Measurement method and system applied to wind power blade and based on stress wave perception

Abstract

The invention belongs to the technical field of equipment monitoring, and discloses a stress wave sensing-based measurement method and system applied to a wind power blade. Comprising the following steps: acquiring stress wave signal data, preprocessing the stress wave signal data, and sequentially performing bipolar cooperative decomposition, error correction and environmental interference suppression on the stress wave signal data to obtain perfect stress wave signal data; performing feature extraction on the perfect stress wave signal data, generating an intermediate feature matrix, performing further feature screening on the intermediate feature matrix, and outputting stress wave signal feature data; constructing a damage assessment model based on the stress wave signal feature data, and performing damage assessment on the wind power blade by using the damage assessment model to obtain a damage assessment record; generating a three-dimensional damage map based on the damage evaluation record, and sending the three-dimensional damage map to a preset wind power equipment safety terminal; equipment monitoring on the wind power blade is realized by using the stress wave.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring. More specifically, the present invention relates to a measurement method and system based on stress wave sensing for wind turbine blades. Background Art

[0002] The patent with the application publication number CN117074215A discloses a damage analysis method based on stress wave segmentation. Through stress wave segmentation, the stress wave obtained by a single impact can be used to analyze the damage situation inside the specimen or structure, avoiding the need for additional damage testing after the structure is impacted, improving the timeliness of damage identification; solving the problem that traditional damage testing requires other equipment to detect damage to the specimen, which will cause different degrees of disturbance to the specimen during this process, resulting in experimental errors; a method for testing and analyzing the dynamic impact damage evolution, amplitude, and spectral attenuation laws of solid materials such as rocks and concrete under in-situ pressure-maintaining conditions. It makes up for the defects of the existing technical methods for dynamic testing of solid materials such as rocks and concrete based on Hopkinson bars in the aspect of dynamic damage testing of materials under in-situ pressure-maintaining conditions.

[0003] With the development of wind turbine generators, as a core component, the health status of wind turbine blades directly affects the operation of the entire wind turbine generator; in order to monitor the health status of wind turbine blades, stress waves are used to detect damage to wind turbine blades; the stress wave data collected based on sensors need to be finely preprocessed, otherwise errors will be generated during the propagation of stress waves due to noise and environmental factor interference, resulting in inaccurate collected data; when extracting features from the collected data, general methods lack the joint analysis of the time domain and frequency domain and are prone to ignoring some features; when constructing a model for damage assessment, the recognition accuracy is reduced due to the model's inability to learn new features, and the evaluation effect is poor; the evaluation results output by the model are not intuitive enough, which is not conducive to technicians' viewing.

[0004] In view of this, the present invention proposes a measurement method and system based on stress wave sensing for wind turbine blades to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention provides the following technical solution: A measurement method based on stress wave sensing for wind turbine blades, comprising:

[0006] S1. Collect stress wave signal data and preprocess the stress wave signal data. By performing bipolar collaborative decomposition, error correction, and environmental interference suppression on the stress wave signal data in sequence, perfect stress wave signal data is obtained;

[0007] S2. Extract features from the improved stress wave signal data, generate an intermediate feature matrix, further screen the intermediate feature matrix, and output the stress wave signal feature data;

[0008] S3. Build a damage assessment model based on the stress wave signal feature data, use the damage assessment model to assess the damage of the wind turbine blade, and obtain the damage assessment record;

[0009] S4. Generate a three-dimensional damage map based on the damage assessment record and send the three-dimensional damage map to a preset wind power equipment safety terminal.

[0010] Furthermore, sensors are symmetrically arranged on the main beam, leading edge, and trailing edge of each wind turbine blade, and the distance between any two sensors is less than or equal to a preset distance threshold; a stress wave generator is used to emit stress waves and the stress wave signal data is collected by the sensors, and the stress wave signal data includes a time-domain waveform signal and a stress wave propagation time difference.

[0011] Furthermore, the method for preprocessing the stress wave signal data includes:

[0012] Use the fast Fourier transform algorithm to convert the time-domain waveform signal into a frequency-domain spectrum signal; perform bipolar collaborative decomposition on the frequency-domain spectrum signal, and at the same time calculate the stress wave propagation speed based on the stress wave propagation time difference, and integrate the bipolar collaborative decomposition result, the stress wave propagation time difference, and the stress wave propagation speed to obtain multi-modal decomposition data; design a propagation dynamic compensation mechanism to correct the error of the multi-modal decomposition data and generate propagation correction data; suppress the environmental interference of the propagation correction data based on the collected multi-physical field parameters to obtain the improved stress wave signal data;

[0013] The method for performing bipolar collaborative decomposition on the frequency-domain spectrum signal includes:

[0014] Using a pre-selected wavelet packet basis function to perform N-layer decomposition on the frequency-domain spectrum signal to generate multi-scale spectrum sub-bands and ensuring that all generated spectrum sub-bands can cover the entire frequency band; calculating the energy proportion of each spectrum sub-band and setting a sub-band energy threshold; when the energy proportion of any spectrum sub-band is greater than or equal to the preset sub-band energy threshold, then determining that spectrum sub-band as a valid sub-band, otherwise determining that spectrum sub-band as a noise sub-band, screening valid sub-bands based on the sub-band energy threshold and separating out noise sub-bands; performing adaptive soft-threshold filtering on the valid sub-bands to obtain valid high-frequency sub-bands; performing empirical mode decomposition on the noise sub-bands to generate M intrinsic mode functions; calculating the stability index of each intrinsic mode function, retaining the intrinsic mode components with a stability index less than the preset stability index threshold, that is, valid mode components; using an energy injection algorithm to extract the energy of the intrinsic mode components with a stability index greater than or equal to the preset stability index threshold and superimposing its energy on the adjacent valid mode components; performing synchronous extraction transformation on the valid high-frequency sub-bands and valid mode components and constructing a time-frequency feature matrix, and integrating the time-frequency feature matrix, stress wave propagation time difference, and stress wave propagation speed to obtain multi-modal decomposition data.

[0015] Further, the method for correcting the error of the multi-modal decomposition data includes:

[0016] Obtaining the basic propagation speed and material coefficient of stress waves in the wind turbine blade by querying a preset wind turbine blade parameter database; constructing a wave speed function based on the basic propagation speed and material coefficient; obtaining the actual propagation speed of stress waves in the wind turbine blade by calculating the function value of the wave speed function;

[0017] Performing ascending sorting on the stress wave propagation time difference based on the time stamp to obtain a time delay sequence and calculating the time delay mean of the time delay sequence; extracting the energy intensity distribution of each sub-band from the multi-modal decomposition data; constructing a wave speed correction equation based on the time delay mean and the energy intensity distribution of each sub-band; using a pre-trained LSTM network model to update the coupling coefficient of the wave speed correction equation and adjusting the actual propagation speed of stress waves in the wind turbine blade based on the updated wave speed correction equation; constructing an attenuation compensation function and correcting the sub-band energy attenuation in the multi-modal decomposition data based on this function; updating the multi-modal decomposition data based on the adjusted actual propagation speed of the stress wave and the corrected sub-band energy;

[0018] The method for suppressing environmental interference on the propagation correction data includes:

[0019] Physical parameter sensors are arranged on each wind power blade to collect temperature, stress, and humidity data of the wind power blade; a multi-physical field coupling factor is constructed based on the temperature, stress, and humidity data of the wind power blade; an adaptive band-stop filter is designed and used to suppress high-frequency noise generated by temperature drift in the frequency domain dimension; a beam null is generated in the interference direction by using the metasurface beamforming method; the weights of each physical parameter in the multi-physical field coupling factor and the stopband parameters of the adaptive band-stop filter are dynamically adjusted by using a multi-objective genetic algorithm.

[0020] Further, the method for feature extraction of the improved stress wave signal data includes:

[0021] The improved stress wave signal data is classified based on the data type, including sub-band energy intensity data, dynamic propagation data, and time-frequency data; the sub-band energy intensity data and time-frequency data are integrated into a time-frequency feature data set, and the dynamic propagation data is used as a dynamic propagation feature data set; the time-frequency feature data sets and dynamic propagation feature data sets from P sensors are stacked in a three-dimensional tensor, and the time sampling points C and feature dimensions D within any preset time window are cross-combined to obtain an initial feature tensor with a dimension of P×C×D within each time window; the initial feature tensor within each time window is subjected to collaborative processing of time-domain synchronization calibration and frequency-domain phase alignment to obtain a calibrated feature tensor; the position of the damage area of the wind power blade is calculated based on the time-frequency feature data set; the Euclidean distance between the position of the damage area of the wind power blade and any one sensor is calculated, and at the same time, the signal-to-noise ratio index of each sensor is calculated. Weights are assigned to each sensor based on the signal-to-noise ratio index of each sensor and the Euclidean distance between the position of the sensor and the damage area of the wind power blade, that is, the calibrated feature tensor is weighted to obtain an intermediate feature matrix.

[0022] Further, the method for further feature screening of the intermediate feature matrix includes:

[0023] The local density of each feature point in the intermediate feature matrix is calculated, and all feature points in the intermediate feature matrix are clustered based on the local density of each feature point. The intermediate feature matrix is divided into two types of sub-matrices, namely a high-density matrix and a low-density matrix; an adaptive kernel function is constructed to map the high-density matrix and the low-density matrix into a high-dimensional space respectively to obtain a high-dimensional space feature matrix.

[0024] Build a classification model, using the random forest model as the basic structure of the classification model; collect historical wind turbine blade damage data and convert this data into a historical feature matrix; query the preset wind turbine blade parameter database to use the key damage features involved in the damage situation corresponding to the historical feature matrix as the training labels of the classification model; use the historical feature matrix to train the classification model until the function value of the loss function of the classification model no longer decreases, obtaining a trained classification model; use the classification model to screen key features from the high-dimensional space feature matrix, and calculate the Gini index of each feature point in the high-dimensional space feature matrix; sort all feature points in descending order based on the Gini index size, and determine the feature points greater than the preset sorting threshold as key feature points; use the mutual information analysis algorithm to remove redundant feature points in the key feature points, and output a damage-sensitive feature set;

[0025] Standardize the damage-sensitive feature set to obtain a standardized feature set; use the principal component analysis algorithm to perform dimensionality reduction and compression on the standardized feature set to obtain stress wave signal feature data.

[0026] Furthermore, the method for building a damage assessment model includes:

[0027] Build a damage coding system that uses a hierarchical coding rule; query the preset wind turbine blade parameter database to determine all specific damage situations of the wind turbine blade and the corresponding features of the damage situations; encode all specific damage situations of the wind turbine blade based on the damage coding system, and combine the code of any specific damage situation with the corresponding features of this damage situation into a feature-code mapping matrix; integrate all feature-code mapping matrices to obtain a feature-code mapping data set;

[0028] Build a damage assessment model and use the residual neural network model as the basic structure of the damage assessment model, including an input layer, a fully connected layer, a feature processing layer, a damage quantification layer, and an output layer; use the stress wave signal feature data as the input data of the damage assessment model; use a dual-channel parallel processing architecture as the basic framework of the feature processing layer in the damage assessment model, including a time-frequency feature channel and a spatial propagation channel; the input data is jointly processed by the time-frequency feature channel and the spatial propagation channel to generate an intermediate feature tensor; the intermediate feature tensor enters the damage quantification layer to calculate a damage index, and the damage index, damage location coordinates, and damage type are output by the output layer; match the damage index, damage location coordinates, and damage type to generate a damage code; dynamically update the existing damage assessment model after each output of the damage code; add a real-time timestamp to all output damage codes and sort them in chronological order to obtain a damage assessment record.

[0029] Furthermore, the method for dynamically updating the existing damage assessment model includes:

[0030] Build an incremental learning architecture, establish a dynamic feature memory bank after the feature processing layer of the damage assessment model, back up the intermediate feature tensors, extract features from the backed-up intermediate feature tensors to obtain a historical damage feature dataset; introduce a dual-model collaborative distillation mechanism in the incremental learning architecture, including a first model and a second model; use the first model to freeze the feature processing layer of the damage assessment model and output a probability tensor of the historical damage feature dataset; the second model copies the parameters of the frozen layer from the first model and constructs an adjustment function based on the probability tensor to adjust the parameters of the unfrozen layer; set the initial learning rate of the incremental learning architecture and adjust the initial learning rate based on the cosine annealing algorithm; set parameter constraints for the damage quantization layer to ensure that the parameter offset is less than a preset offset threshold; calculate the performance metrics of the damage assessment model after adding the incremental learning architecture, and if each performance metric meets the expected value, the damage assessment model update is completed, otherwise, the damage assessment model rollback is triggered.

[0031] Further, the method for generating a three-dimensional damage map based on the damage assessment record includes:

[0032] Extract the damage position coordinates, damage index, and damage type from the damage assessment record and construct a damage spatial distribution matrix; use the Kriging spatial interpolation algorithm to perform three-dimensional grid processing on the damage spatial distribution matrix to obtain the wind turbine blade damage global density field; perform parametric modeling on the wind turbine blade based on the geometric parameters of the wind turbine blade to obtain a three-dimensional wind turbine blade model; map the wind turbine blade damage global density field to the three-dimensional wind turbine blade model to obtain a three-dimensional damage model; render the three-dimensional damage model using colors to obtain a damage rendering model; sort the damage rendering models corresponding to each timestamp in chronological order to obtain a three-dimensional damage map.

[0033] A stress wave perception-based measurement system for wind turbine blades, which is used to implement a stress wave perception-based measurement method for wind turbine blades, includes:

[0034] A data acquisition module, which is used to collect stress wave signal data and preprocess the stress wave signal data to obtain perfect stress wave signal data;

[0035] A feature extraction module, which is used to extract features from the perfect stress wave signal data and output stress wave signal feature data;

[0036] A damage assessment module, which constructs a damage assessment model based on the stress wave signal feature data and uses the damage assessment model to assess the damage of the wind turbine blade to obtain a damage assessment record;

[0037] A visualization generation module generates a three-dimensional damage map based on the damage assessment record and sends the three-dimensional damage map to a preset safety terminal of a wind power device; each module is connected by a wired and / or wireless manner.

[0038] The technical effects and advantages of a measurement method and system based on stress wave sensing applied to a wind turbine blade according to the present invention:

[0039] By collecting stress wave signal data, performing preprocessing and feature extraction, stress wave signal feature data is obtained, and a model is constructed based on this data to evaluate the damage of the wind turbine blade. Finally, a visualized three-dimensional damage map is generated based on the damage assessment record, realizing the process of equipment monitoring of the wind turbine blade based on stress waves; compared with existing experience, the preprocessing of stress wave signal data is more accurate, and at the same time, the influence of noise interference in the data, errors in the propagation process and environmental factors is considered; joint feature extraction is performed from both the time domain and the frequency domain to obtain features related to the damage of the wind turbine blade and construct a classification model to further screen the features, obtaining key features that can represent the damage situation and performing dimensionality reduction compression, improving the processing efficiency; a damage assessment model is constructed and the model is dynamically updated based on an incremental learning architecture, and at the same time, the model is used to identify and evaluate the damage situation, improving the adaptability and processing accuracy of the model; finally, a visualized damage map is generated and sent to a preset safety terminal of the wind power device, and subsequently, technicians take corresponding measures to repair the wind turbine blade based on this three-dimensional damage map. Brief Description of the Drawings

[0040] Figure 1 It is a schematic diagram of a measurement method based on stress wave sensing applied to a wind turbine blade according to the present invention;

[0041] Figure 2 It is a schematic diagram of a measurement system based on stress wave sensing applied to a wind turbine blade according to the present invention. Detailed Embodiments

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0043] Embodiment 1

[0044] Please refer to Figure 1 As shown, a measurement method based on stress wave sensing applied to a wind turbine blade in this embodiment includes:

[0045] S1. Collect stress wave signal data and preprocess the stress wave signal data. By performing bipolar collaborative decomposition, error correction, and environmental interference suppression on the stress wave signal data in sequence, perfect stress wave signal data is obtained;

[0046] S2. Extract features from the perfect stress wave signal data, generate an intermediate feature matrix, and further screen the intermediate feature matrix to output stress wave signal feature data;

[0047] S3. Build a damage assessment model based on the stress wave signal feature data, and use the damage assessment model to assess the damage of the wind turbine blade to obtain a damage assessment record;

[0048] S4. Generate a three-dimensional damage map based on the damage assessment record and send the three-dimensional damage map to a preset wind power equipment safety terminal.

[0049] Sensors are symmetrically arranged on the main beam, leading edge, and trailing edge of each wind turbine blade, and the distance between any two sensors is less than or equal to a preset distance threshold (in this embodiment, the distance threshold is 30 cm); a stress wave generator (a device used to generate controllable stress pulses. In this embodiment, electromagnetic force is used to excite stress waves) is used to emit stress waves and the stress wave signal data is collected by the sensors. The stress wave signal data includes a time-domain waveform signal and a stress wave propagation time difference (each sampling point in the time-domain waveform signal contains a timestamp corresponding to the acquisition time of each sampling point; the stress wave propagation time difference is calculated based on the time difference of multi-sensor timestamps. For example, the timestamp of a certain time-domain waveform signal received by sensor A1 is time1, and the timestamp of the same time-domain waveform signal received by sensor A2 is time2; the absolute value of the difference between time1 and time2 is the stress wave propagation time difference).

[0050] Even if high-precision instruments are used to collect stress wave signal data during the data collection stage, in order to obtain more accurate data in the subsequent processing of stress wave data, preprocessing is an indispensable step; the sensors used in this embodiment are piezoelectric ceramic sensors. For piezoelectric ceramic sensors, the dielectric constant of the piezoelectric material is easily affected by environmental factors, resulting in noise in the received data or reduced accuracy. For example, there are temperature and electromagnetic interferences in the data collection site; therefore, it is necessary to perform denoising processing and environmental interference suppression on the collected data during the preprocessing process to ensure that the stress wave data used in this embodiment is accurate enough to the greatest extent, indirectly improving the accuracy and efficiency of subsequent operations.

[0051] The methods for preprocessing the stress wave signal data include:

[0052] The time-domain waveform signal is converted into a frequency-domain spectrum signal using the fast Fourier transform algorithm (the Fourier transform algorithm is a common algorithm for converting a waveform signal from the time domain to the frequency domain; compared with other types of Fourier transform algorithms, the fast Fourier transform algorithm has lower complexity, occupies less computing resources, and has extremely high accuracy and conversion efficiency; the time characteristics of the frequency-domain spectrum signal obtained by converting the time-domain waveform signal strictly correspond to the original waveform, that is, the timestamps of the two signals are exactly the same); the frequency-domain spectrum signal is subjected to bipolar collaborative decomposition, and at the same time, the stress wave propagation speed is calculated based on the stress wave propagation time difference (the distance between any two sensors is measured and the stress wave propagation speed is calculated in combination with the stress wave propagation time difference of the two sensors at a certain timestamp), and the bipolar collaborative decomposition result, the stress wave propagation time difference, and the stress wave propagation speed are integrated to obtain multi-modal decomposition data; a propagation dynamic compensation mechanism is designed to correct the error of the multi-modal decomposition data to generate propagation correction data; the environmental interference of the propagation correction data is suppressed based on the collected multi-physical field parameters to obtain improved stress wave signal data (bipolar collaborative decomposition is used to denoise the original data, the design of the propagation dynamic compensation mechanism is used to process the error of the denoised data, and environmental interference suppression is used to reduce the influence of environmental factors).

[0053] Bipolar collaborative decomposition refers to the collaborative signal decomposition using wavelet packet basis functions and empirical mode decomposition, where wavelet packet basis functions are used to process high-frequency signals, and empirical mode decomposition is used to process low-frequency signals. The combination of the two ensures the complete processing of the frequency-domain spectrum signal, and at the same time, the synchronous extraction transform algorithm is used to solve the problem of energy dispersion, and finally high-quality multi-modal decomposition data is obtained.

[0054] The methods for performing bipolar collaborative decomposition on the frequency-domain spectrum signal include:

[0055] The frequency-domain spectrum signal is decomposed into N layers (in this embodiment, N = 5) using a pre-selected wavelet packet basis function (such as the Daubechies5 wavelet basis function) to generate multi-scale spectrum sub-bands and ensure that all generated spectrum sub-bands can cover the full frequency band (in this embodiment, the full frequency band refers to the frequency range [1 kHz, 50 MHz]; when using stress waves for equipment monitoring, the sensitive frequency band of equipment damage belongs to the above frequency range, so full frequency band coverage is implemented for comprehensive perception); calculate the energy ratio of each spectrum sub-band and set the sub-band energy threshold; the calculation formula for the energy ratio of each spectrum sub-band is: where E represents the energy ratio of any spectrum sub-band; H j,k(t) represents the wavelet packet coefficient of the k-th spectral sub-band in the j-th layer at time stamp t; when the energy ratio of any spectral sub-band is greater than or equal to the preset sub-band energy threshold, the spectral sub-band is determined as an effective sub-band, otherwise it is determined as a noise sub-band. The effective sub-bands are screened based on the sub-band energy threshold and the noise sub-bands are separated (since the energy of the sub-bands with lower frequencies accounts for a smaller proportion and noise is generally distributed in the low-frequency sub-bands, the sub-bands with an energy ratio less than the preset sub-band energy threshold are determined as noise sub-bands); the effective sub-bands are subjected to adaptive soft threshold filtering to obtain effective high-frequency sub-bands; the calculation formula for adaptive soft threshold filtering is: where, represents the wavelet packet coefficient corresponding to the effective high-frequency sub-band; λ j represents the threshold coefficient that dynamically decays with the decomposition layer j. In this embodiment, λ = 0.8; σ j represents the noise standard deviation of the j-th layer sub-band; the noise sub-bands are subjected to empirical mode decomposition to generate M intrinsic mode functions (in this embodiment, M = 10); the stability index of each intrinsic mode function is calculated, and the calculation formula for the stability index is: where, N0 represents the number of sampling points in a window (all sampling points in each intrinsic mode function are analyzed through a sliding window); f n0 represents the instantaneous frequency of the n0-th sampling point in the window; the intrinsic mode components with a stability index less than the preset stability index threshold are retained, that is, the effective mode components; the energy of the intrinsic mode components with a stability index greater than or equal to the preset stability index threshold is extracted using the energy re-injection algorithm and its energy is superimposed on the adjacent effective mode components (the energy in the un-retained intrinsic mode components is superimposed on the adjacent effective mode components, avoiding energy dispersion and information loss); the effective high-frequency sub-bands and the effective mode components are synchronously extracted and transformed to construct a time-frequency feature matrix (synchronous extraction transformation is a signal processing method that combines time-frequency analysis and mode decomposition, used to extract high-resolution instantaneous frequencies and construct a time-frequency feature matrix; the data processing accuracy is extremely high and the situation of mode mixing is avoided), and the time-frequency feature matrix, the stress wave propagation time difference, and the stress wave propagation speed are integrated to form multi-modal decomposition data (where the time-frequency feature matrix is a feature matrix including the three-dimensional information of the time stamp, frequency, and energy intensity distribution of any sub-band in the multi-modal decomposition data).

[0056] Due to the different materials of wind turbine blades and the different sub-band energy distributions of each frequency in the multi-modal decomposition data, the stress wave propagation speed will change accordingly; during the propagation of stress waves, the energy of stress waves is likely to decay due to factors such as medium absorption, scattering, and path deformation. To ensure the accuracy of subsequent operations, it is necessary to correct the energy decay; therefore, a wave speed correction equation and an attenuation compensation function are constructed respectively to correct the errors in the multi-modal decomposition data.

[0057] The methods for correcting the errors in the multi-modal decomposition data include:

[0058] Obtain the basic propagation speed and material coefficient of stress waves in the wind turbine blade by querying the preset wind turbine blade parameter database (the material coefficient is different for different types of materials of the wind turbine blade, and the propagation speed of stress waves in wind turbine blades of different materials is also different); construct a wave speed function based on the basic propagation speed and material coefficient; the calculation formula of the wave speed function is: Among them, V(f) represents the actual propagation speed of the stress wave frequency band with frequency f in the wind turbine blade with material coefficient α; V0 represents the basic propagation speed of the stress wave in the wind turbine blade with material coefficient α; f1 represents the average frequency of the stress wave; the value of the material coefficient is determined by the specific situation of the wind turbine blade material; the actual propagation speed of the stress wave in the wind turbine blade is obtained by calculating the function value of the wave speed function.

[0059] Ascendingly sort the stress wave propagation time differences based on timestamps to obtain a time delay sequence and calculate the time delay mean of the time delay sequence; extract the energy intensity distribution of each sub-band from the multi-modal decomposition data; construct a wave speed correction equation based on the time delay mean and the energy intensity distribution of each sub-band; the calculation formula of the wave speed correction equation is: Among them, represents the adjusted actual propagation speed; Δt represents the time delay mean; E1 represents the energy ratio of any one sub-band; β represents the coupling coefficient; update the coupling coefficient of the wave speed correction equation using a pre-trained LSTM network model (the LSTM network model is jointly trained by the collected historical data and the enhanced samples generated by online simulation, and the energy intensity distribution and time delay sequence of the stress wave frequency band with frequency f are used as input data to dynamically update the coupling coefficient of the wave speed correction equation, ensuring the timeliness of the wave speed correction equation) and adjust the actual propagation speed of the stress wave in the wind turbine blade based on the updated wave speed correction equation (update the original actual propagation speed to the adjusted actual speed calculated by the wave speed correction equation); construct an attenuation compensation function and correct the sub-band energy attenuation in the multi-modal decomposition data based on this function; the calculation formula of the attenuation compensation function is: Among them, R(f2) represents the compensated frequency attenuation value of the stress wave frequency band with frequency f2; γ0(f2) represents the initial attenuation compensation coefficient of the stress wave frequency band with frequency f2 (this coefficient reflects the inherent attenuation characteristics of the medium and is related to the medium material and environmental factors); a0 represents an adjustment factor (in this embodiment, a0 = 0.2); E2(f2) represents the energy intensity of the stress wave frequency band with frequency f2; Ew(f2) represents the standard energy intensity of the stress wave frequency band with frequency f2 obtained by querying the preset wind turbine blade parameter database; update the multi-modal decomposition data based on the actual propagation speed of the adjusted stress wave and the corrected sub-band energy.

[0060] In the environment where the wind turbine blade operates, it is generally most commonly affected by temperature, stress, and humidity. Therefore, a multi-physical field coupling factor for multi-field collaboration is designed to quantify the degree of influence; at the same time, by designing an adaptive band-stop filter and using the metasurface beamforming method, the interference generated by environmental influences is jointly suppressed from the frequency domain dimension and the spatial dimension; finally, the parameters of the multi-physical field coupling factor and the adaptive band-stop filter are dynamically adjusted through a multi-objective genetic algorithm to maximize the roles that the two methods can play.

[0061] The methods for suppressing environmental interference on the propagation correction data include:

[0062] Physical parameter sensors are arranged on each wind turbine blade to collect temperature, stress, and humidity data of the wind turbine blade; a multi-physical field coupling factor is constructed based on the temperature, stress, and humidity data of the wind turbine blade (the multi-physical field coupling factor is used to quantify the contribution degree of each physical parameter to the signal attenuation of the propagation correction data, which is manifested as a weighted sum of each physical parameter); an adaptive band-stop filter is designed and used to suppress the high-frequency noise generated by temperature drift in the frequency domain dimension (define the stopband depth DE of the adaptive band-stop filter as DE = -40×log 10 (1 + μ); where μ represents the stopband parameter); use the metasurface beamforming method to generate a beam null in the interference direction (the beam null is an extremely low gain area generated by the metasurface beamforming method in the interference source direction to suppress electromagnetic wave interference); dynamically adjust the weights of each physical parameter in the multi-physical field coupling factor and the stopband parameter of the adaptive band-stop filter through a multi-objective genetic algorithm (convert the weights of each physical parameter in the multi-physical field coupling factor and the stopband parameter of the adaptive band-stop filter into a parameter vector; define the genetic algorithm population, and represent any individual in the population as any parameter vector; perform crossover and mutation operations on the population and repeat this process continuously until the maximum number of iterations is reached, and screen the optimal individual based on the Pareto front, that is, the optimal parameter vector).

[0063] In order to more accurately measure the damage condition of wind turbine blades, feature extraction is performed on the improved stress wave signal data. To make the feature extraction effect more comprehensive, the data of all sensors are fused, a feature tensor is constructed to accommodate the features related to damage, the feature tensor is calibrated to eliminate errors, and then feature screening is carried out to obtain accurate and complete stress wave signal feature data.

[0064] The methods for feature extraction of the improved stress wave signal data include:

[0065] Classify the improved stress wave signal data based on the data type (the data type classification standard is the difference in data units. For example, the unit of sub-band energy intensity data is different from that of dynamic propagation data), including sub-band energy intensity data, dynamic propagation data, time-frequency data (time-frequency data includes the corresponding band frequencies collected by any sensor at each timestamp and the time delay sequence belonging to the acquisition time interval; dynamic propagation data includes data such as the actual propagation speed of stress waves and the rate of change of speed); integrate the sub-band energy intensity data and time-frequency data into a time-frequency feature data set, and use the dynamic propagation data as a dynamic propagation feature data set; perform three-dimensional tensor stacking on the time-frequency feature data set and the dynamic propagation feature data set from P sensors (P represents the preset number of installed sensors) (three-dimensional tensor stacking means stacking the feature matrices of three dimensions to form a tensor), cross-combine the time sampling points C and the feature dimension D within any preset time window (where the size of one time window can be changed according to the actual situation. In this embodiment, the size of one time window is 1S; the time sampling point C refers to the number of sampling points within one time window being C; the feature dimension D refers to the sum of the number of feature dimensions for each sub-band within the time window. For example, a certain sub-band contains 3 feature dimensions including sub-band energy intensity distribution, propagation speed, and rate of change of speed, another sub-band contains 5 feature dimensions, and there are only these two sub-bands within one time window, then the feature dimension D represents a total of 8 feature dimensions), to obtain an initial feature tensor with a dimension of P×C×D within each time window; perform collaborative processing of time-domain synchronization calibration and frequency-domain phase alignment on the initial feature tensor within each time window to obtain a calibrated feature tensor (the method for time-domain synchronization calibration is: select one sensor as the main sensor, calculate the time shift amount of other sensor signals relative to the main sensor through the cross-correlation dominant peak tracking algorithm, and perform time shift compensation on each sensor signal based on the time shift amount to obtain compensated sensor signals; the method for frequency-domain phase alignment is: perform short-time Fourier transform on the compensated sensor signals to obtain a compensated time-frequency matrix, and use the instantaneous frequency gradient estimation to correct the phase of the compensated time-frequency matrix); calculate the position of the damaged area of the wind turbine blade based on the time-frequency feature data set (the calculation formula for the position of the damaged area of the wind turbine blade is: Among them, the coordinates of the position of the damaged area of the wind turbine blade are (x, y); the coordinates of sensor a are (x a , y a ), and the coordinates of sensor b are (x b , y b ); represents the actual propagation speed of the stress wave with frequency f; t ab represents the stress wave propagation time difference between sensor a and sensor b); calculate the Euclidean distance between the position of the damaged area of the wind turbine blade and any one sensor, and at the same time calculate the signal-to-noise ratio index of each sensor. Based on the signal-to-noise ratio index of each sensor and the Euclidean distance between the position of this sensor and the damaged area of the wind turbine blade, assign weights to each sensor, that is, weight the calibrated feature tensor (the calculation formula for the weight of each sensor is: Among them, ω p0 represents the weight of the p0th sensor; SNR(p0) represents the signal-to-noise ratio index of the p0th sensor; -γ0(f) represents the initial attenuation compensation coefficient in the f frequency band for the p0th sensor; disp0 represents the Euclidean distance between the p0th sensor and the position of the damaged area; this weight takes into account the signal-to-noise ratio index of each sensor and the Euclidean distance from the damaged area, which is convenient for controlling the contribution ratio of global and local features. For example, the higher the signal-to-noise ratio of the sensor, the greater the energy of the collected stress wave, and the smaller the Euclidean distance between the sensor and the damaged area, the closer the sensor is to the damaged area. Therefore, it is easier to extract damage features from the data of this sensor, and a greater weight should be assigned to such a sensor), and obtain the intermediate feature matrix.

[0066] The intermediate feature matrix contains features of all dimensions. However, due to the large number of features and the presence of redundant features that are repeated or similar, it is necessary to further screen the features of the intermediate feature matrix; first, cluster based on the local density of feature points through a clustering algorithm and map the features to a high-dimensional space, then use a classification model with a random forest model as the basic structure for precise feature screening, and finally use the principal component analysis algorithm to reduce the dimension and compress the screened features, improving the feature screening accuracy and processing efficiency.

[0067] The methods for further feature screening of the intermediate feature matrix include:

[0068] Calculate the local density of each feature point in the intermediate feature matrix (the calculation formula for the local density is: Among them, ρ represents the local density of any feature point pt; ki(pt) represents the K-nearest neighbor set of the feature point pt. In this embodiment, the K-nearest neighbor parameter ki of the K-nearest neighbor set takes a value of 8; dis(pt - pt sum) represents the Euclidean distance between the feature point pt and the sum-th feature point), and clusters all the feature points in the intermediate feature matrix based on the local density of each feature point, dividing the interior of the intermediate feature matrix into two types of sub-matrices, namely the high-density matrix and the low-density matrix (the distance between adjacent feature points in the high-density matrix is small, and this type of matrix is conducive to enhancing local details; the distance between adjacent feature points in the low-density matrix is large, and this type of matrix is conducive to suppressing overfitting); constructs an adaptive kernel function to map the high-density matrix and the low-density matrix to a high-dimensional space respectively, obtaining a high-dimensional space feature matrix (the calculation formula of the adaptive kernel function is: HE(pt - pt sum ) = β1 × HE RB (pt - pt sum ) + β2 × HE RF (pt - pt sum ); where, HE represents the adaptive kernel function; HE RB represents the radial basis kernel function, β1 represents the weight coefficient of HE RB ; HE RF represents the linear kernel function, β2 represents the weight coefficient of HE RF ; in this embodiment, β1 > β2 when processing the high-density matrix, for capturing local details; β1 < β2 when processing the low-density matrix, for capturing the overall trend; β1 + β2 = 1; uses this function to map the features originally presenting a non-linear distribution to a high-dimensional space, making them linearly separable in the high-dimensional space and enhancing the separability of the features).

[0069] Constructs a classification model, using the random forest model as the basic structure of the classification model (the random forest model is an excellent classification model, with strong robustness and fault tolerance, and is suitable for processing high-dimensional data); collects historical wind turbine blade damage data, and converts this data into a historical feature matrix; through querying a preset wind turbine blade parameter database, takes the key damage features involved in the damage situation corresponding to the historical feature matrix as the training labels of the classification model; trains the classification model using the historical feature matrix until the function value of the loss function (such as the cross-entropy loss function) of the classification model no longer decreases, obtaining a trained classification model; uses the classification model to screen key features from the high-dimensional space feature matrix, calculates the Gini index of each feature point in the high-dimensional space feature matrix (the Gini index is an index used to measure the importance of features in the random forest model); sorts all the feature points in descending order based on the Gini index size, and determines the feature points greater than a preset sorting threshold (the sorting threshold is 30%, that is, the top 30% before sorting) as key feature points; uses the mutual information analysis algorithm to remove redundant feature points among the key feature points, and outputs a damage-sensitive feature set (by calculating the mutual information of any two feature points, if it is greater than the preset mutual information threshold, then one of the feature points is removed, reducing the data dimension and improving the efficiency of subsequent processing).

[0070] Standardize the damage-sensitive feature set to obtain a standardized feature set; use the principal component analysis algorithm to reduce the dimension and compress the standardized feature set to obtain stress wave signal feature data (retain the principal components in the standardized feature set whose cumulative variance contribution rate is greater than or equal to 95%. The cumulative variance contribution rate is an index in the principal component analysis algorithm used to measure the amount of information contained in the principal component. The larger this value, the more information the principal component contains, and the more important the corresponding principal component is).

[0071] In order to be able to evaluate the damage condition of the wind turbine blade based on the feature data, a damage assessment model is constructed to identify and evaluate the damage condition of the wind turbine blade; at the same time, a damage coding system is designed to express the evaluation results of the model.

[0072] The ways to construct the damage assessment model include:

[0073] Construct a damage coding system, which uses a hierarchical coding rule (each damage code is divided into three parts, including damage type, damage location, and damage degree; the damage degree is represented by the damage index calculated in the subsequent processing; for example, the damage code G-01-L1 means that a slight crack occurs at the tip of the wind turbine blade. Among them, G means a crack occurs; 01 means the crack location is at the 01 coordinate position, that is, the tip of the wind turbine blade, and the detailed coordinates of the 01 coordinate position can be obtained through calculation; L1 means the damage degree, that is, slight; all parameters in the above damage code can be queried or calculated through a preset wind turbine blade parameter database); query the preset wind turbine blade parameter database to determine all specific damage conditions of the wind turbine blade and the corresponding features of the damage conditions; code all specific damage conditions of the wind turbine blade based on the damage coding system, and combine the code of any specific damage condition with the corresponding features of this damage condition into a feature-code mapping matrix (for example, the damage code G-01-L1 corresponds to the feature that the energy intensity at this position exceeds the preset energy threshold and the stress wave propagation speed and path change suddenly); integrate all feature-code mapping matrices to obtain a feature-code mapping data set.

[0074] Build a damage assessment model and use the residual neural network model as the basic structure of the damage assessment model, including an input layer, a fully connected layer, a feature processing layer, a damage quantification layer, and an output layer; use the stress wave signal feature data as the input data of the damage assessment model; use a dual-channel parallel processing architecture as the basic framework of the feature processing layer in the damage assessment model, including a time-frequency feature channel and a spatial propagation channel; the time-frequency feature channel extracts the time-frequency features of the input data through the convolutional kernels integrated in the residual blocks of the residual neural network, and at the same time introduces a cross-channel attention mechanism (CBAM) to enhance the damage-sensitive frequency bands; the spatial propagation channel obtains the stress wave propagation relationship between each sensor through a graph attention network (GAT), and captures the propagation mode of the stress wave at the damage location (the energy transfer situation of the stress wave between each sensor is reflected by the attention coefficient of the GAT, and at the same time, since the damage location will change the stress wave propagation path, such a propagation mode can also be captured by the GAT); the input data is jointly processed by the time-frequency feature channel and the spatial propagation channel to generate an intermediate feature tensor; the intermediate feature tensor enters the damage quantification layer to calculate the damage index (the calculation formula of the damage index is: LI = μ1×ΔV + μ2×ΔE + μ3×S dis ; where, LI represents the damage index; ΔV represents the attenuation rate of the stress wave propagation speed at the damage location; ΔE represents the energy attenuation rate at the damage location; S dis represents the spatial discreteness of the damage location, which is calculated based on the variance of the detailed coordinates of the damage location; μ1, μ2, and μ3 are the weights of ΔV, ΔE, and S dis respectively, and the particle swarm algorithm is used to optimize the three weights to achieve dynamic adjustment of the weights; set a damage safety threshold, a first damage threshold, and a second damage threshold, and classify the damage degree into four categories based on the three damage thresholds. When the damage index is less than or equal to the damage safety threshold, it is determined to be in a safe degree; when the damage index is greater than the damage safety threshold and less than or equal to the first damage threshold, it is determined to be in a minor damage degree; when the damage index is greater than the first damage threshold and less than or equal to the second damage threshold, it is determined to be in a moderate damage degree; when the damage index is greater than the second damage threshold, it is determined to be in a severe damage degree; in this embodiment, the damage safety threshold is set to 0.1, the first damage threshold is set to 0.3, the second damage threshold is set to 0.7, and the damage index value range is (0,1)), and the damage index, the damage location coordinates, and the damage type are output by the output layer; match the damage index, the damage location coordinates, and the damage type to generate a damage code; dynamically update the existing damage assessment model after each output of the damage code (to ensure the evaluation accuracy and real-time performance of the damage assessment model); add a real-time timestamp to all output damage codes and sort them in chronological order to obtain a damage assessment record (sorting each damage code in chronological order is conducive to observing the change of the damage situation and facilitating the subsequent operation to generate a more accurate three-dimensional damage map).

[0075] Since new damage features may be introduced each time the damage condition of a wind turbine blade is evaluated, if the model has to relearn all features every time it conducts an evaluation in order to perform damage assessment, the efficiency of the model will be quite low. Therefore, incremental learning is chosen to dynamically update the damage assessment model, enabling the model to adapt to new features more quickly. At the same time, to prevent the performance of the model from degrading after the update, a model parameter rollback mechanism is introduced to perform dynamic updates while ensuring the stability of the model.

[0076] The ways to dynamically update the existing damage assessment model include:

[0077] Construct an incremental learning architecture, establish a dynamic feature memory bank after the feature processing layer of the damage assessment model, back up the intermediate feature tensors, and extract features from the backed-up intermediate feature tensors to obtain a historical damage feature dataset (the historical damage feature dataset includes damage-sensitive features and high-confidence damage features, that is, features with a damage index greater than or equal to a preset confidence threshold, which is used to ensure the reliability of the memory features; in this embodiment, the confidence threshold is set to 0.8). Introduce a dual-model collaborative distillation mechanism into the incremental learning architecture, including a first model and a second model (where the first model is used to retain and transfer old features, and the second model is used to receive old features and learn new features). Use the first model to freeze the feature processing layer of the damage assessment model (freezing the feature processing layer is to retain the already identified damage features and ensure the basic feature extraction ability), and output the probability tensor of the historical damage feature dataset (the probability tensor refers to the probability weights of old features and their similar features, which improves the efficiency of the first model in transferring old features). The second model copies the parameters of the frozen layer from the first model (copying the parameters of the first model ensures that the second model has the same extraction ability as the first model), and constructs an adjustment function based on the probability tensor to adjust the parameters of the unfrozen layer (the calculation formula of the adjustment function is: where CA represents the adjustment function; KL represents the KL divergence loss function, which is a loss function for measuring the difference in probability distributions; CE represents the cross-entropy loss function; and respectively represent the weights of the KL divergence loss function and the cross-entropy loss function. In this embodiment ) Set the initial learning rate of the incremental learning architecture, and adjust the initial learning rate based on the cosine annealing algorithm (set the initial learning rate to 30% of the learning rate of the damage assessment model before update; gradually decay based on the cosine annealing algorithm to balance the learning speed of new and old features and prevent the forgetting of old features caused by violent parameter oscillations); set parameter constraints for the damage quantification layer to ensure that the parameter offset is less than the preset offset threshold (in this embodiment, the offset threshold is 0.15, that is, if the parameter offset is less than 0.15, no adjustment is made, otherwise the parameter is restored to the value before the offset, ensuring the stability of the damage quantification layer); calculate the performance indicators of the damage assessment model after adding the incremental learning architecture (such as evaluation accuracy, false alarm rate, and evaluation time). If each performance indicator meets the expected value, the update of the damage assessment model is completed, otherwise, the rollback of the damage assessment model is triggered (since the relevant parameters of the damage assessment model before update are retained in a data buffer each time the damage assessment model is dynamically updated, if the performance indicators do not meet the expected values, the parameters of the damage assessment model are rolled back to the state before update, ensuring the stability of the damage assessment model).

[0078] In order to visualize the damage assessment records generated by the damage assessment model and facilitate technicians to view; therefore, the damage assessment records are converted into three-dimensional damage maps, in which a three-dimensional model of the wind turbine blade is established and the damage areas and their damage degrees are marked with colors, facilitating technicians to perform targeted repairs on the wind turbine blade based on the map.

[0079] The methods for generating three-dimensional damage maps based on damage assessment records include:

[0080] Extract the damage location coordinates, damage index, and damage type from the damage assessment record, and construct a damage spatial distribution matrix (this matrix is used to store the spatial distribution characteristics of damage); use the Kriging spatial interpolation algorithm to perform three-dimensional grid processing on the damage spatial distribution matrix to obtain the global damage density field of the wind turbine blade (the Kriging spatial interpolation algorithm is used to analyze spatial correlation to construct the global damage density field of the wind turbine blade, and this density field is used to represent the global damage distribution state of the wind turbine blade); perform parametric modeling on the wind turbine blade based on the geometric parameters of the wind turbine blade (the above geometric parameters can be obtained by querying a preset wind turbine blade parameter database, such as material parameters and blade specifications, etc.; the parametric modeling method is, for example, using CAD for modeling) to obtain a three-dimensional model of the wind turbine blade; map the global damage density field of the wind turbine blade to the three-dimensional model of the wind turbine blade to obtain a three-dimensional damage model (where the mapping area includes the model surface and the internal structure layer); use color to render the three-dimensional damage model to obtain a damage rendering model (adopt the HSL color gamut to map the damage degree, classify the damage degree corresponding to the color based on the size of the damage index, and the safety degree is represented as green; the slight damage degree is represented as a gradual change from green to blue, and when it completely turns blue, the damage index is at the first damage threshold; the moderate damage degree is represented as a gradual change from blue to yellow, and when it completely turns yellow, the damage index is at the second damage threshold; the severe damage degree is represented as a gradual change from yellow to red); sort the damage rendering models corresponding to each timestamp in chronological order to obtain a three-dimensional damage atlas.

[0081] In this embodiment, by collecting stress wave signal data, performing preprocessing and feature extraction, stress wave signal feature data is obtained, and a model is constructed based on this data to evaluate the damage of the wind turbine blade. Finally, a visual three-dimensional damage atlas is generated based on the damage assessment record, realizing the process of equipment monitoring of the wind turbine blade based on stress waves; compared with existing experience, the preprocessing of stress wave signal data is more accurate, and at the same time, the influence of noise interference in the data, errors in the propagation process, and environmental factors is considered; joint feature extraction is performed from both the time domain and the frequency domain to obtain features related to the damage of the wind turbine blade and construct a classification model to further screen the features, obtaining key features that can represent the damage situation and performing dimensionality reduction and compression to improve the processing efficiency; a damage assessment model is constructed and the model is dynamically updated based on the incremental learning architecture, and at the same time, this model is used to identify and evaluate the damage situation to improve the adaptability and processing accuracy of the model; finally, a visual damage atlas is generated and sent to a preset wind power equipment safety terminal, and subsequently, technicians take corresponding measures to repair the wind turbine blade based on this three-dimensional damage atlas.

[0082] Embodiment 2

[0083] Please refer to Figure 2As shown in the figure, for the parts not described in detail in this embodiment, refer to the description in Embodiment 1. A measurement system based on stress wave perception for wind turbine blades is provided, including:

[0084] A data acquisition module, which is used to acquire stress wave signal data and preprocess the stress wave signal data. By performing bipolar collaborative decomposition, error correction, and environmental interference suppression on the stress wave signal data in sequence, perfect stress wave signal data is obtained;

[0085] A feature extraction module, which is used to extract features from the perfect stress wave signal data and output stress wave signal feature data;

[0086] A damage assessment module, which constructs a damage assessment model based on the stress wave signal feature data and uses the damage assessment model to assess the damage of the wind turbine blade to obtain a damage assessment record;

[0087] A visualization generation module, which generates a three-dimensional damage map based on the damage assessment record and sends the three-dimensional damage map to a preset wind power equipment safety terminal; each module is connected in a wired and / or wireless manner.

[0088] Embodiment 3

[0089] This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it realizes the operation mode of the above-provided measurement method based on stress wave perception for wind turbine blades.

[0090] Since the electronic device introduced in this embodiment is the electronic device adopted for implementing a measurement method based on stress wave perception for wind turbine blades in an embodiment of the present application, based on the measurement method based on stress wave perception for wind turbine blades introduced in an embodiment of the present application, those skilled in the art can understand the specific implementation manner and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device realizes the method in an embodiment of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device adopted for a measurement method based on stress wave perception for wind turbine blades in an embodiment of the present application, it falls within the scope of protection of the present application.

[0091] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0092] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as within the protection scope of the present invention.

Claims

1. A measurement method based on stress wave perception applied to wind turbine blades, characterized in that: include: S1. Collect stress wave signal data and pre-process the stress wave signal data, and obtain perfect stress wave signal data by successively performing bipolar collaborative decomposition, error correction and environmental interference suppression on the stress wave signal data; S2. Extract features from the perfected stress wave signal data, generate an intermediate feature matrix, perform further feature screening on the intermediate feature matrix, and output stress wave signal feature data; S3. construct a damage assessment model based on stress wave signal characteristic data, use the damage assessment model to perform damage assessment on the wind turbine blade, and obtain a damage assessment record; S4. Generate a three-dimensional damage map based on the damage assessment record, and send the three-dimensional damage map to a preset wind power equipment safety terminal.

2. The stress wave sensing measurement method for wind turbine blades according to claim 1 is characterized in that: Sensors are symmetrically arranged on the main beam, leading edge and trailing edge of each wind turbine blade, and the interval between any two sensors is less than or equal to a preset distance threshold; a stress wave generator is used to emit stress waves and stress wave signal data is collected through sensors, wherein the stress wave signal data includes a time domain waveform signal and a stress wave propagation time difference.

3. The stress wave sensing-based measurement method for wind turbine blades according to claim 2 is characterized in that: The method of preprocessing the stress wave signal data includes: The time domain waveform signal is converted into a frequency domain spectrum signal using the fast Fourier transform algorithm; the frequency domain spectrum signal is subjected to bipolar collaborative decomposition, and the stress wave propagation velocity is calculated based on the stress wave propagation time difference, and the bipolar collaborative decomposition results, stress wave propagation time difference and stress wave propagation velocity are integrated to obtain multimodal decomposition data; a propagation dynamic compensation mechanism is designed to correct the multimodal decomposition data error and generate propagation correction data; environmental interference is suppressed on the propagation correction data based on the collected multi-physical field parameters to obtain perfect stress wave signal data; The methods of performing bipolar collaborative decomposition on frequency domain spectrum signals include: The frequency domain spectrum signal is decomposed into N layers using the pre-selected wavelet packet basis function to generate multi-scale spectrum sub-bands and ensure that all the generated spectrum sub-bands can cover the entire frequency band; the energy proportion of each spectrum sub-band is calculated and the sub-band energy threshold is set; when the energy proportion of any spectrum sub-band is greater than or equal to the preset sub-band energy threshold, the spectrum sub-band is determined as a valid sub-band, otherwise the spectrum sub-band is determined as a noise sub-band, and the valid sub-bands are screened based on the sub-band energy threshold and the noise sub-bands are separated; the effective sub-bands are adaptively soft-threshold filtered to obtain the effective high-frequency sub-bands; the noise sub-bands are filtered Empirical mode decomposition is performed to generate M intrinsic mode functions; the stability index of each intrinsic mode function is calculated, and the eigenmode components with stability index less than the preset stability index threshold are retained, namely, the effective mode components; the energy reinjection algorithm is used to extract the energy of the eigenmode components with stability index greater than or equal to the preset stability index threshold, and their energy is superimposed on the adjacent effective mode components; the effective high-frequency sub-bands and effective mode components are synchronously extracted and transformed to construct a time-frequency feature matrix, and the time-frequency feature matrix, stress wave propagation time difference and stress wave propagation velocity are integrated to obtain the multi-modal decomposition data.

4. The stress wave sensing-based measurement method for wind turbine blades according to claim 3 is characterized in that: The method of correcting the multimodal decomposition data error includes: The basic propagation velocity and material coefficient of the stress wave in the wind turbine blade are obtained by querying the preset wind turbine blade parameter database; a wave velocity function is constructed based on the basic propagation velocity and material coefficient; and the actual propagation velocity of the stress wave in the wind turbine blade is obtained by calculating the function value of the wave velocity function; Based on the timestamp, the propagation time difference of the stress wave is sorted in ascending order to obtain the delay sequence and the delay mean of the delay sequence is calculated; the energy intensity distribution of each sub-band is extracted from the multimodal decomposition data; the wave speed correction equation is constructed based on the delay mean and the energy intensity distribution of each sub-band; the coupling coefficient of the wave speed correction equation is updated using the pre-trained LSTM network model and the actual propagation speed of the stress wave in the wind turbine blade is adjusted based on the updated wave speed correction equation; an attenuation compensation function is constructed and the sub-band energy attenuation in the multimodal decomposition data is corrected based on the function; the multimodal decomposition data is updated based on the adjusted actual propagation speed of the stress wave and the corrected sub-band energy; Ways to suppress environmental interference on propagation correction data include: Physical parameter sensors are arranged on each wind turbine blade to collect the temperature, stress and humidity data of the wind turbine blade; a multi-physical field coupling factor is constructed based on the temperature, stress and humidity data of the wind turbine blade; an adaptive band-stop filter is designed and used to suppress the high-frequency noise caused by temperature drift in the frequency domain dimension; a metasurface beamforming method is used to generate beam nulls in the interference direction; and a multi-objective genetic algorithm is used to dynamically adjust the weights of each physical parameter in the multi-physical field coupling factor and the stopband parameters of the adaptive band-stop filter.

5. The stress wave sensing-based measurement method for wind turbine blades according to claim 4 is characterized in that: The method of extracting features from the perfected stress wave signal data includes: The improved stress wave signal data is classified based on data types, including sub-band energy intensity data, dynamic propagation data, and time-frequency data; the sub-band energy intensity data and time-frequency data are integrated into a time-frequency feature data set, and the dynamic propagation data is used as a dynamic propagation feature data set; the time-frequency feature data set and the dynamic propagation feature data set from P sensors are stacked as three-dimensional tensors, and the time sampling points C and the feature dimensions D in any preset time window are cross-combined to obtain an initial feature tensor with a dimension of P×C×D in each time window; the initial feature tensor in each time window is collaboratively processed by time domain synchronous calibration and frequency domain phase alignment to obtain a calibrated feature tensor; the position of the damaged area of ​​the wind turbine blade is calculated based on the time-frequency feature data set; the Euclidean distance between the position of the damaged area of ​​the wind turbine blade and any sensor is calculated, and the signal-to-noise ratio index of each sensor is calculated at the same time, and a weight is assigned to each sensor based on the signal-to-noise ratio index of each sensor and the Euclidean distance between the sensor and the position of the damaged area of ​​the wind turbine blade, that is, the calibration feature tensor is weighted to obtain an intermediate feature matrix.

6. The stress wave sensing-based measurement method for wind turbine blades according to claim 5, characterized in that: The method of further screening the intermediate feature matrix includes: The local density of each feature point in the intermediate feature matrix is ​​calculated and all feature points in the intermediate feature matrix are clustered based on the local density of each feature point. The intermediate feature matrix is ​​divided into two sub-matrices, namely, a high-density matrix and a low-density matrix. An adaptive kernel function is constructed to map the high-density matrix and the low-density matrix to a high-dimensional space, respectively, to obtain a high-dimensional space feature matrix. Construct a classification model and use the random forest model as the basic structure of the classification model; collect historical wind turbine blade damage data and convert the data into a historical feature matrix; query the preset wind turbine blade parameter database and use the key damage features involved in the damage situation corresponding to the historical feature matrix as the training labels of the classification model; use the historical feature matrix to train the classification model until the function value of the loss function of the classification model no longer decreases, and obtain a trained classification model; use the classification model to screen key features from the high-dimensional space feature matrix and calculate the Gini index of each feature point in the high-dimensional space feature matrix; sort all feature points in descending order based on the size of the Gini index, and determine the feature points that are greater than the preset sorting threshold as key feature points; use the mutual information analysis algorithm to remove redundant feature points in the key feature points and output a damage-sensitive feature set; The damage sensitive feature set is standardized to obtain a standardized feature set; the standardized feature set is compressed by reducing the dimension using the principal component analysis algorithm to obtain stress wave signal characteristic data.

7. The stress wave sensing-based measurement method for wind turbine blades according to claim 6, characterized in that: The method of constructing the damage assessment model includes: Construct a damage coding system that uses a hierarchical coding rule; query a preset wind turbine blade parameter database to determine all specific damage conditions of the wind turbine blade and the corresponding features of the damage conditions; encode all specific damage conditions of the wind turbine blade based on the damage coding system, and combine the code of any specific damage condition with the corresponding feature of the damage condition into a feature-code mapping matrix; integrate all feature-code mapping matrices to obtain a feature-code mapping data set; A damage assessment model is constructed and the residual neural network model is used as the basic structure of the damage assessment model, including an input layer, a fully connected layer, a feature processing layer, a damage quantization layer and an output layer; the stress wave signal feature data is used as the input data of the damage assessment model; the dual-channel parallel processing architecture is used as the basic framework of the feature processing layer in the damage assessment model, including a time-frequency feature channel and a space propagation channel; the input data is jointly processed by the time-frequency feature channel and the space propagation channel to generate an intermediate feature tensor; the intermediate feature tensor enters the damage quantization layer to calculate the damage index, and the output layer outputs the damage index, damage location coordinates and damage type; the damage index, damage location coordinates and damage type are matched to generate a damage code; the existing damage assessment model is dynamically updated after each damage code is output; a real-time timestamp is added to all output damage codes, and they are sorted in chronological order to obtain a damage assessment record.

8. The stress wave sensing-based measurement method for wind turbine blades according to claim 7 is characterized in that: The method of dynamically updating the existing damage assessment model includes: An incremental learning architecture is constructed, a dynamic feature memory library is established after the feature processing layer of the damage assessment model, the intermediate feature tensor is backed up and the feature of the backed-up intermediate feature tensor is extracted to obtain a historical damage feature data set; a dual-model collaborative distillation mechanism is introduced into the incremental learning architecture, including a first model and a second model; the first model is used to freeze the feature processing layer of the damage assessment model, and the probability tensor of the historical damage feature data set is output; the second model copies the parameters of the frozen layer from the first model, and constructs an adjustment function based on the probability tensor to adjust the parameters of the unfrozen layer; the initial learning rate of the incremental learning architecture is set, and the initial learning rate is adjusted based on the cosine annealing algorithm; parameter constraints are set for the damage quantization layer to ensure that the parameter offset is less than the preset offset threshold; the performance indicators of the damage assessment model after adding the incremental learning architecture are calculated. If each performance indicator meets the expected value, the update of the damage assessment model is completed, otherwise the damage assessment model rollback is triggered.

9. The stress wave sensing-based measurement method for wind turbine blades according to claim 8 is characterized in that: The method of generating a three-dimensional damage atlas based on the damage assessment record includes: The damage location coordinates, damage index and damage type are extracted from the damage assessment records, and a damage spatial distribution matrix is ​​constructed; the damage spatial distribution matrix is ​​three-dimensionally gridded using the Kriskin space interpolation algorithm to obtain the global density field of wind turbine blade damage; the wind turbine blade is parametrically modeled based on the geometric parameters of the wind turbine blade to obtain a three-dimensional model of the wind turbine blade; the global density field of wind turbine blade damage is mapped to the three-dimensional model of the wind turbine blade to obtain a three-dimensional damage model; the three-dimensional damage model is rendered using color to obtain a damage rendering model; the damage rendering models corresponding to each timestamp are sorted in chronological order to obtain a three-dimensional damage map.

10. A measurement system based on stress wave perception applied to a wind turbine blade, used to implement a measurement method based on stress wave perception applied to a wind turbine blade according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to collect stress wave signal data and pre-process the stress wave signal data to obtain perfect stress wave signal data; A feature extraction module is used to extract features from the perfected stress wave signal data and output stress wave signal feature data; The damage assessment module builds a damage assessment model based on the stress wave signal characteristic data, uses the damage assessment model to conduct damage assessment on the wind turbine blade, and obtains a damage assessment record; The visualization generation module generates a three-dimensional damage map based on the damage assessment record, and sends the three-dimensional damage map to a preset wind power equipment safety terminal; each module is connected by wire and / or wireless means.

Images