Fusion edge wavelet denoising and cloud full waveform inversion of wind turbine tower health monitoring system

Abstract

The application discloses a kind of fusion edge wavelet noise reduction and cloud full waveform inversion wind power tower drum health monitoring system, comprising: edge computing layer: multi-resolution adaptive downsampling is carried out, original signal is decomposed using wavelet transform 8 layers, generate high-frequency detail coefficient and low-frequency approximation coefficient, the high-frequency detail coefficient uses adaptive threshold noise reduction, output downsampling signal of the key band information reservation, parallel multi-band filter, each band signal is independently carried out inverse filtering processing, through D-S evidence theory fusion multi-band feature, simultaneously through nonlinear acoustic index real-time calculation, real-time monitoring nonlinear damage feature and in index threshold value when triggering deep diagnosis process, trigger signal is transmitted to cloud intelligent layer by 5G edge gateway low delay;Cloud intelligent layer, damage type is output by deep learning diagnosis, the system can fuse edge wavelet noise reduction and cloud full waveform inversion wind power tower drum health monitoring.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine tower health monitoring technology, and relates to a wind turbine tower health monitoring system that integrates edge wavelet denoising and cloud-based full waveform inversion. Background Technology

[0002] With the global wind power industry rapidly developing towards larger scale, higher towers, and offshore wind power, the tower, as the main structure bearing the combined effects of complex wind loads, earthquakes, corrosion, and other factors, faces the risk of accumulating hidden damage such as fatigue crack initiation and propagation, bolt connection failure, uneven foundation settlement, and deterioration of anti-corrosion coatings. By deploying intelligent sensor networks such as distributed fiber optic sensors, ultrasonic phased arrays, and tilt sensor arrays, real-time data is collected on tower tilt angle, sway, modal parameters, bolt axial force, flange contact pressure, and vibration amplitude / frequency coupling characteristics. Combined with finite element models and AI-driven damage identification algorithms, closed-loop management from "data perception—transmission—analysis—decision warning" is achieved. This not only reduces the missed detection rate of hidden problems by more than 60% and the discrepancy in detection conclusions by 40% through "early detection and early treatment," compressing the total lifecycle monitoring cost to 40% of traditional manual inspections, but also dynamically corrects traditional design specifications (such as GL). The lack of data support for ultra-high-altitude dynamic response in 2010 prompted the formulation of national standards for mixed-tower safety monitoring and the construction of provincial health monitoring cloud platforms. This enabled multi-project data linkage analysis, upgrading the wind power industry from an "auxiliary maintenance item" to a core safety lifeline, supporting the safety goal of "zero tower collapse and zero casualties" in wind farms, injecting certainty into the sustainable development of the industry, and promoting the integration of interdisciplinary technologies such as materials science, sensor technology, and data analysis. This also propelled the wind power industry towards intelligent, standardized, and high-quality development, meeting the global demand for clean energy transformation and possessing significant economic, safety, environmental, and technological demonstration value.

[0003] In existing technologies, 10x downsampling is used in signal processing. However, 10x downsampling may lose high-frequency damage features, resulting in reduced damage detection sensitivity. In addition, Butterworth bandpass filters are used to suppress low-frequency mechanical vibration and high-frequency electromagnetic interference, but they may filter out low-frequency modes related to damage, affecting the integrity of modal coupling characteristic analysis. Therefore, a wind turbine tower health monitoring system that integrates edge wavelet denoising and cloud-based full waveform inversion is proposed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a wind turbine tower health monitoring system that integrates edge wavelet denoising and cloud-based full waveform inversion. This system can integrate edge wavelet denoising and cloud-based full waveform inversion for wind turbine tower health monitoring.

[0005] To achieve the above objectives, this invention discloses a wind turbine tower health monitoring system that integrates edge wavelet denoising and cloud-based full waveform inversion, comprising: Physical sensing layer: Collects environmental noise time-domain signals by deploying piezoelectric sensors, and the collected data is transmitted to the edge computing layer via an anti-interference CAN bus; Edge computing layer: Multi-resolution adaptive downsampling is performed, and the original signal is decomposed into 8 layers using wavelet transform to generate high-frequency detail coefficients and low-frequency approximation coefficients. The high-frequency detail coefficients are subjected to adaptive threshold noise reduction to retain weak scattering signals related to damage. The low-frequency approximation coefficients are dynamically selected by energy ratio analysis to select downsampling factors and output downsampling signals that retain key frequency band information. Parallel multi-passband filtering is performed, and each frequency band signal is independently subjected to inverse filtering. Multi-frequency band features are fused through DS evidence theory. At the same time, nonlinear acoustic indicators are calculated in real time to monitor nonlinear damage characteristics and trigger a deep diagnosis process when the indicators exceed the threshold. The trigger signal is transmitted to the cloud intelligent layer with low latency through the 5G edge gateway. Cloud-based intelligent layer: Through intelligent signal decomposition, the signal is adaptively decomposed into K center frequency IMFs using VMD and the relevant kurtosis criterion is used to screen damage-related components. After reconstructing the frequency domain impulse response using SVD, the time domain response matrix is ​​obtained through inverse Fourier transform. Noise or environmental excitation is separated from the damage signal. At the same time, damage imaging is generated through full waveform inversion imaging, and the damage type is output through deep learning diagnosis.

[0006] Furthermore, it also includes: Digital twin system: By integrating finite element models and calibrating boundary conditions or material properties in real time, it updates model parameters by combining monitoring data, generates training data to optimize deep learning models by simulating different damage scenarios, and optimizes edge layer preprocessing and diagnostic algorithm parameters based on model-predicted damage evolution trends, thus completing real-time bidirectional mapping between physical entities and virtual models.

[0007] Furthermore, the specific steps for the physical sensing layer to acquire environmental noise time-domain signals are as follows: Deployment: Attach the sensor to the wall of the wind turbine tower; Preloading: Apply preloading stress to ensure that the stress of the adhesive layer meets the stress requirement of ≥10% of the range; Acquisition: Obtaining the time-domain signal x(t); Conditioning: The piezoelectric sensor outputs charge, the charge amplifier outputs voltage, and high-frequency noise is removed by low-pass filtering to eliminate DC drift; Transmission: The conditioned signal is transmitted to the edge node via the CAN bus at a baud rate of 1 Mbps. The frame format includes the sensor ID, timestamp, and 16-bit sample value.

[0008] Furthermore, in the edge computing layer, the specific steps of the multi-resolution adaptive downsampling are as follows: Wavelet basis selection and decomposition hierarchy: High-frequency detail coefficients are generated by performing an 8-level multi-resolution decomposition of the signal using the Daubechies 8th order wavelet. to and low-frequency approximation coefficient : The signal is decomposed into different frequency bands to separate the damage-related weak scattering signal from the noise; High-frequency coefficient adaptive threshold denoising: applying median absolute deviation-based noise standard deviation estimation to high-frequency coefficients. And combined with Donoho's soft threshold formula Adaptive threshold denoising is performed, where N is the coefficient length. High-frequency noise is suppressed while weak scattering signals related to damage are preserved by retaining coefficients with amplitudes greater than the threshold. Low-frequency coefficient energy proportion analysis: analysis of low-frequency approximation coefficients Calculate the proportion of its energy to the total energy of the original signal. Where L is Length, N is the original signal length, and the downsampling factor is dynamically selected based on the energy proportion. If the resolution is >90%, use a 16x downsampling; otherwise, retain the original resolution. Downsampled signal reconstruction and output: Reconstruction and output of downsampled signal The inverse wavelet transform is performed, and the signal is reconstructed with the noise-reduced high-frequency coefficients. The sampling rate of the output signal is adjusted according to the downsampling factor, and then filtered by an anti-aliasing FIR filter to ensure that the Nyquist sampling theorem is satisfied.

[0009] Furthermore, the specific steps of the parallel multi-passband filtering and feature fusion are as follows: Overlapping Filter Bank: By designing a Butterworth bandpass filter bank containing four overlapping frequency bands (5-50kHz, 45-150kHz, 100-500kHz, and 500-1000kHz), the different damage-sensitive frequency bands of the tower structure can be covered and processed in parallel. The Butterworth bandpass filter bank is represented as... ,in, The cutoff frequency is n, and n is the filter order. Parallel filtering: The downsampled signal is filtered independently for each frequency band using Fourier transform and inverse transform to separate and extract the signal components within each frequency band; Inverse filtering and impulse response reconstruction: The frequency domain impulse response of each frequency band is reconstructed through passive inverse filtering. , represented as ,in, For system transfer function, The phase compensation term is used, and the time-domain impulse response is obtained by inverse Fourier transform. Feature extraction and Dempster evidence fusion: Key feature parameters are extracted from the impulse response of each frequency band to achieve a quantitative assessment of the structural health status. Finally, the multi-frequency band features are fused with basic probability assignment and Dempster combination rules through Dempster evidence theory.

[0010] Furthermore, the edge computing layer calculates nonlinear acoustic indicators in real time, monitors nonlinear damage characteristics in real time, and triggers a deep diagnostic process when the indicators exceed a threshold, including: Real-time calculation of nonlinear acoustic parameters: Second or third harmonic energy ratio calculation: The energy ratio of the second or third harmonic is calculated by performing fundamental and harmonic band energy analysis on the reconstructed time-domain impulse response using Fourier transform. The energy ratio is expressed as... , ,in, The second harmonic energy ratio, The third harmonic energy ratio, The main frequency of the signal. () represents the frequency domain impulse response; Frequency modulation coefficient calculation: Short-time Fourier transform is used to extract the instantaneous frequency and calculate the deviation from the fundamental frequency to quantize the frequency modulation coefficient. Represented as The frequency drift is calculated as follows: ; Intermodulation distortion index calculation: Detection of nonlinear products under dual-frequency excitation to calculate the intermodulation distortion index, dual-frequency excitation signal. Intermodulation product frequencies were detected by FFT. Power at the location Ratio to base frequency power: ; Threshold determination: Based on preset conditions and combined with threshold determination logic: Triggering decisions are made based on the following criteria: second harmonic energy ratio >5%, frequency modulation coefficient >1kHz, and intermodulation distortion >3%. Trigger signal generation and transmission: A trigger signal containing sensor ID, timestamp, nonlinear index value and spatial location is generated, and the data packet is transmitted to the cloud using 5G low-latency transmission.

[0011] Furthermore, the specific steps for intelligent signal decomposition in the cloud-based intelligent layer are as follows: VMD Adaptive Mode Decomposition: Based on constrained variational problem construction and ADMM iterative optimization algorithm, the input time domain signal is decomposed into K intrinsic mode functions with center frequencies, and the adaptive mode separation of the signal is completed to obtain the basic characterization of each frequency component. Correlation kurtosis criterion for screening damage components: By calculating the correlation kurtosis between each intrinsic mode function and the original signal, intrinsic mode functions with correlation kurtosis values ​​exceeding a set threshold are selected as damage-related components, retaining weak scattering signals related to damage and suppressing non-damage noise components. SVD Frequency Domain Reconstruction and Noise Reduction: SVD frequency domain reconstruction is implemented. Fourier transform is performed on the selected damage-related intrinsic mode functions to construct a frequency domain matrix. After singular value decomposition, the first M singular values ​​are retained. The frequency domain impulse response is reconstructed and the time domain response matrix is ​​obtained by inverse Fourier transform.

[0012] Furthermore, the specific steps for the cloud-based intelligent layer to generate damage imaging maps through full-waveform inversion imaging are as follows: Initial model construction and prior wave velocity field setting: The initial wave velocity field is constructed by using a healthy state finite element model and the elastic parameters are set as prior references; Forward modeling and residual calculation: Synthetic data is generated by numerically solving the elastic wave equation using the finite-difference time-domain method, and the elastic wave equation is solved: ,in, For displacement field, For stiffness tensor, As the excitation source, simulated synthetic data is generated. Where m is the elastic parameter vector, calculated from observed data. Synthetic data generated by simulation The objective function is defined by the sum of squared residuals, expressed as: ; Gradient Calculation and Parameter Update: Calculate the gradient of the objective function with respect to the model parameters using backpropagation with the adjoint state method. , represented as ,in, The Fréchet derivative of the data with respect to the parameters is obtained through numerical perturbation or analytical methods. The elastic parameter vector is iteratively updated using the L-BFGS optimization algorithm to minimize the residuals, and is expressed as follows: ,in, It is an approximate Hessian matrix. Step size; Elastic parameter perturbation field inversion and damage imaging: The updated elastic parameters obtained by inversion are compared with the initial parameters to calculate the elastic parameter perturbation field at each point in space. The perturbation field is segmented based on a preset threshold to locate the damage area and quantify the degree of damage.

[0013] Furthermore, the specific steps of the cloud-based intelligent layer in diagnosing and outputting the damage type through deep learning are as follows: Multimodal input data preprocessing: Short-time Fourier transform is used to extract time-frequency features from the impulse response time series to capture the 300kHz component energy enhancement signal of microcrack scattering; CNN convolutional layer is used to extract local spatial features from the FWI elastic parameter perturbation map to enhance crack region features and suppress background noise; and max pooling layer is used to reduce the dimensionality of the nonlinear index spectrum to retain key feature regions of second harmonic energy ratio and frequency modulation coefficient. Feature extraction network: The CNN branch achieves the layer-by-layer extraction and enhancement of local features from 64×64×64 to 32×32×128 and then to 16×16×256 through multi-layer convolution and stride adjustment. The GNN branch constructs an adjacency matrix based on the spatial coordinates of the tower sensor and uses graph convolutional layers to aggregate neighborhood information to model the spatial topological relationship between sensors. Multimodal feature fusion and classification: The features extracted by CNN and GNN are concatenated and fused to form fused features through a feature fusion strategy. Then, the fully connected layer classifier combined with the Softmax function outputs the probability of damage type, location coordinates and severity, so as to achieve effective integration of multimodal features and quantitative assessment of damage.

[0014] Furthermore, it also includes: Loss function and training optimization: The model is trained in multiple rounds on a training set with multiple samples by using a hybrid loss function and Adam optimizer. The model training optimization and performance evaluation are completed by combining the division of the training set to 80% and the test set to 20% and the validation metrics of classification accuracy, location regression error and severity regression error.

[0015] The present invention has the following beneficial effects: The wind turbine tower health monitoring system of this invention, which integrates edge wavelet denoising and cloud-based full waveform inversion, improves detection sensitivity by using adaptive threshold denoising of wavelet high-frequency scaling coefficients to avoid the loss of high-frequency damage signals caused by fixed 10x downsampling. The low-frequency approximation coefficients dynamically select downsampling factors based on the dominant frequency to ensure that specific modes related to structural damage are not over-downsampled, thus solving the problem of modal information loss caused by fixed downsampling.

[0016] Furthermore, in this invention, VMD adaptively decomposes the signal into band-limited IMFs, more accurately separating noise, environmental excitation, and damage scattering signals, thus solving the subjectivity problem of SVD threshold setting. Selective reconstruction of key IMFs followed by inverse filtering reduces signal distortion or noise residue caused by inappropriate thresholds, improving the accuracy of impulse response matrix reconstruction. VMD's ability to separate nonlinear harmonic components is superior to SVD, helping to address the bias problem of regularization processing in nonlinear damage feature extraction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0023] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0024] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0027] refer to Figure 1 This invention is a wind turbine tower health monitoring system that integrates edge wavelet denoising and cloud-based full waveform inversion, comprising: Physical sensing layer: Collects environmental noise time-domain signals by deploying piezoelectric sensors, and the collected data is transmitted to the edge computing layer via an anti-interference CAN bus; Edge computing layer: Performs multi-resolution adaptive downsampling, using wavelet transform to decompose the original 2.5MHz signal into 8 layers, generating high-frequency detail coefficients and low-frequency approximation coefficients. The high-frequency coefficients are subjected to adaptive threshold noise reduction (e.g., Donoho thresholding) to retain weak scattering signals related to damage. The low-frequency coefficients are dynamically selected by energy proportion analysis (e.g., 16x downsampling is used when the energy proportion of the A8 layer is >90%, otherwise the original resolution is retained). The output is a downsampled signal that retains key frequency band information, and parallel multi-passband filtering is performed (designing 4 sets of overlapping filter banks (5-50kHz, 45-10 ... (50kHz, 100-500kHz, 500-1000kHz), parallel processing of downsampled signals), each frequency band signal is independently inverse filtered, multi-band features are fused through DS evidence theory, and nonlinear acoustic indicators are calculated in real time (based on the reconstructed time-domain impulse response to calculate the second / third harmonic energy ratio, frequency modulation coefficient and intermodulation distortion index), nonlinear damage characteristics (such as frequency drift caused by "breathing" cracks) are monitored in real time, and a deep diagnostic process is triggered when the index exceeds the threshold (such as second harmonic energy > 5% of the fundamental frequency), and the trigger signal is transmitted to the cloud intelligent layer with low latency through the 5G edge gateway; Cloud-based intelligent layer: Through intelligent signal decomposition, the signal is adaptively decomposed into K center frequency IMFs using VMD and the relevant kurtosis criterion is used to screen damage-related components. After reconstructing the frequency domain impulse response using SVD, the time domain response matrix is ​​obtained through inverse Fourier transform. This separates noise or environmental excitation from the damage signal and improves the accuracy of modal coupling analysis. At the same time, damage imaging is generated through full waveform inversion (FWI) imaging (establishing a priori wave velocity field based on a healthy state finite element model and inverting the elastic parameter perturbation field using the L-BFGS iterative optimization algorithm). Damage imaging is generated through deep learning diagnosis (constructing a multi-channel CNN-GNN fusion network, inputting impulse response time series / FWI elastic parameter perturbation map / nonlinear index spectrum, and outputting damage type / location / severity). Digital twin system: By integrating a high-fidelity finite element model and calibrating boundary conditions or material properties in real time, it updates model parameters by combining monitoring data, generates training data to optimize the deep learning model by simulating different damage scenarios, and optimizes edge layer preprocessing (such as downsampling factor / filter frequency band) and diagnostic algorithm parameters based on the model's prediction of damage evolution trend, thus completing the real-time bidirectional mapping between the physical entity and the virtual model.

[0028] The original signal is transmitted to the edge node via an anti-interference bus, where it undergoes wavelet downsampling, parallel filtering, and nonlinear index calculation. The process is: original signal → preprocessed signal → nonlinear index.

[0029] The trigger signal is transmitted to the cloud through the 5G gateway. The edge layer outputs preprocessed multi-band signals and nonlinear indicators. After receiving the signals, the cloud performs VMD decomposition, FWI imaging, and deep learning diagnosis. The result is: preprocessed signal → feature signal → diagnostic result.

[0030] The digital twin system updates diagnostic results and model parameters. The twin model simulates the damage scenario to generate training data and provides feedback to optimize edge layer preprocessing parameters (such as downsampling factor and filter bandwidth). Diagnostic results → model update → parameter feedback.

[0031] After dynamic calibration of the digital twin model, the edge layer preprocessing algorithm (such as adaptive threshold and filter frequency band) is optimized to improve the accuracy of subsequent monitoring; the cloud diagnostic results are fed back to the physical layer to guide maintenance decisions (such as bolt tightening and coating repair).

[0032] In this embodiment, the specific steps for the physical sensing layer to collect environmental noise time-domain signals are as follows: Deployment: Sensors are bonded to the wind turbine tower wall (the tower circumference C=πD (D is the tower diameter), with 30 sensors arrayed along the circumference, theoretically spaced ΔL=C / 30, but in practice a dense array design is adopted: 10 layers along the tower axis, with 3 sensors arranged on the circumference of each layer (totaling 30 sensors), axial spacing Δz=10mm, and circumferential spacing Δθ=2π / 3 radians (approximately 120° interval)). Preloading: Apply preloading stress to ensure that the adhesive layer stress meets the ≥10% range stress requirement (epoxy resin adhesive (e.g., 3M DP460) shear strength τ≥25 MPa, adhesive layer thickness t≤0.2mm, preloading stress calculation: sensor range Fmax=5000 N (typical value), preloading force Fpreload≥0.1×Fmax=500 N, stress formula: σ=Fpreload / Asensor (Asensor is the sensor contact area, e.g., 20×20 mm²), ensure σ≥1.25 MPa (meeting the ≥10% range stress requirement)). Acquisition: Obtaining the time-domain signal x(t) (sampling frequency) =2.5MHz, sampling interval Δt=1 / =0.4 μs, the time-domain signal is represented as x(t)=Asin(2πft+ (e.g., ambient noise fundamental frequency f=100kHz), after discretization, x[n]=Asin(2πfnΔt+ )); Conditioning: The piezoelectric sensor outputs charge, the charge amplifier outputs voltage, and high-frequency noise is removed by low-pass filtering to eliminate DC drift; Transmission: The conditioned signal is transmitted to the edge node via the CAN bus at a baud rate of 1 Mbps. The frame format includes the sensor ID, timestamp, and 16-bit sample value.

[0033] In this embodiment, the specific steps of the multi-resolution adaptive downsampling in the edge computing layer are as follows: Wavelet basis selection and decomposition hierarchy: The signal is decomposed into high-frequency detail coefficients by using the Daubechies 8th order wavelet (db8) at 8 levels. to and low-frequency approximation coefficient : The signal is decomposed into different frequency bands to separate the damage-related weak scattering signal from the noise; High-frequency coefficient adaptive threshold denoising: applying median absolute deviation (MAD)-based noise standard deviation estimation to high-frequency coefficients. And combined with Donoho's soft threshold formula Adaptive threshold denoising is performed, where N is the coefficient length. High-frequency noise is suppressed while weak scattering signals related to damage are preserved by retaining coefficients with amplitudes greater than the threshold. Low-frequency coefficient energy proportion analysis: analysis of low-frequency approximation coefficients Calculate the proportion of its energy to the total energy of the original signal. Where L=3,125 is The length, N=250,000, is the original signal length, and the downsampling factor is dynamically selected based on the energy proportion. When the bandwidth is >90%, a 16x downsampling is used (retaining a 156.25kHz bandwidth) to reduce the amount of data; otherwise, the original 2.5MHz resolution is retained to ensure that key frequency band information is not lost. Downsampled signal reconstruction and output: Reconstruction and output of downsampled signal The inverse wavelet transform is performed, and the signal is reconstructed with the noise-reduced high-frequency coefficients. The sampling rate of the output signal is adjusted according to the downsampling factor, and then filtered by an anti-aliasing FIR filter to ensure that the Nyquist sampling theorem is satisfied.

[0034] In this embodiment, the specific steps of parallel multi-passband filtering and feature fusion are as follows: Overlapping Filter Bank: By designing a Butterworth bandpass filter bank containing four overlapping frequency bands (5-50kHz, 45-150kHz, 100-500kHz, and 500-1000kHz), it covers and processes different damage-sensitive frequency bands of the tower structure (such as low-frequency overall vibration modes, mid-frequency weld defects, high-frequency surface crack scattering, and ultra-high-frequency corrosion guided waves). The Butterworth bandpass filter bank is represented as follows: ,in, Here, n is the cutoff frequency, and n is the filter order, ensuring that the stopband attenuation is >40dB; Parallel filtering: The downsampled signal is filtered independently for each frequency band using Fourier transform and inverse transform to separate and extract the signal components within each frequency band; Inverse filtering and impulse response reconstruction: The frequency domain impulse response of each frequency band is reconstructed through passive inverse filtering. , represented as ,in, For system transfer function, The phase compensation term is used, and the time-domain impulse response is obtained by inverse Fourier transform. Feature extraction and Dempster evidence fusion: Key feature parameters such as phase velocity, energy decay and second harmonic energy ratio are extracted from the impulse response of each frequency band to achieve quantitative assessment of structural health status. Finally, the multi-frequency band features are fused with basic probability assignment and Dempster combination rules through Dempster evidence theory.

[0035] In this embodiment, the edge computing layer calculates nonlinear acoustic indicators in real time, monitors nonlinear damage characteristics in real time, and triggers a deep diagnostic process when the indicators exceed a threshold, including: Real-time calculation of nonlinear acoustic parameters: Second or third harmonic energy ratio calculation: The energy ratio of the second or third harmonic frequency is calculated by performing fundamental and harmonic band energy analysis on the reconstructed time-domain impulse response using Fourier transform (FFT) (to achieve early warning of nonlinear damage such as "breathing" cracks). The energy ratio is expressed as... , ,in, The second harmonic energy ratio, The third harmonic energy ratio, For the main frequency of the signal, This is the frequency domain impulse response; Frequency modulation coefficient calculation: Short-time Fourier Transform (STFT) extracts the instantaneous frequency and calculates the deviation from the fundamental frequency to quantize the frequency modulation coefficient (achieving the identification function of nonlinear characteristics such as frequency drift). Instantaneous frequency extraction... Represented as The frequency drift is calculated as follows: ; Intermodulation distortion index calculation: Detection of nonlinear products under dual-frequency excitation to calculate the intermodulation distortion index (achieving the evaluation function of material nonlinear enhancement), dual-frequency excitation signal. Intermodulation product frequencies were detected by FFT. Power at the location Ratio to fundamental frequency power: ; Threshold determination: Based on preset conditions and combined with threshold determination logic: Triggering decisions are made based on the following criteria: second harmonic energy ratio >5%, frequency modulation coefficient >1kHz, and intermodulation distortion >3%. Trigger signal generation and transmission: A trigger signal containing sensor ID, timestamp, nonlinear index value, and spatial location is generated, and the data packet is transmitted to the cloud using 5G low-latency transmission (using low-latency frame structure and priority queue scheduling to ensure latency ≤50ms). Finally, the full waveform inversion (FWI) imaging module (after loading the healthy state finite element model, using the passive inverse filter impulse response as the observation data, and using the L-BFGS algorithm to iteratively optimize the elastic parameter perturbation field to generate a high-resolution damage image with a resolution <5cm) and a deep learning diagnostic module (inputting multi-channel data such as impulse response sequence, FWI image, and nonlinear index spectrum, and outputting damage type, location, and severity through CNN-GNN fusion network) are launched in the cloud.

[0036] In this embodiment, the specific steps for the cloud-based intelligent layer to perform intelligent signal decomposition are as follows: VMD Adaptive Mode Decomposition: Based on constrained variational problem construction and ADMM iterative optimization algorithm, the input time domain signal is decomposed into K intrinsic mode functions (IMFs) with center frequencies, and the adaptive mode separation of the signal is completed to obtain the basic characterization of each frequency component; Correlation kurtosis criterion for screening damage components: By using the correlation kurtosis criterion, the correlation kurtosis between each intrinsic mode function (IMF) and the original signal is calculated (by fusing correlation and kurtosis indices), and intrinsic mode functions (IMFs) with correlation kurtosis values ​​exceeding a set threshold (e.g., 0.6) are selected as damage-related components, retaining weak scattering signals related to damage and suppressing non-damage noise components; SVD Frequency Domain Reconstruction and Noise Reduction: SVD frequency domain reconstruction is performed. The selected damage-related intrinsic mode functions (IMFs) are subjected to Fourier transform to construct a frequency domain matrix. After singular value decomposition, the first M singular values ​​are retained (based on an energy percentage threshold such as 95%). The frequency domain impulse response is reconstructed and the time domain response matrix is ​​obtained by inverse Fourier transform.

[0037] In this embodiment, the specific steps of the cloud-based intelligent layer in generating damage imaging maps through full waveform inversion (FWI) imaging are as follows: Initial model construction and prior wave velocity field setting: The initial wave velocity field is constructed through a healthy state finite element model and elastic parameters such as P-wave velocity, S-wave velocity and density are set as prior benchmarks; Forward modeling and residual calculation: Synthetic data is generated by numerically solving the elastic wave equation using the finite-difference time-domain method, and the elastic wave equation is solved: ,in, For displacement field, For stiffness tensor, Using environmental noise as an excitation source, simulated synthetic data is generated. Where m is the elastic parameter vector, calculated from the observed data. Synthetic data generated by simulation The objective function is defined by the sum of squared residuals, expressed as: ; Gradient Calculation and Parameter Update: Calculate the gradient of the objective function with respect to the model parameters using backpropagation with the adjoint state method. , represented as ,in, The Fréchet derivative of the data with respect to the parameters is obtained through numerical perturbation or analytical methods. The elastic parameter vector is iteratively updated using the L-BFGS optimization algorithm to minimize the residuals, expressed as... ,in, It is an approximate Hessian matrix (constructed through historical gradients). Step size; Elastic parameter perturbation field inversion and damage imaging: The updated elastic parameters obtained by inversion are compared with the initial parameters to calculate the elastic parameter perturbation field at each point in space. The perturbation field is segmented based on a preset threshold to locate the damage area and quantify the degree of damage.

[0038] In this embodiment, the specific steps of the cloud-based intelligent layer in diagnosing and outputting the damage type through deep learning are as follows: Multimodal input data preprocessing: Short-time Fourier transform is used to extract time-frequency features from the impulse response time series to capture the 300kHz component energy enhancement signal of microcrack scattering; CNN convolutional layer is used to extract local spatial features from the FWI elastic parameter perturbation map to enhance crack region features and suppress background noise; and max pooling layer is used to reduce the dimensionality of the nonlinear index spectrum to retain key feature regions of second harmonic energy ratio and frequency modulation coefficient. Feature extraction network: The CNN branch uses multi-layer convolution with 64, 128, and 256 3×3 kernels and stride adjustment to extract and enhance local features layer by layer from 64×64×64 to 32×32×128 and then to 16×16×256. The GNN branch constructs an adjacency matrix based on the spatial coordinates of the tower sensor and uses graph convolutional layers to aggregate neighborhood information to model the spatial topological relationship between sensors. Multimodal feature fusion and classification: A feature fusion strategy is used to combine the 16×16×256 dimensional features extracted by CNN with those extracted by GNN. The 128-dimensional features are spliced ​​and fused to form a fused feature. Then, a fully connected layer classifier combined with the Softmax function outputs the damage type probability (crack / corrosion / healthy three-class classification), location coordinates (three-dimensional coordinate regression), and severity (0-10 level regression), realizing the effective integration of multimodal features and quantitative damage assessment. Loss Function and Training Optimization: The model was trained multiple times on a training set of 1000 samples (333 healthy samples, 333 cracked samples, 333 corroded samples, and 1 unknown sample) using a hybrid loss function (a weighted combination of classification cross-entropy loss, location regression mean squared error loss, and severity regression mean squared error loss) and the Adam optimizer (learning rate 0.001, momentum 0.9). The model training was optimized and its performance was evaluated by combining the division of the training set (80%) with the test set (20%) and the validation metrics of classification accuracy, location regression error, and severity regression error.

[0039] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and disclosure of the invention. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0040] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0041] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion, characterized in that, include: Physical sensing layer: Collects environmental noise time-domain signals by deploying piezoelectric sensors, and the collected data is transmitted to the edge computing layer via an anti-interference CAN bus; Edge computing layer: Multi-resolution adaptive downsampling is performed, and the original signal is decomposed into 8 layers using wavelet transform to generate high-frequency detail coefficients and low-frequency approximation coefficients. The high-frequency detail coefficients are subjected to adaptive threshold noise reduction to retain weak scattering signals related to damage. The low-frequency approximation coefficients are dynamically selected by energy ratio analysis to select downsampling factors and output downsampling signals that retain key frequency band information. Parallel multi-passband filtering is performed, and each frequency band signal is independently subjected to inverse filtering. Multi-frequency band features are fused through DS evidence theory. At the same time, nonlinear acoustic indicators are calculated in real time to monitor nonlinear damage characteristics and trigger a deep diagnosis process when the indicators exceed the threshold. The trigger signal is transmitted to the cloud intelligent layer with low latency through the 5G edge gateway. Cloud-based intelligent layer: Through intelligent signal decomposition, the signal is adaptively decomposed into K center frequency IMFs using VMD and the relevant kurtosis criterion is used to screen damage-related components. After reconstructing the frequency domain impulse response using SVD, the time domain response matrix is ​​obtained through inverse Fourier transform. Noise or environmental excitation is separated from the damage signal. At the same time, damage imaging is generated through full waveform inversion imaging, and the damage type is output through deep learning diagnosis.

2. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 1, characterized in that, Also includes: Digital twin system: By integrating finite element models and calibrating boundary conditions or material properties in real time, it updates model parameters by combining monitoring data, generates training data to optimize deep learning models by simulating different damage scenarios, and optimizes edge layer preprocessing and diagnostic algorithm parameters based on model-predicted damage evolution trends, thus completing real-time bidirectional mapping between physical entities and virtual models.

3. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 2, characterized in that, The specific steps for the physical sensing layer to acquire environmental noise time-domain signals are as follows: Deployment: Attach the sensor to the wall of the wind turbine tower; Preloading: Apply preloading stress to ensure that the stress of the adhesive layer meets the stress requirement of ≥10% of the range; Acquisition: Obtaining the time-domain signal x(t); Conditioning: The piezoelectric sensor outputs charge, the charge amplifier outputs voltage, and high-frequency noise is removed by low-pass filtering to eliminate DC drift; Transmission: The conditioned signal is transmitted to the edge node via the CAN bus at a baud rate of 1 Mbps. The frame format includes the sensor ID, timestamp, and 16-bit sample value.

4. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 1, characterized in that, In the edge computing layer, the specific steps of the multi-resolution adaptive downsampling are as follows: Wavelet basis selection and decomposition hierarchy: The signal is decomposed into high-frequency detail coefficients by using the Daubechies 8th order wavelet at 8 levels of multi-resolution decomposition. to and low-frequency approximation coefficients : The signal is decomposed into different frequency bands to separate the damage-related weak scattered signal from the noise; High-frequency coefficient adaptive threshold denoising: applying median absolute deviation-based noise standard deviation estimation to high-frequency coefficients. And combined with Donoho's soft threshold formula Adaptive threshold denoising is performed, where N is the coefficient length. High-frequency noise is suppressed while weak scattering signals related to damage are preserved by retaining coefficients with amplitudes greater than the threshold. Low-frequency coefficient energy proportion analysis: analysis of low-frequency approximation coefficients Calculate the proportion of its energy to the total energy of the original signal. Where L is Length, N is the original signal length, and the downsampling factor is dynamically selected based on the energy proportion. If the resolution is >90%, use a 16x downsampling; otherwise, retain the original resolution. Downsampled signal reconstruction and output: Reconstruction and output of downsampled signal The inverse wavelet transform is performed, and the signal is reconstructed with the noise-reduced high-frequency coefficients. The sampling rate of the output signal is adjusted according to the downsampling factor, and then filtered by an anti-aliasing FIR filter to ensure that the Nyquist sampling theorem is satisfied.

5. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 1, characterized in that, The specific steps of the parallel multi-passband filtering and feature fusion are as follows: Overlapping Filter Bank: By designing a Butterworth bandpass filter bank containing four overlapping frequency bands (5-50kHz, 45-150kHz, 100-500kHz, and 500-1000kHz), the different damage-sensitive frequency bands of the tower structure can be covered and processed in parallel. The Butterworth bandpass filter bank is represented as... ,in, The cutoff frequency is n, and n is the filter order. Parallel filtering: The downsampled signal is filtered independently for each frequency band using Fourier transform and inverse transform to separate and extract the signal components within each frequency band; Inverse filtering and impulse response reconstruction: The frequency domain impulse response of each frequency band is reconstructed through passive inverse filtering. , represented as ,in, For system transfer function, The phase compensation term is used, and the time-domain impulse response is obtained by inverse Fourier transform. Feature extraction and Dempster evidence fusion: Key feature parameters are extracted from the impulse response of each frequency band to achieve a quantitative assessment of the structural health status. Finally, the multi-frequency band features are fused with basic probability assignment and Dempster combination rules through Dempster evidence theory.

6. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 1, characterized in that, The edge computing layer calculates nonlinear acoustic indicators in real time, monitors nonlinear damage characteristics in real time, and triggers a deep diagnostic process when the indicators exceed a threshold, including: Real-time calculation of nonlinear acoustic parameters: Second or third harmonic energy ratio calculation: The energy ratio of the second or third harmonic is calculated by performing fundamental and harmonic band energy analysis on the reconstructed time-domain impulse response using Fourier transform. The energy ratio is expressed as... , ,in, The second harmonic energy ratio, The third harmonic energy ratio, The main frequency of the signal. () represents the frequency domain impulse response; Frequency modulation coefficient calculation: Short-time Fourier transform is used to extract the instantaneous frequency and calculate the deviation from the fundamental frequency to quantize the frequency modulation coefficient. Represented as The frequency drift is calculated as follows: ; Intermodulation distortion index calculation: Detection of nonlinear products under dual-frequency excitation to calculate the intermodulation distortion index, dual-frequency excitation signal. Intermodulation product frequencies were detected by FFT. Power at Ratio to base frequency power: ; Threshold determination: Based on preset conditions and combined with threshold determination logic: Triggering decisions are made based on the following criteria: second harmonic energy ratio >5%, frequency modulation coefficient >1kHz, and intermodulation distortion >3%. Trigger signal generation and transmission: A trigger signal containing sensor ID, timestamp, nonlinear index value and spatial location is generated, and the data packet is transmitted to the cloud using 5G low-latency transmission.

7. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 1, characterized in that, The specific steps for intelligent signal decomposition in the cloud-based intelligent layer are as follows: VMD Adaptive Mode Decomposition: Based on constrained variational problem construction and ADMM iterative optimization algorithm, the input time domain signal is decomposed into K intrinsic mode functions with center frequencies, and the adaptive mode separation of the signal is completed to obtain the basic characterization of each frequency component. Correlation kurtosis criterion for screening damage components: By calculating the correlation kurtosis between each intrinsic mode function and the original signal, intrinsic mode functions with correlation kurtosis values ​​exceeding a set threshold are selected as damage-related components, retaining weak scattering signals related to damage and suppressing non-damage noise components. SVD Frequency Domain Reconstruction and Noise Reduction: SVD frequency domain reconstruction is implemented. Fourier transform is performed on the selected damage-related intrinsic mode functions to construct a frequency domain matrix. After singular value decomposition, the first M singular values ​​are retained. The frequency domain impulse response is reconstructed and the time domain response matrix is ​​obtained by inverse Fourier transform.

8. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 7, characterized in that, The specific steps by which the cloud-based intelligent layer generates damage imaging maps through full-waveform inversion imaging are as follows: Initial model construction and prior wave velocity field setting: The initial wave velocity field is constructed by using a healthy state finite element model and the elastic parameters are set as prior references; Forward modeling and residual calculation: Synthetic data is generated by numerically solving the elastic wave equation using the finite-difference time-domain method, and the elastic wave equation is solved: ,in, For displacement field, For stiffness tensor, As the excitation source, simulated synthetic data is generated. Where m is the elastic parameter vector, calculated from the observed data. Synthetic data generated by simulation The objective function is defined by the sum of squared residuals, expressed as: ; Gradient Calculation and Parameter Update: Calculate the gradient of the objective function with respect to the model parameters using backpropagation with the adjoint state method. , represented as ,in, The Fréchet derivative of the data with respect to the parameters is obtained through numerical perturbation or analytical methods. The elastic parameter vector is iteratively updated using the L-BFGS optimization algorithm to minimize the residuals, expressed as... ,in, It is an approximate Hessian matrix. Step size; Elastic parameter perturbation field inversion and damage imaging: The updated elastic parameters obtained by inversion are compared with the initial parameters to calculate the elastic parameter perturbation field at each point in space. The perturbation field is segmented based on a preset threshold to locate the damage area and quantify the degree of damage.

9. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 7, characterized in that, The specific steps by which the cloud-based intelligent layer diagnoses and outputs the damage type through deep learning are as follows: Multimodal input data preprocessing: Short-time Fourier transform is used to extract time-frequency features from the impulse response time series to capture the 300kHz component energy enhancement signal of microcrack scattering; CNN convolutional layer is used to extract local spatial features from the FWI elastic parameter perturbation map to enhance crack region features and suppress background noise; and max pooling layer is used to reduce the dimensionality of the nonlinear index spectrum to retain key feature regions of second harmonic energy ratio and frequency modulation coefficient. Feature extraction network: The CNN branch achieves the layer-by-layer extraction and enhancement of local features from 64×64×64 to 32×32×128 and then to 16×16×256 through multi-layer convolution and stride adjustment. The GNN branch constructs an adjacency matrix based on the spatial coordinates of the tower sensor and uses graph convolutional layers to aggregate neighborhood information to model the spatial topological relationship between sensors. Multimodal feature fusion and classification: The features extracted by CNN and GNN are concatenated and fused to form fused features through a feature fusion strategy. Then, the fully connected layer classifier combined with the Softmax function outputs the probability of damage type, location coordinates and severity, so as to achieve effective integration of multimodal features and quantitative assessment of damage.

10. The wind turbine tower health monitoring system integrating edge wavelet denoising and cloud-based full waveform inversion as described in claim 9, characterized in that, Also includes: Loss function and training optimization: The model is trained in multiple rounds on a training set with multiple samples by using a hybrid loss function and Adam optimizer. The model training optimization and performance evaluation are completed by combining the division of the training set to 80% and the test set to 20% and the validation metrics of classification accuracy, location regression error and severity regression error.

Images