Damage spatial identification and positioning technology based on lamb wave in strong noise environment

Abstract

This invention discloses a damage spatial identification and localization technology based on lamb wave in a strong noise environment, comprising the following steps: (1) Signal acquisition; piezoelectric ceramic sensor arrays are uniformly arranged on the carbon fiber composite laminate to be tested and the standard carbon fiber composite laminate, respectively, with the number of sensors being M. The first to the (M-1)th sensors are used as excitation sensors to generate lamb waves, and the other sensors are used as receiving sensors to acquire the lamb wave response signals; (2) Signal denoising and reconstruction; (3) Damage horizontal localization; the horizontal localization of the damage is performed using a tomographic imaging algorithm to determine the two-dimensional coordinates of the damage; (4) Damage depth localization; the layer number of the damage is determined using a convolutional neural network to determine the depth of the damage; (5) Visualization of the spatial location of the damage. The method of this invention is simple and can intuitively display the location of the damage in three-dimensional space in a strong noise environment.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing of carbon fiber composite laminates, and relates to a damage spatial identification and localization technology based on lamb wave in a strong noise environment. Background Technology

[0002] Carbon fiber composite laminates are widely used in aerospace, public transportation, and shipbuilding due to their advantages such as high strength, high stiffness, low density, and corrosion resistance. However, during manufacturing and actual use, carbon fiber composite laminates are inevitably subjected to external forces, leading to damage such as delamination, fiber breakage, and matrix cracking. Delamination damage is one of the most dangerous types of damage to carbon fiber composite laminates, severely affecting their stiffness and strength. Since delamination damage often occurs within the carbon fiber composite laminate, it is difficult to detect visually. Therefore, accurately locating the delamination damage is crucial to prevent further degradation of the structural performance and serious accidents.

[0003] With the development of new sensing technologies, composite material structure health monitoring technology based on lambda wave technology has been widely used in damage detection due to its advantages such as good long-distance propagation capability, low energy attenuation, and suitability for large-area detection. The basic principle of lambda wave damage detection is as follows: a lambda wave excitation signal of a certain frequency is sent to the structure under test. Damage in the structure will change the signal propagation and be reflected in the received response signal. Damage is identified by extracting the change features caused by the damage contained in the response signal. To visually present the location of damage, researchers have studied damage imaging methods based on lambda waves. Lambda wave-based imaging methods such as time delay and time reversal typically require calculating information such as wave packet arrival time, wave velocity, and sensor distance in the response signal. Due to the multimodal and dispersion characteristics of lambda waves, their propagation velocity is not constant, increasing the difficulty of damage localization using these methods. To overcome the problem of inaccurate damage localization caused by factors such as wave packet arrival time and wave velocity, researchers have proposed a lamb wave-based tomographic imaging technique. This technique identifies and locates damage by calculating the difference between the signal before and after damage, without needing to calculate parameters such as wave velocity and wave arrival time in the signal, and has great application prospects.

[0004] Patent CN114910562A proposes a damage localization imaging method based on lamb wave signal spectrum, achieving damage localization on a two-dimensional plane in carbon fiber composite laminates. Patent CN111521691A proposes a composite material lamb wave damage imaging method based on time-reversed weighted distribution, achieving effective identification of multi-source damage in composite structural plates; however, this method also focuses on damage localization on a two-dimensional plane. Patent CN110376282A proposes a lamb wave damage localization method based on elliptic probability and Bayesian estimation. It uses the elliptic trajectory method and probabilistic damage reconstruction method, fusing arrival time characteristic values ​​and correlation coefficient characteristic values ​​through Bayesian estimation to image and localize damage; however, this method also achieves damage imaging on a two-dimensional plane. The literature (Impact damage assessment in orthotropic CFRP laminates using nonlinear Lamb wave: Experimental and numerical investigations. CompositeStructures, 2020, 236.) proposes using nonlinear lamb waves to assess damage in carbon fiber composite plates, and using the presence of second harmonics in the nonlinear lamb wave to determine whether damage exists in the carbon fiber composite plate. The paper (Impact Damage Detection in Patch-Repaired CFRP Laminates Using Nonlinear Lamb Waves. Sensors (Basel), 2020.21(1).) proposes using nonlinear lamb waves to detect impact damage in carbon fiber laminates, which also involves using the second harmonic of the nonlinear lamb wave to determine whether damage exists. The paper (Locating and Imaging FiberBreaks in CFRP Using Guided Wave Tomography and Eddy Current Testing. Sensors (Basel), 2022.22(19).) proposes a detection method combining tomographic imaging algorithm and eddy current detection technology for locating and evaluating fiber fractures in carbon fiber composite structures, but this paper also achieves damage imaging and localization on a two-dimensional plane.

[0005] In summary, existing research on damage detection of carbon fiber composite laminates based on lambda wave technology mainly focuses on "whether damage exists" and "two-dimensional planar location of damage," with almost no research on damage depth localization. The fundamental reason is that lambda wave excitation requires high precision, and while the thickness of carbon fiber composite laminates is typically in the millimeter range, their length and width can be in the meter range. While horizontal damage localization is relatively easier, determining damage depth within the millimeter range is much more difficult. Currently, methods that can effectively identify damage depth include B-scan and phased array methods, but these methods require specialized detection equipment. Due to the anisotropy of carbon fiber composite laminates and their thin thickness, the lambda wave damage signals acquired at different damage depths contain various nonlinear influencing factors, making it difficult to locate the damage depth using conventional characterization methods. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art and provide a damage spatial identification and localization technology based on lamb waves in a strong noise environment.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A damage spatial identification and localization technology based on lambda waves in a high-noise environment includes the following steps:

[0009] (1) Signal acquisition;

[0010] Piezoelectric ceramic sensor arrays are uniformly arranged on the carbon fiber composite laminate to be tested and the standard carbon fiber composite laminate, respectively. The number of sensors is M. The first to the (M-1)th sensors are used as excitation sensors to generate lamb waves, and the other sensors are used as receiving sensors to collect lamb wave response signals.

[0011] The only difference between the standard carbon fiber composite laminate and the carbon fiber composite laminate under test is that the standard carbon fiber composite laminate does not contain any damage; the number, arrangement, and numbering of sensors on the standard carbon fiber composite laminate and the carbon fiber composite laminate under test are the same.

[0012] (2) Signal denoising and reconstruction;

[0013] Denoising and reconstruction are performed on each lambda wave response signal separately;

[0014] (3) Location of the level of damage;

[0015] The denoised and reconstructed signal of the carbon fiber composite plate under test is input into the tomographic imaging algorithm to perform damage imaging, and the two-dimensional coordinate position of the damage in the coordinate system is determined. The origin of the coordinate system is a vertex of the carbon fiber composite plate under test, the X-axis of the coordinate system is parallel to a horizontal side of the carbon fiber composite plate under test, and the Y-axis of the coordinate system is parallel to a vertical side of the carbon fiber composite plate under test.

[0016] (4) Damage depth localization;

[0017] (4.1) Point O is the point on the upper surface of the carbon fiber composite laminate to be tested that coincides with the orthographic projection of a certain damage center. Two sensors connected by a line passing through point O are arranged on the upper surface of the carbon fiber composite laminate to be tested. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted.

[0018] (4.2) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and then input into a trained convolutional neural network, which outputs a damage depth prediction label.

[0019] The training steps for a convolutional neural network are as follows:

[0020] (i) Construct damage of different depths in a standard carbon fiber composite laminate. The point on the upper surface of the standard carbon fiber composite laminate that coincides with the orthographic projection of a certain damage center is denoted as point O. Two sensors connected by a line passing through point O are arranged on the upper surface of the standard carbon fiber composite laminate. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted.

[0021] (ii) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and Gaussian white noise of different degrees is added to the two-dimensional time-frequency image to increase the number of samples;

[0022] (iii) Label the two-dimensional time-frequency images with different damage depths with the true damage depth to obtain the sample dataset;

[0023] (iv) Divide the sample dataset into a training set and a test set. The training set is used for training the convolutional neural network model, and the test set is used for testing the convolutional neural network model.

[0024] (v) Input the training set into the convolutional neural network model for feature extraction and classification until the error of the damage function of the convolutional neural network model is less than the set target value or the maximum number of training iterations is reached.

[0025] (5) Visualization of the spatial location of the damage;

[0026] Using Matlab software as a platform, the horizontal and depth information of the damage is input into the plotting function to draw the spatial location of the damage.

[0027] As a preferred technical solution:

[0028] In the above-mentioned damage spatial identification and localization technology based on lamb waves in a strong noise environment, in step (1), the piezoelectric ceramic sensor array is a circular piezoelectric ceramic sensor array.

[0029] As described above, in a high-noise environment, a damage spatial identification and localization technology based on lamb waves is used. In step (1), the excitation sensor is connected to an arbitrary waveform generator through a high-voltage power amplifier. The arbitrary waveform generator amplifies the specified excitation signal through the high-voltage power amplifier and loads it onto the piezoelectric ceramic sensor to emit a lamb wave signal. The receiving sensor is connected to a multi-channel oscilloscope to collect the lamb wave response signal.

[0030] The damage spatial identification and localization technology based on lambda waves in a strong noise environment, as described above, has the following specific steps in step (2):

[0031] (2.1) Using envelope entropy as the fitness function, the improved gray wolf optimization algorithm based on dimensional learning hunting search strategy is used to optimize the mode number N and penalty term α of the variational mode decomposition method;

[0032] (2.2) Select the variational mode decomposition method with optimal parameters to perform mode decomposition on the signal. The formula used for mode decomposition is as follows:

[0033] imf ik =VMD(x i (t)), k=1,...,N;

[0034] res = x i (t)-imf ik k = 1, ..., N;

[0035] In the formula, imf ik It is the k-th mode component obtained after mode decomposition; res is the original signal minus all the decomposed mode components; x i (t) is the signal to be decomposed into modes; N is the number of modal components;

[0036] (2.3) Calculate the correlation coefficient between each modal component and the original signal, using the following formula:

[0037]

[0038] In the formula, P ik For correlation coefficient; imf ik It is the k-th modal component obtained after modal decomposition; x i(t) is the signal to be decomposed into modes; Cov(imf) ik ,x i (t) is imf ik and x i The covariance of (t); D(imf ik ) is IMF ik The variance of D(x); i (t) is x i The variance of (t);

[0039] (2.4) Wavelet threshold denoising and reconstruction, the formula is as follows:

[0040] reimf ip =wdenoise(oimf ip ,lev,wavelet,method,thresholdrelu,noiseestimate);

[0041] reres=wdenoise(res,lev,wavelet,method,thresholdrelu,noiseestimate);

[0042] rex i (t)=reimf ip +reres;

[0043] In the formula, reimf ip for oimf ip The result after wavelet thresholding and denoising of the components; oimf ip The modal component with the highest correlation coefficient; reres is the result of res after wavelet thresholding and denoising reconstruction; rex i (t) represents the denoised and reconstructed signal; lev represents the wavelet decomposition level; wavelet represents the wavelet basis function; method represents the denoising method; thresholdrelu represents the threshold rule; and noiseestimate represents the noise estimate.

[0044] The damage spatial identification and localization technology based on lambda waves in a strong noise environment, as described above, has the following specific steps in step (2.1):

[0045] (a) Set initialization parameters;

[0046] Let G = 30, I = 15, α = [50, 15000], N = [2, 8], where G represents the population size and I represents the maximum number of iterations;

[0047] (b) The wolf pack positions are initialized using the following formula, where the position of the gray wolf represents the number of modes N and the penalty term α of the variational mode decomposition method;

[0048] X ij =l j +rand j [0,1)×(u j -l j ), i∈[1,G],j∈[1,D];

[0049] In the formula, X ij This represents the position of the i-th wolf in dimension j; l j rand represents the lower bound values ​​of α and N. j [0,1) represents generating data within the interval [0,1) in dimension j; u j α represents the upper limit of N; D represents the dimension of the problem. Since the parameters to be optimized in variational mode decomposition are α and N, D is set to 2.

[0050] The i-th wolf uses vector X at its position in the t-th iteration. i (t)={x i1 ,x i2 ,...,x iD} indicates that the entire wolf pack is stored in a matrix Pop;

[0051] (c) Search for a solution;

[0052] In the dimensionality-based hunting search strategy, besides updating X from the matrix Pop, i A candidate position X of (t) i-GWO In addition to (t+1), X is also generated using the following formula. i Another candidate position X of (t) i-DLH (t+1):

[0053] X i-DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t));

[0054] In the formula, X i-DLH,d (t+1) represents candidate X i-DLH The position of (t+1) in the d-th dimension, where d = 1 or 2; X i,d (t) represents the position of the i-th wolf in the d-th dimension; X r,d (t) represents the position of a wolf selected from matrix Pop (i.e., the wolf pack initialization position formula in (b), the entire wolf pack position) in the d-th dimension; X n,d(t) represents the position of a wolf selected from the vicinity of the i-th wolf in the d-th dimension; rand×(X n,d (t)-X r,d (t) represents randomly selecting X from the vicinity of the i-th wolf. n,d (t) and randomly select X from the entire wolf pack r,d (t), and simultaneously perform the difference calculation;

[0055] (d) Iterative solution;

[0056] The following formula compares two candidate positions X. i-DLH (t+1) and X i-GWO The fitness value at (t+1) is used to select the best candidate position X. i (t+1);

[0057]

[0058] In the formula, f represents the envelope entropy fitness function;

[0059] (e) Obtain the optimal parameters;

[0060] Determine X i Is the fitness of (t+1) less than X? i The fitness of (t) is determined; if so, X is set to... i (t) is replaced with X i (t+1); conversely, X i (t) remains unchanged;

[0061] After performing steps (b) to (e) on all wolves, the algorithm iteration count I is increased by 1, and the process returns to step (c). Steps (c) to (e) are then executed repeatedly until the preset maximum number of iterations is reached.

[0062] When lambda waves are used for nondestructive testing of structures, the signal inevitably contains noise. Researchers have conducted extensive studies on extracting effective information from the original signal. Commonly used denoising algorithms include wavelet transform, empirical mode decomposition (EMD), ensemble EMD, and variational mode decomposition (VMD). Wavelet transform denoising is affected by the choice of wavelet function. Furthermore, EMD and ensemble EMD exhibit significant mode aliasing during signal decomposition. VMD can effectively avoid mode aliasing and has good denoising performance. However, VMD's application in damage detection of carbon fiber composite laminates requires further investigation. Additionally, the effectiveness of VMD is greatly influenced by parameter values, necessitating optimization of the VMD parameters.

[0063] This invention utilizes an improved gray wolf optimization algorithm based on a dimensional learning-based hunting search strategy to optimize the number of modes and penalty terms in variational mode decomposition methods. Existing technologies (such as patent CN114861697A) directly use the gray wolf optimization algorithm to optimize the number of modes and penalty terms in variational mode decomposition methods. While the gray wolf optimization algorithm alone can easily obtain local optima, it is difficult to obtain global optima. Compared with the gray wolf optimization algorithm in existing technologies, this invention adds a hunting search strategy based on dimensional learning (i.e., steps (c) and (d)). The hunting search strategy based on dimensional learning can effectively help the gray wolf optimization algorithm find the global optima.

[0064] In this invention, the envelope entropy function is used as the fitness function. The smaller the calculated fitness value, the better the optimization effect. Figure 16 As shown, by comparing the VMD optimization results of the Improved Gray Wolf Optimization (IGWO), Gray Wolf Optimization (GWO), Whale Optimization Algorithm (WOA), and Particle Swarm Optimization (PSO) algorithms, it can be found that the IGWO algorithm (i.e., the improved Gray Wolf Optimization algorithm based on the dimensional learning hunting search strategy of this invention) can obtain the minimum fitness value. Therefore, it can be concluded that the dimensional learning-based hunting search strategy can effectively help the Gray Wolf Optimization algorithm find the global optimum.

[0065] In the above-mentioned damage spatial identification and localization technology based on lamb waves in a strong noise environment, in step (2.4), the wavelet threshold adopts a hard threshold processing method.

[0066] The damage spatial identification and localization technology based on lamb waves in a strong noise environment, as described above, has the following specific steps in step (3):

[0067] (3.1) Calculate the SDC of the path from excitation sensor i to receiver sensor j on the carbon fiber composite plate under test. ij The value, the path from excitation sensor i to receiver sensor j, is the connection between the i-th excitation sensor and the j-th receiver sensor, and the formula is as follows:

[0068]

[0069] In the formula, Rf ij When the i-th excitation sensor on the standard carbon fiber composite laminate is working, the lambda wave response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the health signal; Y ijWhen the i-th excitation sensor on the carbon fiber composite plate under test is working, the lambda response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the damage signal; Coν(Rf) ij ,Y ij ) is Rf ij and Y ij covariance; σ Rfij For Rf ij variance; σ Yij For Y ij The variance;

[0070] (3.2) Calculate the damage probability S of the elliptical region on the carbon fiber composite plate under test with excitation sensor i and receiving sensor j as foci. ij (x, y), the formula is as follows:

[0071]

[0072]

[0073] In the formula, β is the area of ​​the elliptical region; (x,y) is the coordinate of any point within the elliptical region in the coordinate system described; (x ik ,y ik (x) represents the coordinates of the excitation sensor i on the carbon fiber composite plate under test in the coordinate system; (x) represents the coordinates of the sensor i on the carbon fiber composite plate under test in the coordinate system. jk ,y jk () represents the coordinates of sensor j on the carbon fiber composite plate to be tested in the coordinate system.

[0074] (3.3) Calculate the value of any point (x) within the detection area composed of all the elliptical regions on the carbon fiber composite plate to be tested. p ,y p The damage probability distribution P(x) p ,y p The formula is as follows:

[0075]

[0076] The two-dimensional coordinate position of the damage in the coordinate system is determined based on the damage probability distribution.

[0077] In the above-described damage spatial identification and localization technology based on lamb waves in a strong noise environment, in steps (4.1) and (i), the distance between the two sensors connected by lines passing through point O and point O is the same.

[0078] In the above-described damage spatial identification and localization technology based on lamb waves in a strong noise environment, in steps (4.1) and (i), the distance between the two sensors connected by lines passing through point O and point O is 150mm.

[0079] The principle of this invention is as follows:

[0080] To address the problem of difficulty in locating damage depth using lambda wave technology in existing technologies, this invention utilizes the powerful nonlinear learning capabilities of deep learning to deeply extract features of damage at different depths, thereby achieving damage depth location based on lambda wave technology.

[0081] Existing technologies include two-dimensional planar damage localization using lambda waves and deep learning. This requires deploying numerous sensors to collect damage signals along different paths and using geometric topology methods to locate the damage. However, directly using these technologies to detect damage depth presents the following problems:

[0082] ① The two-dimensional planar location of the damage must be known in order to determine the depth of the damage, but in reality, the location of the damage is unknown;

[0083] ② Taking a sensor ring array as an example, damage signals from different paths were collected for damage at different depths and input into a convolutional neural network for damage depth localization. However, the results were not ideal because the collected signals contained reflections from the damage and reflections from the plate boundary, which included too much variable information. In addition, the phases of the signals collected from different paths were also different, adding another variable. Therefore, the experimental results were not ideal.

[0084] This invention first uses a tomographic imaging algorithm to locate the two-dimensional planar position of unknown damage. Based on the obtained two-dimensional coordinates of the damage, the damage signal on the path is extracted using the method in the damage depth localization step of this invention. This signal is then input into a trained damage depth localization model based on a convolutional neural network (i.e., a convolutional neural network) to locate the damage depth, effectively overcoming the aforementioned problem ①. In addition, this invention extracts the damage scattering signal (which is a segment of the lambda wave response signal) from the damage signal as training material for the damage depth localization model, avoiding the influence of redundant information and achieving accurate damage localization, effectively overcoming the aforementioned problem ②.

[0085] Specifically, regarding problem ①, the sensor array required by the tomographic imaging algorithm used in this invention can be adjusted according to the shape of the carbon fiber composite laminate under test, so that the detection range of the sensor array covers the test area of ​​the carbon fiber composite laminate as much as possible. Therefore, the tomographic imaging algorithm can basically achieve damage localization on a two-dimensional plane.

[0086] Regarding problem ②, the lambda wave response signal acquired by the sensor often contains scattered waves caused by damage and reflected waves caused by the boundary of the carbon fiber composite laminate under test, which contains complex signal information. Furthermore, the lambda wave response signals acquired along different paths have different phase differences, resulting in significant differences in the signal content acquired within the same time period. Due to these reasons, directly using deep learning to achieve damage depth identification is challenging. Therefore, this invention achieves damage depth localization through the following steps:

[0087] (1) Construct the dataset required for model training; prepare a standard carbon fiber composite plate and process layered damage of different depths at different positions on the standard carbon fiber composite plate; with each damage as the center, place sensors above and below it, and collect the lamb wave response signals on the corresponding path using a one-to-one transmission and one-to-reception method; extract the damage scattering signal of a specified length from the collected lamb response signals (the scattering wave signals corresponding to damages of different depths have obvious differences); use continuous wavelet transform to convert the extracted damage scattering signals into two-dimensional time-frequency images for model training; since deep learning training requires sufficient data, this invention adds different degrees of Gaussian white noise to the two-dimensional time-frequency images to increase the number of samples; input the dataset into a convolutional neural network to train the damage depth localization model and obtain the trained damage depth localization model;

[0088] (2) Predict the actual depth of the damage; Based on the determination of the two-dimensional coordinate position of the damage in the coordinate system using the tomographic imaging algorithm, two sensors are arranged above and below each damage as the center, and the lamb wave response signal of each damage is collected in a one-to-one manner; the damage scattering wave signal is extracted from the lamb wave response signal of each damage, and a two-dimensional time-frequency image is generated using continuous wavelet transform; the generated two-dimensional time-frequency image is input into the trained damage depth localization model to obtain the depth of each damage.

[0089] This invention innovatively proposes using the damage scattering signal in the lamb wave response signal as the input of the model. Experiments have verified that there are significant differences in the scattering signals corresponding to damage at different depths, which is more conducive to model training and improving the model's localization accuracy.

[0090] Beneficial effects:

[0091] (1) Compared with most metal plates, carbon fiber composite laminates are anisotropic and have a more complex structure, which increases the difficulty of research. Expensive instruments, such as thermal imagers and phased array imaging systems, are professional damage detection instruments, which are generally expensive and difficult to use on a large scale. The present invention provides a damage spatial identification and localization technology based on lamb wave in a strong noise environment. It can easily identify damage depth by using inexpensive piezoelectric ceramic sensors combined with convolutional neural network models. At the same time, it provides a new method for damage depth identification of carbon fiber composite laminates, without the need for expensive and complex sensors for three-dimensional damage visualization, effectively reducing the cost of damage detection.

[0092] (2) In this invention, tomographic imaging and convolutional neural networks are used to determine the horizontal position and depth of the damage. Combined with Matlab software, the spatial position of the damage is visualized, which can present the location of the damage more intuitively.

[0093] (3) This invention provides a new method for spatial identification and localization of damage based on a combination of lamb wave tomography algorithm and convolutional neural network, providing a new method for spatial identification, localization and visualization of damage in carbon fiber composite laminates;

[0094] (4) This invention can solve the problem of damage spatial identification and localization under strong noise. The original signal collected is input into the denoising algorithm, and the algorithm will automatically output the denoised reconstructed signal with the highest signal-to-noise ratio without the need for manual adjustment of the corresponding parameters.

[0095] (5) The present invention can realize multi-damage localization, that is, the present invention can locate multiple damages at different locations on a two-dimensional plane. As for locating multiple damages at the same location on a two-dimensional plane at different depths, further research is needed. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the piezoelectric ceramic sensor arrangement during two-dimensional damage localization in this invention;

[0097] Figure 2 This is a schematic diagram of the two-dimensional planar distribution of different layered damage in this invention;

[0098] Figure 3 This is a schematic diagram showing the distribution of different layered damage depth directions in this invention; in the figure, (a) to (g) are schematic diagrams showing the positions of damage 1# to 7# in the carbon fiber composite plate to be tested, respectively.

[0099] Figure 4 This is a comparison diagram of signals before and after damage to the S1-S5 path in this invention;

[0100] Figure 5This is a flowchart of the spatial localization of damage to carbon fiber composite laminates under strong noise in this invention; in the figure, (a) to (f) are respectively signal acquisition, noise loading, signal denoising, two-dimensional plane localization of damage, localization of damage depth direction, and visualization of damage spatial localization.

[0101] Figure 6 The figure shows the SNR curves of the S1-S5 path damage signal reconstructed based on variational mode decomposition and wavelet threshold denoising in this invention. In the figure, (a) is the signal SNR curve reconstructed using the hard threshold method, and (b) is the signal SNR curve reconstructed using the soft threshold method. It can be seen from the figure that the hard threshold processing effect is better.

[0102] Figure 7 The figure shows the SNR curves of the S1-S5 path health signal reconstructed based on variational mode decomposition and wavelet threshold denoising in this invention. In the figure, (a) is the signal SNR curve reconstructed using the hard threshold method, and (b) is the signal SNR curve reconstructed using the soft threshold method. It can be seen from the figure that the hard threshold processing effect is better.

[0103] Figure 8 This is a comparison diagram of the reconstruction of the S1-S5 path damage signal based on variational mode decomposition and wavelet threshold denoising in this invention;

[0104] Figure 9 This is a comparison diagram of the reconstruction of the S1-S5 path health signal based on variational mode decomposition and wavelet threshold denoising in this invention;

[0105] Figure 10 This is a schematic diagram showing the sensor arrangement when acquiring the response signal of damage at a certain depth in this invention.

[0106] Figure 11 This is a time-domain signal diagram obtained from damage at different depths collected by the sensor in this invention;

[0107] Figure 12 This is a diagram of damage scattering signals extracted from damage time-domain signals at different depths collected by the sensor in this invention.

[0108] Figure 13 These are some time-frequency images of damage used in the training dataset of the damage depth model in this invention; from left to right, each layer in the figure represents a time-frequency image of the damage at that layer without noise, with noise added at a strength of 0.04, with noise added at a strength of 0.06, and with noise added at a strength of 0.10; adding different strengths of noise to different time-frequency images of damage increases the amount of data used to train the convolutional neural network;

[0109] Figure 14The figure shows the visualization of the damage depth model training result based on convolutional neural network in this invention. In the figure, (a) is the visualization of the result after one iteration, and (b) is the visualization of the result after ten iterations. The horizontal axis represents the distance, and the vertical axis represents the similarity.

[0110] Figure 15 This is a three-dimensional visualization of the spatial localization of damage #1 after convolution training in this invention. In the figure, layers 1-2 represent the damage located between the first and second layers of the carbon fiber composite plate under test, layers 2-3 represent the damage located between the second and third layers of the carbon fiber composite plate under test, and layers 5-6 represent the damage located between the fifth and sixth layers of the carbon fiber composite plate under test.

[0111] Figure 16 A comparison chart of VMD parameter optimization results for different optimization algorithms is shown. As can be seen from the chart, the IGWO algorithm (i.e., the improved gray wolf optimization algorithm based on dimensional learning hunting search strategy of this invention) can obtain the minimum fitness value. Detailed Implementation

[0112] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0113] A damage spatial identification and localization technology based on lambda waves in a high-noise environment includes the following steps:

[0114] (1) Signal acquisition;

[0115] A circular piezoelectric ceramic sensor array is uniformly arranged on the carbon fiber composite laminate under test and the standard carbon fiber composite laminate, respectively. The number of sensors is M. The first to the (M-1)th sensors are used as excitation sensors to generate lamb waves, and the other sensors are used as receiving sensors to collect lamb wave response signals.

[0116] The excitation sensor is connected to an arbitrary waveform generator via a high-voltage power amplifier. The arbitrary waveform generator amplifies the specified excitation signal through the high-voltage power amplifier and loads it onto the piezoelectric ceramic sensor to emit a lamb wave signal. The receiving sensor is connected to a multi-channel oscilloscope to acquire the lamb wave response signal.

[0117] The only difference between the standard carbon fiber composite laminate and the carbon fiber composite laminate under test is that the standard carbon fiber composite laminate is free of damage; the number, arrangement and numbering of sensors on the standard carbon fiber composite laminate and the carbon fiber composite laminate under test are the same.

[0118] (2) Signal denoising and reconstruction;

[0119] Denoising and reconstruction are performed on each lambda wave response signal; the specific steps are as follows:

[0120] (2.1) Using envelope entropy as the fitness function, the improved gray wolf optimization algorithm based on dimensionality learning hunting search strategy is used to optimize the mode number N and penalty term α of the variational mode decomposition method; the specific steps are as follows:

[0121] (a) Set initialization parameters;

[0122] Let G = 30, I = 15, α = [50, 15000], N = [2, 8], where G represents the population size and I represents the maximum number of iterations;

[0123] (b) The wolf pack positions are initialized using the following formula, where the position of the gray wolf represents the number of modes N and the penalty term α of the variational mode decomposition method;

[0124] X ij =l j +rand j [0,1)×(u j -l j ), i∈[1,G],j∈[1,D];

[0125] In the formula, X ij This represents the position of the i-th wolf in dimension j; l j rand represents the lower bound values ​​of α and N. j [0,1) represents generating data within the interval [0,1) in dimension j; u j α represents the upper limit of N; D represents the dimension of the problem. Since the parameters to be optimized in variational mode decomposition are α and N, D is set to 2.

[0126] The i-th wolf uses vector X at its position in the t-th iteration. i (t)={x i1 ,x i2 ,...,x iD} indicates that the entire wolf pack is stored in a matrix Pop;

[0127] (c) Search for a solution;

[0128] In the dimensionality-based hunting search strategy, besides updating X from the matrix Pop, i A candidate position X of (t) i-GWO In addition to (t+1), X is also generated using the following formula. i Another candidate position X of (t) i-DLH (t+1):

[0129] X i-DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t));

[0130] In the formula, X i-DLH,d (t+1) represents candidate X i-DLH The position of (t+1) in the d-th dimension, where d = 1 or 2; X i,d (t) represents the position of the i-th wolf in the d-th dimension; X r,d (t) represents the position of a wolf selected from matrix Pop (i.e., the wolf pack initialization position formula in (b), the entire wolf pack position) in the d-th dimension; X n,d (t) represents the position of a wolf selected from the vicinity of the i-th wolf in the d-th dimension; rand×(X n,d (t)-X r,d (t) represents randomly selecting X from the vicinity of the i-th wolf. n,d (t) and randomly select X from the entire wolf pack r,d (t), and simultaneously perform the difference calculation;

[0131] (d) Iterative solution;

[0132] The following formula compares two candidate positions X. i-DLH (t+1) and X i-GWO The fitness value at (t+1) is used to select the best candidate position X. i (t+1);

[0133]

[0134] In the formula, f represents the envelope entropy fitness function;

[0135] (e) Obtain the optimal parameters;

[0136] Determine X i Is the fitness of (t+1) less than X? i The fitness of (t) is determined; if so, X is set to... i (t) is replaced with X i (t+1); conversely, X i (t) remains unchanged;

[0137] After performing steps (b) to (e) on all wolves, the algorithm iteration count I is increased by 1, and the process returns to step (c). Steps (c) to (e) are then executed repeatedly until the preset maximum number of iterations is reached.

[0138] (2.2) Select the variational mode decomposition method with optimal parameters to perform mode decomposition on the signal. The formula used for mode decomposition is as follows:

[0139] imf ik =VMD(x i (t)), k=1,...,N;

[0140] res = x i (t)-imf ik k = 1, ..., N;

[0141] In the formula, imf ik It is the k-th mode component obtained after mode decomposition; res is the original signal minus all the decomposed mode components; x i (t) is the signal to be decomposed into modes; N is the number of modal components;

[0142] (2.3) Calculate the correlation coefficient between each modal component and the original signal, using the following formula:

[0143]

[0144] In the formula, P ik For correlation coefficient; imf ik It is the k-th modal component obtained after modal decomposition; x i (t) is the signal to be decomposed into modes; Cov(imf) ik ,x i (t) is imf ik and x i The covariance of (t); D(imf ik ) is IMF ik The variance of D(x); i (t) is x i The variance of (t);

[0145] (2.4) Wavelet threshold denoising and reconstruction, the formula is as follows:

[0146] reimf ip =wdenoise(oimf ip ,lev,wavelet,method,thresholdrelu,noiseestimate);

[0147] reres=wdenoise(res,lev,wavelet,method,thresholdrelu,noiseestimate);

[0148] rex i (t)=reimfip +reres;

[0149] In the formula, reimf ip for oimf ip The result after the component is denoised and reconstructed using wavelet thresholding (hard thresholding is used in this invention); oimf ip The modal component with the highest correlation coefficient; reres is the result of res after denoising and reconstruction using wavelet thresholding (hard thresholding is used in this invention); rex i (t) represents the denoised and reconstructed signal; lev represents the wavelet decomposition level; wavelet represents the wavelet basis function; method represents the denoising method; thresholdrelu represents the threshold rule; noiseestimate represents the noise estimate;

[0150] (3) Location of the level of damage;

[0151] The denoised and reconstructed signal of the carbon fiber composite laminate under test is input into a tomographic imaging algorithm for damage imaging. The two-dimensional coordinate position of the damage is determined in a coordinate system. The origin of the coordinate system is a vertex of the carbon fiber composite laminate under test. The X-axis of the coordinate system is parallel to a horizontal edge of the carbon fiber composite laminate under test, and the Y-axis is parallel to a vertical edge of the carbon fiber composite laminate under test. The specific steps are as follows:

[0152] (3.1) Calculate the SDC of the path from excitation sensor i to receiver sensor j on the carbon fiber composite plate under test. ij The value, the path from excitation sensor i to receiver sensor j, is the connection between the i-th excitation sensor and the j-th receiver sensor, and the formula is as follows:

[0153]

[0154] In the formula, Rf ij When the i-th excitation sensor on the standard carbon fiber composite laminate is working, the lambda wave response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the health signal; Y ij When the i-th excitation sensor on the carbon fiber composite plate under test is working, the lambda response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the damage signal; Coν(Rf) ij ,Y ij ) is Rf ij and Y ij covariance; For Rf ij The variance; For Y ij The variance;

[0155] (3.2) Calculate the damage probability S of the elliptical region on the carbon fiber composite plate under test with excitation sensor i and receiving sensor j as foci. ij (x, y), the formula is as follows:

[0156]

[0157]

[0158] In the formula, β is the area of ​​the elliptical region; (x,y) is the coordinate of any point within the elliptical region in the coordinate system described; (x ik ,y ik (x) represents the coordinates of the excitation sensor i on the carbon fiber composite plate under test in the coordinate system; (x) represents the coordinates of the sensor i on the carbon fiber composite plate under test in the coordinate system. jk ,y jk () represents the coordinates of sensor j on the carbon fiber composite plate to be tested in the coordinate system.

[0159] (3.3) Calculate the value of any point (x) within the detection area composed of all the elliptical regions on the carbon fiber composite plate to be tested. p ,y p The damage probability distribution P(x) p ,y p The formula is as follows:

[0160]

[0161] The two-dimensional coordinate position of the damage in the coordinate system is determined based on the damage probability distribution.

[0162] The damage probability distribution will eventually be presented in the form of an image, for example... Figure 5 The image displayed in step 4 shows the horizontal coordinate of the carbon fiber composite plate along the X-axis and the vertical coordinate of the carbon fiber composite plate along the Y-axis. The brightest red area in the image is the predicted two-dimensional plane area of ​​the damage. The red area can be displayed separately by thresholding. The center coordinate of the red area is determined according to the horizontal and vertical coordinates, which gives the two-dimensional coordinate position of the damage in the coordinate system.

[0163] (4) Damage depth localization;

[0164] (4.1) Point O is the point on the upper surface of the carbon fiber composite laminate to be tested that coincides with the orthographic projection of a certain damage center. Two sensors connected by a line passing through point O are arranged on the upper surface of the carbon fiber composite laminate to be tested. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted.

[0165] (4.2) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and then input into a trained convolutional neural network, which outputs a damage depth prediction label.

[0166] The training steps for a convolutional neural network are as follows:

[0167] (i) Construct damage of different depths in a standard carbon fiber composite laminate. The point on the upper surface of the standard carbon fiber composite laminate that coincides with the orthographic projection of a certain damage center is denoted as point O. Two sensors connected by a line passing through point O are arranged on the upper surface of the standard carbon fiber composite laminate. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted.

[0168] (ii) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and Gaussian white noise of different degrees is added to the two-dimensional time-frequency image to increase the number of samples;

[0169] (iii) Label the two-dimensional time-frequency images with different damage depths with the true damage depth to obtain the sample dataset;

[0170] (iv) Divide the sample dataset into a training set and a test set. The training set is used for training the convolutional neural network model, and the test set is used for testing the convolutional neural network model.

[0171] (v) Input the training set into the convolutional neural network model for feature extraction and classification until the error of the damage function of the convolutional neural network model is less than the set target value or the maximum number of training iterations is reached.

[0172] (5) Visualization of the spatial location of the damage;

[0173] Using Matlab software as a platform, the horizontal and depth information of the damage is input into the plotting function to draw the spatial location of the damage.

[0174] Example 1

[0175] A damage spatial identification and localization technology based on lambda waves in a high-noise environment, such as Figure 5 As shown, it includes the following steps:

[0176] (1) Signal acquisition;

[0177] Two undamaged carbon fiber composite laminates with dimensions of 600mm x 600mm x 2.56mm and a layup pattern of [45° / -45° / 0° / 90°]s were selected.

[0178] One carbon fiber composite laminate was used as the standard carbon fiber composite laminate. Another carbon fiber composite laminate was adjusted by pre-embedding polytetrafluoroethylene (PTFE) films at different horizontal positions and depths within the laminate to simulate delamination damage in the carbon fiber composite laminate. A total of seven different delamination damage categories were created, labeled 1#, 2#...7#. The two-dimensional planar distribution of the different delamination damage types is shown below. Figure 2 As shown, the distribution locations of damage at different depths are as follows: Figure 3 As shown; the adjusted carbon fiber composite laminate is used as the carbon fiber composite laminate to be tested;

[0179] According to such Figure 1 The schematic diagram shows piezoelectric ceramic sensors arranged on the carbon fiber composite laminate to be tested. Twelve PZT sensors are evenly distributed on a circle with a diameter of 300 mm at the center of the upper surface of the carbon fiber composite laminate to be tested, and these 12 PZT sensors are sequentially labeled as S1, S2...S12 in clockwise order.

[0180] The S1 piezoelectric ceramic sensor is used as the excitation sensor to generate a lamb wave, and the other 11 piezoelectric ceramic sensors are used as receiving sensors to receive the corresponding lamb wave response signals. Then, the remaining piezoelectric ceramic sensors are started to generate lamb waves in sequence by clockwise excitation signals, and the other 11 sensors receive the corresponding lamb wave response signals, until the S11 piezoelectric ceramic sensor generates a lamb wave.

[0181] Acquire lamb wave response signals from standard carbon fiber composite laminates (non-damaged carbon fiber composite laminates);

[0182] The excitation signal generated by the excitation sensor is a five-peak sine wave signal modulated by a Hamming window with a center frequency of 170kHz. The number, arrangement and numbering of the sensors on the standard carbon fiber composite plate and the carbon fiber composite plate under test are the same. The excitation sensor is connected to an arbitrary waveform generator through a high-voltage power amplifier. The receiving sensor is connected to a multi-channel oscilloscope to acquire the lamb wave response signal.

[0183] (2) Signal denoising and reconstruction;

[0184] Denoising and reconstruction are performed on each lambda wave response signal; the specific steps are as follows:

[0185] (2.1) Using envelope entropy as the fitness function, the improved gray wolf optimization algorithm based on dimensionality learning hunting search strategy is used to optimize the mode number N and penalty term α of the variational mode decomposition method; the specific steps are as follows:

[0186] (a) Set initialization parameters;

[0187] Let G = 30, I = 15, α = [50, 15000], N = [2, 8], where G represents the population size and I represents the maximum number of iterations;

[0188] (b) The wolf pack positions are initialized using the following formula, where the position of the gray wolf represents the number of modes N and the penalty term α of the variational mode decomposition method;

[0189] X ij =l j +rand j [0,1)×(u j -l j ), i∈[1,G],j∈[1,D];

[0190] In the formula, X ij This represents the position of the i-th wolf in dimension j; l j rand represents the lower bound values ​​of α and N. j [0,1) represents generating data within the interval [0,1) in dimension j; u j α represents the upper limit of N; D represents the dimension of the problem. Since the parameters to be optimized in variational mode decomposition are α and N, D is set to 2.

[0191] The i-th wolf uses vector X at its position in the t-th iteration. i (t)={x i1 ,x i2 ,...,x iD} indicates that the entire wolf pack is stored in a matrix Pop;

[0192] (c) Search for a solution;

[0193] In the dimensionality-based hunting search strategy, besides updating X from the matrix Pop, i A candidate position X of (t) i-GWO In addition to (t+1), X is also generated using the following formula. i Another candidate position X of (t) i-DLH (t+1):

[0194] X i-DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t));

[0195] In the formula, X i-DLH,d (t+1) represents candidate X i-DLH The position of (t+1) in the d-th dimension, where d = 1 or 2; X i,d (t) represents the position of the i-th wolf in the d-th dimension; X r,d(t) represents the position of the wolf selected from matrix Pop in the d-th dimension; X n,d (t) represents the position of a wolf selected from the vicinity of the i-th wolf in the d-th dimension; rand×(X n,d (t)-X r,d (t) represents randomly selecting X from the vicinity of the i-th wolf. n,d (t) and randomly select X from the entire wolf pack r,d (t), and simultaneously perform the difference calculation;

[0196] (d) Iterative solution;

[0197] The following formula compares two candidate positions X. i-DLH (t+1) and X i-GWO The fitness value at (t+1) is used to select the best candidate position X. i (t+1);

[0198]

[0199] In the formula, f represents the envelope entropy fitness function;

[0200] (e) Obtain the optimal parameters;

[0201] Determine X i Is the fitness of (t+1) less than X? i The fitness of (t) is determined; if so, X is set to... i (t) is replaced with X i (t+1); conversely, X i (t) remains unchanged;

[0202] After performing steps (b) to (e) on all wolves, the algorithm iteration count I is increased by 1, and the process returns to step (c). Steps (c) to (e) are then executed repeatedly until the preset maximum number of iterations is reached.

[0203] (2.2) Select the variational mode decomposition method with optimal parameters to perform mode decomposition on the signal. The formula used for mode decomposition is as follows:

[0204] imf ik =VMD(x i (t)), k=1,...,N;

[0205] res = x i (t)-imf ik k = 1, ..., N;

[0206] In the formula, imf ikIt is the k-th mode component obtained after mode decomposition; res is the original signal minus all the decomposed mode components; x i (t) is the signal to be decomposed into modes; N is the number of modal components (number of modes);

[0207] (2.3) Calculate the correlation coefficient between each modal component and the original signal, using the following formula:

[0208]

[0209] In the formula, P ik For correlation coefficient; imf ik It is the k-th modal component obtained after modal decomposition; x i (t) is the signal to be decomposed into modes; Cov(imf) ik ,x i (t) is imf ik and x i The covariance of (t); D(imf ik ) is IMF ik The variance of D(x); i (t) is x i The variance of (t);

[0210] (2.4) Wavelet threshold denoising and reconstruction, the formula is as follows:

[0211] reimf ip =wdenoise(oimf ip ,lev,wavelet,method,thresholdrelu,noiseestimate);

[0212] reres=wdenoise(res,lev,wavelet,method,thresholdrelu,noiseestimate);

[0213] rex i (t)=reimf ip +reres;

[0214] In the formula, reimf ip for oimf ip The result after the component is denoised and reconstructed using wavelet thresholding (hard thresholding is used in this invention); oimf ip The modal component with the highest correlation coefficient; reres is the result of res after denoising and reconstruction using wavelet thresholding (hard thresholding is used in this invention); rex i(t) represents the denoised and reconstructed signal; lev represents the wavelet decomposition level, set to [1, 13]; wavelet represents the wavelet basis function, set to [sym8, db8, fk22, boir6.8, coif5]; method represents the denoising method; thresholdrelu represents the threshold rule, set to [hard, soft]; noiseestimate represents the noise estimate.

[0215] Taking damage #1 as an example, Figure 4 This describes a comparison of the excitation signal from the S1 piezoelectric ceramic sensor and the lambda response signal received by the S5 piezoelectric ceramic sensor on the carbon fiber composite laminate under test and the standard carbon fiber composite laminate. On the standard carbon fiber composite laminate, there is no scattered wave in the lambda response signal collected by the S5 piezoelectric ceramic sensor. On the carbon fiber composite laminate under test, there is a significant scattered wave in the lambda response signal collected by the S5 piezoelectric ceramic sensor caused by damage #1.

[0216] The reconstructed SNR curve of the signal along the S1-S5 path on the carbon fiber composite plate under test (where the S1 piezoelectric ceramic sensor acts as the excitation sensor and the S5 piezoelectric ceramic sensor acts as the receiving sensor) is shown below. Figure 6 As shown in (a); the SNR curves of the S1-S5 path signal reconstruction on the standard carbon fiber composite laminate are as follows. Figure 7 As shown in (a);

[0217] In addition, this embodiment adds strong noise (3dB) to the lamb wave response signals collected along paths S1-S5 on the fiber composite laminate under test and the standard fiber composite laminate to simulate noise interference in the real environment; and the lamb wave response signals with added strong noise are then denoised and reconstructed through the above steps.

[0218] The reconstructed lambd wave response signal of the S1-S5 path on the fiber composite layer under test is compared with its noise-free path lambd wave response signal, as follows: Figure 8 As shown; the reconstruction results of the lamb wave response signals along the S1-S5 path on the standard fiber composite laminate are compared with the lamb wave response signals along the path without noise, as shown. Figure 9 As shown; via Figure 8 , Figure 9 The results show that after denoising by variational mode decomposition and wavelet thresholding, the reconstructed signal fits the noise-free signal very well.

[0219] (3) Location of the level of damage;

[0220] The denoised and reconstructed signal of the carbon fiber composite laminate under test is input into a tomographic imaging algorithm for damage imaging. The two-dimensional coordinate position of the damage is determined in a coordinate system. The origin of the coordinate system is a vertex of the carbon fiber composite laminate under test. The X-axis of the coordinate system is parallel to a horizontal edge of the carbon fiber composite laminate under test, and the Y-axis is parallel to a vertical edge of the carbon fiber composite laminate under test. The specific steps are as follows:

[0221] (3.1) Calculate the SDC of the path from excitation sensor i to receiver sensor j on the carbon fiber composite plate under test. ij The value, the path from excitation sensor i to receiver sensor j, is the connection between the i-th excitation sensor and the j-th receiver sensor, and the formula is as follows:

[0222]

[0223] In the formula, Rf ij When the i-th excitation sensor on the standard carbon fiber composite laminate is working, the lambda wave response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the health signal; Y ij When the i-th excitation sensor on the carbon fiber composite plate under test is working, the lambda response signal collected by the j-th receiving sensor is denoised and reconstructed, and this signal is used as the damage signal; Coν(Rf) ij ,Y ij ) is Rf ij and Y ij covariance; For Rf ij The variance; For Y ij The variance;

[0224] (3.2) Calculate the damage probability S of the elliptical region on the carbon fiber composite plate under test with excitation sensor i and receiving sensor j as foci. ij (x, y), the formula is as follows:

[0225]

[0226]

[0227] In the formula, β is the area of ​​the elliptical region; (x,y) is the coordinate of any point within the elliptical region in the coordinate system described; (x ik ,y ik (x) represents the coordinates of the excitation sensor i on the carbon fiber composite plate under test in the coordinate system; (x) represents the coordinates of the sensor i on the carbon fiber composite plate under test in the coordinate system. jk ,y jk () represents the coordinates of sensor j on the carbon fiber composite plate to be tested in the coordinate system.

[0228] (3.3) Calculate the value of any point (x) within the detection area composed of all the elliptical regions on the carbon fiber composite plate to be tested. p ,y p The damage probability distribution P(x) p ,y p The formula is as follows:

[0229]

[0230] The two-dimensional coordinate position of the damage in the coordinate system is determined based on the damage probability distribution.

[0231] (4) Damage depth localization;

[0232] (4.1) such as Figure 10 As shown, point O is denoted as the point on the upper surface of the carbon fiber composite laminate under test that coincides with the orthographic projection of a certain damage center. Two sensors connected by a line passing through point O are arranged on the upper surface of the carbon fiber composite laminate under test. The distance between the two sensors and point O is 150mm. A one-to-one transmitter and one-to-receiver method is used to collect the lamb wave response signals (lamb wave response signals of each path) of each damage. The collected lamb wave response signals of the seven types of damage are as follows: Figure 11 As shown; the acquired lambda wave response signal includes damage scattering signal and plate boundary reflection signal. The damage scattering signal is extracted, and the scattering signal length is set to 1024. The scattering time-domain signal is as follows. Figure 12 As shown;

[0233] (4.2) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and then input into a trained convolutional neural network, which outputs a damage depth prediction label.

[0234] The training steps for a convolutional neural network are as follows:

[0235] (i) Construct damage of different depths in a standard carbon fiber composite laminate. The point on the upper surface of the standard carbon fiber composite laminate that coincides with the orthographic projection of a certain damage center is denoted as point O. Two sensors connected by a line passing through point O are arranged on the upper surface of the standard carbon fiber composite laminate. The distance between the two sensors connected by a line passing through point O and point O is 150mm. The lamb wave response signal of the damage is collected by a one-to-one transmission and one-to-reception method, and then the damage scattering signal is extracted.

[0236] (ii) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform. Different levels of Gaussian white noise are added to the two-dimensional time-frequency image to increase the number of samples. Some training samples are shown below. Figure 13 As shown;

[0237] (iii) Label the two-dimensional time-frequency images with different damage depths with the true damage depth to obtain the sample dataset;

[0238] (iv) Divide the sample dataset into a training set and a test set. The training set is used for training the convolutional neural network model, and the test set is used for testing the convolutional neural network model.

[0239] (v) Input the training set into the convolutional neural network model for feature extraction and classification until the error of the convolutional neural network model's damage function is less than the set target value or the maximum number of training iterations is reached. During the training process, the TSNE method is used to visualize the model training results, as shown in the visualization results. Figure 14 As shown; via Figure 14 It can be seen that after 10 iterations, the model can effectively identify damage at different depths, and there are clear boundaries between damage at different depths.

[0240] During training, it's important to note that for the same damage, there are multiple training samples, each corresponding to different noise environments. The convolutional neural network model also outputs multiple damage depth prediction labels. Therefore, the damage depth prediction label that appears most frequently should be used as the corresponding damage depth prediction label. For example, the true damage depth label for damage 1# is "layers 1-2". Using continuous wavelet transform, the damage scattering signal corresponding to damage 1# is converted into a two-dimensional time-frequency image. Five different levels of Gaussian white noise are added to the two-dimensional time-frequency image, resulting in six training samples. These six training samples are then input into the convolutional neural network model, as shown below. Figure 15 As shown, the convolutional neural network model outputs labels "1-2 layers", "1-2 layers", "1-2 layers", "1-2 layers", "2-3 layers", and "5-6 layers". The most frequently occurring label for predicting damage depth is "1-2 layers". Therefore, "1-2 layers" is used as the label for predicting damage depth corresponding to damage #1.

[0241] (5) Visualization of the spatial location of the damage;

[0242] Using Matlab software as a platform, the horizontal and depth information of the damage is input into the plotting function to draw the spatial location of the damage;

[0243] Calculate the horizontal localization error for each injury using the following formula:

[0244]

[0245] Where e is the horizontal positioning error, (x r ,y r (x) represents the actual damage coordinates; p ,y p() is the predicted damage coordinate location;

[0246] The horizontal positioning errors for the seven types of damage were 6.43 mm, 0.67 mm, 4.24 mm, 5.41 mm, 4.66 mm, 3.67 mm, and 3.61 mm, respectively.

[0247] In the damage depth identification and localization, the depth identification accuracy of damages 1# to 7# is 100%. It should be noted that there are only two identification rates: 100% and 0%. If the damage depth prediction label and the damage depth actual label are consistent, the identification is 100%; otherwise, the identification rate is 0%.

Claims

1. A damage spatial identification and localization technology based on lambda waves in a high-noise environment, characterized in that, Includes the following steps: (1) Signal acquisition; Piezoelectric ceramic sensor arrays are uniformly arranged on the carbon fiber composite laminate to be tested and the standard carbon fiber composite laminate, respectively. The number of sensors is M. The first to the (M-1)th sensors are used as excitation sensors to generate lamb waves, and the other sensors are used as receiving sensors to collect lamb wave response signals. The only difference between the standard carbon fiber composite laminate and the carbon fiber composite laminate to be tested is that the standard carbon fiber composite laminate does not contain any damage. The number, arrangement, and numbering of sensors on the standard carbon fiber composite laminate and the carbon fiber composite laminate under test are the same. (2) Signal denoising and reconstruction; Denoising and reconstruction are performed on each lambda wave response signal separately; (3) Location of the level of damage; The denoised and reconstructed signal of the carbon fiber composite plate under test is input into the tomographic imaging algorithm to perform damage imaging, and the two-dimensional coordinate position of the damage in the coordinate system is determined. The origin of the coordinate system is a vertex of the carbon fiber composite plate under test, the X-axis of the coordinate system is parallel to a horizontal side of the carbon fiber composite plate under test, and the Y-axis of the coordinate system is parallel to a vertical side of the carbon fiber composite plate under test. (4) Damage depth localization; (4.1) Point O is the point on the upper surface of the carbon fiber composite laminate to be tested that coincides with the orthographic projection of a certain damage center. Two sensors connected by a line passing through point O are arranged on the upper surface of the carbon fiber composite laminate to be tested. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted. (4.2) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and then input into a trained convolutional neural network, which outputs a damage depth prediction label. The training steps for a convolutional neural network are as follows: (i) Construct damage of different depths in a standard carbon fiber composite laminate. The point on the upper surface of the standard carbon fiber composite laminate that coincides with the orthographic projection of a certain damage center is denoted as point O. Two sensors connected by a line passing through point O are arranged on the upper surface of the standard carbon fiber composite laminate. The lamb wave response signal of the damage is collected by one transmitter and one receiver, and the damage scattering signal is extracted. (ii) The damage scattering signal is converted into a two-dimensional time-frequency image using continuous wavelet transform, and Gaussian white noise of different degrees is added to the two-dimensional time-frequency image to increase the number of samples; (iii) Label the two-dimensional time-frequency images with different damage depths with the true damage depth to obtain the sample dataset; (iv) Divide the sample dataset into a training set and a test set. The training set is used for training the convolutional neural network model, and the test set is used for testing the convolutional neural network model. (v) Input the training set into the convolutional neural network model for feature extraction and classification until the error of the damage function of the convolutional neural network model is less than the set target value or the maximum number of training iterations is reached. (5) Visualization of the spatial location of the damage; Using Matlab software as a platform, the horizontal and depth information of the damage is input into the plotting function to draw the spatial location of the damage.

2. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 1, characterized in that, In step (1), the piezoelectric ceramic sensor array is a circular piezoelectric ceramic sensor array.

3. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 1, characterized in that, In step (1), the excitation sensor is connected to an arbitrary waveform generator through a high-voltage power amplifier; the receiving sensor is connected to a multi-channel oscilloscope to acquire the lamb wave response signal.

4. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 1, characterized in that, The specific steps for step (2) are as follows: (2.1) Using envelope entropy as the fitness function, the improved gray wolf optimization algorithm based on dimensional learning hunting search strategy is used to optimize the mode number N and penalty term α of the variational mode decomposition method; (2.2) Select the variational mode decomposition method with optimal parameters to perform mode decomposition on the signal. The formula used for mode decomposition is as follows: imf ik = VMD(x i (t)), k = 1,..., N; res=x i (t)-imf ik ,k=1,...,N; In the formula, imf ik It is the kth modal component obtained after modal decomposition; res is the result of subtracting all the modal components obtained from the decomposition from the original signal; x i (t) is the signal to be decomposed into modes; (2.3) Calculate the correlation coefficient between each modal component and the original signal, using the following formula: In the formula, P ik Cov(imf) is the correlation coefficient. ik ,x i (t) is imf ik and x i The covariance of (t); D(imf ik ) is IMF ik The variance of D(x); i (t) is x i The variance of (t); (2.4) Wavelet threshold denoising and reconstruction, the formula is as follows: reimf ip =wdenoise(oimf ip ,lev,wavelet,method,thresholdrelu,noiseestimate); reres=wdenoise(res,lev,wavelet,method,thresholdrelu,noiseestimate); rex i (t)=reimf ip +reres; In the formula, reimf ip for oimf ip The result after wavelet thresholding and denoising reconstruction of the components; oimf ip The modal component with the highest correlation coefficient; reres is the result of res after wavelet thresholding and denoising reconstruction; rex i (t) represents the denoised and reconstructed signal; lev represents the wavelet decomposition level; wavelet represents the wavelet basis function; method represents the denoising method; thresholdrelu is the threshold rule; noiseestimate is the noise estimation.

5. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 4, characterized in that, The specific steps of step (2.1) are as follows: (a) Set initialization parameters; Let G = 30, I = 15, α = [50, 15000], N = [2, 8], where G represents the population size and I represents the maximum number of iterations; (b) The wolf pack positions are initialized using the following formula, where the position of the gray wolf represents the number of modes N and the penalty term α of the variational mode decomposition method; X ij =l j +rand j [0,1)×(u j -l j ),i∈[1,G],j∈[1,D]; In the formula, X ij This represents the position of the i-th wolf in dimension j; l j rand represents the lower bound values ​​of α and N. j [0,1) represents generating data within the interval [0,1) in dimension j; u j α and N represent the upper limits of α and N; D represents the dimension of the problem, and D takes a value of 2. The i-th wolf uses vector X at its position in the t-th iteration. i (t)={x i1 ,x i2 ,...,x iD } indicates that the entire wolf pack is stored in a matrix Pop; (c) Search for a solution; In the dimensionality-based hunting search strategy, besides updating X from the matrix Pop, i A candidate position X of (t) i-GWO In addition to (t+1), X is also generated using the following formula. i Another candidate position X of (t) i-DLH (t+1): X i-DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t)); In the formula, X i-DLH,d (t+1) represents candidate X i-DLH The position of (t+1) in the d-th dimension, where d = 1 or 2; X i,d (t) represents the position of the i-th wolf in the d-th dimension; X r,d (t) represents the position of the wolf selected from matrix Pop in the d-th dimension; X n,d (t) represents the position of a wolf selected from the vicinity of the i-th wolf in the d-th dimension; rand×(X n,d (t)-X r,d (t) represents randomly selecting X from the vicinity of the i-th wolf. n,d (t) and randomly select X from the entire wolf pack r,d (t), and simultaneously perform the difference calculation; (d) Iterative solution; The following formula compares two candidate positions X. i-DLH (t+1) and X i-GWO The fitness value at (t+1) is used to select the best candidate position X. i (t+1); In the formula, f represents the envelope entropy fitness function; (e) Obtain the optimal parameters; Determine X i Is the fitness of (t+1) less than X? i The fitness of (t), if so, then X i (t) is replaced with X i (t+1); conversely, X i (t) remains unchanged; After performing steps (b) to (e) on all wolves, the algorithm iteration count I is increased by 1, and the process returns to step (c). Steps (c) to (e) are then executed repeatedly until the preset maximum number of iterations is reached.

6. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 4, characterized in that, In step (2.4), the wavelet threshold is processed using a hard thresholding method.

7. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 1, characterized in that, The specific steps for step (3) are as follows: (3.1) Calculate the SDC of the path from excitation sensor i to receiver sensor j on the carbon fiber composite plate under test. ij The value, the path from excitation sensor i to receiver sensor j, is the connection between the i-th excitation sensor and the j-th receiver sensor, and the formula is as follows: In the formula, Rf ij When the i-th excitation sensor is working on the standard carbon fiber composite laminate, the signal obtained by denoising and reconstructing the lambda wave response signal collected by the j-th receiving sensor is as follows: Y ij When the i-th excitation sensor on the carbon fiber composite plate under test is working, the lamb wave response signal collected by the j-th receiving sensor is obtained by denoising and reconstructing. Coν(Rf ij ,Y ij ) is Rf ij and Y ij covariance; σ Rfij For Rf ij variance; σ Yij For Y ij The variance; (3.2) Calculate the damage probability S of the elliptical region on the carbon fiber composite plate under test with excitation sensor i and receiving sensor j as foci. ij (x, y), the formula is as follows: In the formula, β is the area of ​​the elliptical region; (x,y) is the coordinate of any point within the elliptical region in the coordinate system described; (x ik ,y ik (x) represents the coordinates of the excitation sensor i on the carbon fiber composite plate under test in the coordinate system; (x) represents the coordinates of the sensor i on the carbon fiber composite plate under test in the coordinate system. jk ,y jk () represents the coordinates of sensor j on the carbon fiber composite plate to be tested in the coordinate system. (3.3) Calculate the value of any point (x) within the detection area composed of all the elliptical regions on the carbon fiber composite plate to be tested. p ,y p The damage probability distribution P(x) p ,y p The formula is as follows: The two-dimensional coordinate position of the damage in the coordinate system is determined based on the damage probability distribution.

8. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 1, characterized in that, In steps (4.1) and (i), the two sensors whose lines pass through point O are at the same distance from point O.

9. The damage spatial identification and localization technology based on lambda waves in a strong noise environment according to claim 8, characterized in that, In steps (4.1) and (i), the distance between the two sensors whose connecting lines pass through point O and point O is 150mm.

Images