Nonlinear ultrasonic detection signal noise reduction and feature enhancement method and system

Abstract

The invention discloses a nonlinear ultrasonic detection signal noise reduction and feature enhancement method and system, and relates to the technical field of nonlinear ultrasonic detection.The method comprises the steps that CEEMDAN decomposition is conducted on detected nonlinear ultrasonic signals of concrete damage, and a plurality of intrinsic mode components IMFs are obtained; classifying and integrating the IMFs obtained by decomposition by using a sample entropy value to obtain a component IN-IMFs after classification and integration; performing variational mode decomposition on the IN-IMFs after classification and integration to obtain a group of intrinsic mode functions IMFs; and dividing the intrinsic mode function IMFs into a training data set and a test data set according to a time step length, inputting the training data set into a gating cycle unit model for training, and outputting a noise reduction and feature enhancement effect result of each IN-IMF. According to the invention, noise interference can be reduced for ultrasonic and nonlinear ultrasonic detection of damage of materials such as rocks and concrete, and nonlinear characteristics of detection signals in nonlinear ultrasonic can be enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of nonlinear ultrasonic detection, and in particular to a method and system for denoising and enhancing a nonlinear ultrasonic detection signal. Background Art

[0002] There are many concrete buildings in existing engineering constructions. Many concrete buildings will inevitably age and crack due to long-term use. Failure to detect related problems in time leads to large economic losses and casualties. Therefore, it is of great significance to detect cracks in existing concrete structures in advance. However, the current mainstream non-destructive testing methods, such as ultrasonic testing, can only detect macro cracks, but not micro cracks.

[0003] In order to overcome the problem of microcrack detection, a nonlinear ultrasonic detection method has been developed. However, the nonlinear ultrasonic signal in the nonlinear ultrasonic detection method is very sensitive and small, and is often hidden in the system noise, which causes strong interference to the measurement precision and accuracy. This will lead to failure in the identification of microcrack states and the prediction of microcrack development. At the same time, nonlinear ultrasonic detection signals are often nonlinear, non-stationary, and have a low signal-to-noise ratio. Most of the existing methods are aimed at denoising ultrasonic detection signals, while there are still few methods for denoising nonlinear ultrasonic detection signals. In addition, the nonlinear characteristics of nonlinear ultrasonic detection signals change in the same overall direction as the detected damage, but are volatile. In summary, it is necessary to propose a denoising and feature enhancement method for nonlinear ultrasonic detection signals of concrete.

[0004] On the other hand, there are many methods for denoising ultrasonic detection signals in the non-destructive field. Initially, noise reduction was performed by setting relevant filters, but this method requires adjusting the filter-related parameters and has poor noise reduction effect. An effective signal processing method has been developed from the perspective of time series processing. The time series is decomposed into a series of modes, and processing the decomposed modes can achieve the effect of signal processing. The initial effective time series processing method is wavelet transform, which uses wavelet basis functions and combines the window function of short-time Fourier transform to select a certain proportion to improve processing performance; but for different types of signals, it is not easy to select a suitable wavelet basis function.

[0005] Empirical mode decomposition (EMD) can adaptively decompose the original signal into multiple intrinsic mode functions without manual parameter adjustment, thus solving the drawbacks of wavelet transform. However, this method is prone to mode aliasing, that is, similar time scales are distributed in different intrinsic mode components (IMFs), resulting in large errors in signal processing results.

[0006] Variational mode decomposition (VMD) has a rigorous mathematical theory derivation, which can decompose the signal into multiple single components and reduce the generation of mixed components. Therefore, it has become another method to solve the problems of EMD mode aliasing, endpoint effect and lack of theoretical support. It is worth noting that VMD requires setting the initial guess, which may affect the accuracy and stability of the final result; determining the number of modes is also a challenge for VMD, which will also affect the accuracy of the decomposition results. Finally, since VMD is based on an optimization algorithm, there may be a local minimum problem that leads to inaccurate decomposition results. Therefore, more effective and advanced signal processing methods need to be explored.

[0007] Neural networks can learn effective features of time series data from the original time series. If neural networks are used to learn the features of noisy time series, since neural networks cannot extract effective information from the noise component, neural networks can be used to extract the features of real information masked by noise. At the same time, neural networks can learn the nonlinear relationship between the original signal and the noisy signal from the training data and establish a mapping. Compared with previous methods, they can automatically learn insensitive features in the noise to achieve the purpose of noise reduction and avoid artificial selection and adjustment of parameters. Therefore, it has great potential advantages to integrate traditional signal processing methods with deep learning and propose signal processing methods that can effectively improve the level of noise reduction. At the same time, the existing noise reduction methods only study the effect of signal noise reduction, and the research objects include ultrasonic signals, audio signals, vibration signals, and medical ECG signals. However, the existing methods have not yet enhanced the characteristics of the processed signals.

[0008] Therefore, there is an urgent need for a nonlinear ultrasonic detection signal processing solution that can take into account both signal noise reduction and feature enhancement. Summary of the invention

[0009] The purpose of the present invention is to provide a method and system for noise reduction and feature enhancement of nonlinear ultrasonic detection signals, which can significantly reduce the noise interference of ultrasonic and nonlinear ultrasonic detection of damage to materials such as rocks and concrete, and can also enhance the nonlinear characteristics of detection signals in nonlinear ultrasound.

[0010] To achieve the above object, the present invention provides the following solutions:

[0011] A method for noise reduction and feature enhancement of nonlinear ultrasonic detection signals comprises the following steps:

[0012] S1. Perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs;

[0013] S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs;

[0014] S3, performing variational mode decomposition on the classified and integrated IN-IMFs to obtain further decomposition results of each IN-IMF - intrinsic mode functions IMFs;

[0015] S4. Divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into the gated recurrent unit model for training, and output each IN-IMF denoising and feature enhancement result.

[0016] Furthermore, the S1, performing CEEMDAN decomposition on the nonlinear ultrasonic signal of concrete damage to obtain multiple intrinsic mode components IMFs, specifically includes the following steps:

[0017] S101, first at time n, a white noise sequence ω with a signal-to-noise ratio of ξ0 is generated. n [t] (n = 1, 2, ..., N) is added to the original data sequence x[t] to obtain the data sequence x for decomposition n [t]:

[0018] x n [t]=x[t]+ξ0ω n [t],n=1,2,…,N (1)

[0019] Where t represents different time points in the data, n represents the nth time white noise is added; N represents the total number of times white noise is added, which is also the total number of data points collected; ξ0 represents the signal-to-noise ratio, which is used to determine the standard deviation of white noise;

[0020] S102, respectively x n [t] Perform N EMD decompositions to obtain N IMF1 i [t] Results, taking the average value, we get the first IMF as The formula is:

[0021]

[0022] EMD1[·] means EMD decomposition of the sequence in the brackets;

[0023] Then the first residual is calculated as:

[0024] S103, adding an adaptive noise term to the first residual r1[t] to form a new data sequence; wherein the adaptive noise term is a white noise ω with a signal-to-noise ratio of ξ1 n [t], as a new carrier to perform empirical mode decomposition on the new data sequence, and convert the signal r1[t]+ξ1EMD1[ω n[t]]Use EMD method to further decompose and obtain the second component IMF and residual:

[0025]

[0026] S104, repeat step S102 and step S103 to obtain the kth residual and the kth IMF component:

[0027]

[0028] Where EMDk[·] represents the kth IMF component obtained by EMD decomposition, k = 2, 3…K, K is the total amount of IMFs finally obtained;

[0029] S105, repeat step S104 until the residual cannot be further decomposed by EMD, terminate the CEEMDAN algorithm, obtain K IMF components, and the final residual is R[t]; the final decomposition result of the original data sequence is:

[0030]

[0031] Furthermore, the S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs, specifically includes the following steps:

[0032] S201, let the dimension of sample data be m, which is also the length of comparison sequence, the total number of data points be M, let the sample data be {y m (i),y m (i+1),…,y m (Mm-1)}, i = 1, 2, ... N, the threshold is the similarity tolerance r, forming an m-dimensional vector Y m (i);

[0033] Y m (i)=[y(i),y(i+1),…,y(i+m-1)],1≤i≤Mm (9)

[0034] S202, let vector Y m (i) and Y m (j) The distance between them is d[Y m (i),Y m (j)], which is the maximum absolute value of the difference between corresponding elements of the vectors:

[0035] d[Y m (i),Y m (j)]=max[y(i+k)-y(j+k)] (10)

[0036] Where, k=1,2,…,m-1,1≤i,j≤Mm;

[0037] S203. Under the condition of threshold r, calculate Y m (i) and Y m The number of distances d[Ym(i), Ym(j)] between (j) is less than or equal to the threshold r, denoted as B i , which means that the number of B that satisfies d[Ym(i),Ym(j)]≤r(1≤i≤Mm,j≠i) is i , and define the probability that two vectors match m points within the threshold range as B m (r), the expression is as follows:

[0038]

[0039]

[0040] S204, increase the dimension to m+1, and calculate Y according to step S203 m +1(i) and Y m +1(j) the distance d[Y m +1(i),Y m +1(j)] is less than or equal to the threshold r, denoted as A i , which means that d[Y m +1(i), Y m +1(j)]≤r(1≤i≤Mm,j≠i) is the number of i , and define the probability that two vectors match m+1 points within the threshold range as A m (r), the expression is as follows:

[0041]

[0042] S205. Solve the sample entropy S(m,r) of the sequence:

[0043]

[0044] In the actual signal, M is a finite value, so the sample entropy S(m,r,M) is estimated as follows:

[0045]

[0046] S206. Sort, classify and integrate the IMFs according to the sequence complexity and disorder degree represented by the sample entropy obtained in step S205 to obtain the classified and integrated components IN-IMFs.

[0047] Furthermore, the step S3, performing variational mode decomposition on the classified and integrated IN-IMFs to obtain a further decomposition result of each IN-IMF, namely a set of intrinsic mode functions IMFs, specifically comprises the following steps:

[0048] S301. Set input: signal to be decomposed The number of modes to be decomposed K;

[0049] S302, initialization: each mode and its center frequency The initial value of is set to 0, and the Grange multiplier Set the initial value to 0, set n = 0;

[0050] S303, let n = n + 1, and use the following formula to update each mode Center frequency and the Lagrange multiplier

[0051]

[0052] Where k = 1, 2, ..., K; γ represents the noise tolerance. When the signal contains strong noise, set γ = 0 to achieve better denoising effect. represents the Fourier transform of f(t); Indicates u k (t) The nth iteration value in the Fourier domain;

[0053] S304, judging the iteration termination condition according to the following formula, if it is satisfied, outputting the K patterns obtained by decomposition, otherwise returning to step S303 to continue iteratively updating and decomposing;

[0054]

[0055] S305. The whole process is solved in the frequency domain. Perform inverse Fourier transform to obtain IMFs.

[0056] Furthermore, in S4, the training data set is input into the gated recurrent unit model for training, which specifically includes the following steps:

[0057] 1) Calculate the reset gate value: According to the input and the system state at the previous moment, determine how much information in the previous stage needs to be forgotten, as shown in the following formula:

[0058] r t =σ(W r x t +U r h t-1 +b r ) (twenty two)

[0059] Among them, r t ∈[0,1] D is the reset gate, σ is the Logistic function, whose output interval is (0, 1), W r and U r are all network training parameter matrices, x t is the input at the current moment, h t-1 is the system state at the previous moment, b r is the bias term;

[0060] 2) Calculate the intermediate state value:

[0061]

[0062] Among them, ⊙ is the Hadamard product operator of the vector, which performs a dot multiplication operation, indicating the multiplication of each element; W h and U h are all network training parameter matrices; b h is the bias term; is the intermediate state value; Depends on the state h of the previous moment t-1 ; When r t = 0, candidate state Only with the current input x t It is related to the historical state; when r t =1, candidate state and the current input x t And the historical status h t-1 Related;

[0063] 3) Calculate the update gate value according to formula (24): The update gate determines how much information to retain in the previous state and how much new information to accept from the intermediate state based on the input and the system state at the previous moment. This process does not undergo nonlinear transformation.

[0064] z t =σ(W z x t +U z h t-1 +b z ) (twenty four)

[0065] Among them, z t ∈[0,1] D is the update gate; W z and U z are all network training parameter matrices; b z is the bias term;

[0066] 4) Calculate the output value:

[0067]

[0068] Among them, h t is the state vector output at time t. When zt is 0, the current state h t and the previous state h t-1 There is a nonlinear functional relationship between them. When z t When it is 1, the current state h t and the previous state h t-1 There is a linear function relationship between them.

[0069] The present invention also provides a nonlinear ultrasonic detection signal noise reduction and feature enhancement system, which is applied to the nonlinear ultrasonic detection signal noise reduction and feature enhancement method, comprising:

[0070] CEEMDAN decomposition module, used to perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs;

[0071] The classification and integration module is used to classify and integrate the decomposed IMFs using the sample entropy value to obtain the classified and integrated components IN-IMFs;

[0072] The variational mode decomposition module is used to perform variational mode decomposition on the classified and integrated IN-IMFs to obtain a set of intrinsic mode functions IMFs;

[0073] The result output module is used to divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into the gated recurrent unit model for training, and output each IN-IMF denoising and feature enhancement result.

[0074] Furthermore, the system also includes a nonlinear ultrasonic detection device for detecting damage to concrete and obtaining a nonlinear ultrasonic signal of concrete damage; the nonlinear ultrasonic detection device includes a signal generator, a signal amplifier, an oscilloscope, a transmitting transducer and a receiving transducer.

[0075] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the nonlinear ultrasonic detection signal denoising and feature enhancement method and system provided by the present invention, the IMFs obtained by decomposing the original signal using CEEMDAN, are classified into three types of IN-IMFs according to sample entropy, and the VMD decomposition results of each IN-IMF are subjected to denoising by applying the GRU model, and finally the denoising results of each IN-IMF are synthesized into the final comprehensive denoising result by superposition (ADD data) or GRU model (GRU data). The denoising results are evaluated by using indicators such as signal-to-noise ratio SNR (dB), root mean square error RMSE, waveform similarity coefficient NCC, and the denoising effect is obvious. At the same time, the denoising method of the present invention enhances the nonlinear characteristics of the detection signal and makes the change trend more obvious. To a certain extent, the method weakens the fluctuation of the nonlinear coefficient of the original nonlinear ultrasonic signal with the change of damage, and reduces the reverse increase or reverse decrease in the process of changing the nonlinear coefficient. Ultrasonic and nonlinear ultrasonic detection of damage to materials such as rock and concrete can reduce noise interference and enhance the nonlinear characteristics of the detection signal in nonlinear ultrasound. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0077] Figure 1 It is a flow chart of the method for denoising and enhancing features of nonlinear ultrasonic detection signals according to an embodiment of the present invention;

[0078] Figure 2 A diagram of a concrete sample with prefabricated cracks of different lengths in an embodiment of the present invention;

[0079] Figure 3 The CEEMDAN decomposition results of the embodiment of the present invention are as follows;

[0080] Figure 4 The result of integrating IMFs according to sample entropy in an embodiment of the present invention;

[0081] Figure 5 The final result of integrating IMFs into IN-IMFs in the embodiment of the present invention;

[0082] FIG6 is a VMD decomposition result of IN-IMFs components according to an embodiment of the present invention, wherein (a) is the VMD decomposition result of IN-IMF1, (b) is the VMD decomposition result of IN-IMF2, and (c) is the VMD decomposition result of IN-IMF3;

[0083] FIG7 shows the denoising of the IN-IMFs components according to the method of an embodiment of the present invention, wherein (a) the denoising of IN-IMF1, (b) the denoising training of GRU of IN-IMF1, (c) the denoising of IN-IMF2, (d) the denoising training of GRU of IN-IMF2, (e) the denoising of IN-IMF3, and (f) the denoising training of GRU of IN-IMF3;

[0084] FIG8 is a diagram showing the overall denoising of the original data according to an embodiment of the present invention, wherein (a) is the denoising corresponding to the superimposed summary of the overall denoising (superimposed data), and (b) is the GRU model summary of the overall denoising (GRU data);

[0085] FIG9 is a damage analysis of the nonlinear ultrasonic detection signal before and after noise reduction in an embodiment of the present invention, wherein (a) is the variation trend of the nonlinear coefficient of the specimen group, and (b) is the variation trend of the nonlinear coefficient of the specimen group before and after noise reduction. DETAILED DESCRIPTION

[0086] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0087] The purpose of the present invention is to provide a method and system for noise reduction and feature enhancement of nonlinear ultrasonic detection signals, which can reduce the noise interference of ultrasonic and nonlinear ultrasonic detection of damage to materials such as rocks and concrete, and can also enhance the nonlinear characteristics of detection signals in nonlinear ultrasound.

[0088] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0089] like Figure 1 As shown, a nonlinear ultrasonic detection signal noise reduction and feature enhancement method provided by an embodiment of the present invention includes the following steps:

[0090] S1. Perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs. In this embodiment, the number of trials used by CEEMDAN to generate noise is set to 10, and the number of modes after decomposition, num_clusters, is set to 3. Figure 3 As shown;

[0091] S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs;

[0092] S3, performing variational mode decomposition on the classified and integrated IN-IMFs to obtain further decomposition results of each IN-IMF - a set of intrinsic mode functions IMFs;

[0093] S4. Divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into a gated recurrent unit model (i.e., GRU) for training, and output each IN-IMF denoising and feature enhancement result.

[0094] In this embodiment, the S1, performing CEEMDAN decomposition on the nonlinear ultrasonic signal of concrete damage to obtain multiple intrinsic mode components IMFs, specifically includes the following steps:

[0095] S101, first at time n, a white noise sequence ω with a signal-to-noise ratio of ξ0 is generated. n [t] (n = 1, 2, ..., N) is added to the original data sequence x[t] to obtain the data sequence x for decomposition n [t]:

[0096] x n [t]=x[t]+ξ0ω n [t],n=1,2,…,N (1)

[0097] Where t represents different time points in the data, n represents the nth time white noise is added; N represents the total number of times white noise is added, which is also the total number of data points collected; ξ0 represents the signal-to-noise ratio, which is used to determine the standard deviation of white noise;

[0098] S102, respectively x n [t] Perform N EMD decompositions to obtain N IMF1 i [t] Results, taking the average value, we get the first IMF as As shown in formula (2), the significance of taking the average value is that the added white noise is zero after averaging, so the influence of white noise is eliminated by taking the average value.

[0099]

[0100] EMD1[·] means EMD decomposition of the sequence in the brackets;

[0101] Then the first residual is calculated as:

[0102] S103, adding an adaptive noise term to the first residual r1[t] to form a new data sequence; wherein the adaptive noise term is a white noise sequence ω with a signal-to-noise ratio of ξ1 n [t], as a new carrier to perform empirical mode decomposition on the new data sequence, and convert the signal r1[t]+ξ1EMD1[ω n [t]]Use EMD method to further decompose and obtain the second component IMF and residual:

[0103]

[0104] S104, repeat step S102 and step S103 to obtain the kth residual and the kth IMF component:

[0105]

[0106] Among them, EMD k [·] represents the kth IMF component obtained by EMD decomposition, k = 2, 3…K, K is the total amount of IMFs finally obtained;

[0107] S105, repeat step S104 until the residual cannot be further decomposed by EMD, terminate the CEEMDAN algorithm, obtain K IMF components, and the final residual is R[t]; the final decomposition result of the original data sequence is:

[0108]

[0109] Due to the IMF uniqueness problem in the EMD-like decomposition method, that is, in the EMD decomposition process, different decomposition results of the adjustment parameters can produce multiple different IMFs; too many decomposition times will reduce the physical meaning of each IMF in the decomposition result, on the other hand, too few decomposition times cannot obtain pure IMF components with a single trend. The IMFs obtained after CEEMDAN decomposition have different types of frequencies, trends, etc. If the many IMFs obtained by decomposition can be classified and integrated into several types, such as three large types of integrated IMFs (recorded as IN-IMF), the subsequent denoising process will be clearer and more specific.

[0110] In addition, since sample entropy has a good ability to characterize the complexity of noise sequences, this indicator is introduced to measure the distinction between IMF types.

[0111] In this embodiment, the sample entropy of each IMF including the remainder is calculated, and the complexity of calculating each component after decomposition is as follows: Figure 4 That is, the step S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs, specifically includes the following steps:

[0112] S201, let the dimension of sample data be m, which is also the length of comparison sequence, the total number of data points be M, let the sample data be {y m (i),y m (i+1),…,y m (Mm-1)}, i = 1, 2, ... N, the threshold is the similarity tolerance r, forming an m-dimensional vector Y m (i);

[0113] Y m (i)=[y(i),y(i+1),…,y(i+m-1)],1≤i≤Mm (9)

[0114] In this embodiment, the reconstruction dimension m is set to 1, and the threshold size r is set to 0.1;

[0115] S202, let vector Y m (i) and Y m (j) The distance between them is d[Y m (i),Y m (j)], which is the maximum absolute value of the difference between corresponding elements of the vectors:

[0116] d[Y m (i),Y m (j)]=max[|y(i+k)-y(j+k)|] (10)

[0117] Where, k=1,2,…,m-1,1≤i,j≤Mm;

[0118] S203. Under the condition of threshold r, calculate Y m (i) and Y m The number of distances d[Ym(i), Ym(j)] between (j) is less than or equal to the threshold r, denoted as B i , which means that the number of B that satisfies d[Ym(i),Ym(j)]≤r(1≤i≤Mm,j≠i) is i , and define the probability that two vectors match m points within the threshold range as B m (r), the expression is as follows:

[0119]

[0120] S204, increase the dimension to m+1, and calculate Y according to step S203 m +1(i) and Y m +1(j) the distance d[Y m +1(i),Y m +1(j)] is less than or equal to the threshold r, denoted as A i, which means that d[Y m +1(i), Y m +1(j)]≤r(1≤i≤Mm,j≠i) is the number of i , and define the probability that two vectors match m+1 points within the threshold range as A m (r), the expression is as follows:

[0121]

[0122] S205. Solve the sample entropy S(m,r) of the sequence:

[0123]

[0124] In the actual signal, M is a finite value, so the sample entropy S(m,r,M) is estimated as follows:

[0125]

[0126] S206. Sort, classify and integrate the IMFs according to the sequence complexity and disorder degree represented by the sample entropy obtained in step S205 to obtain the classified and integrated components IN-IMFs.

[0127] In this embodiment, the K-Means clustering algorithm is used to classify the IMFs samples into several integrated intrinsic mode functions according to their entropy values. The k value in the K-Means clustering algorithm, i.e., n_clusters, is set to 3. Figure 5 As shown in the figure, IMF0 and IMF1 are more complex, contain more noise, and have more potential trends. The essential trend is less obvious due to the masking of more noise. At the same time, the sample entropy values ​​of the remaining IMF components and Residue are lower, with less volatility and complexity, making it easier to see the trend of the component itself.

[0128] Taking the above into consideration, the IMF0-IMF9 and Residue components obtained after CEEMDAN decomposition are integrated into three new integrated intrinsic mode functions (IN-IMF), namely, the high-frequency component IN-IMF1 (IMF0-IMF1), the low-frequency component IN-IMF2 (IMF2-IMF5), and the trend component IN-IMF3 (IMF6-IMF7, Residue), as shown in Figure 2. Figure 5As shown in the figure. As can be seen from the figure, the amplitude range from small to large is IN-IMF1, IN-IMF2, and IN-IMF3, and IN-IMF1 is more complex than the other two IN-IMFs, and at the same time, it is more volatile, contains more noise, and has more potential change trends. IN-IMF2 contains fewer noise components and begins to gradually show the fluctuation trend of the data, while IN-IMF3 contains the least noise components among the three IN-IMFs, and displays the most complete fluctuation trend of the original data. In general, the high-frequency component IN-IMF1 of the integrated three IN-IMFs components contains more noise, and the trend component IN-IMF3 represents the approximate fluctuation of the original data. If it is directly input into the gated recurrent unit model (GRU), the model cannot be effectively learned, and the expected noise reduction effect cannot be achieved. Therefore, it is also necessary to use variational mode decomposition (VMD) for preprocessing decomposition of GRU model learning.

[0129] VMD can adaptively process nonlinear and non-stationary information. VMD can perform multi-resolution analysis, so that the three integrated IN-IMFs components can show more detailed information. At the same time, the decomposition results of VMD can be used as the training set and test set for the subsequent GRU model division. VMD decomposition is performed on the three integrated IN-IMFs components, and the results are shown in Figure 6. The specific parameter settings of the VMD decomposition process are as follows: the balance parameter alpha of the data fidelity constraint is set to 2000, the number of decomposition layers K, that is, the number of modal decompositions, is set to 10, the convergence fault tolerance criterion tol is set to 1e-7, and other parameters are default values. After VMD decomposition, the data is input into the GRU model for training and learning.

[0130] In this embodiment, the step S3, performing variational mode decomposition (VMD) on the classified and integrated IN-IMFs to obtain a further decomposition result of each IN-IMF, namely a set of intrinsic mode functions IMFs, specifically includes the following steps:

[0131] Variational mode decomposition ultimately decomposes the input real signal f into a set of intrinsic mode functions IMFs, namely u k The decomposition result is required to be: the intrinsic mode function can not only represent the input signal characteristics, but also has a certain sparsity. Taking advantage of the quadratic penalty term and Lagrange multiplier, the augmented Lagrangian function is introduced to transform the constrained variational model into an unconstrained variational model, as shown below:

[0132]

[0133] The IMF components, center frequencies, and Lagrangian operators are continuously iterated and updated through the alternating direction multiplier algorithm to solve the saddle point of the unconstrained variational problem in the frequency domain, i.e., the optimal solution. When the iterative termination condition is met, the entire cycle ends. The main steps are as follows:

[0134] S301. Set input: signal to be decomposed The number of modes to be decomposed K; in subsequent calculations, K is taken as 10 and α is taken as 2000.

[0135] S302, initialization: each mode and its center frequency The initial value of is set to 0, and the Grange multiplier Set the initial value to 0, set n = 0;

[0136] S303, let n = n + 1, and use the following formula to update each mode Center frequency and the Lagrange multiplier and Similar to the following formula:

[0137]

[0138] Where k = 1, 2, ..., K; γ represents the noise tolerance. When the signal contains strong noise, set γ = 0 to achieve better denoising effect. represents the Fourier transform of f(t); Indicates u k (t) The nth iteration value in the Fourier domain;

[0139] S304, judging the iteration termination condition, i.e., the convergence condition, according to the following formula, if it is satisfied, outputting the K patterns obtained by decomposition, otherwise returning to step S303 to continue iteratively updating and decomposing;

[0140]

[0141] S305. The above is the basic process of variational mode decomposition. The whole process is solved in the frequency domain. The IMF can be obtained by performing inverse Fourier transform.

[0142] The VMD decomposition results of the three IN-IMFs components are input into the GRU model for training, and the data features of each VMD component are learned. Finally, the denoising results for each IN-IMFs component and the loss curves of each component are output, as shown in Figure 7. The Loss represents the loss value of the training set, and the Val_Loss represents the loss value of the test set. The GRU parameter settings are timestep set to 30, epochs set to 100, the activation function is Tanh, the optimizer optimizer uses Adam parameters, and the Loss curve calculation index uses the root mean square error Mean SquaredError (mse). The VMD decomposition results of IN-IMFs are used as the object of GRU data set division. The training set and test set are divided by time step. The proportions of the divided training set and test set are 80% and 20% respectively. In the model training process, callback functions such as EarlyStopping and ReduceLROnPlateau are used to control the training process of GRU.

[0143] Since the recurrent neural network (RNN) can mine the timing information and semantic information in the time series data, and can handle the data gradient explosion problem and memory capacity problem, the algorithm is very effective in processing time series data. The improved algorithm gated recurrent unit model (GRU) is now adopted. In this embodiment, in S4, the training data set is input into the gated recurrent unit model for training (i.e., the specific process of GRU), which specifically includes the following steps:

[0144] 1) Calculate the reset gate value: According to the input and the system state at the previous moment, determine how much information in the previous stage needs to be forgotten, as shown in the following formula:

[0145] r t =σ(W r x t +U r h t-1 +b r ) (twenty two)

[0146] Among them, r t ∈[0,1] D is the reset gate, σ is the Logistic function, whose output interval is (0, 1), W r and U r are all network training parameter matrices, x t is the input at the current moment, h t-1 is the system state at the previous moment, b r is the bias term;

[0147] 2) Calculate the intermediate state value. It should be noted that the dot product of the reset gate and the system state at the previous moment determines the propagation ratio of the previous stage information:

[0148]

[0149] Among them, ⊙ is the Hadamard product operator of the vector, which performs a dot multiplication operation, indicating the multiplication of each element; W h and U h are all network training parameter matrices; b h is the bias term; is the intermediate state value; the reason for using the tanh function is that its derivative has a relatively large value range, which can alleviate the gradient disappearance problem. Depends on the state h of the previous moment t-1 ; When r t = 0, candidate state Only with the current input x t It is related to the historical state; when r t =1, candidate state and the current input x t And the historical status h t-1 Related;

[0150] 3) Calculate the update gate value according to formula (24): The update gate determines how much information to retain in the previous state and how much new information to accept from the intermediate state based on the input and the system state at the previous moment. This process does not undergo nonlinear transformation. The update reset gate has a similar form, but the matrix parameters and functions are different, as shown below:

[0151] z t =σ(W z x t +U z h t-1 +b z ) (twenty four)

[0152] Among them, z t ∈[0,1] D is the update gate; W z and U z are all network training parameter matrices; b z is the bias term;

[0153] 4) Calculate the output value:

[0154]

[0155] Among them, h t is the state vector output at time t. When zt is 0, the current state h t and the previous state h t-1There is a nonlinear functional relationship between them. When z t When it is 1, the current state h t and the previous state h t-1 There is a linear function relationship between them. A simple analysis shows that when z t =0, r=1, the GRU model degenerates into a simple recurrent network model. t =0, r=0, the current state h t Only with the current input x t Related, and historical status h t-1 It doesn't matter. t =1, the current state h t =h t-1 Equal to the state h at the previous moment t-1 , and the current input x t Not relevant.

[0156] The denoising of the three IN-IMFs components is summarized into the overall denoising of the original data, and two methods are used for summarization, as shown in Figure 8. Method 1 is to superimpose and summarize the denoising of the three IN-IMFs components to obtain the final overall denoising of the original data (called ADD data), as shown in Figure 9(a). Method 2 is to fit the denoising of the three IN-IMFs components through GRU model learning and training, and finally summarize it into the overall denoising of the original data, as shown in Figure 9(b). Figure (c) is the loss curve of the GRU model summary training. The loss value gradually decreases and stabilizes at a smaller value, indicating that the effect is good.

[0157] One of the reasons for combining CEEMDAN and VMD with GRU is to reduce overfitting during GRU training. On the one hand, the input signal data is decomposed into a series of IMFs through CEEMDAN decomposition and VMD decomposition, which is equivalent to expanding the volume of the detection data set, thereby preventing overfitting. On the other hand, IMFs of different complexities are integrated according to sample entropy, so that IMFs with similar complexity are integrated together, avoiding IMFs with too large differences in model learning complexity, and preventing mutual noise interference in the respective learning results. Thus, the effect of preventing overfitting is achieved.

[0158] Finally, the noise reduction effect of the method is evaluated. The indicators to be used include signal-to-noise ratio SNR (dB), root mean square error RMSE, and waveform similarity coefficient NCC. These indicators are generally used to judge the noise reduction effect based on the relationship between the original noisy signal and the noise-reduced signal. As shown below:

[0159]

[0160] The better values ​​of SNR, RMSE and NCC calculation results are 11.3, 0.13 and 0.95 respectively.

[0161] As shown in Figure 9, the nonlinear coefficients of the original signal and the noise reduction signal of the detection signal are calculated, and the trend of the nonlinear coefficient changing with the damage is analyzed. In general, the trend of the nonlinear coefficient is expressed in the form of error bands. In the figure, the blue error band represents the trend of the original data, the red error band represents the trend of the GRU data, and the purple error band represents the trend of the ADD data.

[0162] Now the nonlinear coefficient trends calculated by the original data, GRU data, and ADD data are plotted in the same figure, such as Figure (a) added to Figure (c). First, compared with the original data, the nonlinear coefficient change trend in the GRU data and ADD data is more obvious as the crack damage changes, that is, the nonlinear characteristics of the damage are more prominent. At the same time, in the nonlinear coefficient change trend of the GRU data and ADD data, the change range of the ADD data is slightly larger than the change range of the GRU data. Specifically, the change ranges of the GRU data and ADD data are 18.7 times and 27.9 times the change range of the original data, respectively. This shows that the ability of the ADD data to enhance nonlinear features is slightly stronger than that of the GRU data. In the original data of Figure (a), in the length specimen, there is a slight reverse growth trend in the crack length range of 1-2cm and 3-4cm. Corresponding to Figure (c), the reverse growth trend has been eliminated. At the same time, the GRU data has a more stable growth trend relative to the ADD data, and the growth of the ADD data is volatile, and the standard deviation is gradually increasing.

[0163] In general, whether it is ADD data or GRU data, the final denoising results enhance the nonlinear characteristics of the detection signal and make the change trend more obvious. At the same time, the fluctuation of the nonlinear coefficient of the original nonlinear ultrasonic signal is weakened to a certain extent, and the inverse increase or decrease of the nonlinear coefficient with the change of damage is reduced.

[0164] Example 2

[0165] The embodiment of the present invention provides a nonlinear ultrasonic detection signal noise reduction and feature enhancement system, which is applied to the nonlinear ultrasonic detection signal noise reduction and feature enhancement method described in Example 1, including:

[0166] CEEMDAN decomposition module, used to perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs;

[0167] The classification and integration module is used to classify and integrate the decomposed IMFs using the sample entropy value to obtain the classified and integrated components IN-IMFs;

[0168] The variational mode decomposition module is used to perform variational mode decomposition on the classified and integrated IN-IMFs to obtain a set of intrinsic mode functions IMFs;

[0169] The result output module is used to divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into the gated recurrent unit model for training, and output each IN-IMF denoising and feature enhancement result.

[0170] In this embodiment, the system also includes a nonlinear ultrasonic detection device for detecting damage to concrete and obtaining a nonlinear ultrasonic signal of concrete damage; the nonlinear ultrasonic detection device includes a signal generator DG1022U, a signal amplifier ATA-2042, an oscilloscope TBS1072B, a transmitting transducer, and a receiving transducer. The signal frequency emitted by the signal generator is set to 40kHz. The purpose of precast concrete damage is achieved by inserting a customized steel sheet during concrete pouring and pulling it out before it solidifies. The lengths of precast concrete crack damage are 1cm, 2cm, 3cm, 4cm, 5cm, and 6cm, respectively. The specimen groups are as follows: Figure 2 shown.

[0171] In summary, the denoising method provided by the present invention integrates CEEMDAN algorithm, VMD algorithm and GRU model. The IMFs obtained by decomposing the original signal by CEEMDAN are classified into three types of IN-IMFs according to sample entropy, and the VMD decomposition results of each IN-IMF are subjected to denoising by applying the GRU model. Finally, the denoising results of each IN-IMF are synthesized into the final comprehensive denoising result by superposition (ADD data) or GRU model (GRU data). The denoising results are evaluated by indicators such as signal-to-noise ratio SNR (dB), root mean square error RMSE, and waveform similarity coefficient NCC, and the results show that the denoising effect is obvious. At the same time, the proposed denoising method enhances the nonlinear characteristics of the detection signal and makes the change trend more obvious. To a certain extent, the method weakens the fluctuation of the nonlinear coefficient of the original nonlinear ultrasonic signal with the change of damage, and reduces the reverse increase or reverse decrease in the process of nonlinear coefficient change. Ultrasonic and nonlinear ultrasonic detection of damage to materials such as rock and concrete can reduce noise interference and enhance the nonlinear characteristics of the detection signal in nonlinear ultrasound.

[0172] The remaining technical features in this embodiment can be flexibly selected by those skilled in the art according to actual conditions to meet different specific practical needs. However, it is obvious to those skilled in the art that it is not necessary to adopt these specific details to implement the present invention. In other examples, in order to avoid confusing the present invention, the well-known components, structures or parts are not specifically described, which are all within the technical protection scope defined by the technical solution claimed for protection in the claims of the present invention.

[0173] Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the scope of protection of the claims attached to the present invention. In the above description, a large number of specific details are set forth in order to provide a thorough understanding of the present invention. However, it is obvious to those of ordinary skill in the art that these specific details are not necessary to practice the present invention. In other examples, in order to avoid confusing the present invention, well-known technologies, such as specific construction details, operating conditions and other technical conditions, are not specifically described.

[0174] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for noise reduction and feature enhancement of nonlinear ultrasonic detection signals, characterized in that: The following steps are involved: S1. Perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs; S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs; S3, performing variational mode decomposition on the classified and integrated IN-IMFs to obtain further decomposition results of each IN-IMF - intrinsic mode functions IMFs; S4. Divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into the gated recurrent unit model for training, and output each IN-IMF denoising and feature enhancement result.

2. The method for denoising and enhancing nonlinear ultrasonic detection signals according to claim 1, characterized in that: The step S1, performing CEEMDAN decomposition on the nonlinear ultrasonic signal of concrete damage to obtain a plurality of intrinsic mode components IMFs, specifically comprises the following steps: S101, first at time n, a white noise sequence ω with a signal-to-noise ratio of ξ0 is generated. n [t] (n = 1, 2, ..., N) is added to the original data sequence x[t] to obtain the data sequence x for decomposition n [t]: x n [t]=x[t]+ξ0ω n [t],n=1,2,…,N (1) Where t represents different time points in the data, n represents the nth time white noise is added; N represents the total number of times white noise is added, which is also the total number of data points collected; ξ0 represents the signal-to-noise ratio, which is used to determine the standard deviation of white noise; S102, respectively x n [t] Perform N EMD decompositions to obtain N IMF1 i [t] Results, taking the average value, we get the first IMF as The formula is: EMD1[·] means EMD decomposition of the sequence in the brackets; Then the first residual is calculated as: S103, adding an adaptive noise term to the first residual r1[t] to form a new data sequence; wherein the adaptive noise term is a white noise ω with a signal-to-noise ratio of ξ1 n [t], as a new carrier to perform empirical mode decomposition on the new data sequence, and convert the signal r1[t]+ξ1EMD1[ω n [t]]Use EMD method to further decompose and obtain the second component IMF and residual: S104, repeat step S102 and step S103 to obtain the kth residual and the kth IMF component: Among them, EMD k [·] represents the kth IMF component obtained by EMD decomposition, k = 2, 3…K, K is the total amount of IMFs finally obtained; S105, repeat step S104 until the residual cannot be further decomposed by EMD, terminate the CEEMDAN algorithm, obtain K IMF components, and the final residual is R[t]; the final decomposition result of the original data sequence is:

3. The method for denoising and enhancing nonlinear ultrasonic detection signals according to claim 1, characterized in that: S2, using the sample entropy value to classify and integrate the decomposed IMFs to obtain the classified and integrated components IN-IMFs, specifically includes the following steps: S201, let the dimension of sample data be m, which is also the length of comparison sequence, the total number of data points be M, let the sample data be {y m (i),y m (i+1),…,y m (Mm-1)}, i = 1, 2, ... N, the threshold is the similarity tolerance r, forming an m-dimensional vector Y m (i); Y m (i)=[y(i),y(i+1),…,y(i+m-1)],1≤i≤M-m (9) S202, let vector Y m (i) and Y m (j) The distance between them is d[Y m (i),Y m (j)], which is the maximum absolute value of the difference between corresponding elements of the vectors: d[Y m (i),Y m (j)]=max[y(i+k)-y(j+k)] (10) Where, k=1,2,…,m-1,1≤i,j≤Mm; S203. Under the condition of threshold r, calculate Y m (i) and Y m The number of distances d[Ym(i), Ym(j)] between (j) is less than or equal to the threshold r, denoted as B i , which means that the number of B that satisfies d[Ym(i),Ym(j)]≤r(1≤i≤Mm,j≠i) is i , and define the probability that two vectors match m points within the threshold range as B m (r), the expression is as follows: S204, increase the dimension to m+1, and calculate Y according to step S203 m +1(i) and Y m +1(j) the distance d[Y m +1(i),Y m +1(j)] is less than or equal to the threshold r, denoted as A i , which means that d[Y m +1(i), Y m +1(j)]≤r(1≤i≤Mm,j≠i) is the number of i , and define the probability that two vectors match m+1 points within the threshold range as A m (r), the expression is as follows: S205. Solve the sample entropy S(m,r) of the sequence: In the actual signal, M is a finite value, so the sample entropy S(m,r,M) is estimated as follows: S206. Sort, classify and integrate the IMFs according to the sequence complexity and disorder degree represented by the sample entropy obtained in step S205 to obtain the classified and integrated components IN-IMFs.

4. The method for denoising and enhancing nonlinear ultrasonic detection signals according to claim 1, characterized in that: S3, performing variational mode decomposition on the classified and integrated IN-IMFs to obtain further decomposition results of each IN-IMF, namely, intrinsic mode functions IMFs, specifically includes the following steps: S301. Set input: signal to be decomposed The number of modes to be decomposed K; S302, initialization: each mode and its center frequency The initial value of is set to 0, and the Grange multiplier Set the initial value to 0, set n = 0; S303, let n = n + 1, and use the following formula to update each mode Center frequency and the Lagrange multiplier Where k = 1, 2, ..., K; γ represents the noise tolerance. When the signal contains strong noise, set γ = 0 to achieve better denoising effect. represents the Fourier transform of f(t); Indicates u k (t) The nth iteration value in the Fourier domain; S304, judging the iteration termination condition according to the following formula, if it is satisfied, outputting the K patterns obtained by decomposition, otherwise returning to step S303 to continue iteratively updating and decomposing; S305. The whole process is solved in the frequency domain. Perform inverse Fourier transform to obtain IMFs.

5. The method for denoising and enhancing nonlinear ultrasonic detection signals according to claim 1, characterized in that: In S4, the training data set is input into the gated recurrent unit model for training, which specifically includes the following steps: 1) Calculate the reset gate value: According to the input and the system state at the previous moment, determine how much information in the previous stage needs to be forgotten, as shown in the following formula: r t =σ(W r x t +U r h t-1 +b r ) (22) Among them, r t ∈[0,1] D is the reset gate, σ is the Logistic function, whose output interval is (0, 1), W r and U r are all network training parameter matrices, x t is the input at the current moment, h t-1 is the system state at the previous moment, b r is the bias term; 2) Calculate the intermediate state value: Among them, ⊙ is the Hadamard product operator of the vector, which performs a dot multiplication operation, indicating the multiplication of each element; W h and U h are all network training parameter matrices; b h is the bias term; is the intermediate state value; Depends on the state h of the previous moment t-1 ; When r t = 0, candidate state Only with the current input x t It is related to the historical state; when r t =1, candidate state and the current input x t And the historical status h t-1 Related; 3) Calculate the update gate value according to formula (24): The update gate determines how much information to retain in the previous state and how much new information to accept from the intermediate state based on the input and the system state at the previous moment. This process does not undergo nonlinear transformation. z t =σ(W z x t +U z h t-1 +b z ) (24) Among them, z t ∈[0,1] D is the update gate; W z and U z are all network training parameter matrices; b z is the bias term; 4) Calculate the output value: Among them, h t is the state vector output at time t. When zt is 0, the current state h t and the previous state h t-1 There is a nonlinear functional relationship between them. When z t When it is 1, the current state h t and the previous state h t-1 There is a linear function relationship between them.

6. A nonlinear ultrasonic detection signal denoising and feature enhancement system, used to implement the nonlinear ultrasonic detection signal denoising and feature enhancement method according to any one of claims 1 to 5, characterized in that: include: CEEMDAN decomposition module, used to perform CEEMDAN decomposition on the nonlinear ultrasonic signal of the detected concrete damage to obtain multiple intrinsic mode components IMFs; The classification and integration module is used to classify and integrate the decomposed IMFs using the sample entropy value to obtain the classified and integrated components IN-IMFs; The variational mode decomposition module is used to perform variational mode decomposition on the classified and integrated IN-IMFs to obtain a set of intrinsic mode functions IMFs; The result output module is used to divide the intrinsic mode functions IMFs into a training data set and a test data set according to the time step, input the training data set into the gated recurrent unit model for training, and output each IN-IMF denoising and feature enhancement result.

7. The nonlinear ultrasonic detection signal noise reduction and feature enhancement system according to claim 6, characterized in that: The system also includes a nonlinear ultrasonic detection device for detecting concrete damage and obtaining a nonlinear ultrasonic signal of concrete damage; the nonlinear ultrasonic detection device includes a signal generator, a signal amplifier, an oscilloscope, a transmitting transducer and a receiving transducer.

Images