Composite material wind power blade damage detection method

Abstract

The invention discloses a composite material wind power blade damage detection method, and aims to solve the problems of difficulty in quantitative analysis and low distinguishing precision of various damage types in composite material wind power blade damage detection in the prior art. Comprising the following steps that acoustic emission signals in the operation process of the wind power blade are obtained in real time and preprocessed, and the acoustic emission signals comprise parameter data and waveform data; performing unsupervised clustering on the preprocessed parameter data to obtain a preliminary clustering result; training the preprocessed waveform data by adopting an improved multi-branch convolutional neural network, and outputting an accurate classification result; the unsupervised clustering result is verified by using the supervised learning classification result, and the accuracy of the clustering result is evaluated; and fusing the quantitative index accumulated energy, the damage type and the frequency, and constructing a comprehensive damage evaluation model to determine the damage degree. According to the method, the unsupervised clustering method and the supervised learning method are combined, and accurate identification and quantification of the damage are realized.

Description

Technical Field

[0001] The present invention relates to the field of material damage monitoring and identification, and in particular to a method for detecting damage to composite material wind turbine blades. Background Art

[0002] Composite wind turbine blades are prone to damage during their service life due to long-term exposure to harsh environments such as strong winds, rain, and lightning strikes. As blade size increases, their weight and mechanical loads also increase accordingly, making the blades more susceptible to stress concentration, leading to problems such as microcracks and fatigue damage. Unlike metal materials, the multi-layered structure of composite materials often makes damage appear hidden and complex, which increases the difficulty of damage detection. Therefore, real-time monitoring and precise quantification of damage in composite wind turbine blades is of great engineering significance and can effectively improve the safety and operational efficiency of wind turbines.

[0003] Currently, most methods for monitoring damage in composite wind turbine blades using acoustic emission technology rely on either unsupervised or supervised learning to identify damage types, but each has its own shortcomings. Unsupervised learning is not accurate enough in analyzing acoustic emission data, especially when the differences between damage features are not obvious. The clustering results are often not highly discriminative and fail to accurately reflect the actual physical damage mechanism. Damage identification methods that rely solely on supervised learning require the establishment of a high-quality annotated dataset, including accurate annotation of each damage type in the acoustic emission data. However, in practice, acoustic emission data annotation is difficult and costly, and obtaining a large number of samples of real damage data is particularly challenging in the field of composite wind turbine blades.

[0004] Although some studies have attempted to apply machine learning to the processing of acoustic emission signals, most methods lack quantitative analysis of different degrees of damage and cannot be effectively promoted and applied on a larger scale; for example, the non-destructive testing system and method for wind turbine blades with publication number CN103901111A. Summary of the Invention

[0005] The present invention solves the problems of difficulty in quantitative analysis and low accuracy in distinguishing multiple damage types in the existing technology of damage detection of composite material wind turbine blades. It proposes a damage detection method for composite material wind turbine blades, which combines unsupervised clustering and supervised learning methods to achieve accurate identification and quantification of damage.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solution: a method for detecting damage to composite material wind turbine blades, comprising the following steps: S1, real-time acquisition of acoustic emission signals during the operation of wind turbine blades and pre-processing of the signals, the acoustic emission signals including parameter data and waveform data; S2, performing unsupervised clustering on the preprocessed parameter data to obtain preliminary clustering results; S3, uses an improved multi-branch convolutional neural network to train the preprocessed waveform data and output accurate classification results; S4, using the supervised learning classification results to verify the unsupervised clustering results and evaluate the accuracy of the clustering results; S5, the quantitative indicators cumulative energy, damage type and frequency are integrated to construct a comprehensive damage assessment model to determine the damage extent.

[0007] This technical solution provides a composite material wind turbine blade damage detection method that combines unsupervised clustering and supervised learning. It uses unsupervised learning to perform preliminary clustering to discover the natural distribution structure of acoustic emission signals, and further verifies and accurately classifies the clustering results through supervised learning, thereby achieving accurate identification of wind turbine blade damage and establishing a quantitative analysis model. It is highly innovative and practical.

[0008] The present invention is further configured as follows: Step S1 includes: S11, installing an acoustic emission sensor at a key position of the wind turbine blade, and the acoustic emission sensor detects the generated acoustic emission signal in real time; S12, performing bandpass filtering and wavelet denoising on the parameter data of the acoustic emission signal, and then performing standardization on the parameter data; S13, extracting a time-frequency spectrum diagram, a power spectrum density diagram, and a wavelet transform feature diagram of the waveform data.

[0009] In this technical solution, the acoustic emission signals of the wind turbine blades are specifically acquired through acoustic emission sensors installed at different key parts of the wind turbine blades. The acoustic emission sensors can acquire parameter data and waveform data. After acquiring the above data, the parameter data is denoised and standardized, and the waveform data is feature extracted.

[0010] The present invention is further configured as follows: Step S2 includes: The HDBSCAN clustering algorithm is used to perform unsupervised clustering on the preprocessed acoustic emission parameter data. After clustering, the distribution characteristics and central eigenvalues ​​of the acoustic emission parameters in each cluster are analyzed to preliminarily determine the damage type that each cluster may correspond to.

[0011] In this technical solution, the HDBSCAN clustering algorithm is used in the unsupervised clustering processing of acoustic emission parameter data.

[0012] The present invention is further configured as follows: the step of unsupervised clustering includes: S21, calculate the distance matrix between data points based on the parameter data, and estimate the core distance and reachability distance of each data point according to its density; S22, construct a density-based tree hierarchy, divide the clusters on this basis, obtain multiple clusters with higher density, and mark the noise points; S23, performing multiple iterative cluster analysis on the acoustic emission parameter data, using the minimum sample number parameter for tuning, and determining the optimal clustering result.

[0013] In this technical solution, the damage type that each cluster may correspond to can be preliminarily determined through the above process.

[0014] The present invention is further configured as follows: the network architecture of the improved multi-branch convolutional neural network includes three parallel feature extraction branches, a feature fusion module and a classification module.

[0015] In this technical solution, an improved multi-branch convolutional neural network algorithm is used to train acoustic emission waveform feature maps and identify damage. The input of the supervised learning model is a variety of feature maps extracted through preprocessing, including time-frequency spectrum maps, power spectrum density maps, and wavelet transform feature maps.

[0016] The present invention is further configured as follows: Step S4 includes: S41, using the adjusted Rand index or normalized mutual information metric to compare supervised learning classification results with unsupervised clustering results; S42, if the supervised learning classification result and the unsupervised clustering result are highly consistent, the unsupervised clustering result is considered accurate; if the supervised learning classification result and the unsupervised clustering result are not consistent, the cause of the clustering error is found by analyzing the specific difference samples, and the clustering model is optimized or the hyperparameters of the unsupervised clustering are readjusted.

[0017] In this technical solution, the verification of supervised learning results and the comparative analysis of unsupervised clustering help evaluate the accuracy of unsupervised clustering. By comparing the consistency of supervised learning classification results with unsupervised clustering results, and using the relevant indicators ARI and NMI to evaluate both, the reliability of the clustering results is verified.

[0018] The present invention is further configured as follows: Step S5 includes: Through the classification results of the acoustic emission waveform feature map using a supervised learning model, various damage types are identified and quantitative numerical weights are assigned to different damage types. Combined with the output of the multi-branch CNN model, the number of damage events occurring in each time period and their corresponding damage types are counted to obtain the frequency of damage events.

[0019] In this technical solution, the above steps need to be performed before the quantitative indicators are integrated.

[0020] The present invention is further configured as follows: Step S3 includes constructing a data set, and the data set construction includes: The waveform data of the acoustic emission signal was processed to obtain the time-scale coefficient matrix describing the multi-scale characteristics of the signal. The classification results obtained by unsupervised clustering were used as the initial labels. The acoustic emission events were paired with the corresponding damage types in chronological order. The complete dataset was divided into training set, validation set, and test set in a ratio of 7:2:1.

[0021] In this technical solution, each training sample contains the above three feature representations and their corresponding damage type labels.

[0022] The present invention is further configured as follows: Step S5 further includes: The quantitative indicators are integrated to construct a comprehensive damage assessment model, which combines the results of cumulative energy, damage type and frequency to calculate an overall damage score to evaluate the current health status of the material.

[0023] The present invention is further configured as follows: when the overall damage score of the blade is monitored to exceed a first threshold, the operation of the blade will enter an observation state, and its damage changes will be monitored regularly; when the overall damage score of the blade is monitored to exceed a second threshold, the blade will suspend operation and undergo a comprehensive inspection.

[0024] In this technical solution, the status of the blade is determined based on the overall damage score. During actual operation, if the overall damage score of the blade exceeds the first threshold or the second threshold, corresponding measures will be taken for monitoring or inspection.

[0025] The present invention can bring the following beneficial effects: 1. A multi-branch convolutional neural network that combines unsupervised clustering of acoustic emission parameter data with supervised learning of acoustic emission waveform features can efficiently identify and classify different damage types. This multi-level damage identification method significantly improves recognition accuracy. Unsupervised learning clusters parameter features to quickly identify potential patterns and clusters in the data, providing a foundation for subsequent supervised learning. Supervised learning is then used to more accurately analyze and classify waveform features rich in information, improving the accuracy and reliability of damage identification. The combination of the two can quickly detect typical damage in various composite wind turbine blades and adapt to different damage patterns. 2. Using multiple acoustic emission feature maps (time-frequency spectrum, power spectrum density, and wavelet transform feature maps) as input for supervised learning can more accurately analyze the damage mechanism of materials. Compared with traditional single-feature analysis methods, multi-feature fusion can more comprehensively reflect the diversity and complexity of damage, making the damage classification results more physically meaningful. 3. The damage quantification analysis step of the present invention not only qualitatively determines the damage type but also provides a quantitative damage assessment based on the cumulative energy, damage type, and frequency of the acoustic emission signal. This allows for automated assessment and prediction of the extent of material damage, significantly improving the accuracy and automation of damage monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a composite material wind turbine blade damage detection method of the present application.

[0027] Figure 2 This is a schematic diagram of acoustic emission parameter data clustering for a composite material wind turbine blade damage detection method in this application.

[0028] Figure 3 This is a schematic diagram of a multi-branch CNN architecture for acoustic emission signal analysis in a composite material wind turbine blade damage detection method of the present application. DETAILED DESCRIPTION

[0029] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] The multi-layer structure of composite wind turbine blades often makes damage hidden and complex, making damage detection more difficult.

[0031] Acoustic emission (AE) technology, as a non-destructive testing (NDT) method, has been widely used in material damage monitoring and structural health monitoring. It has significant application value in damage monitoring of composite wind turbine blades. The complex structure and multiple damage mechanisms of composite wind turbine blades, such as delamination, cracks, and fiber breakage, make accurate damage monitoring and quantification difficult. Traditional damage detection methods, such as ultrasonic testing and X-ray imaging, while offering high accuracy, have limitations in large-scale, real-time monitoring and cannot meet the requirements for online monitoring of wind turbine blades during operation.

[0032] Research at home and abroad has made some progress in using acoustic emission technology to monitor damage in composite wind turbine blades. Most methods rely on parameter characteristics of the acoustic emission signal (such as ring count, energy, and amplitude) to qualitatively identify damage. However, these parameters only roughly characterize the damage state and are difficult to quantitatively describe the severity of the damage. Furthermore, acoustic emission signals are often contaminated by background noise, making it difficult for a single parameter to reflect complex damage mechanisms. Therefore, improving detection accuracy and damage classification accuracy by combining multiple acoustic emission characteristics (such as spectral characteristics and time-frequency analysis) with advanced algorithms (such as machine learning) remains a challenge.

[0033] In the analysis of acoustic emission data, existing methods mostly rely on unsupervised or supervised learning to identify damage types, but each has its own shortcomings. When using only unsupervised learning methods, damage types are usually classified based on feature clustering. Common algorithms include K-means, hierarchical clustering, and density clustering. However, unsupervised learning is not accurate enough in acoustic emission data analysis. Especially when the differences between damage features are not obvious, the clustering results are often not highly discriminative and cannot accurately reflect the actual physical damage mechanism. In addition, the classification results obtained by unsupervised clustering lack clear physical meaning, making it difficult to directly use them to determine damage level or quantify health status.

[0034] Damage identification methods that rely solely on supervised learning require the establishment of a high-quality annotated dataset, including accurate annotations of each damage type in the acoustic emission data. However, in practice, acoustic emission data annotation is difficult and costly, especially in the field of composite wind turbine blades, where obtaining a large number of real-world damage data samples is extremely challenging. Furthermore, supervised learning relies on a predefined number of categories, but determining an appropriate number of damage categories in unknown or complex damage patterns is challenging, making it difficult to guarantee the effectiveness of the classification model.

[0035] Although some studies have attempted to apply machine learning to acoustic emission signal processing, most methods lack quantitative analysis of different damage levels, hindering their widespread application. Furthermore, existing approaches are still insufficient in correlating clustering results with physical damage characteristics, making it difficult to fully reveal damage mechanisms. Therefore, a more systematic acoustic emission analysis method is needed to accurately identify and quantify damage and address the shortcomings of existing technologies.

[0036] Example 1 In view of the above technical problems, this embodiment proposes a composite material wind turbine blade damage detection method, referring to Figures 1 to 3 , which mainly includes the following steps.

[0037] Step S1 : Acquire and pre-process the acoustic emission signal of the wind turbine blade in real time during operation. The acoustic emission signal includes parameter data and waveform data.

[0038] The above-mentioned step S1 mainly includes the following sub-steps.

[0039] Step S11 : installing an acoustic emission sensor at a key position of a wind turbine blade, and detecting an acoustic emission signal in real time through the installed acoustic emission sensor.

[0040] In step S12, the parameter data of the acoustic emission signal are subjected to bandpass filtering and wavelet denoising in sequence, and then the parameter data are subjected to standardization.

[0041] Step S13: extracting the time-frequency spectrum, power spectrum density and wavelet transform characteristic diagram of the waveform data.

[0042] In this technical solution, the acoustic emission signals of the wind turbine blades are specifically acquired through acoustic emission sensors installed at different key parts of the wind turbine blades. The acoustic emission sensors can acquire parameter data and waveform data. After acquiring the above data, the parameter data is denoised and standardized, and the waveform data is feature extracted.

[0043] Specifically, the key parts of a wind turbine blade include but are not limited to the leading edge, trailing edge, root and wingtip areas of the blade.

[0044] The aforementioned parameter data include but are not limited to amplitude, energy, duration, peak frequency and centroid frequency.

[0045] The above data sampling frequency is set above 1 MHz to ensure that the complete damage characteristic signal is captured.

[0046] The data acquisition adopts the trigger mode, which records data only when the acoustic emission event exceeds the set threshold to reduce the interference of background noise.

[0047] The above data preprocessing process mainly includes data denoising, standardization and feature extraction.

[0048] Firstly, the acoustic emission signal is bandpass filtered to limit the signal bandwidth to the range of 20kHz to 1MHz, and the wavelet denoising method is used to reduce the influence of high-frequency noise.

[0049] Then, the acoustic emission parameter data were normalized and converted into a 0-1 interval or a standard normal distribution.

[0050] For waveform data, feature images such as time spectrum graph, power spectrum density graph and wavelet transform feature graph are extracted.

[0051] Step S2: performing unsupervised clustering on the parameter data pre-processed in step S1 to preliminarily classify the damage status of the wind turbine blades and obtain a preliminary clustering result.

[0052] More specifically, step S2 employs the HDBSCAN clustering algorithm to perform unsupervised clustering on the preprocessed acoustic emission parameter data. This algorithm automatically determines the number of clusters based on data density and effectively distinguishes between various damage types. The clustering results enable preliminary classification of wind turbine blade damage states, providing a dataset for subsequent supervised learning.

[0053] In this technical solution, the HDBSCAN clustering algorithm is used in the unsupervised clustering processing of acoustic emission parameter data.

[0054] The unsupervised clustering process begins with a preprocessed parameter dataset, including standardized multidimensional parameter features such as duration, peak frequency, and centroid frequency. The HDBSCAN clustering algorithm adaptively determines the number of clusters based on local density variations in the data, eliminating the need to pre-set the number of clusters. It also effectively identifies noise points and excludes them from the clustering results, thereby improving clustering accuracy.

[0055] The above unsupervised clustering process includes the following steps.

[0056] In step S21 , a distance matrix between data points is calculated based on the parameter data, and then the core distance and reachability distance of each data point are estimated based on its density.

[0057] Step S22: construct a density-based tree hierarchy, divide the clusters based on it, obtain multiple clusters with higher density, and mark the noise points at the same time.

[0058] In step S23, in order to improve the stability and accuracy of clustering, multiple iterative clustering analyses are performed on the input acoustic emission parameter data, and different minimum sample number parameters are used for tuning to ultimately determine the optimal clustering result.

[0059] After clustering is complete, the distribution characteristics and central eigenvalues ​​of the acoustic emission parameters in each cluster are analyzed to preliminarily determine the damage type that each cluster may correspond to. For the treatment of noise points, you can choose to ignore them or treat them as special damage signals for further analysis.

[0060] Step S3: Use an improved multi-branch convolutional neural network to train the preprocessed waveform data and output accurate classification results.

[0061] Step S3 includes a data set construction process, which specifically includes: The waveform data of the acoustic emission signal was processed to obtain a time-scale coefficient matrix describing the multi-scale characteristics of the signal. The classification results obtained by unsupervised clustering were used as initial labels. Acoustic emission events were paired with corresponding damage types in chronological order. The complete dataset was divided into training, validation, and test sets in a ratio of 7:2:1. Each training sample contained the three feature representations described above and its corresponding damage type label.

[0062] More specifically, the short-time Fourier transform (SFT) is first used to obtain the time-frequency characteristics of the acoustic emission signal. A Hanning window function is used during the transformation, with a window length of 256 points and a time window overlap of 50%. This ultimately results in a two-dimensional time-frequency spectrum that reflects the signal's time-frequency characteristics.

[0063] The Welch method was used to calculate the power spectral density of the acoustic emission signal. The segment length was set to 512 points, the overlap ratio was set to 50%, and each segment of data was windowed using a Hanning window. The one-dimensional power spectral density curve representing the frequency domain energy distribution was obtained by averaging multiple periodograms.

[0064] The time-scale characteristics of acoustic emission signals are obtained based on continuous wavelet transform. The Morlet wavelet is selected as the mother wavelet function, and the scale range is set to 1-64 to cover the main frequency components of the signal. Finally, a time-scale coefficient matrix that can describe the multi-scale characteristics of the signal is obtained.

[0065] An improved multi-branch convolutional neural network algorithm is then used to train the acoustic emission waveform feature maps and identify damage. The input of the supervised learning model is a variety of feature maps extracted through preprocessing, including time-frequency spectrum maps, power spectrum density maps, and wavelet transform feature maps.

[0066] refer to Figure 3 ,For the network architecture of the above-mentioned improved multi-branch convolutional neural network, it mainly includes three parallel feature extraction branches, a feature fusion module and a classification module.

[0067] In more detail, the three feature extraction branches are the time-frequency spectrum branch, the power spectrum density branch, and the wavelet coefficient branch.

[0068] In this technical solution, an improved multi-branch convolutional neural network algorithm is used to train acoustic emission waveform feature maps and identify damage. The input of the supervised learning model is a variety of feature maps extracted through preprocessing, including time-frequency spectrum maps, power spectrum density maps, and wavelet transform feature maps.

[0069] For the above-mentioned time-frequency branch, it consists of two convolution blocks. Each convolution block contains two convolution layers with a convolution kernel size of 3×3. The two layers of the first convolution block use 32 and 64 convolution kernels respectively, and the two layers of the second convolution block use 64 and 128 convolution kernels respectively. Each convolution layer is followed by batch normalization and ReLU activation function, and a 2×2 maximum pooling layer is used for dimensionality reduction.

[0070] For the above-mentioned power spectral density branch, it adopts a one-dimensional convolution structure, which includes three convolution layers, using one-dimensional convolution kernels with kernel sizes of 7, 5, and 3 respectively, and the number of convolution kernels is 32, 64, and 64. After each convolution layer, a maximum pooling operation with a step size of 2 is used. Finally, 1×1 convolution is used for dimensionality reduction and global average pooling is used to obtain the feature vector.

[0071] For the wavelet coefficient branch, its structure is similar to that of the time-frequency spectrum branch, but a dilated convolution is introduced in the convolution operation to increase the receptive field. It consists of two convolution blocks. The first block uses 32 and 64 convolution kernels, and the second block uses 64 and 128 convolution kernels. It is also equipped with batch normalization, ReLU activation function and maximum pooling layer.

[0072] The feature fusion module described above first uses a channel-attention mechanism to weight the features extracted by the three branches, achieving adaptive feature selection. A 1×1 convolutional layer then learns the correlations between different features, with the number of output channels set to 256, effectively fusing features from multiple sources.

[0073] The classification module described above consists of two fully connected layers, with 512 and 128 neurons, respectively. A dropout layer is added after the first fully connected layer, with a dropout rate of 0.5 to prevent overfitting. Finally, a Softmax classifier is used to output a probability distribution of damage types, where the number of output nodes equals the total number of predefined damage categories.

[0074] Step S4: Use the supervised learning classification results to verify the unsupervised clustering results and evaluate the accuracy of the clustering results.

[0075] The above-mentioned step S4 includes the following sub-steps.

[0076] In step S41 , the adjusted Rand index (ARI) or normalized mutual information (NMI) is used to compare the supervised learning classification result and the unsupervised clustering result.

[0077] In step S42, if the classification results of supervised learning are highly consistent with the results of unsupervised clustering, it means that unsupervised clustering has high accuracy and reliability in the classification of damage types. In addition, the results of supervised learning can further identify possible noise points or abnormal points in unsupervised clustering, providing a basis for subsequent data correction. If there is a large difference between the two results, it may mean that during the clustering process, some damage types were not correctly classified due to unclear feature differences or noise interference. At this point, the cause of the clustering error can be found by analyzing specific difference samples, and then the clustering model can be optimized or the hyperparameters of the unsupervised clustering can be readjusted to improve the clustering effect.

[0078] In this technical solution, the verification of supervised learning results and the comparative analysis of unsupervised clustering help evaluate the accuracy of unsupervised clustering. By comparing the consistency of supervised learning classification results with unsupervised clustering results, and using the relevant indicators ARI and NMI to evaluate both, the reliability of the clustering results is verified.

[0079] In step S5, the quantitative index cumulative energy, damage type and frequency are integrated to construct a comprehensive damage assessment model to determine the damage extent.

[0080] The above-mentioned step S5 mainly includes the following sub-steps.

[0081] Through the classification results of the acoustic emission waveform feature map using a supervised learning model, various damage types are identified and quantitative numerical weights are assigned to different damage types. Combined with the output of the multi-branch CNN model, the number of damage events occurring in each time period and their corresponding damage types are counted to obtain the frequency of damage events.

[0082] The quantitative indicators are integrated to construct a comprehensive damage assessment model, which combines the results of cumulative energy, damage type and frequency to calculate an overall damage score to evaluate the current health status of the material.

[0083] In more detail, when the overall damage score of the blade is monitored to exceed the first threshold, the operation of the blade will enter the observation state, and its damage changes will be monitored regularly; when the overall damage score of the blade is monitored to exceed the second threshold, the blade will suspend operation and undergo a comprehensive inspection.

[0084] In this embodiment, blades can be classified into the following states based on their overall damage score: normal (D < 100), mild damage (100 ≤ D < 500), moderate damage (500 ≤ D < 1000), severe damage (1000 ≤ D < 2000), and critical (D ≥ 2000). In actual operation, if the overall damage score D of a blade exceeds 500 (i.e., moderate damage), the blade will enter observation mode and its damage changes will be regularly monitored. If the score exceeds 1000 (severe damage), it will be recommended to suspend operation and conduct a comprehensive inspection.

[0085] Example 2 Based on Example 1, this example describes the technical solution of Example 1 in more detail.

[0086] The technical solution of Example 1 mainly includes five steps: data acquisition and preprocessing, unsupervised clustering of acoustic emission parameter data, supervised learning of acoustic emission waveform data, verification of supervised learning results and comparison of unsupervised clustering results, and damage quantification analysis. These steps correspond to steps S1 to S5 of Example 1, respectively.

[0087] For the data acquisition and preprocessing process, the acoustic emission sensors are specifically arranged at key locations of the wind turbine blades, usually the root, middle and tip areas.

[0088] In this example, a highly sensitive acoustic emission sensor was selected with a frequency response range of 20 kHz to 2 MHz to cover the primary frequency range of acoustic emission signals. The sampling frequency was set to 5 MHz to ensure clear time-domain signal detail. The acquisition system was set to continuous acquisition mode, recording all acoustic emission events and background noise for subsequent data processing.

[0089] Among them, data preprocessing mainly includes the following three sub-processes.

[0090] 1. Signal Denoising: The collected raw acoustic emission signals may be contaminated by environmental noise, such as mechanical vibration and electromagnetic interference. Therefore, the signal is first filtered using a bandpass filter with a frequency range of 20kHz to 1MHz to remove background noise and high-frequency electromagnetic interference. Wavelet denoising is then used to further remove white noise from the signal, using the db4 wavelet base and a five-level decomposition. The denoised signal is reconstructed to retain valid acoustic emission events.

[0091] 2. Feature Extraction: Extract various time-domain and frequency-domain features from the denoised AE signal, including but not limited to amplitude, energy, duration, peak frequency, and centroid frequency. Frequency-domain features can be calculated using a fast Fourier transform (FFT), while energy features are obtained by integrating the square of the signal amplitude. The features extracted from each AE event are used in subsequent cluster analysis.

[0092] 3. Standardization: To eliminate the influence of different feature dimensions, all extracted features need to be standardized. In this embodiment, Z-score standardization is used. The mean of each feature is subtracted from the feature value and then divided by the standard deviation, so that the standardized data has a mean of 0 and a standard deviation of 1. Standardization can improve the accuracy and stability of cluster analysis.

[0093] For the unsupervised clustering process of acoustic emission parameter data, refer to Figure 2 , which specifically adopts the HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm to perform unsupervised cluster analysis on acoustic emission parameter data.

[0094] More specifically, the preprocessed parameter data is used as input, with the main features including duration, peak frequency, and centroid frequency. The HDBSCAN clustering parameters are then set, including the following key parameters: ① Minimum cluster size: The minimum cluster size is set to 500, indicating that each cluster contains at least 500 data points. This parameter setting can effectively remove isolated points and noise, improving clustering stability. ② Minimum number of samples: The minimum number of samples is set to 50 to determine the smoothness of the local density estimate. A smaller minimum number of samples can identify more cluster details, while a larger minimum number of samples can reduce noise points. ③ Cosine distance is used as a similarity metric. In high-dimensional space, cosine distance can better reflect the similarity of acoustic emission parameters.

[0095] After completing the parameter setting, the above-mentioned HDBSCAN clustering algorithm is executed, which mainly includes the following steps: ① Construct a similarity matrix: The similarity matrix is ​​obtained based on the cosine distance calculation to represent the density connectivity between data points. ② Construct a minimum spanning tree: Construct a density-connected minimum spanning tree through the similarity matrix, and sort the data points according to density. ③ Cluster division and noise identification: According to the structure of the minimum spanning tree, the data points are clustered, and isolated points that do not meet the density requirements are identified and marked as noise. ④ Calculate the number of core samples and allocation probability of each cluster: For each cluster, calculate the number of core samples and assign the probability of belonging to each cluster to each data point for subsequent damage identification.

[0096] Finally, the clustering results are visualized and analyzed, creating a cluster distribution map. Different colors represent different clusters, and points with cluster number -1 represent data points identified as noise. Unsupervised clustering allows for natural grouping of acoustic emission data, providing a reference for supervised learning and damage identification.

[0097] For the supervised learning process of acoustic emission waveform data, this embodiment uses an improved multi-branch convolutional neural network (Multi-branch CNN) for supervised learning to further identify and quantify damage to composite wind turbine blades.

[0098] First, feature extraction of the acoustic emission signal was performed. Three feature representations were extracted for each acoustic emission event: a time-frequency spectrum obtained using a short-time Fourier transform (SFT), a power spectral density estimated using the Welch method, and a time-scale coefficient plot obtained using a continuous wavelet transform (CWT). A Hanning window function was used to extract the time-frequency spectrum, with a window length of 256 points and a 50% overlap. The power spectral density was calculated using the Welch method with a segment length of 512 points and a 50% overlap. The Morlet wavelet was used as the mother wavelet function for the CWT, with a scale range of 1-64 to cover the main frequency components of the signal.

[0099] Next, all feature data were preprocessed and standardized. The amplitude of the time-frequency spectrum was normalized to the range [0, 1], the power spectrum density was logarithmically transformed and normalized, and the wavelet coefficients were amplitude normalized. The processed dataset was divided into training, validation, and test sets in a ratio of 7:2:1.

[0100] Next, a multi-branch CNN architecture was designed, consisting of three parallel feature extraction branches. The time-frequency branch utilizes two convolutional blocks, each containing two convolutional layers (kernel size 3×3). The first block uses 32 and 64 kernels, while the second uses 64 and 128 kernels. Each convolutional layer is followed by a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer. The power spectral density branch uses three one-dimensional convolutional layers with kernel sizes of 7, 5, and 3, respectively, and channels of 32, 64, and 64, respectively. Max pooling with a stride of 2 is used in each layer. The wavelet coefficient branch has a similar structure to the time-frequency branch, but employs dilated convolution to increase the receptive field.

[0101] In the feature fusion module, a channel-wise attention mechanism is first used to calculate the weights of the three branch features, enabling adaptive feature selection. A 1×1 convolutional layer (with 256 output channels) is then used to learn the correlations between features. The classification module consists of two fully connected layers (512 and 128 neurons), equipped with a dropout layer (dropout rate 0.5) to prevent overfitting. Finally, a softmax classifier is used to output the probability of the damage type.

[0102] Finally, the model was trained and optimized. The Adam optimizer was used with an initial learning rate of 0.001 and a cosine annealing schedule (50 epochs). The loss function combined the categorical cross entropy loss and the feature consistency loss, with a weighting ratio of 7:3. After each epoch, model performance was evaluated on the validation set, monitoring changes in classification accuracy, precision, recall, and F1 score. Early stopping was triggered when validation set performance stopped improving, saving the optimal model parameters. The training process lasted 200 epochs with a batch size of 64.

[0103] Through the above steps, the multi-branch CNN network can fully utilize the expression of acoustic emission signals in different feature domains to achieve effective recognition and accurate quantification of complex damage patterns.

[0104] The verification process of supervised learning results and the comparison process of unsupervised clustering results are carried out through comparative analysis of the two to further verify the rationality of clustering and the accuracy of the supervised learning model.

[0105] First, for the acoustic emission data classification task, unsupervised clustering (HDBSCAN) provided preliminary clustering results, while supervised learning used the trained model to classify the acoustic emission signatures in the test set. The purpose of this comparative analysis was to examine the consistency and differences between the two methods in damage identification and to analyze the physical significance of these differences. For this comparative analysis, the following evaluation metrics were used to quantify the differences in the results: 1. Adjusted Rand Index (ARI) The Adjusted Rand Index (ARI) assesses the consistency between clustering results and true classifications. It is the ratio of the first parameter (Index) to the second parameter (Expected Index), where the first parameter is the difference between the actual number of correct comparisons (Index) and the expected number of random comparisons (Expected Index). The second parameter is the difference between the maximum possible number of comparisons (Max Index) and the expected number of random comparisons (Expected Index). The ARI ranges from -1 to 1, with values ​​closer to 1 indicating a higher similarity between the clustering results and the supervised learning classification results.

[0106] 2. Normalized Mutual Information (NMI). Normalized Mutual Information (NMI) compares the degree of information sharing between clustering and classification results. It is the ratio of the third parameter to the fourth parameter. The third parameter is twice the mutual information (MI(U, V)), and the fourth parameter is the sum of the clustering entropy (H(U)) and the classification entropy (H(V)). The value of NMI ranges from 0 to 1, with values ​​closer to 1 indicating a higher correlation between the two. MI(U, V) measures the amount of shared information between cluster labels and classification labels.

[0107] The mutual information MI(U, V) is the cumulative product of the fifth parameter and the sixth parameter, where the fifth parameter is the ratio of the number of elements that belong to both Ui and Vj to the total number of data N. The total number of data N is multiplied by the number of elements that belong to both Ui and Vj, and then divided by the product of the number of elements of Ui and the number of elements of Vj. The logarithm of the above ratio is then taken to obtain the mutual information MI(U, V). U and V represent the sets of clustering results and supervised classification results, respectively. The higher the mutual information, the stronger the similarity between the clustering and supervised learning results.

[0108] By comparing these two metrics, we can evaluate the consistency of unsupervised clustering and supervised learning in acoustic emission damage classification. If the supervised learning results are highly consistent with the unsupervised fern results, it means that the initial cluster classification analysis has good physical significance. In addition, inconsistent classification results may reveal new material damage mechanisms or potential unrecognized damage characteristics, which can be further analyzed.

[0109] In the damage quantification process, clustering and supervised learning analysis results are used to further quantify the damage of composite wind turbine blades. Based on acoustic emission characteristic parameters and waveform features, the severity and location of the damage are linked to the characteristics of the acoustic emission event, achieving a quantitative assessment of material damage.

[0110] First, based on the results of clustering and classification, a damage level is assigned to each acoustic emission event. The damage level is set based on the cumulative energy of the acoustic emission event, the damage type, and the numerical range of the frequency. Events with higher cumulative energy and higher frequency are generally associated with more severe damage and are therefore given higher damage weights in the quantification process. If N acoustic emission events occur within a certain time period, and the damage score of each event is D1, D2, ..., DN, the following relationship can be used to quantify the score of each event: The damage score Di for the i-th acoustic emission event is equal to the sum of the product of the damage frequency Ai and the weight coefficient ω1, and the product of the cumulative energy Ei and the weight coefficient ω2. ω1 and ω2 are weight coefficients that are optimized based on experimental data. For example, depending on the severity of the damage type, ω1 can adopt the following values: matrix cracking (0.3), adhesive layer shear failure (0.4), fiber / matrix interface debonding (0.5), interlayer delamination (0.7), and fiber breakage (0.9). ω2 can be set to 0.05.

[0111] The output probability of the damage classification model is regarded as the confidence level of the damage category, and the overall damage quantification index is calculated based on the distribution of the confidence level. For example, the damage probability of each damage type can be calculated, and the damage quantification value of the blade as a whole can be obtained by weighted averaging. Assume that the classification model of the i-th acoustic emission event outputs the confidence level of the damage type Pi,j, where Pi,j represents the confidence level that the event belongs to the j-th damage type. For the damage frequency Ai, the confidence level of the damage type Pi,j and the corresponding weight ωj are weighted and calculated to obtain the weighted damage score of the event in terms of damage type and frequency. That is, the product of the damage frequency Ai and the weight ωi is equal to the cumulative sum of the confidence level of the damage type Pi,j, Aj and the corresponding weight ωj.

[0112] Finally, the quantified damage values ​​are linked to the actual structural parameters and experimentally measured mechanical properties of the material, establishing a correlation model between damage characteristics and material properties. This analysis enables a quantitative prediction of the extent of damage and provides a basis for assessing the health of wind turbine blades. The overall damage score, D, is the average sum of the damage scores, Di, for each acoustic emission event over the blade's service life, T.

[0113] Based on the overall damage score, blades can be classified into the following states: normal (D < 100), mild damage (100 ≤ D < 500), moderate damage (500 ≤ D < 1000), severe damage (1000 ≤ D < 2000), and critical (D ≥ 2000). In actual operation, when the overall damage score D of a blade exceeds 500 points (i.e., moderate damage), the blade will enter observation mode and its damage changes will be regularly monitored. If the score exceeds 1000 points (severe damage), it will be recommended to suspend operation and conduct a comprehensive inspection.

[0114] Compared with the prior art, the content of this embodiment has the following advantages.

[0115] 1. By combining unsupervised clustering with supervised learning, a multi-level damage identification and quantitative analysis process is constructed. Unsupervised cluster analysis using the HDBSCAN clustering algorithm can discover natural patterns in acoustic emission data and adapt to the distribution of various complex damage types. The introduction of supervised learning further improves model accuracy, enabling verification of clustering results and classification of damage types. Compared to traditional single-model approaches, this phased, multi-level approach better captures the complex characteristics of the data and improves the accuracy of damage identification.

[0116] 2. By using multiple acoustic emission feature maps (such as time-frequency spectrograms, power spectral density, and wavelet transform feature maps) as model input, the model's ability to identify diverse damage patterns is enhanced. Compared to methods that use only raw signals or single features, feature fusion can more comprehensively reflect the material's damage state, increase the model's adaptability, and improve its robustness to complex damage patterns.

[0117] 3. By using the HDBSCAN clustering algorithm, the optimal number of clusters can be automatically determined without manual setting. This advantage solves the defect of traditional clustering methods that require manual setting of the number of clusters, significantly improves the algorithm's automation and adaptability, and makes the model more flexible in application.

[0118] 4. The quantification method of acoustic emission signals combines qualitative judgment and quantitative evaluation to improve the accuracy of damage assessment. Compared with traditional quantification methods that rely on a single feature, the improvements in this embodiment make the damage quantification results more physically meaningful and valuable.

[0119] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A composite material wind turbine blade damage detection method, characterized in that: The following steps are involved: S1, real-time acquisition of acoustic emission signals during the operation of wind turbine blades and pre-processing of the signals, the acoustic emission signals including parameter data and waveform data; S2, performing unsupervised clustering on the preprocessed parameter data to obtain preliminary clustering results; S3, uses an improved multi-branch convolutional neural network to train the preprocessed waveform data and output accurate classification results; S4, using the supervised learning classification results to verify the unsupervised clustering results and evaluate the accuracy of the clustering results; S5, the quantitative indicators cumulative energy, damage type and frequency are integrated to construct a comprehensive damage assessment model to determine the damage extent.

2. A composite material wind turbine blade damage detection method according to claim 1, characterized in that: The step S1 comprises: S11, installing an acoustic emission sensor at a key position of the wind turbine blade, and the acoustic emission sensor detects the generated acoustic emission signal in real time; S12, performing bandpass filtering and wavelet denoising on the parameter data of the acoustic emission signal, and then performing standardization on the parameter data; S13, extracting a time-frequency spectrum diagram, a power spectrum density diagram, and a wavelet transform feature diagram of the waveform data.

3. A composite material wind turbine blade damage detection method according to claim 1 or 2, characterized in that: The step S2 comprises: The HDBSCAN clustering algorithm is used to perform unsupervised clustering on the preprocessed acoustic emission parameter data. After clustering, the distribution characteristics and central eigenvalues ​​of the acoustic emission parameters in each cluster are analyzed to preliminarily determine the damage type that each cluster may correspond to.

4. A composite material wind turbine blade damage detection method according to claim 3, characterized in that: The steps of the unsupervised clustering include: S21, calculate the distance matrix between data points based on the parameter data, and estimate the core distance and reachability distance of each data point according to its density; S22, construct a density-based tree hierarchy, divide the clusters on this basis, obtain multiple clusters with higher density, and mark the noise points; S23, performing multiple iterative cluster analysis on the acoustic emission parameter data, using the minimum sample number parameter for tuning, and determining the optimal clustering result.

5. A composite material wind turbine blade damage detection method according to claim 1 or 4, characterized in that: The network architecture of the improved multi-branch convolutional neural network includes three parallel feature extraction branches, a feature fusion module and a classification module.

6. A composite material wind turbine blade damage detection method according to claim 1, 2 or 4, characterized in that: The step S4 comprises: S41, using the adjusted Rand index or normalized mutual information metric to compare supervised learning classification results with unsupervised clustering results; S42, if the supervised learning classification result and the unsupervised clustering result are highly consistent, the unsupervised clustering result is considered accurate; if the supervised learning classification result and the unsupervised clustering result are not consistent, the cause of the clustering error is found by analyzing the specific difference samples, and the clustering model is optimized or the hyperparameters of the unsupervised clustering are readjusted.

7. A composite material wind turbine blade damage detection method according to claim 1 or 2, characterized in that: The step S5 comprises: Through the classification results of the acoustic emission waveform feature map using a supervised learning model, various damage types are identified and quantitative numerical weights are assigned to different damage types. Combined with the output of the multi-branch CNN model, the number of damage events occurring in each time period and their corresponding damage types are counted to obtain the frequency of damage events.

8. The damage detection method for composite material wind turbine blades according to claim 1, characterized in that: The step S3 includes data set construction, which includes: The waveform data of the acoustic emission signal was processed to obtain the time-scale coefficient matrix describing the multi-scale characteristics of the signal. The classification results obtained by unsupervised clustering were used as the initial labels. The acoustic emission events were paired with the corresponding damage types in chronological order. The complete dataset was divided into training set, validation set, and test set in a ratio of 7:2:

1.

9. The composite material wind turbine blade damage detection method according to claim 7, characterized in that: The step S5 further includes: The quantitative indicators are integrated to construct a comprehensive damage assessment model, which combines the results of cumulative energy, damage type and frequency to calculate an overall damage score to evaluate the current health status of the material.

10. A composite material wind turbine blade damage detection method according to claim 9, characterized in that: When the overall damage score of the blade is monitored to exceed the first threshold, the blade operation will enter the observation state and its damage changes will be monitored regularly; when the overall damage score of the blade is monitored to exceed the second threshold, the blade operation will be suspended for a comprehensive inspection.

Images