Laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning

Abstract

The invention discloses a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning, which relates to the field of nondestructive testing, and comprises the following steps: detecting a tested structure by adopting laser detection and calculating a frequency dispersion curve of the tested structure to obtain an analysis frequency band of a detection signal, and further constructing a basic detection model; acquiring a signal data set of laser detection of the detected structure, calculating to obtain a multi-source loss function, constraining the loss function into the basic detection model to obtain a damage depth migration detection model, and performing iterative training on the damage depth migration detection model until the number of iterations reaches a set threshold value, or obtaining an interpretable deep learning detection model when the model reaches a convergence condition, using the interpretable deep learning detection model to identify a laser ultrasonic guided wave detection sample of the detected structure, and outputting a corresponding label of the detection sample. By the adoption of the method, the problems that a traditional laser ultrasonic guided wave detection signal detection model is poor in interpretability and insufficient in applicability are solved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing, and in particular to a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning. Background Technology

[0002] Non-destructive testing based on ultrasonic guided waves has attracted increasing attention due to its wide monitoring range. Excitation methods for ultrasonic guided waves include electromagnetic ultrasound, piezoelectric ultrasound, and laser ultrasound. Among these, laser ultrasonic guided wave testing technology is considered a highly attractive damage detection technique due to its advantages such as non-contact operation, visualization, long working distance, and high sensitivity.

[0003] Among various laser detection methods, laser spatial light modulation excitation can achieve focusing at any position and has great potential in laser detection. However, the propagation mechanism of laser ultrasonic guided wave field in complex structures is still unclear, and the characteristics of the detection signal are complex and difficult to distinguish.

[0004] Deep learning directly obtains the mapping relationship between ultrasonic guided wave signals and structural states from the detection data, and constructs a damage detection model by quantifying the high-dimensional feature distance of the detection signals under different states of the tested object. Because it avoids the complex prior knowledge requirements caused by guided wave dispersion and multimodal characteristics, it has attracted much attention in the field of laser ultrasonic guided wave nondestructive testing. However, current deep learning detection methods still mainly focus on model design, and their integration with the nondestructive testing mechanism is not deep enough, making it difficult to guarantee the stability and reliability of the models.

[0005] To address these issues, there is an urgent need for laser ultrasonic guided wave damage detection methods based on interpretable deep transfer learning. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning. This overcomes the limitation of existing laser ultrasonic detection signal feature mining techniques in mining complex vibration signals, thereby improving the detection accuracy and damage identification precision of laser ultrasonic damage detection. The specific steps are as follows:

[0007] A laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning includes the following steps:

[0008] S1: Laser detection is used to detect the structure under test and calculate the dispersion curve of the structure under test to obtain the analysis frequency band of the detection signal. The frequency band with relatively simple modal component type and weak frequency shift phenomenon in the dispersion curve is selected as the frequency band for feature extraction of the detection signal, i.e., the analysis frequency band of the detection signal. The frequency band signal is further analyzed by filtering.

[0009] S2: Construct a guided wave waveform dictionary using an atomic library, extract guided wave packet information from the analysis frequency band signal based on the guided wave waveform dictionary, and then construct a basic detection model;

[0010] S3: Obtain the signal dataset of laser detection of the structure under test, calculate the multi-source loss function based on the signal dataset of laser detection, and constrain the loss function into the basic detection model to obtain the damage depth transfer detection model.

[0011] S4: Iteratively train the damage depth transfer detection model until the number of iterations reaches a set threshold, or the model reaches the convergence condition to obtain an interpretable deep learning detection model.

[0012] S5: Use an interpretable deep learning detection model to identify laser ultrasonic guided wave detection samples of the structure under test, and output the corresponding labels of the detection samples.

[0013] Preferably, the laser detection method in S1 includes two types: single-point detection and modulation detection.

[0014] The excitation energy for single-point detection is a Gaussian beam, and the detection signal is used to determine whether the position of the excitation point is damaged.

[0015] Modulation detection enables damage detection at any specified location by optimizing the excitation shape.

[0016] Preferably, in S1, the dispersion curve of the structure under test is calculated by numerical solution, and the frequency of the target wave velocity change is selected as the analysis frequency band of the detection signal.

[0017] Preferably, the specific content of constructing the guided wave waveform dictionary through the atomic library in S2 is as follows:

[0018] The atomic library is constructed based on the signal characteristics of the guided wave signal;

[0019] Let the excitation signal be u(t), the waveguide propagation distance be x, and the waveguide signal after propagation x be w(x,t). Then the waveguide signal at x=0 is:

[0020] w(x)=w(x,t)| x＝0 =u(t);

[0021] In the frequency domain, it can be represented as:

[0022]

[0023] According to the steady-state phase method, x = x p The Fourier transform of the signal at point can be expressed as:

[0024]

[0025] Where k(ω) represents the wavenumber function with frequency ω. Taking the inverse Fourier transform of the above equation, we get x = x p The time-domain signal at that location.

[0026]

[0027] Given the initial guided wave signal and the dispersion relation of the guided wave in the structure, the guided wave waveform after the initial signal has propagated any distance can be obtained. This waveform can be expressed in terms of propagation distance and wavenumber corresponding to the mode as (x p ,k(ω)), that is:

[0028] φ(x p ,k(ω))=w(x p ,t);

[0029] By setting the detection distance and accuracy parameters one by one, a series of guided wave dispersion signals can be obtained as atoms. Given a set X of equally spaced propagation distances;

[0030] X={x n |x n =x0+nΔx,n∈N,Δx=x1-x0};

[0031] The dispersive sub-dictionary D is a series of signals obtained by propagating the excitation signal in mode m over a distance X. m It can be represented as:

[0032] D m ={φ(x1,k m (ω)),φ(x2,k m (ω)),...,φ(x n ,k m (ω))};

[0033] Given a set of wavenumber functions for M modes {k1(ω),k2(ω),...,k M (ω)}, then the resulting dispersion dictionary D sum It can be represented as:

[0034] D sum ={D1,D2,...,D M};

[0035] Each sub-dictionary consists of n+1 atomic signals, and there are a total of M sub-dictionaries, then dictionary D sum There are a total of (n+1)M atomic signals.

[0036] Preferably, the specific content of extracting guided wave packet information from the analysis frequency band signal based on the guided wave waveform dictionary, and then constructing the basic detection model, is as follows:

[0037] The basic detection model is established using a one-dimensional convolutional neural network.

[0038] Through the atomic dictionary D sum Extract (n+1)M guided wave packet atomic signals, determine the time-domain signal sampling frequency, and obtain the data length and envelope information of the atomic signals;

[0039] The data length and value of the convolution kernel of the convolutional layer of the basic detection model are determined based on the wave packet data length.

[0040] The matrix parameters for determining the kernel weights are based on the envelope shape.

[0041] Preferably, the one-dimensional convolutional neural network includes a feedforward layer, a convolutional layer, and a pooling layer;

[0042] The output features of the convolutional layer are:

[0043]

[0044] In the formula, Convolution(x) represents the convolutional layer, x represents the input feature of the convolutional layer, and Ker represents the kernel weights. The kernel data lengths of the first two convolutional layers are related to the atomic dictionary D. sum The (n+1)M guided wave packets in the diagram have consistent atomic signals, and b is the deviation term. The convolution between Ker and x is calculated, and σ(·) is the activation function that generates a non-linear mapping between the input and output. It is mainly used to introduce non-linearity into the data, and here it is the ReLU activation function.

[0045] The output characteristics of the pooling layer are:

[0046] h = Pooling(z);

[0047] In the formula, z is the output feature of the convolutional layer, and Pooling(z) is the max pooling layer, which is used to calculate the maximum value of the elements within the pooling window.

[0048] Preferably, the laser detection signal dataset of the structure under test is obtained, and the specific content of the multi-source loss function is calculated based on the laser detection signal dataset as follows:

[0049] Acquire the signal dataset of laser detection of the structure under test, and extract key frequency band information through filtering;

[0050] The signal dataset from laser detection is input into the detection model to extract high-dimensional features of the signal in the dataset.

[0051] The signals acquired under the target detection conditions are labeled as different source domains, and the multi-source loss function is calculated.

[0052] The damage depth migration detection model is obtained by optimizing the basic detection model through loss function feedback, and the model is confirmed by a preset number of iterations.

[0053] Preferably, the multi-source domain damage function of the damage depth migration detection model is:

[0054] E(θ f ,θ y ,θ d ) = L ay (θ f ,θ y )+L ad (θ f ,θ d );

[0055] In the formula, E is the total loss value, and L ay For structural health status label identification loss, L ad For domain label recognition loss, θ y and θ d These are the sample conditional classifiers G y Domain Discriminator G d The parameter vector in;

[0056] Domain label recognition loss L ad The expression is:

[0057]

[0058] In the formula, The loss term is the label of the j-th source domain, where k is the number of source domains and the classification loss is for a single source domain. Defined as:

[0059]

[0060] In the formula, n is the total number of samples, d is the actual domain label, and λ is the domain label. d The parameters control the weights between the sample state label and the domain label loss during the learning process.

[0061] Preferably, the structural health status label recognition loss is based on a label error weighting factor, expressed as:

[0062]

[0063] If the first n samples in the j-th source domain are defined as normal, then the state label prediction loss for the j-th source domain samples is... for:

[0064]

[0065] In the formula, the feature vector is generated by the label predictor G. yMapping to labels y, θ y Let α be the parameter vector of the mapping. l Indicates the label error weighting factor, l y This is the sample state label prediction loss. The structural health state label loss term is calculated using the negative log probability loss, and its calculation formula is as follows:

[0066]

[0067] Its output is:

[0068] G y (G f (x i ); V2,c2)=f(V2G f (x i )+c2);

[0069] In the formula, V2 and c2 are calculation parameters, and f is the activation function softmax.

[0070] Preferably, the model's output labels include domain labels and sample state labels. The sample state includes normal signals and damaged signals, meaning that the sample state prediction loss includes normal signal prediction loss and damaged signal prediction loss.

[0071] A laser ultrasonic guided wave damage detection device based on interpretable deep transfer learning, used to implement a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning, comprising:

[0072] Nanosecond pulsed laser, spatial light modulator, laser receiving probe, laser ultrasound receiving module, and computer;

[0073] The laser exciter is connected to the laser excitation probe, and the laser ultrasonic receiver module is connected to the computer.

[0074] Among them, the laser exciter is used to excite the laser signal and excite the object under test through the spatial light modulator;

[0075] The laser detection probe is used to acquire the vibration signal of the device under test by laser ultrasonic testing and transmit it to the laser ultrasonic receiving module through optical fiber;

[0076] The laser ultrasound receiving module is used to receive the vibration signal from the laser ultrasound detection probe and send it to the computer.

[0077] The computer is used to receive vibration signals from laser ultrasonic testing, perform damage detection on the vibration signals, and output the damage detection results.

[0078] In summary, the laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning of the present invention has the following advantages compared with traditional techniques:

[0079] This invention first clarifies the key frequency band information of the object under test by calculating the dispersion curve and constructs the guided wave atom library information; secondly, it constructs a feature extraction module for a one-dimensional deep migration detection model using the atom library information; thirdly, it sets the objective function of the multi-source domain migration detection model and aligns different signal feature spaces, solving the problems of poor interpretability and insufficient applicability of traditional laser ultrasonic guided wave detection signal detection models, and improving the recognition accuracy of laser guided wave detection data.

[0080] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning in the embodiment.

[0082] Figure 2 This is a diagram of the damage depth migration detection model architecture in the embodiment;

[0083] Figure 3 This is a schematic diagram of the laser ultrasonic damage detection device in the embodiment;

[0084] Figure 4 The image shown is of a U-shaped bend with simulated damage in the embodiment.

[0085] Explanation of icon numbers:

[0086] 1. Laser exciter; 2. Scanning frame; 3. Laser ultrasonic receiver module; 4. Laser excitation probe; 5. Laser detection probe; 6. Computer. Detailed Implementation

[0087] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0088] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0089] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0090] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0091] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0092] This invention provides a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0093] S1: Laser detection is used to inspect the structure under test and calculate the value of the copper pipe under test (such as U-bends). Figure 4 The dispersion curve of the structure (as shown) is used to obtain the analysis frequency band of the detection signal.

[0094] Furthermore, the laser detection method in S1 is a single-point detection method: the excitation energy for single-point detection is a Gaussian beam, and the detection signal is used to determine whether the position of the excitation point is damaged.

[0095] Furthermore, in S1, the dispersion curve of the copper pipe structure under test is calculated using a numerical solution method, and the frequency of the target wave velocity change is selected as the analysis frequency band of the detection signal. Here, the three main guided wave modes in the pipe are mainly considered: the first-order torsional mode T0, the bending mode B0, and the longitudinal mode L0.

[0096] S2: Construct a guided wave waveform dictionary using an atomic library, extract guided wave packet information from the analysis frequency band signal based on the guided wave waveform dictionary, and then construct a basic detection model.

[0097] Furthermore, the specific content of constructing the guided wave waveform dictionary through the atomic library in S2 is as follows:

[0098] The atomic library is constructed based on the signal characteristics of the guided wave signal;

[0099] Let the excitation signal be u(t), the waveguide propagation distance be x, and the waveguide signal after propagation x be w(x,t). Then the waveguide signal at x=0 is:

[0100] w(x)=w(x,t)| x＝0 =u(t);

[0101] In the frequency domain, it can be represented as:

[0102]

[0103] According to the steady-state phase method, x = x p The Fourier transform of the signal at point can be expressed as:

[0104]

[0105] Where k(ω) represents the wavenumber function with frequency ω. Taking the inverse Fourier transform of the above equation, we get x = x p The time-domain signal at that location,

[0106]

[0107] Given the initial guided wave signal and the dispersion relation of the guided wave in the structure, the guided wave waveform after the initial signal has propagated any distance can be obtained. This waveform can be expressed in terms of propagation distance and wavenumber corresponding to the mode as (x p ,k(ω)), that is: φ(x p ,k(ω))=w(x p ,t);

[0108] Here, the detection distance is set to 50mm and the positional accuracy to 2mm, allowing the acquisition of a series of guided wave dispersion signals as atoms. Given a set of equally spaced propagation distances X; X = {x n |x n =x0+nΔx,n∈N,Δx=x1-x0};

[0109] The dispersive sub-dictionary D is composed of a series of signals obtained by propagating the excitation signal in mode m over a distance of 50 mm. m It can be represented as:

[0110] D m ={φ(x1,k m (ω)),φ(x2,k m (ω)),...,φ(x 25 ,k m (ω))};

[0111] Given a set of wavenumber functions for a certain mode {k1(ω),k2(ω),...,k n (ω)}, then the total dispersion dictionary D of the three guided wave modes, namely the first-order torsional mode T0, bending mode B0, and longitudinal mode L0, is obtained. sum It can be represented as: D sum ={D1,D2,D3}.

[0112] Each sub-dictionary consists of 26 atomic signals, and there are 3 sub-dictionaries in total. Therefore, dictionary D... sum There are a total of 78 atomic signals.

[0113] Furthermore, the specific content of extracting guided wave packet information from the frequency band signal based on the guided wave waveform dictionary, and then constructing the basic detection model, is as follows:

[0114] The basic detection model is established using a one-dimensional convolutional neural network.

[0115] Through the atomic dictionary D sumExtract (n+1)M guided wave packet atomic signals, determine the time-domain signal sampling frequency, and obtain the data length and envelope information of the atomic signals;

[0116] The data length and value of the convolution kernel of the convolutional layer of the basic detection model are determined based on the wave packet data length.

[0117] The matrix parameters for determining the kernel weights are based on the envelope shape.

[0118] Furthermore, the one-dimensional convolutional neural network includes a feedforward layer, a convolutional layer, and a pooling layer, and the output features of the convolutional layer are:

[0119] In the formula, Convolution(x) represents the convolutional layer, x represents the input feature of the convolutional layer, and Ker represents the kernel weights. The kernel data lengths of the first two convolutional layers are related to the atomic dictionary D. sum The (n+1)M guided wave packets in the diagram have consistent atomic signals, and b is the deviation term. The convolution between Ker and x is calculated, and σ(·) is the activation function that generates a nonlinear mapping between the input and output. It is mainly used to introduce nonlinearity into the data, and here it is the ReLU activation function.

[0120] The output characteristic of the pooling layer is: h = Pooling(z);

[0121] In the formula, z is the output feature of the convolutional layer, and Pooling(z) is the max pooling layer, which is used to calculate the maximum value of the elements within the pooling window.

[0122] The one-dimensional convolutional network consists of six layers. The kernel size of the first layer is the same as the length of the atomic signal, and it has 78 channels. The convolutional layers and pooling layers of the last five layers are configured similarly. The specific network structure of the copper pipe detection network in this invention is shown in Table 1.

[0123] Table 1 Network Structure

[0124]

[0125]

[0126] This approach allows for the rapid extraction of modal features from the network in the first convolutional layer, ensuring that the features of each modal unit are quickly and fully extracted. This significantly accelerates the network's convergence speed and improves its feature extraction capabilities.

[0127] S3: Obtain the signal dataset of laser detection of the structure under test, calculate the multi-source loss function based on the signal dataset of laser detection, and constrain the loss function into the basic detection model to obtain the damage depth transfer detection model.

[0128] The specific content of the multi-source loss function calculated based on the laser detection signal dataset of the structure under test is as follows: the laser detection signal dataset of the structure under test is obtained, and the laser is set into multiple source domains according to the incident angle of the laser. Key frequency band information is extracted by filtering.

[0129] This application uses a U-shaped tube for experiments, and the laser incident angle error of the U-shaped tube in the experiment is 90°±15.0°.

[0130] The experiment included two labels: normal and damaged. The damaged pipes were of the same size and located at random positions on the circumference.

[0131] The acquired dataset is as follows: 1) Training set: A total of 9 U-shaped pipes were used as the experimental research objects, of which 1 was normal and 8 had cracks near the bend of the pipe. Detection signals were collected from the pipes under laser excitation at incident angles of 90°, 75°, and 105°, with 20 samples collected at each location. The collected data were labeled as different source domains according to different locations. Therefore, the experiment included three source domains.

[0132] 2) Test set: A total of 6 U-shaped pipes were used for testing, of which 2 were normal and 4 had cracks and damage near the bends of the pipes.

[0133] Each pipe collects detection signals at two excitation locations: one with an incident angle of 82° and the other with an incident angle of 98°. Twenty samples are collected at each location.

[0134] To achieve consistency between positive and negative samples, the number of normal samples is increased to match that of damaged samples by adding 5% Gaussian noise. The number of samples collected for each laser incident angle in this invention is shown in Table 2.

[0135] Table 2. Number of samples collected for each laser incident angle in the case study.

[0136]

[0137] The signal dataset from laser detection is input into the detection model to extract high-dimensional features of the signals in the dataset.

[0138] The signals acquired under the target detection conditions are labeled as different source domains, and the multi-source loss function is calculated.

[0139] The damage depth migration detection model is obtained by optimizing the basic detection model through loss function feedback, and the model is confirmed by a preset number of iterations.

[0140] During the training phase of the detection model, the goal is to minimize the source domain label prediction loss, which ensures good prediction performance of the label predictor on the overall samples of the source domain. The loss of the proposed method is divided into domain label recognition loss and structural health status label recognition loss.

[0141] Furthermore, the multi-source domain damage function of the damage depth migration detection model is:

[0142] E(θ f ,θ y ,θ d ) = L ay (θ f ,θ y )+L ad (θ f ,θ d ).

[0143] In the formula, E is the total loss value, and L ay For structural health status label identification loss, L ad For domain label recognition loss, θ y and θ d These are the sample conditional classifiers G y Domain Discriminator G d The parameter vector in.

[0144] As can be seen from the formula, minimizing the target object is to obtain the saddle point parameters of the function. The calculation method is as follows:

[0145]

[0146]

[0147] Domain label recognition loss L ad The expression is:

[0148]

[0149] In the formula, The loss term is the label of the j-th source domain, where k is the number of source domains and the classification loss is for a single source domain. It can be defined as:

[0150]

[0151] In the formula, n is the total number of samples, d is the actual domain label, and λ is the domain label. d The parameters control the weights between the sample state label and the domain label loss during the learning process. Since the parameters in the equation cannot simultaneously achieve the desired weights through G... f Minimize L d and through G dMaximize L d Therefore, in G d Adding a minus sign at the beginning makes the optimization direction consistent.

[0152] Furthermore, the structural health status label recognition loss is based on a label error weighting factor, expressed as:

[0153]

[0154] To adjust the prediction loss contribution of normal and damaged signal tags, this application proposes a tag error weighting factor α. l If the first n samples in the j-th source domain are defined as normal, then the state label prediction loss for the j-th source domain samples is... for:

[0155]

[0156] In the formula, the feature vector is generated by the label predictor G. y Mapping to labels y, θ y Let α be the parameter vector of the mapping. l Indicates the label error weighting factor, l y This is the sample state label prediction loss. The structural health state label loss term is calculated using the negative log probability loss, and its calculation formula is as follows:

[0157] Its output is: G y (G f (x i ); V2,c2)=f(V2G f (x i )+c2).

[0158] In the formula, V2 and c2 are calculation parameters, and f is the activation function softmax.

[0159] Furthermore, the sample state prediction loss includes normal signal prediction loss and damaged signal prediction loss.

[0160] The weights for different source domains are determined using adaptive weight coefficients in the formula. During model training, the training objective is to minimize the total loss function value through stochastic gradient descent. After training, the model's performance is tested using test samples from the target domain.

[0161] S4: Iteratively train the damage depth transfer detection model until the number of iterations reaches a set threshold, or the model reaches the convergence condition to obtain an interpretable deep learning detection model.

[0162] S5: An interpretable deep learning detection model is used to identify laser ultrasonic guided wave detection samples of the tested structure and output the corresponding labels for the detected samples. The simulated damage inside the pipe is crack damage, and the labels are divided into two categories: normal samples and damaged samples. The effectiveness of the method is not explained, but the proposed method is compared with current classical methods. The identification results (test signal detection accuracy) of different methods are shown in Table 3.

[0163] Table 3. Signal detection accuracy (%) tested using different methods

[0164] Migration task S→T1 S→T2 average This application method 97.5 96.2 96.9 LeNet5 90.0 86.2 88.1 ResNet 93.7 88.7 91.2 DAN 88.7 90.0 89.4

[0165] A laser ultrasonic guided wave damage detection device based on interpretable deep transfer learning is provided to implement a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning, such as... Figure 3 As shown, it includes: a nanosecond pulsed laser, a spatial light modulator, a laser receiving probe, a laser ultrasound receiving module 3, and a computer 6.

[0166] The laser exciter 1 is connected to the laser excitation probe 4, and the laser ultrasonic receiver module 3 is connected to the computer 6.

[0167] Among them, the laser exciter 1 is used to excite the laser signal and excite the object under test through the spatial light modulator.

[0168] The laser detection probe 5 is used to acquire the vibration signal of the device under test by laser ultrasonic testing and transmit it to the laser ultrasonic receiving module 3 through optical fiber.

[0169] The laser ultrasound receiving module 3 is used to receive the vibration signal of laser ultrasound detection sent by the laser detection probe 5 and send it to the computer 6.

[0170] Computer 6 is used to receive vibration signals from laser ultrasonic testing, perform damage detection on the vibration signals, and output the damage detection results.

[0171] An electronic device includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, characterized in that the processor executes the computer program to implement the steps of a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning.

[0172] The electronic device of the present invention can execute the non-contact laser ultrasonic damage detection method based on deep transfer learning of the present invention, and can execute any combination of implementation steps of the method embodiments, and has the corresponding functions and beneficial effects of the method.

[0173] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a laser ultrasonic guided wave damage detection method based on interpretable deep transfer learning.

[0174] The computer storage medium of the present invention stores instructions or programs that can execute the non-contact laser ultrasonic damage detection method based on deep transfer learning of the present invention. When the instructions or programs are run, any combination of implementation steps of the method embodiment can be executed, and the method has the corresponding functions and beneficial effects.

[0175] The technical solution of this invention can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0176] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A laser-ultrasonic guided wave damage detection method based on explainable deep transfer learning, characterized in that, The method comprises the following steps: S1: detecting the measured structure by laser detection and calculating the dispersion curve of the measured structure to obtain an analysis frequency band of the detection signal; S2: constructing a guided wave waveform dictionary from an atomic library, extracting guided wave packet information in the analysis frequency band signal according to the guided wave waveform dictionary, and then constructing a basic detection model; S3: obtaining a signal data set of the measured structure by laser detection, calculating a multi-source loss function according to the signal data set of the laser detection, and constraining the loss function into the basic detection model to obtain a damage depth transfer detection model; S4: iteratively training the damage depth transfer detection model until the number of iterations reaches a set threshold or the model reaches a convergence condition to obtain an interpretable deep learning detection model; S5: using the interpretable deep learning detection model for laser ultrasonic guided wave detection of the measured structure to identify the recognition of the measured structure, and outputting the corresponding label of the detection sample.

2. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 1, characterized in that, The laser detection method in S1 includes single-point detection and modulation detection: The excitation energy of single-point detection is a Gaussian beam, and the detection signal is used to judge whether the position of the excitation point is damaged; The modulation detection realizes damage detection at any specified position by optimizing the excitation shape.

3. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 1, characterized in that, The specific content of constructing a guided wave waveform dictionary from an atomic library in S2 is: The atomic library is constructed according to the signal characteristics of the guided wave signal; The preset excitation signal is u(t), the guided wave propagation distance is x, and the guided wave signal after propagation x is w(x, t), t is the time domain guided wave signal; The guided wave signal at x=0 is: w(x) = w(x, t) x＝0 = u(t); The guided wave signal at x=0 is expressed in the frequency domain as: According to the steady phase method, the signal Fourier transform at any point p, x = x p is expressed as: where k(ω) represents the wave number function at frequency ω, ω is the angular frequency, i is the imaginary unit, x p the position to which the guided wave signal propagates; Taking inverse Fourier transform of the above equation, we get the time domain signal at x = x p : Given the initial guided wave signal and the dispersion relation of the guided wave in the structure, the guided wave waveform after the initial signal propagates any distance is obtained, and the waveform is expressed as (x p ,k(ω)) corresponding to the propagation distance and the wave number of the mode, i.e. φ(x p ,k(ω)) = w(x p ,t); Where φ is a scalar potential, The detection distance and accuracy parameters are set one by one to obtain a series of guided wave dispersion signals as atoms, and a set of equal-interval propagation distances X is given: X = {x p | x p = x0+ nΔx, p e N, Δx = x1- x0} ; where p is an arbitrary discrete point position, and D is a dispersion sub-dictionary consisting of a series of signals resulting from the propagation of the excitation signal X over a distance X in mode m m is represented as: D m = {φ(x1,k m (ω)),φ(x2,k m (ω)),...,φ(x n ,k m (ω))} ; Given a set of wavenumber functions {k1(ω), k2(ω),..., kM(ω)} for M modalities, the resulting dispersion dictionary D n sum is represented as:​ D sum = {D1, D2,..., D M}; where each sub-dictionary consists of n+1 atomic signals, and there are M sub-dictionaries, then the dictionary D sum contains (n+1)M atomic signals.

4. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 3, characterized in that, The specific content of extracting guided wave packet information in the analysis frequency band signal according to the guided wave waveform dictionary and then constructing a basic detection model is: The basic detection model selects a one-dimensional convolutional neural network; By the atomic dictionary D sum Extracting (n+1)M guided-wave packet atomic signals, after determining the time-domain signal sampling frequency, the data length and envelope information of the atomic signal are obtained; The data length and value of the convolution kernel of the convolution layer of the basic detection model are determined according to the packet data length; The matrix parameters of the convolution kernel weight are determined according to the envelope shape.

5. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 1, characterized in that, The one-dimensional convolutional neural network includes a feedforward layer, a convolution layer, and a pooling layer; The output feature of the convolution layer is: wherein Convolution(x) is a convolution layer, x is an input feature of the convolution layer, Ker is a convolution kernel weight, b is a bias term, is a convolution calculation between Ker and x, and σ(·) is an activation function that generates a non-linear mapping between an input and an output; The output feature of the pooling layer is: h=Pooling(z); In the formula, z is the output feature of the convolution layer, Pooling(z) is the maximum pooling layer, and the function is to calculate the maximum value of the elements in the pooling window.

6. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 1, characterized in that, The specific content of obtaining a signal data set of the measured structure by laser detection and calculating a multi-source loss function according to the signal data set of the laser detection is: Obtain the signal data set of the measured structure by laser detection and extract the key frequency band information by filtering; Input the signal data set of the laser detection into the detection model to extract the high-dimensional features of the signals in the signal data set; Mark the signals obtained by the target detection condition as different source domains, and calculate the multi-source loss function; Optimize the basic detection model by the loss function to obtain the damage depth transfer detection model, and realize model confirmation by the pre-set number of iterations.

7. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 6, characterized in that, The multi-source domain damage function of the damage depth transfer detection model is: E(θ f ,θ y ,θ d ) = L ay (θ f ,θ y ) + L ad (θ f ,θ d ); where E is the total loss value, L ay is the structural health state label identification loss, L ad is the domain label identification loss, θ y and θ d are the parameter vectors in the sample condition classifier G y and the domain discriminator G d respectively. Domain label identification loss L ad The expression is: wherein is the j-th source domain label loss term, k is the number of source domains, and the single source domain classification loss is defined as: where N is the total number of samples, d is the actual domain label, λ d is a parameter to control the weight between the sample state label loss and the domain label loss in the learning process, is the signal of the i-th sample in the j-th source domain, is the domain label loss value in the j-th source domain.

8. The laser ultrasonic guided wave damage detection method based on explainable deep transfer learning according to claim 6, characterized in that, The structural health state label recognition loss is based on a label error weight factor, and the expression is: Define the first n samples in the jth source domain as normal, then the jth source domain sample state label prediction loss is: where G f is the feature vector predicted by the tag predictor, G y maps to the label y, θ y denotes the parameter vector of the mapping, α l denotes the label error weight factor, l y is the sample state label prediction loss, the structural health state label loss term is calculated by the negative logarithmic probability loss, and the calculation formula is: The output is: G y (G f (x i ) ; V2, c2) = f(V2G f (x i )+c2) ; In the formula, V2 and c2 are calculation parameters, and f is an activation function softmax.

9. A laser-ultrasonic guided wave damage detection device based on explainable deep transfer learning, configured to implement the laser-ultrasonic guided wave damage detection method based on explainable deep transfer learning according to any one of claims 1-8, characterized in that, Comprise: A nanosecond pulse laser, a spatial light modulator, a laser receiving probe, a laser ultrasonic receiving module and a computer; The laser exciter is connected with the laser excitation probe, and the laser ultrasonic receiving module is connected with the computer; The laser exciter is used to excite the laser signal and excite it to the measured object through the spatial light modulator; The laser detection probe is used to obtain the vibration signal of the laser ultrasonic detection of the device to be measured and transmit it to the laser ultrasonic receiving module through the optical fiber; The laser ultrasonic receiving module is used to receive the vibration signal of the laser ultrasonic detection sent by the laser detection probe and send it to the computer; The computer is used to receive the vibration signal of the laser ultrasonic detection, and perform damage detection on the vibration signal and output the damage detection result.

Images