A Defect Detection Method for Wind Turbine Blades Based on Terahertz Time-Domain Spectroscopy

Abstract

This invention discloses a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy, belonging to the field of intelligent fault detection technology. The method involves controlling the probe of a terahertz time-domain spectroscopy system to scan the wind turbine blade along a planned detection path and collecting raw spectral data. The raw spectral data is preprocessed to obtain first waveform data. Defect features are extracted from the first waveform data, and candidate defect regions are determined based on these features. The time-domain characteristics of the first waveform data within the candidate defect regions are analyzed to determine the location and geometric dimensions of the defects. This method enables precise location, quantitative assessment, and type identification of defects in wind turbine blades.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault detection technology, specifically relating to a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy. Background Technology

[0002] As a core component of wind energy conversion, the internal structural health of wind turbine blades directly affects the operational safety and lifespan of the entire turbine. Therefore, developing efficient and accurate non-destructive testing (NDT) technologies is crucial. Terahertz time-domain spectroscopy, due to its strong penetrating power to non-metallic materials and sensitivity to changes in internal structure, is considered a potential means of achieving NDT of internal blade defects. This technology acquires time-domain waveform data containing rich internal information by emitting terahertz pulses to the blade and receiving its reflected echoes.

[0003] However, existing terahertz-based detection methods, when dealing with large composite material components such as wind turbine blades, often rely on manual experience to interpret waveforms. Due to the complex structure and vast scanning area of ​​the blades, the resulting time-domain spectral data is extremely large. This data not only contains defect signals but also a large amount of complex fluctuations caused by inherent material inhomogeneities, multi-layer interface reflections, and environmental noise. This makes it difficult for manual analysis to systematically and efficiently separate the weak characteristic signals that truly represent defects from massive, high-dimensional continuous waveforms, let alone achieve quantitative extraction of defect information.

[0004] This situation leads to two closely related core technical challenges. The primary challenge lies in the "submergence" of defect characteristics, meaning that the time-domain response characteristics of defects (such as weak anomalous reflection peaks, amplitude variations, or waveform broadening) are submerged in strong background signals and random noise, making them difficult to observe and capture directly. This difficulty in effectively separating and highlighting characteristic signals leads to the second challenge: the "parameterization dilemma" of defect information. It is impossible to clearly extract quantifiable characteristic parameters directly related to the physical properties of defects (such as location, size, and type) from the original waveform. For example, the exact time delay of reflection peaks, amplitude variations, or specific values ​​of pulse widths hinder the establishment of a stable and accurate mathematical model between defect types and signal characteristics.

[0005] Therefore, how to automatically and intelligently identify and extract key feature parameters that can clearly characterize different defects (such as bubbles, inclusions, and delamination) from the massive and complex continuous waveform data generated by terahertz time-domain spectral scanning of wind turbine blades, and form standardized feature expressions to achieve accurate location, quantitative assessment, and type discrimination of defects, has become a key issue in improving the level and reliability of intelligent non-destructive testing of blades. Summary of the Invention

[0006] The purpose of this invention is to overcome the problem of inaccurate location, quantitative assessment and type identification of defects in wind turbine blades, and to propose a wind turbine blade defect detection method based on terahertz time-domain spectroscopy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy, comprising the following steps: The probe of the terahertz time-domain spectroscopy system is controlled to scan the wind turbine blades along the planned detection path to collect raw spectral data; The original spectral data is preprocessed to obtain the first waveform data; Extract defect features from the first waveform data, and determine candidate defect regions based on the defect features; Analyze the temporal characteristics of the first waveform data within the defect candidate region to determine the location and geometric dimensions of the defect.

[0008] Furthermore, the planned detection path is obtained through the following methods: Identify the key detection areas of the blade, obtain the three-dimensional model of the blade, and generate an adaptive scanning path based on the structural characteristics of the key detection areas.

[0009] Furthermore, the raw spectral data undergoes preprocessing, including: A background signal model of the blade is established, and background interference in the original spectral data is adaptively removed based on the background signal model to obtain the first waveform data.

[0010] Further, the defect features of the first waveform data are extracted, including: Multi-scale window analysis is performed on the first waveform data to identify abnormal signal points, extract potential signal points submerged by features, and cluster the abnormal signal points and potential signal points to obtain defect features. Defect candidate regions are determined based on defect features, including: Multi-dimensional risk scoring is performed on the clustered signal point clusters, and candidate defect regions are determined based on the risk scores.

[0011] Furthermore, multi-scale window analysis is performed on the first waveform data to identify anomalous signal points, specifically: The first waveform data is scanned point by point to identify abnormal reflection peaks and points with significant amplitude changes, thus obtaining a preliminary set of abnormal points; In the initial set of anomalies, data on points where amplitude changes exceed a preset threshold are obtained, and the abnormal signal points are determined by combining the morphological characteristics of the abnormal reflection peaks. Extract latent signal points from the feature-submerged data, and cluster the abnormal signal points and latent signal points to obtain defect features, specifically: Extract potential signal points from the abnormal feature points to obtain a list of potential signal points; For the list of potential signal points, cluster analysis is used to group the potential signal points to obtain the signal point grouping results; Based on the signal point grouping results, preliminary candidate defect regions are determined.

[0012] Furthermore, the temporal characteristics of the first waveform data within the defect candidate region are analyzed, including: The spectral kurtosis and variational mode decomposition of the first waveform data are performed to obtain multiple frequency band time-domain waveforms, and the feature vectors of the frequency band time-domain waveforms are extracted.

[0013] Further, the location and geometry of the defect are determined, including: The feature vector is input into the defect detection model, and the local and global features are fused by a multi-path encoder. The location and geometric dimensions of the defect are then decoupled and output.

[0014] Secondly, the present invention provides a wind turbine blade defect detection system based on terahertz time-domain spectroscopy, comprising: The data acquisition module is used to control the probe of the terahertz time-domain spectroscopy system to scan the wind turbine blades along the planned detection path and acquire raw spectral data. The data preprocessing module is used to preprocess the raw spectral data to obtain the first waveform data; The defect region determination module is used to extract defect features from the first waveform data and determine candidate defect regions based on the defect features. The defect location determination module is used to analyze the temporal characteristics of the first waveform data within the defect candidate region to determine the location and geometric dimensions of the defect.

[0015] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects: This invention proposes a wind turbine blade defect detection method based on terahertz time-domain spectroscopy, effectively solving the problem of accurate localization, quantitative assessment, and type identification of wind turbine blade defects in existing technologies. This invention achieves complete coverage and efficient detection of key blade regions by combining a three-dimensional blade model with mechanical properties to design an adaptive scanning path. It constructs a systematic and fully automated data processing workflow from raw spectral data acquisition, preprocessing, feature extraction to defect localization and quantitative analysis, significantly reducing reliance on manual experience and improving the objectivity and efficiency of the detection. Adaptive filtering, wavelet denoising, background signal modeling and removal techniques effectively suppress various interferences such as environmental noise, mechanical vibration, and inherent material reflections, improving the system's stability and reliability under complex operating conditions. Furthermore, it utilizes multiple threshold judgments, multi-scale window analysis, and other techniques to effectively suppress environmental noise, mechanical vibration, and inherent material reflections, enhancing the system's stability and reliability under complex operating conditions. By employing methods such as submerged signal extraction and cluster analysis, weak signals characterizing defects can be effectively identified and separated from complex background noise and material inhomogeneities, improving the sensitivity and accuracy of defect detection. Combining time-domain feature analysis, spectral kurtosis calculation, and variational mode decomposition, quantitative feature parameters directly related to the physical properties of defects are extracted, providing reliable data support for precise defect location, size assessment, and type identification. Furthermore, by fusing local and global features through a multi-path encoder and combining a decoupled output mechanism, intelligent identification and classification of single and composite defects are achieved, enhancing the comprehensiveness and accuracy of fault diagnosis. This method can be directly integrated into existing terahertz detection equipment and adapted to different blade models and defect types through parameterized configuration, demonstrating good engineering applicability and technological promotion value. It provides effective technical support for intelligent operation and maintenance and life management of wind turbine blades. Attached Figure Description The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a flowchart of the wind turbine blade defect detection method based on terahertz time-domain spectroscopy according to the present invention.

[0018] Figure 2 This is a simplified structural diagram of the wind turbine blade defect detection system based on terahertz time-domain spectroscopy of the present invention.

[0019] Figure 3 This is an electronic device diagram of the wind turbine blade defect detection method based on terahertz time-domain spectroscopy according to the present invention.

[0020] Figure 4 This is a schematic diagram of the wind turbine blade defect detection method based on terahertz time-domain spectroscopy in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] Example 1 See Figure 1 A method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy includes the following steps: A detection path is constructed, and the transmitting and receiving probes of the terahertz time-domain spectroscopy system are controlled to scan along a preset detection area of ​​the wind turbine blade to obtain raw spectral data. The raw spectral data is preprocessed to obtain first waveform data. Defect features are extracted from the first waveform data, and candidate defect regions are determined based on the extracted defect features. Time-domain feature analysis is performed on each first waveform data within the candidate defect region to obtain the specific location and size information of the defect. This invention, through multiple threshold judgments, multi-scale window analysis, feature submerged signal extraction, and cluster analysis, can effectively identify and separate weak signals characterizing defects from complex background noise and material inhomogeneity, improving the sensitivity and accuracy of defect detection.

[0023] Specifically, as shown below: S1. Construct a detection path and control the transmitting and receiving probes of the terahertz time-domain spectroscopy system to scan along the preset detection area of ​​the wind turbine blade to obtain the raw spectral data.

[0024] Using finite element analysis models and historical failure data, stress concentration areas and areas with a high incidence of typical defects were identified. Specifically, the main beam cap area bears the primary bending load and is prone to delamination and fiber breakage; the leading edge area is susceptible to surface damage and moisture intrusion due to wind and sand erosion and rainwater impact; the trailing edge area has a relatively weak structure and is prone to debonding and crack propagation; the adhesive interface between the shear web and skin is prone to debonding defects; and the blade root bolt connection area is a high-stress concentration area.

[0025] Obtain a precise 3D CAD model of the wind turbine blade or establish a point cloud model through 3D laser scanning. Define the blade coordinate system as follows: X-axis along the span (from blade heel to tip), Y-axis along the chord (from leading edge to trailing edge), and Z-axis along the thickness direction. Employ parametric surface modeling technology to discretize the blade surface into a finite number of detection mesh elements. The size of each detection mesh can be determined based on the detection resolution and efficiency requirements.

[0026] In this embodiment, a terahertz probe is mounted on the end effector of a six-degree-of-freedom robotic arm, employing an equidistant offset scanning path for linear critical areas such as the main beam cap. Using the centerline of the main beam cap as the reference path, the path extends equidistantly to both sides, forming parallel clusters of scanning lines. The scanning line spacing d is determined based on the terahertz beam diameter D and the detection overlap rate η1: d = D × (1 - η1), where η1 is taken as 20%-30% to ensure full coverage. For curved areas such as the leading and trailing edges, an adaptive isoparametric scanning path is used. Based on the UV parameter lines of the curved surface, the path density is adaptively adjusted according to the curvature changes. Path points are denser in areas with smaller curvature radii (such as the leading edge arc region) and sparser in areas with gentler curvature (such as the chordal region in the middle of the blade). For the annular region of the bonding interface, a spiral progressive scanning path is used. A spiral path that gradually expands outward from the bonding line is generated to ensure complete coverage of the potential defect expansion area.

[0027] During the scanning process, the transmitting probe emits terahertz pulse waves toward the blade, and the receiving probe receives the terahertz pulse signals that penetrate the blade or are reflected back from the interface of defects inside the blade, thereby obtaining the raw spectral data.

[0028] In this embodiment, a photoconductive antenna is used as the terahertz emission source, with a spectral range set to 0.1-3 THz. Below 0.1 THz, the penetration depth is large but the spatial resolution is low; above 3 THz, attenuation is severe in composite materials. By adjusting the laser power and bias voltage, the pulse energy is controlled within the range of 0.1-10 μJ. A high-energy mode is used for thick cross-section regions, and a low-energy mode is used for thin regions to reduce interference from multiple reflections. The repetition frequency is set to 80 MHz, synchronized with the mode-locked laser. A high repetition frequency is beneficial for improving the signal-to-noise ratio, but the bandwidth limitation of the data acquisition system must be considered. The receiving probe uses photoconductive sampling or electro-optic sampling technology to convert the terahertz timing signal into an electrical signal.

[0029] S2. Preprocess the original spectral data to obtain the first waveform data.

[0030] The preprocessing of the raw spectral data sequence includes: baseline drift correction, outlier detection and handling, time axis calibration, noise suppression and signal enhancement, and removal of background signal interference. In this embodiment, asymmetric least squares method is used for baseline estimation and correction.

[0031] The optimization problem is described as follows: ; In the formula, Table baseline estimation; Indicates the weighting coefficient; Represents the raw spectral data; λ Indicates the smoothing parameter;N Indicates the number of sampling points.

[0032] in, ; In the formula, p This represents an asymmetric parameter.

[0033] Corrected original spectral data: .

[0034] The outlier detection and processing process includes: calculating the median and absolute median difference of each raw spectral data point within its nearest window. In this embodiment, 11 medians are used. When the absolute value of the difference between the raw spectral data and the median is greater than the median difference... k If the value is more than twice the normal value, it is considered an outlier and replaced by the neighborhood median or linear interpolation. Due to mechanical vibration and environmental disturbances, the time zero point of different measurement points may have slight shifts. A cross-correlation algorithm is used to align all time zero points: first, a high-quality reference waveform is selected. b ref For each raw spectral data b i Calculate the cross-correlation function between the waveform and the reference waveform, find the time offset that maximizes the cross-correlation function, and shift the waveform according to the time offset.

[0035] Cross-correlation functions include: ; In the formula, τ Indicates time offset; ω ( t ) represents the weight function.

[0036] in, ; In the formula, Indicates the peak position of the reference waveform; This indicates the control weight width, which is set to 1 / 2 of the pulse width in this embodiment.

[0037] By employing wavelet threshold transform, selecting appropriate wavelet functions and decomposition levels, the noisy signal is decomposed into wavelet components of different frequencies. Then, a threshold and threshold function are selected for threshold processing. Finally, wavelet reconstruction is performed, and the filtered sub-signals are merged to obtain a noise-removed signal.

[0038] In this embodiment, the adaptive threshold used includes: ; In the formula, Indicates the first j Noise standard deviation estimation of layer wavelet coefficients.

[0039] This embodiment employs an adaptive filtering method to remove background signal interference. In terahertz scanning of wind turbine blades, the background signal may primarily originate from the uniform attenuation characteristics of the blade substrate material itself and the fixed reflection of the standard protective coating. By statistically analyzing waveform data from multiple regions considered "intact," an average background waveform or the energy distribution range of the background signal can be calculated.

[0040] Specifically, scan data from the area at the blade root without historical damage records can be selected, and the mean curve of its time-domain waveform can be calculated. This curve can serve as a typical representative of the background signal. Determining the distribution range of the main interference sources means analyzing the spatial variation pattern of the background signal. For example, due to the gradual change in thickness of the blade from the root to the tip, the attenuation capability of its material for terahertz waves also changes accordingly, resulting in a gradient distribution of the background signal amplitude from high to low. By plotting a contour map of the background signal amplitude across the entire scan area, the distribution range of this strong background signal caused by the structure itself can be visually determined, such as identifying situations where the signal amplitude at the blade main spar cap is consistently more than 30% higher than in other areas. Adaptive filtering is used to dynamically eliminate background signals with the known distribution patterns described above. The advantage of adaptive filtering is that its filter parameters can be automatically adjusted according to the characteristics of the local background signal, rather than using globally fixed parameters.

[0041] A sliding window approach is used to traverse the original spectral data. Within each window, the current local waveform is compared with the predicted background signal model for that location based on the location information. The prediction model can be built based on the aforementioned gradient map. The filter automatically calculates an optimal weight to subtract the predicted background component from the original spectral data.

[0042] In the signal after initial filtering by the adaptive filter, a series of low-amplitude but scattered signal points may be observed at specific depth locations (such as the time delay corresponding to the bonding interface between the core material and the skin). These points do not conform to the concentrated distribution characteristics of typical defects (such as delamination and bubbles), but constitute residual interference. After detecting such interference, a statistically based smoothing constraint is applied to the signal amplitude near this depth for local adjustment. For example, if the signal amplitude of a certain location deviates from the average of its eight neighboring points by more than two standard deviations, and its spatial distribution is isolated, its amplitude is adjusted to the average level of the surrounding points, thereby obtaining an adjusted intermediate signal. This helps reduce misjudgment noise caused by random non-homogeneity. The signal separation results are filtered by preset time-domain amplitude thresholds and signal width thresholds.

[0043] Specifically, from the refined signal set, only components with amplitudes exceeding three standard deviations of the normal background fluctuation range and pulse widths between 0.5 and 5 picoseconds are retained. This width range corresponds to the possible size of actual physical defects (such as micro-delamination), while excessively narrow pulses may originate from noise, and excessively wide pulses may originate from structural boundary reflections. This step effectively eliminates interfering components that do not conform to typical defect characteristics, thereby identifying a refined signal set highly suspected of being a real anomaly. Finally, core components are extracted from this refined set, and signal missing components are identified. Here, core components refer to the set of signal points that are spatially continuous or regularly distributed and may characterize the real defect region. For example, a set of signal points with amplitudes exceeding the limit is found to be continuously distributed along the blade span for 10 cm, but there is a 2 cm interval in the middle where no exceeding points are detected. This may be due to signal missing caused by scanning resolution limitations or local noise masking. In this case, interpolation methods are used to supplement the signal, such as linear interpolation, to generate reasonable estimates for this 2 cm interval based on the amplitude trends of the effective signal points before and after, thereby obtaining a spatially continuous and complete signal structure. The first waveform data generated based on this complete structure can more clearly and completely reflect the characteristics of potential defects inside the blade after eliminating various interferences to the greatest extent, laying a reliable foundation for subsequent accurate diagnosis and evaluation.

[0044] S3. Extract defect features from the first waveform data and determine the defect candidate region based on the extracted defect features.

[0045] Time-domain analysis techniques are used to scan the first waveform data point by point, identifying anomalous reflection peaks and significant amplitude variations, resulting in a preliminary set of anomalies. From this preliminary set, data on points with amplitude variations exceeding a preset threshold are extracted. These points are then labeled with the morphological characteristics of the anomalous reflection peaks to determine the labeled anomalous feature points. When selecting points from the preliminary set with amplitude variations exceeding the preset threshold, the morphological characteristics of the reflection peaks are also used for labeling. These morphological characteristics include the sharpness or duration of the peak; for example, a sharp peak with a short duration is more likely to be associated with a defect. This labeling method helps focus on more representative signal anomalies. For the labeled anomalous feature points, potential signal points that may be associated with defect feature flooding are extracted by comparing the signal strength and variation trends of adjacent points, resulting in a potential signal point list. Based on this list, cluster analysis is used to group the points, determining the spatial clustering between them. If the distance between points is less than a preset threshold, they are grouped together, resulting in the signal point grouping results. Based on the signal point grouping results, preliminary candidate defect regions are determined.

[0046] Specifically, in this embodiment, the abnormal reflection peak needs to simultaneously meet the following conditions, and the determination method includes: Amplitude determination: ; In the formula, A p Indicates peak amplitude; α This represents the relative threshold coefficient, with a value ranging from 0.05 to 0.15. s i ( n ) represents the first waveform data.

[0047] Significance judgment: ; In the formula, P ( n 0) indicates peak significance; β The significance coefficient is denoted as , ranging from 0.03 to 0.08. For candidate peak points, the peak significance is... ,in , These are the locations of the two most recent valley points.

[0048] Width determination: ; In the formula, W min =0.5, W max =5.0, which is the width limit; FWHM This indicates half the width and height.

[0049] in, ; ; In the formula, n R , n L Indicates the left and right boundaries; A 1 / 2 Indicates half-high value; Indicates the sampling time interval.

[0050] Time delay judgment: ; In the formula, t p Indicates the candidate peak time; t min Indicates the peak time; The time threshold is set to 2.0 ps.

[0051] Methods for determining significant amplitude changes include: ; In the formula, , These represent the average energy values ​​of two sliding windows. In this embodiment, the sliding window includes... MOne sampling point; s 1. s 2 represents the energy variance of the two sliding windows respectively; n 1. n 2 represents the window size. In this embodiment, n 1= n 2= M .

[0052] Meanwhile, in this embodiment, the sliding window adopts a small-scale window (such as...). M =10), mesoscale window (e.g.) M =50) and large-scale windows (such as M =200), when anomalies are detected on both scales, it is judged as a point of significant amplitude change.

[0053] The methods for determining potential signal points in feature-submerged areas include: Intensity assessment: ; In the formula, Indicates signal strength; This represents the intensity threshold, which is set to 1.5 in this embodiment.

[0054] Correlation assessment: ; In the formula, This represents the average correlation coefficient; The value represents the correlation threshold, which is set to 0.7 in this embodiment.

[0055] in, ; ; In the formula, Represents the waveform of candidate points s c Waveforms of neighboring points s i The correlation coefficient; the superscript - indicates that the average is taken; Indicates the distance between candidate points as r All points within the range.

[0056] Trend judgment: In the formula, Indicates the degree of trend change; This represents the trend threshold, which is set to 2.0 in this embodiment. It represents the standard deviation of the first-order coefficients of the neighborhood points.

[0057] Feature consistency judgment: ; In the formula, Sim ( ) indicates a feature similarity measure; Fc , F defect These respectively represent morphological characteristics and typical defect characteristics; This represents the consistency threshold, which is set to 0.8 in this embodiment.

[0058] A potential flooding signal must satisfy at least two of the above criteria.

[0059] Cluster analysis methods include: Each potential signal point is represented by a feature vector: ; In the formula, ( x i , y i () represents spatial coordinates; I i Table normalized intensity; C i Indicates the confidence score; T i Indicates the defect type encoding; F i Indicates morphological characteristic indicators; superscript T This indicates transpose.

[0060] Calculate potential signal points i With potential signal points j Distance between: ; In the formula, , , These represent the spatial distance weight, feature distance weight, and type distance weight, respectively, which are set to 0.6, 0.3, and 0.1 in this embodiment. , , These represent spatial distance, feature distance, and type distance, respectively.

[0061] in, ; ; ; In the formula, This represents the weighting coefficient, which is set to 1.

[0062] Treat each potential signal point as a cluster and calculate the distance matrix between all clusters. Find the cluster pair with the smallest distance using the distance matrix. When the distance is less than a set merging threshold, merge the cluster pair into a new cluster; update the distance matrix until there are no more clusters to merge.

[0063] The methods for calculating the distance between clusters include: ; In the formula, C a , C b Representing potential signal points i and potential signal points j The cluster it belongs to.

[0064] Methods for determining preliminary candidate defect regions include: High risk is determined directly: In the formula, R k Cluster C k Risk score; R high This represents a high-risk threshold, which is set to 70 in this embodiment. ; ; ; ; ; ; ; In the formula, , , , These represent intensity risk, spatial risk, temporal risk, and confidence level risk, respectively. , , , Individually, they include intensity risk weight, spatial risk weight, temporal risk weight, and confidence level risk. Indicates average intensity; T high , T mid These represent the upper and lower limits of intensity, respectively. f compact ( A k ) represents the compactness function; g shape ( E k ) represents the shape function; TC k Indicates time consistency; Indicates the average confidence level within the cluster; I iIndicates potential signal points i The strength of the cluster it belongs to; The standard deviation represents the peak time; Indicates potential signal points i The precise time location of the peak value of the defect echo.

[0065] Strength exceeding limit determination: In the formula, Indicates average intensity; I crit Indicates critical strength; R min This represents the minimum risk threshold, which is set to 40 in this embodiment.

[0066] Large-scale intermediate risk assessment: In the formula, N large Indicates the maximum cluster threshold; R mid This indicates the lower limit of the risk score.

[0067] If any of the above conditions are met, it is identified as a candidate region for defect.

[0068] Time-domain feature analysis is performed on each first waveform data within the defect candidate region to obtain the specific location and size information of the defect.

[0069] The method for performing time-domain feature analysis on the first waveform data includes: firstly, performing Laplace wavelet transform on the first waveform data, extracting the significant spectral concentration region at each time step based on the wavelet energy distribution, and performing frequency band segmentation by combining the spectral kurtosis function; then performing variational mode decomposition and inverse Laplace wavelet transform to obtain the frequency band time-domain waveform; and finally, inputting the frequency band time-domain waveform into the defect detection model to obtain the specific location and size information of the defect.

[0070] The methods for calculating spectral kurtosis include: ; In the formula, X ( ω E() represents the Fourier transform result of the first waveform data; E() represents the expected operation.

[0071] The variable mode decomposition process includes: decomposing the first waveform data into... O Modal components u o ( t ), o =1,..., O The following variational optimization model was constructed: ; In the formula, Indicates the first o The center frequency of each modal component; Indicates time t Find the first derivative.

[0072] To solve the above variational optimization model, a broadband penalty factor is introduced. and Lagrange multipliers An augmented objective model is constructed using augmented Lagrange multipliers: ; The first term is the modal bandwidth term; the second term is the reconstruction error term; and the third term is the Lagrange constraint term. The solution is iteratively applied in the frequency domain using the alternating direction multiplier method until the convergence condition is met, ultimately outputting a set of quasi-orthogonal modal components. u o ( t ).

[0073] In this embodiment, in each iteration, according to the first... n The second-order energy moment Wn of each modal spectrum adaptively adjusts its broadband penalty factor: ; in, ; ; In the formula, c Indicates the adjustment coefficient; ε Represents the numerically stable term; Represents the complex spectrum of a mode in the frequency domain; Indicates the fusion weights; This represents the ridge frequency extracted by Laplacian wavelet convolution; This represents the frequency centroid calculated during the variational mode decomposition iteration.

[0074] Multidimensional statistics and temporal features of the frequency band time-domain waveform are extracted as feature vectors and input into the defect detection model. First, the signal passes through a multi-path encoder. In this embodiment, the multi-path encoder consists of stacked coding blocks with identical structures. Each coding block includes two branches: a local attention branch and a global attention branch. The local attention branch uses a sliding window mechanism to perform local attention calculations on the input, while the global attention branch uses aggregated tokens and fully connected attention to obtain global dependencies.

[0075] Finally, the output of the multipath encoder is: Z=Concat(LocalAttn(H(n)),GlobalAttn(H(n))); where H(n) represents the input of the nth layer encoding block; Concat represents the concatenation operation on the feature dimension; LocalAttn represents the local attention branch operation; GlobalAttn represents the global attention branch operation.

[0076] Then, a multi-label vector is generated using a decoupler to obtain the fault detection results: The location and size of the fault are determined based on the coordinates of the first waveform data corresponding to the frequency band time domain waveform.

[0077] ; In the formula, Indicate whether it is a compound fault; Indicates whether or not the first i Classified as a single fault, 1 represents a compound fault, and 0 represents a non-compound fault; C This represents the number of all possible single fault categories; G i Indicates the first i Weight matrix for fault types; h CIS Represents the global representation vector of the output; m i This indicates an adjustment to the bias term.

[0078] This invention, by constructing a systematic data processing workflow, from raw spectral data acquisition, preprocessing, feature extraction to defect localization and quantitative analysis, achieves fully automated processing of massive temporal spectral data, significantly reducing reliance on human experience and improving the objectivity and efficiency of detection. Through multiple threshold judgments, multi-scale window analysis, feature-submerged signal extraction, and cluster analysis, it can effectively identify and separate weak signals characterizing defects from complex background noise and material inhomogeneities, improving the sensitivity and accuracy of defect detection. By employing temporal feature analysis, spectral kurtosis calculation, and variational mode decomposition, it extracts quantitative feature parameters directly related to the physical properties of defects (such as location, size, and type), providing reliable data support for accurate defect localization, size assessment, and type identification. By combining the blade's three-dimensional model with its mechanical properties, it designs adaptive scanning paths (such as equidistant offset, isoparametric adaptive, and helical progressive), ensuring complete coverage and efficient detection of key areas of the blade (such as the main beam cap, leading edge, trailing edge, and bonding interface). By employing adaptive filtering, wavelet denoising, and background signal modeling and removal techniques, this invention effectively suppresses various types of interference, including environmental noise, mechanical vibration, and inherent material reflections, thereby improving the system's stability and reliability under complex operating conditions. Through a multi-path encoder that fuses local and global features and combines this with a decoupled output mechanism, intelligent identification and classification of single and compound defects are achieved, enhancing the comprehensiveness and accuracy of fault diagnosis. The method provided by this invention can be directly integrated into existing terahertz detection equipment and can be adapted to different blade models and defect types through parameterized configuration, demonstrating good engineering applicability and technological promotion value.

[0079] In summary, this invention has significant advantages in improving the automation, accuracy, quantification capabilities, and engineering applicability of wind turbine blade defect detection, providing effective technical support for intelligent operation and maintenance and life management of blades.

[0080] Example 2 See Figure 2 A wind turbine blade defect detection system based on terahertz time-domain spectroscopy includes: The data acquisition module is used to control the probe of the terahertz time-domain spectroscopy system to scan the wind turbine blades along the planned detection path and acquire raw spectral data. The data preprocessing module is used to preprocess the raw spectral data to obtain the first waveform data; The defect region determination module is used to extract defect features from the first waveform data and determine candidate defect regions based on the defect features. The defect location determination module is used to analyze the temporal characteristics of the first waveform data within the defect candidate region to determine the location and geometric dimensions of the defect.

[0081] Example 3 See Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy.

[0082] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, read-only optical discs, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for detecting defects in a wind turbine blade based on terahertz time-domain spectroscopy, characterized in that, Includes the following steps: The probe of the terahertz time-domain spectroscopy system is controlled to scan the wind turbine blades along the planned detection path to collect raw spectral data; The original spectral data is preprocessed to obtain the first waveform data; Extract the defect features from the first waveform data, and determine the defect candidate region based on the defect features; The temporal characteristics of the first waveform data within the defect candidate region are analyzed to determine the location and geometric dimensions of the defect.

2. The terahertz time-domain spectroscopy based wind turbine blade defect detection method according to claim 1, characterized in that, The planned detection path is obtained in the following way: The key detection areas of the blade are identified, a three-dimensional model of the blade is obtained, and an adaptive scanning path is generated based on the structural characteristics of the key detection areas.

3. The terahertz time-domain spectroscopy based wind turbine blade defect detection method according to claim 1, characterized in that, The preprocessing of the raw spectral data includes: A background signal model of the blade is established, and background interference in the original spectral data is adaptively removed based on the background signal model to obtain the first waveform data.

4. The terahertz time-domain spectroscopy based wind turbine blade defect detection method according to claim 1, characterized in that, The defect features extracted from the first waveform data include: Multi-scale window analysis is performed on the first waveform data to identify abnormal signal points, extract potential signal points that are submerged by features, and cluster the abnormal signal points and potential signal points to obtain defect features. The method of determining the defect candidate region based on defect features includes: A multi-dimensional risk score is performed on the clustered signal point clusters, and the defect candidate regions are determined based on the risk score.

5. The terahertz time-domain spectroscopy based wind turbine blade defect detection method according to claim 4, characterized in that, The step of performing multi-scale window analysis on the first waveform data to identify abnormal signal points specifically involves: The first waveform data is scanned point by point to identify abnormal reflection peaks and points with significant amplitude changes, thus obtaining a preliminary set of abnormal points. In the initial set of anomalies, data on points where the amplitude change exceeds a preset threshold are obtained, and the abnormal signal points are determined by combining the morphological characteristics of the abnormal reflection peaks. The process of extracting submerged potential signal points and clustering the abnormal signal points and potential signal points to obtain defect features is as follows: Extract potential signal points from the abnormal feature points to obtain a list of potential signal points; The potential signal point list is grouped using cluster analysis to obtain the signal point grouping results. Based on the grouping results of the signal points, preliminary candidate defect regions are determined.

6. The method for detecting defects in wind turbine blades based on terahertz time-domain spectroscopy according to claim 1, characterized in that, The temporal characteristics of the first waveform data within the candidate defect region are analyzed, including: The first waveform data is subjected to spectral kurtosis calculation and variational mode decomposition to obtain multiple frequency band time-domain waveforms, and the feature vectors of the frequency band time-domain waveforms are extracted.

7. The terahertz time-domain spectroscopy based wind turbine blade defect detection method according to claim 6, characterized in that, Determining the location and geometry of the defect includes: The feature vector is input into the defect detection model, and the local and global features are fused by a multi-path encoder. The location and geometric dimensions of the defect are then decoupled and output.

8. A wind turbine blade defect detection system based on terahertz time domain spectroscopy, characterized by, include: The data acquisition module is used to control the probe of the terahertz time-domain spectroscopy system to scan the wind turbine blades along the planned detection path and acquire raw spectral data. The data preprocessing module is used to preprocess the raw spectral data to obtain the first waveform data; The defect region determination module is used to extract defect features from the first waveform data and determine defect candidate regions based on the defect features. The defect location determination module is used to analyze the temporal characteristics of the first waveform data within the defect candidate region to determine the location and geometric dimensions of the defect.

9. An electronic device, comprising: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wind turbine blade defect detection method based on terahertz time-domain spectroscopy as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the wind turbine blade defect detection method based on terahertz time-domain spectroscopy as described in any one of claims 1-7.

Images