entire simulated test block is scanned, and the A scan data and B scan data are obtained; the scanning step is 12 step / mm, the scanning resolution is 1mm, and the typical A scan view, B scan view and C scan view are shown in Figure 5 As shown in the figure, data were collected for 9 simulated test blocks, and 11,457 A-scan images and 201 B-scan images were obtained for each sample.

[0099] Step 2: Feature extraction and clustering training based on A-scan signals

[0100] The A-scan signal acquired by the ultrasonic phased array is a time-domain waveform containing components such as defect echoes, interface reflection waves, and noise. By normalizing the signal, performing wavelet packet transform decomposition, and calculating E(N) and SNR, defect feature extraction and noise suppression are achieved. Feature vectors are constructed for the A-scan signal, clustering is performed using the K-means algorithm, and C-scan images are reconstructed based on the labels, enabling rapid and accurate identification, location, and quantification of wind turbine blade defects.

[0101] (1) Normalized signal: Use the direct traversal method to scan all sampling points of the A-scan data signal x(t) one by one, and record the current maximum value x max and the minimum value x min , let the original signal be x(t)=[x1,x2,…,x N ], initialize: x max =x1,x min =x1, then i=2 to N:x max =max(x max , x i ), x min =min(x min ,x i ); for each sampling point value x i The normalized value x is calculated by the following formula / i , the signal interval range is [0,1], and all normalized values ​​are x(n):

[0102]

[0103] (2) Wavelet packet decomposition: The normalized numerical signal is recursively decomposed to the jth layer using wavelet packet transform to generate 2 j low-frequency and high-frequency sub-bands;

[0104] First-level decomposition (initial decomposition): The normalized signal is processed using low-pass and high-pass filters corresponding to the orthogonal wavelet basis functions. Low-pass filtering is used to extract the low-frequency components of the signal, representing the overall trend of the signal; high-pass filtering is used to extract the high-frequency components of the signal, representing the local details and mutation information of the signal. After decomposition, two frequency bands are obtained: low-frequency sub-band A1 and high-frequency sub-band D1.

[0105] a. Low-pass filtering (approximate component, A1)

[0106] The normalized value is x(n), the low-pass filter coefficient is h(n), and the low-frequency component A1 is calculated as:

[0107]

[0108] Where: k is the index of the decomposed signal, h(2k-n) represents the convolution operation after even sampling and flipping the filter system;

[0109] b. High-pass filtering (detail component, D1)

[0110] The normalized value is x(n), the high-pass filter coefficient is g(n), and the high-frequency component D1 is calculated as:

[0111]

[0112] Where: g(n) is usually orthogonal to h(n), and g(2k-n) represents the convolution operation after even sampling and flipping the filter system;

[0113] Second-layer decomposition (recursive processing): The same low-pass filter h(n) and high-pass filter g(n) as in the first layer are applied to the low-frequency subband (A1) and high-frequency subband (D1) of the first layer respectively to ensure the consistency of the decomposition rules. After decomposition, four frequency bands are obtained: A2, D21, D22, and D23.

[0114] a. Layer 1 low-frequency subband (A1) decomposition

[0115] The A1 signal is A1(m), which is subjected to low-pass and high-pass filtering to obtain the second-layer low-frequency sub-band A2 and high-frequency sub-band D21:

[0116] The calculation formula of A2 is:

[0117]

[0118] The D21 calculation formula is:

[0119]

[0120] b. Decomposition of the first high-frequency subband (D1)

[0121] The D1 signal is D1(m), which is subjected to low-pass and high-pass filtering to obtain the second-layer low-frequency sub-band D22 and high-frequency sub-band D23:

[0122] The D22 calculation formula is:

[0123]

[0124] The D23 calculation formula is:

[0125]

[0126] Recursive decomposition from layer k to layer j (k = 2, 3, ..., j-1): Each layer repeats the "low-pass + high-pass filtering" decomposition operation on all low-frequency and high-frequency sub-bands of the previous layer, and each decomposition splits each sub-band into 2 sub-bands, where the kth layer has 2 k sub-bands, and the k+1th layer has 2 k+1 sub-bands;

[0127] j-th layer decomposition: After j times of recursive decomposition, 2j sub-bands are finally generated. Each sub-band corresponds to a different frequency band range and covers the full frequency band of the original signal.

[0128] (3) Feature vector construction: The energy proportion E(n), signal-to-noise ratio SNR and comprehensive index F are integrated to construct a feature vector. The feature vector is expressed as

[0129] a. Calculate the target sub-band energy ratio E(n):

[0130] a.1 For the coefficient X(j,n) of the nth subband in the jth layer, calculate the subband energy:

[0131]

[0132] Where M is the subband coefficient length;

[0133] a.2 Calculate the total energy of all sub-bands:

[0134]

[0135] a.3 Energy proportion E(n) is:

[0136]

[0137] b. Calculate the signal-to-noise ratio (SNR):

[0138]

[0139] in, X is the original noisy signal; It is the result of normalized signal decomposition by wavelet packet, sub-band denoising and inverse transformation reconstruction;

[0140] c. Calculate comprehensive indicators:

[0141]

[0142] Where: α is the weight coefficient, which is 0.6, SNR max is the maximum SNR value of all filtered signals;

[0143] (4) Determine the number of clusters K and the standardization of eigenvectors: According to the detection properties of wind turbine blades, the data are divided into defective and non-defective data, that is, K = 2, and the eigenvector [E(n) i , SNR i 、F i ] to ensure that features of different dimensions have consistent weights in distance calculation, specifically:

[0144] a. Calculate the mean and standard deviation of each feature

[0145] First, collect the E(n) values ​​of all signals, denoted as E1, E2…, E m , where k is the total number of signals; calculate the mean and standard deviation of E(n), the mean formula and standard deviation formula are:

[0146]

[0147]

[0148] Similarly, we process the SNR and F features to get μ SNR , σ SNR 、μ F , σ F .

[0149] b. Standardize each eigenvalue

[0150] E(n) normalization:

[0151]

[0152] SNR normalization:

[0153]

[0154] F feature normalization:

[0155]

[0156] c. Generate standardized feature vector

[0157] The feature vector of each sample is [E(n) i , SNR i 、F i ] is converted to [E / (n) i , SNR / i 、F / i ], at this time, the mean of each feature is 0, the standard deviation is 1, the dimension is unified, and the standardized feature vector set is {E / (n), SNR / 、F / };

[0158] (5) Using K-means algorithm to set cluster centers

[0159] From the normalized feature vector set {E / (n), SNR / 、F / } randomly select k vectors as the initial cluster centers At this time, the initial cluster centers are random samples and cannot reflect the true distribution of the clusters;

[0160] a. Sample allocation (based on initial center)

[0161] For each eigenvector [E(n) i , SNR i 、F i ] Calculate its relationship with all initial centers distance, [E(n) i , SNR i 、F i ]Assigned to the cluster with the nearest center In, S j represents the jth cluster, which contains all feature vector samples assigned to the jth cluster center; j corresponds to cluster centers 1, 2, ..., k;

[0162] b. Calculate the mean within the cluster and update the center

[0163] For each Calculate the mean vector of all its samples as the new center after the first iteration at this time It is not a random sample, but the centroid of the cluster sample, which is more representative than the initial center. The formula is:

[0164]

[0165] c. Repeat allocation and update until convergence

[0166] Based on the new center Reassign all samples to the nearest cluster , calculate the average value of each cluster again to get Similarly, when the center of two adjacent iterations no longer changes, the iteration stops, and the final C t It is the cluster center;

[0167] (6) Label assignment and C-scan image reconstruction based on K-means algorithm: Set label L i =1 means defective, L i= 0 means no defect, each eigenvector [E(n) i , SNR i 、F i ] After clustering, it belongs to a certain category j and its corresponding label L i Set to 1 or 0 and retain its spatial coordinates (x i ,y i );

[0168] The A-scan signal of each element of the phased array probe constitutes three-dimensional data [x, y, A(x, y, t)], where (x, y) is the spatial position and A(x, y, t) is the time domain signal of the position; the label L of each feature vector is i Corresponding to its spatial coordinates (x i ,y i ), that is, each position (x, y) gets a category label, defect or non-defect;

[0169] Set the color mapping rule: if the label of the position (x, y) is a defect, the pixel color is set to red; if the label of the position (x, y) is not a defect, the pixel color is set to a light color; arrange all pixels in order of spatial coordinates to generate a C-scan reconstructed image. The C-scan reconstructed image is a two-dimensional plane image, and the color can reflect the defect distribution.

[0170] (7) Defect location and quantification:

[0171] Extract the pixel coordinates (x, y) of the cluster label "defect" from the C-scan reconstructed image to determine the plane position of the defect. Combined with the time t of the A-scan signal and the material sound velocity c, calculate the defect depth d = (c × t) / 2;

[0172] The image width is calculated from the number of acoustic beams and the chip spacing. The total width = (number of acoustic beams - 1) × chip spacing. The distance between the centers of adjacent acoustic beams in the width direction is the pixel width, and the actual scanning step length in the length direction is the pixel length. The actual physical size of each pixel in the C-scan reconstructed image is pixel width × pixel length. The number of pixels marked as defect areas in the C-scan reconstructed image is counted. The actual defect size = number of pixels N × area of ​​a single pixel. The final defect size is expressed as area × depth, which is the final K-means model.

[0173] The original image of the typical defect simulation sample and the C-scan image reconstructed by clustering algorithm are shown as follows: Figure 6 Table 3 shows the comparison between the actual defect size and the predicted size of defects 1, 2, 3, 4 and 5. The predicted error of the defect size of the defective specimens is less than 3.0%.

[0174] Figure 5 Typical defect C-scan images: (a) macroscopic image; (b) K-means reconstructed image;

[0175] (c) 20mm deep defect; (d) 30mm deep defect

[0176] Table 3 Quantitative evaluation of inclusion and delamination defect specimens

[0177]

[0178]

[0179] An ultrasonic phased array device (phased array ultrasonic instrument) is used to scan the wind turbine blades with a low-frequency linear array probe to obtain A-scan and B-scan signals. The A-scan signal is input into the K-means model and the image recognition result is output. The result will clearly identify the possible defects in the blade root and the location and size of the defects.

[0180] The method also includes step 3: model training and defect characterization based on B-scan signals, which involves learning the signals of the B-scan data of defective wind turbine blade samples and characterizing the defects. Specifically,

[0181] (1) Defect data annotation and dataset division

[0182] Use Labelmg annotation software to select the defect area in the ultrasonic B-scan data image and associate the corresponding defect category labels, including normal, wrinkle, inclusion, and delamination;

[0183] The specific steps are:

[0184] Press the left mouse button in the upper left corner of the defect area in the image, drag it to the lower right corner, and release it to form a rectangular bounding box. Make sure the selection covers the visible part of the defect. After the selection is complete, associate the corresponding defect category label in the pop-up label window;

[0185] Defect categories are divided into four categories: normal, wrinkles, inclusions, and delamination. (1) Normal: There are no abnormal echo signals in the image, the grayscale distribution is uniform, there are no obvious boundaries or areas with sudden changes in shape, and it is consistent with the ultrasonic imaging characteristics of the normal tissue structure of the material; (2) Wrinkles: The image appears as a continuous or discontinuous linear high echo area, which is mostly wavy or folded in shape, with relatively clear boundaries, and often appears in the bending or stress concentration parts of the material; (3) Inclusions: Isolated mass-like high echo areas appear in the image, with irregular boundaries, uneven internal grayscale, and may be accompanied by acoustic shadows around them; (4) Delamination: The image appears as a straight line or curve parallel to the material surface, extending along the interface between layers, accompanied by the "peeling" morphology of interlayer separation;

[0186] After the annotation is completed, a YOLOV format annotation file (.txt) is generated. All coordinates and dimensions are expressed in normalized pixel values, ranging from 0 to 1. After annotation, all annotated samples are randomly shuffled and divided into training set, validation set, and test set according to the first 70% samples, the middle 20% samples, and the last 10% samples, that is, the training set: validation set: test set = 7:2:1;

[0187] Defect image annotation files and LabelImg software page can be found at Figure 7 As shown;

[0188] (2) Yolov8 model training and model optimization

[0189] The training set is input into the Yolov8 model for training. After every 10 rounds of training, the mean average precision (mAP) is calculated using the validation set. If mAP is less than 0.7, the neck network in the Yolov8 model parameters is optimized and adjusted, a shallow feature detection layer is added, and the model is continuously trained. If the validation set mAP does not improve for 15 consecutive rounds, the current optimal model is saved and training is terminated.

[0190]

[0191] Where mAP is the mean average precision, AP is the average precision, C is the total number of categories in the dataset, and AP(c) is the average precision of the cth category. Average precision AP is the area of ​​a single category on the precision-recall curve;

[0192] (3) Deployment Inference

[0193] Model pruning optimization is used to reduce the model size while maintaining model accuracy. The Network Slimming algorithm is used to prune based on the BN layer scaling factor γ, with a weight of 0.001 applied to γ, and channels with γ < 0.1 are pruned, compressing the model volume by 30% to 50%. The pruning formula is:

[0194]

[0195] Where: γ is the scaling factor; N is the number of samples in the current batch; x i is the input feature; μ B and is the mean and variance of the current batch; γ and β are learnable scaling factors and offsets; ∈ is a small constant that prevents the denominator from being zero;

[0196] (4) Effect evaluation

[0197] The optimized YOLOv8 model is input into the test set to output defect recognition and classification results. The model effectiveness is evaluated using three metrics: precision, recall, and mAP.

[0198]

[0199] Among them, TP is a true positive, FP is a false positive, and FN is a false negative;

[0200] Under the relevant environment configuration, train 100 rounds to obtain the final YOLOv8 model; Figure 7 The graph of the change of loss and average accuracy during model training with the increase of training rounds shows that the model performance increases rapidly at the beginning of training and then gradually stabilizes. After 100 rounds of training, the prediction accuracy of the model is very high; the prediction results of B-scan images at different positions of typical defective samples are shown in Figure 8 shown.

[0201] An ultrasonic phased array device (phased array ultrasonic instrument) is used to scan the wind turbine blades with a low-frequency linear array probe to obtain A-scan and B-scan signals. The B-scan data is input into the improved YOLOv8 model, and the image recognition results are output. The results will clearly identify the possible defects and categories in the blade roots.

[0202] This application has achieved good technical results through practical application, and the application examples are as follows:

[0203] The method of the present invention is used to detect the blades of a wind turbine in a certain wind farm. The data collected on site are as follows: Figure 10 After data collection, it was found that there was a wrinkle defect on the C010 blade, the defect size was 31mm×10mm, the detailed information of the defect is shown in Table 4, and the identification images of the defect quantitative and qualitative are shown in Figure 11 、 12 As shown. The blade defects after the fan is removed from the rack are shown Figure 12 shown.

[0204] Table 4 Detailed information of defects detected on site

[0205]

Claims

1. An intelligent wind turbine blade defect identification method based on phased array detection, characterized in that: The following steps are involved: Step 1: Use a low-frequency linear array probe to test a simulated wind turbine blade sample and obtain A-scan, B-scan, and C-scan views. Specifically: (1) Determine the probe model The probe model is determined to be 0.5L64-1.5×22-MW5-F2.5-D3; (2) Preparation of simulated test blocks 9 simulated wind turbine blade specimens with inclusions, delaminations, and wrinkles were processed and numbered; (3) Step test block production To establish the corresponding relationship between the defect depth and size of wind turbine blade simulation specimens and the detection signal characteristics and provide a quantitative identification basis, a stepped test block with dimensions of 100 mm in length × width × 50 mm and thicknesses of 10 mm, 20 mm, 30 mm, and 40 mm was designed and manufactured. The sides of the four stepped test blocks were connected end to end in ascending order along the length direction, with the bottom surfaces flush, forming a stepped shape. (4) Basic parameter settings After connecting the probe to the instrument, calibrate the device's zero point and sound velocity using the direct contact method on a stepped test block using chemical paste as a coupling agent. Then, place the probe on a defect-free area of ​​the simulated specimen and adjust the bottom wave to 80% of the screen's full scale, which serves as the test sensitivity. (5) Scanning process Nine simulated test blocks were tested using ultrasonic phased array technology, with data collected using a C-scan method. The length of the simulated test block was along the X-axis, and the width was along the Y-axis. The intersection of the X-axis and Y-axis at the lower left corner of the simulated test block was at point (0,0). The probe first scanned the simulated test block from point (0,0) along the positive X-axis until it reached the edge of the simulated test block. The probe then scanned from the edge of the simulated test block along the negative X-axis, with a 15% overlap with the previous scan path, until it reached the edge of the simulated test block. This process was repeated until the entire simulated test block was scanned, resulting in A-scan and B-scan data. Step 2: Feature extraction and clustering training based on A-scan signals (1) Normalized signal: Use the direct traversal method to scan all sampling points of the A-scan data signal x(t) one by one, and record the current maximum value x max and the minimum value x min , let the original signal be x(t)=[x1,x 2,…, x N ], initialize: x max =x1,x min =x1, then i=2 to N:x max =max(x max , x i ), x min =min(x min ,x i ); for each sampling point value x i The normalized value x is calculated by the following formula / i , the signal interval range is [0,1], and all normalized values ​​are x(n): (2) Wavelet packet decomposition: The normalized numerical signal is recursively decomposed to the jth layer using wavelet packet transform to generate 2 j low-frequency and high-frequency sub-bands; First-layer decomposition: The normalized signal is processed using low-pass filters and high-pass filters corresponding to the orthogonal wavelet basis functions; After decomposition, two frequency bands are obtained, low-frequency sub-band A1 and high-frequency sub-band D1; a. Low-pass filtering The normalized value is x(n), the low-pass filter coefficient is h(n), and the low-frequency component A1 is calculated as: Where: k is the index of the decomposed signal, h(2k-n) represents the convolution operation after even sampling and flipping the filter system; b. High-pass filtering The normalized value is x(n), the high-pass filter coefficient is g(n), and the high-frequency component D1 is calculated as: Where: g(n) is usually orthogonal to h(n), and g(2k-n) represents the convolution operation after even sampling and flipping the filter system; Second-layer decomposition: The same low-pass filter h(n) and high-pass filter g(n) as in the first layer are applied to the low-frequency subband (A1) and high-frequency subband (D1) of the first layer respectively to ensure the consistency of the decomposition rules. After decomposition, four frequency bands are obtained: A2, D21, D22, and D23. a. Decomposition of the first layer low-frequency subband (A1) The A1 signal is A1(m), which is subjected to low-pass and high-pass filtering to obtain the second-layer low-frequency sub-band A2 and high-frequency sub-band D21: The calculation formula for A2 is: The D21 calculation formula is: b. Decomposition of the first high-frequency subband (D1) The D1 signal is D1(m), which is subjected to low-pass and high-pass filtering to obtain the second-layer low-frequency sub-band D22 and high-frequency sub-band D23: The calculation formula for D22 is: The D23 calculation formula is: Recursive decomposition from layer k to layer j (k = 2, 3, ..., j-1): Each layer repeats the "low-pass + high-pass filtering" decomposition operation on all low-frequency and high-frequency sub-bands of the previous layer, and each decomposition splits each sub-band into 2 sub-bands, where the kth layer has 2 k sub-bands, and the k+1th layer has 2 k+1 sub-bands; j-th layer decomposition: After j times of recursive decomposition, 2j sub-bands are finally generated. Each sub-band corresponds to a different frequency band range and covers the full frequency band of the original signal. (3) Feature vector construction: The energy proportion E(n), signal-to-noise ratio SNR and comprehensive index F are integrated to construct a feature vector. The feature vector is expressed as a. Calculate the target sub-band energy ratio E(n): a.1 For the coefficient X(j,n) of the nth subband in the jth layer, calculate the subband energy: Where M is the subband coefficient length; a.2 Calculate the total energy of all sub-bands: a.3 Energy proportion E(n) is: b. Calculate the signal-to-noise ratio (SNR): in, X is the original noisy signal; It is the result of normalized signal decomposition by wavelet packet, sub-band denoising and inverse transformation reconstruction; c. Calculate comprehensive indicators: Where: α is the weight coefficient, which is 0.6, SNR max is the maximum SNR value of all filtered signals; (4) Determine the number of clusters K and the standardization of eigenvectors: According to the detection properties of wind turbine blades, the data are divided into defective and non-defective data, that is, K = 2, and the eigenvector [E(n) i , SNR i 、F i ] to ensure that features of different dimensions have consistent weights in distance calculation, specifically: a. Calculate the mean and standard deviation of each feature First, collect the E(n) values ​​of all signals, denoted as E1, E2…, E m , where k is the total number of signals; calculate the mean and standard deviation of E(n), the mean formula and standard deviation formula are: Similarly, we process the SNR and F features to get μ SNR , σ SNR 、μ F , σ F . b. Standardize each eigenvalue E(n) normalization: SNR normalization: F feature normalization: c. Generate standardized feature vector The feature vector of each sample is [E(n) i , SNR i 、F i ] is converted to [E / (n) i , SNR / i 、F / i ], at this time, the mean of each feature is 0, the standard deviation is 1, the dimension is unified, and the standardized feature vector set is {E / (n), SNR / 、F / }; (5) Using K-means algorithm to set cluster centers From the normalized feature vector set {E / (n), SNR / 、F / } randomly select k vectors as the initial cluster centers At this time, the initial cluster centers are random samples and cannot reflect the true distribution of the clusters; a. Sample allocation For each eigenvector [E(n) i , SNR i 、F i ] Calculate its relationship with all initial centers distance, [E(n) i , SNR i 、F i ]Assigned to the cluster with the nearest center In, S j represents the jth cluster, which contains all feature vector samples assigned to the jth cluster center; j corresponds to cluster centers 1, 2, ..., k; b. Calculate the mean within the cluster and update the center For each Calculate the mean vector of all its samples as the new center after the first iteration at this time It is not a random sample, but the centroid of the cluster sample, which is more representative than the initial center. The formula is: c. Repeat allocation and update until convergence Based on the new center Reassign all samples to the nearest cluster , calculate the average value of each cluster again to get Similarly, when the center of two adjacent iterations no longer changes, the iteration stops, and the final C t It is the cluster center; (6) Label assignment and C-scan image reconstruction based on K-means algorithm: Set label L i =1 means defective, L i = 0 means no defect, each eigenvector [E(n) i , SNR i 、F i ] After clustering, it belongs to a certain category j and its corresponding label L i Set to 1 or 0 and retain its spatial coordinates (x i ,y i ); The A-scan signal of each element of the phased array probe constitutes three-dimensional data [x, y, A(x, y, t)], where (x, y) is the spatial position and A(x, y, t) is the time domain signal of the position; the label L of each feature vector is i Corresponding to its spatial coordinates (x i ,y i ), that is, each position (x, y) gets a category label, defect or non-defect; Set the color mapping rule: if the label of the position (x, y) is a defect, the pixel color is set to red; if the label of the position (x, y) is not a defect, the pixel color is set to a light color; arrange all pixels in order of spatial coordinates to generate a C-scan reconstructed image. The C-scan reconstructed image is a two-dimensional plane image, and the color can reflect the defect distribution. (7) Defect location and quantification: Extract the pixel coordinates (x, y) of the cluster label "defect" from the C-scan reconstructed image to determine the plane position of the defect. Combined with the time t of the A-scan signal and the material sound velocity c, calculate the defect depth d = (c × t) / 2; The image width is calculated from the number of acoustic beams and the chip spacing. The total width = (number of acoustic beams - 1) × chip spacing. The distance between the centers of adjacent acoustic beams in the width direction is the pixel width, and the actual scanning step length in the length direction is the pixel length. The actual physical size of each pixel in the C-scan reconstructed image is pixel width × pixel length. The number of pixels marked as defect areas in the C-scan reconstructed image is counted. The actual defect size = number of pixels N × area of ​​a single pixel. The final defect size is expressed as area × depth, which is the final K-means model. Ultrasonic phased array equipment is used to scan wind turbine blades using a low-frequency linear array probe to obtain A-scan and B-scan signals. The A-scan signal is input into the K-means model and the image recognition result is output. The result will clearly identify possible defects in the blade root and the location and size of the defects.

2. The method for intelligently identifying defects in wind turbine blades based on phased array detection according to claim 1, characterized in that: The 9 simulation test blocks described in step 1 are: Two test blocks with dimensions of 300mm x 300mm x 50mm in length, width and height, containing inclusions. Each test block has two inclusion defects. The two inclusion defects on one test block are 33mm x 27mm x 20mm and 35mm x 33mm x 40mm respectively. The two inclusion defects on the other test block are 47mm x 41mm x 20mm and 56mm x 33mm x 30mm respectively. Three test blocks with a length, width and height of 300mm × 300mm × 55mm and a single delamination defect. The delamination defect dimensions of the three test blocks are 49mm × 47mm × 30mm, 138mm × 130mm × 30mm, and 113mm × 107mm × 30mm, respectively. The four test blocks containing one wrinkle defect have a length, width and height of 300mm×300mm×55mm, 300mm×300mm×60mm, 300mm×300mm×40mm and 300mm×300mm×65mm respectively; the corresponding wrinkle defect aspect-to-height ratios are 0.03, 0.06, 0.1 and 0.14 respectively.

3. The intelligent identification method for wind turbine blade root defect detection based on phased array detection according to claim 1 is characterized in that: The method also includes step 3: model training and defect characterization based on B-scan signals, which involves learning the signals of the B-scan data of defective wind turbine blade samples and characterizing the defects. Specifically, (1) Defect data annotation and dataset division Use Labelmg annotation software to select the defect area in the ultrasonic B-scan data image and associate the corresponding defect category labels, including normal, wrinkle, inclusion, and delamination; After the annotation is completed, a YOLOV format annotation file (.txt) is generated. All coordinates and dimensions are expressed in normalized pixel values, ranging from 0 to 1. After annotation, all annotated samples are randomly shuffled and divided into training set, validation set, and test set according to the first 70% samples, the middle 20% samples, and the last 10% samples, that is, the training set: validation set: test set = 7:2:1; (2) Yolov8 model training and model optimization The training set is input into the Yolov8 model for training. After every 10 rounds of training, the mean average precision (mAP) is calculated using the validation set. If mAP is less than 0.7, the neck network in the Yolov8 model parameters is optimized and adjusted, a shallow feature detection layer is added, and the model is continuously trained. When the validation set mAP does not improve for 15 consecutive rounds, save the current optimal model and terminate the training; Where mAP is the mean average precision, AP is the average precision, C is the total number of categories in the dataset, and AP(c) is the average precision of the cth category. Average precision AP is the area of ​​a single category on the precision-recall curve; (3) Deployment Inference Model pruning optimization is used to reduce the model size while maintaining model accuracy. The Network Slimming algorithm is used to prune based on the BN layer scaling factor γ, with a weight of 0.001 applied to γ, and channels with γ < 0.1 are pruned, compressing the model volume by 30% to 50%. The pruning formula is: Where: γ is the scaling factor; N is the number of samples in the current batch; x i is the input feature; μ B and is the mean and variance of the current batch; γ and β are learnable scaling factors and offsets; ∈ is a small constant that prevents the denominator from being zero; (4) Effect evaluation The optimized YOLOv8 model is input into the test set to output defect recognition and classification results. The model effectiveness is evaluated using three metrics: precision, recall, and mAP. Among them, TP is a true positive, FP is a false positive, and FN is a false negative; Under the relevant environment configuration, train 100 rounds to obtain the final YOLOv8 model; Ultrasonic phased array equipment is used to scan wind turbine blades with a low-frequency linear array probe to obtain A-scan and B-scan signals. The B-scan data is input into the improved YOLOv8 model, and the image recognition results are output. The results will clearly identify the possible defects and categories in the blade roots.

Images