Non-contact Laser Ultrasound Damage Detection Method and System Based on Deep Transfer Learning

Abstract

The present invention discloses a non-contact laser ultrasonic damage detection method based on deep transfer learning, comprising the following steps: S1: Obtain the vibration signals of laser ultrasonic detection, divide the vibration signals into ablation signals and thermoelastic signals, and label the ablation signals; S2: Respectively extract the low-frequency component features of the ablation signals and thermoelastic signals through wavelet decomposition to obtain the feature maps of the ablation signals and thermoelastic signals; S3: Construct a damage detection model, input the feature map of the thermoelastic signal into the laser energy mapping network, map the thermoelastic signal to the ablation signal, and obtain the feature map of the mapped thermoelastic signal; S4: Input the feature maps of the ablation signal and the mapped thermoelastic signal into the feature extraction network simultaneously to extract the features of the ablation signal and the mapped thermoelastic signal; S5: Iteratively train the damage detection model until the number of iterations reaches the set threshold; S6: Use the damage detection model to perform damage detection on the thermoelastic signal of the device to be measured and output the damage detection result.

Description

Technical Field

[0001] The present invention relates to the field of non-destructive testing, and more specifically, to a non-contact laser ultrasonic damage detection method and system based on deep transfer learning. Background Art

[0002] Non-destructive testing based on ultrasonic guided waves has received increasing attention due to its wide monitoring range. The excitation methods of ultrasonic guided waves include electromagnetic ultrasound, piezoelectric ultrasound, and laser ultrasound, etc. Among them, the laser ultrasonic guided wave detection technology is considered to be an extremely attractive damage detection technology due to its non-contact, visualization, long working distance, high sensitivity and other advantages.

[0003] Among various laser detection methods, laser scanning imaging detection has been widely used because it can realize damage visualization. Commonly used laser wave field signal processing methods include adjacent scan point difference analysis, standing wave energy and wavenumber frequency domain analysis, etc. With the development of artificial intelligence, the mapping relationship between ultrasonic guided wave signals and structural states can be directly obtained from big data. These data-driven methods can avoid complex prior knowledge caused by frequency dispersion and multimode characteristics. In order to further improve the intelligence of data-driven monitoring technology, deep transfer learning has emerged to solve the generality and generalization problems of monitoring models in different scenarios. Deep transfer learning can use the existing knowledge in the source domain to solve related target domain problems.

[0004] According to the different energy intensities, laser excitation can be divided into thermoelastic mechanism and ablation mechanism. The laser energy intensity excited by the thermoelastic mechanism is relatively small, and the laser irradiation energy absorbed by the surface of the device to be tested will not melt the structure. The structure will generate elastic stress due to the rapid expansion of volume caused by the increase in temperature. The signal generated by thermoelastic excitation is mainly surface wave, and its internal damage detection ability is poor. The laser energy intensity excited by the ablation mechanism is extremely high. However, under ablation excitation, due to the rapid rise in temperature, the surface of the structure will melt. Although ablation excitation detection is not a non-destructive testing method, internal damage can be detected by the high-energy radiation pressure generated under ablation excitation. The thermo-expansion effect and radiation pressure effect exist simultaneously, and the contribution to the response signal depends on the laser energy intensity. However, it is difficult to explore the specific relationship between thermoelastic signals and ablation signals. Summary of the Invention

[0005] The object of the present invention is to overcome the problem of poor internal damage detection ability of the existing laser thermoelastic excitation non-contact detection, and propose a non-contact laser ultrasonic damage detection method and system based on deep transfer learning, which improves the detection accuracy and damage recognition accuracy of laser ultrasonic damage detection.

[0006] In order to achieve the above object, the technical solutions adopted by the present invention are as follows:

[0007] A non-contact laser ultrasonic damage detection method based on deep transfer learning, comprising the following steps:

[0008] S1: Obtain the vibration signal of laser ultrasonic detection through the laser detection method, divide the vibration signal into ablation signals and thermo-elastic signals according to the difference in laser excitation energy, and label the ablation signals. The ablation signals are the source domain, and the thermo-elastic signals are the target domain;

[0009] S2: Perform wavelet decomposition on the ablation signals and thermo-elastic signals respectively, extract the low-frequency component features of the ablation signals and thermo-elastic signals, and obtain the feature maps of the ablation signals and thermo-elastic signals;

[0010] S3: Construct a damage detection model, input the feature map of the thermo-elastic signal into the laser energy mapping network of the damage detection model, map the feature space of the thermo-elastic signal to the feature space of the ablation signal, and obtain the feature map of the mapped thermo-elastic signal;

[0011] S4: Input the feature maps of the ablation signal and the mapped thermo-elastic signal into the feature extraction network of the damage detection model at the same time, and extract the features of the ablation signal and the mapped thermo-elastic signal respectively;

[0012] S5: Set the objective function of the damage detection model, perform iterative training on the damage detection model until the number of iterations reaches the threshold set by the damage detection model;

[0013] S6: Obtain the trained damage detection model, use the damage detection model to detect the damage of the thermo-elastic signal of the device to be measured, and output the damage detection result.

[0014] Further, in step S1, the laser detection method includes two methods: single-point detection and scanning imaging detection. The relative positions of the detection point and the excitation point in single-point detection are fixed, and the detection signal is used to judge whether the excitation point is damaged; Scanning imaging detection realizes the detection of the scanning area by quickly moving the excitation point, and performs imaging according to whether each excitation point is determined to be damaged.

[0015] Further, in step S3, the laser energy mapping network is selected to be established by a one-dimensional convolutional neural network. The one-dimensional convolutional neural network includes a feedforward layer, a convolutional layer and a pooling layer. The output feature of the convolutional layer is:

[0016]

[0017] In the formula, x is the input feature of the convolutional layer, Ker is the convolutional kernel weight, b is the bias term, is the convolutional calculation between Ker and x, and σ is the activation function that generates a non-linear mapping between the input and the output;

[0018] The output feature of the pooling layer is:

[0019] h = Pooling(z)

[0020] Where z is the output feature of the convolutional layer.

[0021] Furthermore, in step S4, the feature extraction network is a Resnet network. The feature extraction network includes multiple residual units. The input of the feature extraction network is x, the learned feature is defined as H(x), the residual is F(x) = H(x) - x, and the initially learned feature is F(x) + x. The output x l+1 is expressed as

[0022] x l+1 = f(h(x l ) + F(x l , W l ))

[0023] Where x l represents the input of the l-th residual unit, W l is the weight matrix of the l-th residual unit, F is the residual function, h(x l ) = x l represents the identity mapping, and f is the Relu activation function.

[0024] Furthermore, the learned features from shallow l to deep L are as follows:

[0025]

[0026] Where x l represents the output of the shallow residual unit, x L represents the output of the deep residual unit, x i represents the input feature of the i-th residual unit, W i represents the weight matrix of the i-th residual unit, F is the residual function, and L represents the total number of units of the deep residual unit.

[0027] Furthermore, in step S5, the objective function of the damage detection model is obtained by weighted summation of the margin loss function L Margin and the MMD loss function L MMD . The purpose of the margin loss function L Margin is to minimize the classification error in a supervised manner based on the labeled source domain data. The margin loss function is used to narrow the intra-class distance and widen the inter-class distance. The purpose of the MMD loss function is to minimize the distribution difference between the source domain and the target domain. The objective function is:

[0028] L EMTN = αL Margin + βL MMD

[0029] Where α is the margin loss function LMargin The weight coefficient, and β is the MMD loss function L MMD 's weight coefficient.

[0030] Furthermore, the marginal loss function L Margin is:

[0031]

[0032] In the formula, is the index of the sample label, indicates that the sample belongs to the c-th class, indicates that the sample does not belong to the c-th class; is the predicted label value of the source domain sample, and the regularization parameter λ is the weighted penalty factor of the missing class loss; m- is the upper edge of the predicted label value of the source domain sample of, and m+ is the lower edge of the predicted label value of the source domain sample of. If the sample belongs to the c-th class, then is not less than m+, and if the sample does not belong to the c-th class, then is not greater than m-;

[0033] The MMD loss function L MMD is:

[0034]

[0035] In the formula, N S is the total number of samples in the source domain, N T is the total number of samples in the target domain, is the output feature map of the i-th sample in the source domain, is the output feature map of the j-th sample in the target domain, and f is the Relu activation function.

[0036] A non-contact laser ultrasonic damage detection device based on deep transfer learning, comprising: a laser exciter, a laser excitation probe, a laser ultrasonic receiving module, a laser detection probe and a computer. The laser exciter is connected to the laser excitation probe, and the laser ultrasonic receiving module is respectively connected to the laser detection probe and the computer;

[0037] Among them, the laser exciter is used to excite laser signals and transmit them to the laser excitation probe through optical fibers; the laser excitation probe is used to perform laser excitation on the device to be measured; 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 optical fibers; 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 thermoelastic signal in the vibration signal and output the damage detection result.

[0038] An electronic device includes a memory and a processor. A computer program is stored in the memory and can run on the processor. When the processor executes the computer program, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning in any one of the above are implemented.

[0039] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning in any one of the above are implemented.

[0040] Compared with the prior art, the present invention constructs a damage detection model. Through the laser energy mapping network, the feature space of the thermoelastic signal is mapped to the feature space of the ablation signal to obtain the mapped thermoelastic signal, so that the thermoelastic signal obtains the internal damage detection accuracy approximate to that of the ablation signal, and the damage detection accuracy of the laser thermoelastic excitation non-destructive testing method is improved; by setting the objective function of the damage detection model to align the feature spaces of the mapped thermoelastic signal and the ablation signal, the problem that the completely non-destructive non-contact detection method of laser thermoelastic excitation has poor ability to detect internal damage of devices is solved, and the damage recognition accuracy of laser thermoelastic non-destructive testing is improved. Description of the Drawings

[0041] Figure 1 It is a schematic flow chart of the laser ultrasonic damage detection method in the embodiment.

[0042] Figure 2 It is a schematic flow chart of the laser ultrasonic damage detection method in the embodiment.

[0043] Figure 3 It is a schematic diagram of the laser ultrasonic damage detection device in the embodiment.

[0044] Figure 4 It is a physical diagram of the damaged aluminum plate in the embodiment.

[0045] Figure 5 It is a schematic comparison diagram of the detection accuracy of test signals by different methods in the embodiment.

[0046] Figure 6 It is a schematic comparison diagram of the detection results by different methods under different source domain laser energies in the embodiment.

[0047] Figure 7 It is a schematic diagram of the laser scanning imaging results by different methods in the embodiment. (a) is the imaging result of the laser energy mapping transfer (EMTN) method, (b) is the imaging result of the classification method based on the Resnet feature extraction network, (c) is the imaging result of the joint distribution adaptation (JDA) method, and (d) is the imaging result of the transfer component analysis (TCA) method.

[0048] Figure 8Schematic diagram of the imaging results of threshold damage detection by laser scanning imaging with different methods in the embodiments. (a) Threshold imaging result of the laser energy mapping and transfer (EMTN) method; (b) Threshold imaging result of the classification method based on the Resnet feature extraction network; (c) Threshold imaging result of the joint distribution adaptation (JDA) method; (d) Threshold imaging result of the transfer component analysis (TCA) method.

[0049] Explanation of the reference numerals in the drawings:

[0050] Laser exciter - 1; Scanning frame - 2; Laser ultrasonic receiving module - 3; Laser excitation probe - 4; Laser detection probe - 5; Computer - 6; Internal damage position of the aluminum plate - 7. Detailed implementation manners

[0051] The non-contact laser ultrasonic damage detection method and system based on deep transfer learning of the present invention will be further described below with reference to the drawings and specific embodiments.

[0052] Please refer to Figure 1 and Figure 2 The present invention discloses a non-contact laser ultrasonic damage detection method based on deep transfer learning, including the following steps;

[0053] S1: Obtain the vibration signals of laser ultrasonic detection through the laser detection method, divide the vibration signals into ablation signals and thermo-elastic signals according to the difference in laser excitation energy, and label the ablation signals. The ablation signals are the source domain, and the thermo-elastic signals are the target domain.

[0054] S2: Perform wavelet decomposition on the ablation signals and thermo-elastic signals respectively, and extract the low-frequency component features of the ablation signals and thermo-elastic signals to obtain the feature maps of the ablation signals and thermo-elastic signals.

[0055] S3: Construct a damage detection model, input the feature map of the thermo-elastic signal into the laser energy mapping network of the damage detection model, map the feature space of the thermo-elastic signal to the feature space of the ablation signal, and obtain the feature map of the mapped thermo-elastic signal.

[0056] S4: Input the feature maps of the ablation signal and the mapped thermo-elastic signal into the feature extraction network of the damage detection model at the same time, and extract the features of the ablation signal and the mapped thermo-elastic signal respectively.

[0057] S5: Set the objective function of the damage detection model, and perform iterative training on the damage detection model until the number of iterations reaches the threshold set by the damage detection model.

[0058] S6: Obtain the trained damage detection model, use the damage detection model to detect the damage of the thermo-elastic signal of the device to be tested, and output the damage detection result.

[0059] Deep transfer learning has been effectively verified in machine vision and fault diagnosis. The transfer learning method can improve the generalization ability of the hidden features captured by deep learning. The laser ultrasonic damage detection method proposed in the present invention can improve the internal damage detection accuracy of the thermoelastic signal by aligning the thermoelastic signal feature space with the ablation signal feature space. The non-contact laser ultrasonic damage detection method based on deep transfer learning of the present invention is a detection method based on laser energy mapping transfer (EMTN).

[0060] Specifically, in step S1, the laser detection method includes two methods: single-point detection and scanning imaging detection. The relative positions of the detection point and the excitation point in single-point detection are fixed, and the detection signal is used to determine whether the excitation point is damaged; scanning imaging detection realizes the detection of the scanning area by quickly moving the excitation point, and performs imaging according to whether each excitation point is determined to be damaged.

[0061] In step S2, the low-frequency component features of the original vibration signal are extracted by wavelet decomposition, and the laser thermoelastic excitation detection signal is mapped to the laser ablation excitation detection signal. The original vibration signal is decomposed into three layers by using the Meyer wavelet basis, and the low-frequency component features are retained.

[0062] In step S3, in order to maintain the features of the vibration signal, a one-dimensional convolutional neural network is selected to establish the laser energy mapping network. The one-dimensional convolutional neural network includes a feedforward layer, a convolutional layer and a pooling layer. The output feature of the convolutional layer is:

[0063]

[0064] In the formula, x is the input feature of the convolutional layer, Ker is the convolutional kernel weight, b is the bias term, is the convolutional calculation between Ker and x, and σ is the activation function that generates a non-linear mapping between the input and the output;

[0065] The output feature of the pooling layer is:

[0066] h = Pooling(z)

[0067] In the formula, z is the output feature of the convolutional layer.

[0068] In the laser energy mapping network, the laser thermoelastic excitation detection signal is used as the input signal, and the laser ablation excitation detection signal is used as the output signal. The feature space mapping from the laser thermoelastic excitation detection signal to the laser ablation excitation detection signal is realized, and the mapped thermoelastic signal can be obtained through this step. Compared with the laser thermoelastic excitation detection signal, the mapped thermoelastic signal has a higher signal-to-noise ratio and more radiation pressure characteristics. The accuracy and efficiency of the calculation of the laser energy mapping network are affected by the convolution kernel size and the network depth. In this embodiment, the laser energy mapping network has 8 convolutional layers and 4 max-pooling layers, and the convolution kernels have two sizes of 3×1 and 2×1.

[0069] In step S4, the feature extraction network is a Resnet network. The feature extraction network includes multiple residual units. The input of the feature extraction network is x, the learned feature is defined as H(x), the residual is F(x) = H(x) - x, and the initially learned feature is F(x) + x. When the residual is 0, the output feature of the cumulative layer is equal to the input feature. In fact, the input feature will learn new features in the cumulative layer because the residual will not be 0. The residual unit can be expressed as the output x of the l-th residual unit l+1 Expressed as

[0070] x l+1 = f(h(x l ) + F(x l , W l ))

[0071] In the formula, x l represents the input of the l-th residual unit, W l is the weight matrix of the l-th residual unit, F is the residual function, h(x l ) = x l represents the identity mapping, and f is the Relu activation function.

[0072] The learned features from the shallow layer l to the deep layer L are as follows:

[0073]

[0074] In the formula, x l represents the output of the shallow residual unit, x L represents the output of the deep residual unit, x i represents the input feature of the i-th residual unit, W i represents the weight matrix of the i-th residual unit, F is the residual function, and L represents the total number of units of the deep residual unit.

[0075] In step S5, the objective function of the damage detection model is obtained by weighted addition of the margin loss function L Margin and the MMD loss function L MMD . The objective function is:

[0076] L EMTN = αL Margin + βL MMD

[0077] In the formula, α is the weight coefficient of the marginal loss function L Margin and β is the weight coefficient of the MMD loss function L MMD .

[0078] The purpose of the marginal loss function L Margin is to minimize the classification error in a supervised manner based on the labeled source domain data. The marginal loss function is used to reduce the intra-class distance and increase the inter-class distance. The marginal loss function L Margin is as follows:

[0079]

[0080] In the formula, is the index of the sample label, indicating that the sample belongs to the c-th class, indicating that the sample does not belong to the c-th class; is the predicted label value of the source domain sample, the regularization parameter λ is the weighted penalty factor for the missing class loss; m- is the upper margin of the predicted label value of the source domain sample, m+ is the lower margin of the predicted label value of the source domain sample. If the sample belongs to the c-th class, then is not less than m+, and if the sample does not belong to the c-th class, then is not greater than m-. According to experience, m+ = 0.9, m- = 0.1, and λ = 0.25.

[0081] The purpose of the MMD loss function is to minimize the distribution difference between the source domain and the target domain. The MMD loss function L MMD is as follows:

[0082]

[0083] In the formula, N S is the total number of samples in the source domain, N T is the total number of samples in the target domain, is the output feature map of the i-th sample in the source domain, is the output feature map of the j-th sample in the target domain, and f is the Relu activation function.

[0084] After model training, a damage detection model is obtained. And the test signal can be detected. Through this model, it can be identified whether the signal is damaged.

[0085] To verify the effectiveness of the non-contact laser ultrasonic damage detection method based on deep transfer learning of the present invention, the present invention also provides a non-contact laser ultrasonic damage detection device based on deep transfer learning, as Figure 3 shown. The non-contact laser ultrasonic damage detection device based on deep transfer learning includes a laser exciter 1, a scanning frame 2, a laser ultrasonic receiving module 3, a laser excitation probe 4, a laser detection probe 5, and a computer 6. The laser excitation probe 4 is fixed on the scanning frame 2, and laser scanning excitation is achieved through the movement of the scanning frame 2. The laser exciter 1 is connected to the laser excitation probe 4 through an optical fiber, the laser ultrasonic receiving module 3 is connected to the laser detection probe 5 through an optical fiber, and at the same time, the laser ultrasonic receiving module 3 is connected to the computer 6 through a signal transmission line.

[0086] Among them, the laser exciter 1 is used to excite a laser signal and transmit it to the laser excitation probe 4 through an optical fiber; the laser excitation probe 4 is used to perform laser excitation on the device to be tested; the laser detection probe 5 is used to acquire the vibration signal of the laser ultrasonic detection of the device to be tested and transmit it to the laser ultrasonic receiving module 3 through an optical fiber; the laser ultrasonic receiving module 3 is used to receive the vibration signal of the laser ultrasonic detection sent by the laser detection probe 5 and send it to the computer 6; the computer 6 is used to receive the vibration signal of the laser ultrasonic detection, perform damage detection on the thermoelastic signal in the vibration signal, and output the damage detection result.

[0087] The laser exciter 1 uses a Nd:YAG pulsed laser emitter, and the laser ultrasonic receiving module detects the surface displacement. The scanning frame 2 is used to fix the laser excitation probe 4 and can realize the flexible movement of the laser excitation probe 4. The laser detection point is controlled by the laser detection probe 5. The laser detection probe 5 in this experiment is fixed. The distance between the detection point and the scanning area is 55 mm. The signal received by the laser detection system is saved by a data acquisition card and processed in the computer. An aluminum plate with an artificial internal crack damage is used as the test object. The size of the aluminum plate is 500 mm × 500 mm × 3 mm, as Figure 4 shown. There are four aluminum plates of the same size. Among them, one aluminum plate is normal, and the other three aluminum plates have damages with different depths in the middle of the back of the aluminum plate. The three crack depths are 1 mm, 1.5 mm, and 2 mm respectively. The crack surface size of the three aluminum plates is 20 mm × 2 mm. The sampling length is 6k, and the sampling frequency is 5 MHz.

[0088] To illustrate that the laser energy mapping transfer method (EMTN) proposed by the present invention can improve the detection accuracy of thermoelastic signals, verification was carried out through experiments. In the experiments, three different intensities of thermoelastic laser excitation energies were set. The laser thermoelastic excitation energies were 3 mJ, 10 mJ, and 17 mJ respectively. The ablation excitation energy was set at 31 mJ. To illustrate the advantages of the selected laser ablation excitation energy, 24 mJ and 38 mJ were used as comparisons. For easy understanding, the six laser excitation energies were labeled. The three thermoelastic excitation energies of 3 mJ, 10 mJ, and 17 mJ were respectively labeled as A1, A2, and A3, and the three ablation excitation energies of 24 mJ, 31 mJ, and 38 mJ were respectively labeled as B1, B2, and B3.

[0089] The laser detection method proposed by the present invention includes two modes: single-point detection and scanning imaging detection. In single-point detection, the positions of the detection point and the excitation point are fixed, and a model can be established using a small number of samples. However, it is not suitable for large-scale detection. Fast damage imaging can be achieved through scanning imaging detection. However, at different excitation point positions, the detection signals will show obvious differences. Therefore, in laser scanning imaging detection, a training model needs to be established using the signals at all scanning positions, and the required training samples are much larger than those in single-point detection. Next, the performance of the laser energy mapping transfer method (EMTN) proposed by the present invention was verified through the data obtained from two laser detection experiments. And the detection results of this method were compared with other advanced transfer methods. Next, the results of the two experiments are given.

[0090] Mode 1: Results of laser single-point damage detection. The samples collected in the single-point damage detection experiment are shown in Table 1. There are four cases of aluminum plates detected in the experiment, including one normal aluminum plate and three different damaged aluminum plates. Six kinds of laser energy level signals were collected. 80 samples were collected at each point. Then, according to the gross error theory, 4 gross error samples were divided. Half of them were randomly selected as the training set, and the other half were the test set. Therefore, there are 152 training samples and 152 test samples for each laser energy level.

[0091] Table 1. Samples collected by lasers of each energy in Experiment 1

[0092]

[0093] After obtaining the detection samples, the low-frequency component features of the original vibration signal were extracted through wavelet decomposition, and the laser thermoelastic excitation detection signal was mapped to the laser ablation excitation detection signal. The original vibration signal was decomposed into three layers using the Meyer wavelet basis, the low-frequency component features were retained, and they were converted into a two-dimensional picture format to obtain the detection signal feature map.

[0094] In the laser energy mapping network, the feature map of the laser thermoelastic excitation detection signal is used as the input signal, and the feature map of the laser ablation excitation detection signal is used as the target signal to achieve the feature space mapping from the feature map of the laser thermoelastic excitation detection signal to the feature map of the laser ablation excitation detection signal, and the feature map of the mapped thermoelastic signal is obtained. Compared with the laser thermoelastic excitation detection signal, the mapped thermoelastic signal has a higher signal-to-noise ratio and more radiation pressure characteristics. In this embodiment, the laser energy mapping network has 8 convolutional layers and 4 max pooling layers, and the convolutional kernels have two sizes of 3×1 and 2×1. The mapping losses of different energy mapping tasks are shown in Table 2. It can be seen from the table that when the input laser energy of the mapping is close to the output laser energy, the mapping loss is small.

[0095] Table 2. Loss values of different laser signal energy mapping tasks

[0096]

[0097] Next, feature learning is performed through the Resnet network to learn and extract deeply discriminative feature maps from the input signal. And the target function composed of the margin loss function L Margin and the MMD loss function L MMD is used to train the model, and a damage detection model is obtained. The test signal is detected. Through this model, it can be identified whether the signal is a damage signal.

[0098] To illustrate the superiority of the proposed method in laser energy mapping migration, the proposed method is compared with the Transfer Component Analysis (TCA) and Joint Distribution Adaptation (JDA) methods. At the same time, the detection results under different laser energies are compared. TCA is a classic method in the field of transfer learning. However, TCA cannot align the feature conditional distributions. In this case, JDA is proposed to solve this problem. JDA is a domain adaptation method that adapts the target domain signal to the source domain signal according to the probability distribution. Its purpose is to reduce the differences in the marginal probability distribution and conditional probability distribution between the source domain and the target domain. The damage detection results of different methods are shown in Table 3.

[0099] Table 3. Detection accuracy of test signals by different methods

[0100]

[0101] It can be seen from Table 3 that the detection accuracy of EMTN is higher than that of the other two methods. Take the task B1→A1 B1 as an example to explain the naming rule. Among them, A1 B1 means that the feature map of the detection signal under the A1 laser energy excitation is mapped to the feature map of the detection signal under the B1 laser energy excitation through the laser energy mapping network, and the feature map of the mapped A1 laser energy signal is obtained, B1→A1B1 It represents the migration from the B1 laser energy signal feature map to the mapped A1 laser energy signal feature map. Compared with the detection accuracies of signals from different source domains, the detection accuracy of the B2 laser energy is the highest. The comparison results of the three methods are as Figure 5 shown. Next, the result comparison is divided into two parts.

[0102] a) Compare the detection accuracies of the EMTN and Resnet networks. To illustrate that using EMTN can improve the detection accuracy of damage in the laser thermoelastic excitation detection signal. Here, the signals with the same laser excitation energy are directly classified after feature extraction through the Resnet network and compared with EMTN. It should be noted that since the laser intensity received on the surface of the detected object is affected by many parameters, the excitation energies of the training signal and the test signal are not exactly the same. The comparison between the EMTN and Resnet networks demonstrates the superiority of this method. The test accuracies under different laser excitation energies are shown in Table 4. It can be seen from Table 4 that as the laser energy increases, the damage detection accuracy is higher. Comparing the detection results of the EMTN and Resnet networks, when the B2 laser energy signal is selected as the source domain, the detection accuracy of the laser thermoelastic excitation detection signal using EMTN is better than that of the Resnet network.

[0103] Table 4. Detection accuracies under different laser excitation energies

[0104]

[0105] b) Compare the detection accuracies of different source domains. Average the detection accuracies of different migration tasks within the same source domain. The results are as Figure 6 shown. In the figure, the average accuracies of the EMTN, TCA, and JDA methods are 84.85%, 62.48%, and 49.48% respectively. The detection accuracies of the EMTN migration tasks for B1→A, B2→A, and B3→A are 84.08%, 91.71%, and 78.97% respectively. Using the selected source domain laser energy (B2) as the source domain, the detection accuracy is the highest. The average accuracy of the laser thermoelastic excitation detection signal is 85.96% (Table 4). The detection results show that the detection accuracy of this method has increased by 5.75%. Among them, the detection accuracy of B2→A2 has increased by 7.88% compared with the detection accuracy of the A2 laser energy signal, which is the most significant improvement in detection accuracy among the migration tasks.

[0106] Mode 2: Laser scanning damage imaging detection results. To illustrate the application value of this method in damage imaging detection, a laser scanning experiment was conducted. The scanning area was 50×50 mm, the scanning interval was 1 mm, and there were 2500 detection points in total. To avoid the influence of different excitation positions on the detection results, the detection signal was intercepted from the start of vibration. In the scanning imaging detection experiment, two states, damage and normal, were set. The detection signal of the ordinary aluminum plate excited by B2 laser energy was used as the source domain sample. In Case 1, the three damage signals detected under B2 laser energy excitation were used as the source domain damage samples. The aluminum plate to be measured was detected under 10 mJ laser energy excitation as the target domain sample. The number of samples is shown in Table 5.

[0107] Table 5. Samples collected in the experiment of Case 2

[0108]

[0109] The detection results of Case 1 show that using EMTN can improve the detection accuracy of damage in the laser thermoelastic excitation detection signal. The selected source domain B2 laser energy is appropriate. Case 2 verifies the detection accuracy of scanning imaging detection. By synthesizing the sample detection results of each scanning point into an image, rapid visualization of damage detection can be achieved. The detection signal of A2 laser energy is set as the target domain, and the detection signal of B2 laser energy is set as the source domain.

[0110] The laser excitation position is controlled by the scanning frame, and the distance between adjacent excitation points is 1 mm. The detection points are fixed. To illustrate that this method can improve the detection accuracy of the laser thermoelastic excitation detection signal, the detection results of the laser thermoelastic excitation detection signal are compared. In the detection of the laser thermoelastic excitation detection signal, the detection signal of the ordinary aluminum plate scanned by A2 laser energy is used as the normal sample, and the single-point damage signal of A2 laser energy is used as the damage sample. The damage identification method is modeled and detected through the Resnet network. At the same time, the TCA and JDA methods are used for comparison. The detection accuracies of EMTN, Resnet network, JDA, and TCA are 99.95%, 99.6%, 98.86%, and 94.60% respectively.

[0111] Next, damage imaging is performed according to the detection label values of the scanning points. The imaging results are as Figure 7 shown. The high-bright points in the figure indicate that the detection result is damaged, and the actual damage position is indicated by a dotted box. It can be seen from the figure that all three methods can detect the damage position, and the damage size detected by EMTN is closest to the actual damage. However, there are false detections in the normal area in the imaging results of these three methods.

[0112] In the detected damage image, a single detected damage point (a damage point surrounded by normal points) may be a false detection or a minor damage. In this work, such small-sized damages do not need to be detected, so these single damage points are removed. The label value 0.5 is used as the damage threshold. The final threshold imaging result is as Figure 8 shown. It can be seen from the figure that the damage size detected by the proposed method is the most accurate. The results show that this method can improve the damage scanning imaging detection accuracy of the laser thermoelastic excitation detection signal. The detection accuracy of this method is higher than that of other advanced methods.

[0113] In summary, the present invention constructs a damage detection model. Through the laser energy mapping network, the feature space of the thermoelastic signal is mapped to the feature space of the ablation signal to obtain the mapped thermoelastic signal, so that the thermoelastic signal obtains the internal damage detection accuracy similar to that of the ablation signal, improving the damage detection accuracy of the laser thermoelastic excitation non-destructive testing method; by setting the objective function of the damage detection model to align the feature spaces of the mapped thermoelastic signal and the ablation signal, the problem of poor internal damage detection ability of the completely non-destructive non-contact detection method of laser thermoelastic excitation is solved, and the damage recognition accuracy of the laser thermoelastic non-destructive testing is improved.

[0114] The present invention also discloses an electronic device, including a memory and a processor. A computer program is stored in the memory and can run on the processor. When the processor executes the computer program, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning described in any one of the above are implemented. 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, can execute any combination of the implementation steps of the method embodiments, and has the corresponding functions and beneficial effects of the method.

[0115] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning described in any one of the above are implemented. The computer-readable storage medium of the present invention stores instructions or programs that can execute the non-contact laser ultrasonic damage detection method of the present invention. When the instructions or programs are run, any combination of the implementation steps of the method embodiments can be executed, and the corresponding functions and beneficial effects of the method are provided.

[0116] The technical solution of the present invention can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to enable a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs and other various media that can store program codes.

[0117] More specific examples (nonexhaustive list) of computer-readable media include the following: electrical connection parts (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which a program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0118] The above description is a detailed description of the preferred and feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications made under the technical spirit disclosed by the present invention shall fall within the scope of the patent covered by the present invention.

Claims

1. A non-contact laser ultrasonic damage detection method based on deep transfer learning, characterized in that, it includes the following steps: S1: Obtain the vibration signals of laser ultrasonic detection through the laser detection method, divide the vibration signals into ablation signals and thermoelastic signals according to the difference in laser excitation energy, and label the ablation signals. The ablation signals are the source domain, and the thermoelastic signals are the target domain; S2: Perform wavelet decomposition on the ablation signals and thermoelastic signals respectively, and extract the low-frequency component features of the ablation signals and thermoelastic signals to obtain the feature maps of the ablation signals and thermoelastic signals; S3: Construct a damage detection model, input the feature map of the thermoelastic signal into the laser energy mapping network of the damage detection model, map the feature space of the thermoelastic signal to the feature space of the ablation signal, and obtain the feature map of the mapped thermoelastic signal; S4: Input the feature maps of the ablation signal and the mapped thermoelastic signal into the feature extraction network of the damage detection model at the same time, and extract the features of the ablation signal and the mapped thermoelastic signal respectively; S5: Set the objective function of the damage detection model, and perform iterative training on the damage detection model until the number of iterations reaches the threshold set by the damage detection model; S6: Obtain the trained damage detection model, use the damage detection model to detect the damage of the thermoelastic signal of the device to be measured, and output the damage detection result; In step S5, the objective function of the damage detection model is composed of the margin loss function L Margin and the MMD loss function L MMD obtained by weighted summation. The purpose of the margin loss function L Margin is to minimize the classification error in a supervised manner based on the labeled source domain data. The margin loss function is used to reduce the intra-class distance and increase the inter-class distance. The purpose of the MMD loss function is to minimize the distribution difference between the source domain and the target domain. The objective function is as follows: L EMTN = αL Margin + βL MMD where α is the weight coefficient of the marginal loss function L Margin and β is the weight coefficient of the MMD loss function L MMD ; Marginal loss function L Margin is as follows: Wherein, is an index of the sample label, indicating that the sample belongs to the c-th class, indicating that the sample does not belong to the c-th class; is the predicted label value of the source domain sample, and the regularization parameter λ is the weighted penalty factor of the missing class loss; m- is the upper edge of the predicted label value of the source domain sample , and m+ is the lower edge of the predicted label value of the source domain sample . If the sample belongs to the c-th class, then is not less than m+, and if the sample does not belong to the c-th class, then is not greater than m-; MMD loss function L MMD is as follows: Where N S is the total number of samples in the source domain, and N T is the total number of samples in the target domain. is the output feature map of the i-th sample in the source domain, and h j T is the output feature map of the j-th sample in the target domain. f is the Relu activation function.

2. The non-contact laser ultrasonic damage detection method based on deep transfer learning according to claim 1, characterized in that, in step S1, the laser detection method includes two methods: single-point detection and scanning imaging detection. The relative positions of the detection point and the excitation point in single-point detection are fixed, and the detection signal is used to judge whether the excitation point is damaged; Scanning imaging detection realizes the detection of the scanning area by quickly moving the excitation point, and performs imaging according to whether each excitation point is determined to be damaged.

3. The non-contact laser ultrasonic damage detection method based on deep transfer learning according to claim 1, characterized in that, in step S3, the laser energy mapping network is selected to be established by a one-dimensional convolutional neural network. The one-dimensional convolutional neural network includes a feedforward layer, a convolutional layer and a pooling layer. The output feature of the convolutional layer is: Where x is the input feature of the convolutional layer, Ker is the convolutional kernel weight, b is the bias term, is the convolution calculation between Ker and x, and σ is the activation function that generates a non-linear mapping between the input and output; The output feature of the pooling layer is: h = Pooling(z) where z is the output feature of the convolutional layer.

4. The non-contact laser ultrasonic damage detection method based on deep transfer learning according to claim 1, characterized in that, In step S4, the feature extraction network is a Resnet network. The feature extraction network includes multiple residual units. The input of the feature extraction network is x, the learned feature is defined as H(x), the residual is F(x) = H(x) - x, the initially learned feature is F(x) + x, and the output x of the l-th residual unit l+1 is expressed as x l+1 = f(h(x l ) + F(x l , W l )) where x l represents the input of the l-th residual unit, W l is the weight matrix of the l-th residual unit, F is the residual function, h(x l ) = x l represents the identity mapping, and f is the Relu activation function.

5. The non-contact laser ultrasonic damage detection method based on deep transfer learning according to claim 1, characterized in that, the learning features from shallow l to deep L are as follows: where x l represents the output of the shallow residual unit, x L represents the output of the deep residual unit, x i represents the input feature of the i-th layer residual unit, W i represents the weight matrix of the i-th layer residual unit, F is the residual function, and L represents the total number of units in the deep residual unit.

6. A non-contact laser ultrasonic damage detection device applying the non-contact laser ultrasonic damage detection method according to any one of claims 1 to 5, characterized in that, it includes: a laser exciter, a laser excitation probe, a laser ultrasonic receiving module, a laser detection probe and a computer. The laser exciter is connected to the laser excitation probe, and the laser ultrasonic receiving module is respectively connected to the laser detection probe and the computer; Among them, the laser exciter is used to excite a laser signal and transmit it to the laser excitation probe through an optical fiber; the laser excitation probe is used to perform laser excitation on the device under test; the laser detection probe is used to acquire the vibration signal of the laser ultrasonic detection of the device under test and transmit it to the laser ultrasonic receiving module through an 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, detect damage to the thermoelastic signal in the vibration signal, and output the damage detection result.

7. An electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory and the computer program is executable on the processor, characterized in that when the processor executes the computer program, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by the processor, the steps of the non-contact laser ultrasonic damage detection method based on deep transfer learning according to any one of claims 1 to 5 are implemented.

Images