Microcrack detection method based on superposition spectrogram and neural network

Abstract

The invention discloses a microcrack detection method based on a superposition spectrogram and a neural network, and belongs to the technical field of defect detection. The method comprises the following steps: firstly, arranging a plurality of acquisition sensors and excitation sensors in a microcrack damage area of a to-be-detected plate, symmetrically transmitting S0 modal Lamb wave excitation signals on the upper and lower surfaces of the plate, processing and discretizing the signals received by the acquisition sensors, and generating a frequency spectrum data matrix and a superposed spectrogram; and training a neural network model based on the frequency spectrum data matrix and the superposed spectrogram to detect the coordinates, the angle and the length of the microcrack. According to the method, the superposed spectrogram formed by superposing the frequency spectrum data of all the sensors in the aspect of characteristics is larger in contained information amount and higher in visualization capability, the characteristics of different micro-cracks can be predicted, experimental results show that the method can generate the characteristic information of the position coordinates, the direction and the length of the micro-cracks at the same time, and the method can be used for predicting the characteristics of the micro-cracks at the same time. And the efficiency and the accuracy of microcrack detection of the plate-shaped structure are improved.

Description

Technical Field

[0001] The invention relates to a micro-crack detection method based on superimposed frequency spectrum and neural network, belonging to the technical field of defect detection. Background Art

[0002] Metal sheet materials have the advantages of high strength, light weight, and corrosion resistance. They are widely used in large or full-size structures such as bridges, buildings, ships, aircraft, wind turbines, and pipelines. During long-term use, metal sheet materials may be affected by factors such as impact, wear, and corrosion, resulting in damage such as holes and cracks. These damages will reduce the performance and safety of the structure and even cause catastrophic consequences. Therefore, it is particularly important to detect and evaluate damage in the structure in a timely manner. In recent years, structural health monitoring (SHM) and non-destructive testing (NDT) technologies have been widely studied and applied to the detection of metal sheet materials.

[0003] In the nondestructive testing of metal plate materials, Lamb wave testing technology has significant advantages over traditional ultrasonic testing technology. First, Lamb wave technology can detect on one side of the material, while traditional ultrasonic testing usually needs to be performed on both sides of the material, which makes Lamb wave technology more suitable for areas that are difficult to access or can be accessed on one side; second, Lamb waves can propagate a long distance in thin plate materials, which means that a larger area can be detected from a single location; third, Lamb wave testing is very sensitive to surface and near-surface damage, especially when the damage extends along the surface of the material; in addition, Lamb waves can also provide more detailed information, such as delamination and corrosion between material layers, which is difficult to achieve with traditional ultrasonic testing. In addition, Lamb wave testing technology also shows advantages in real-time monitoring. It can be used to continuously monitor the health of the structure, which helps to detect problems in time and take preventive measures, while traditional ultrasonic testing is more used for regular or when needed testing. This continuous monitoring capability makes Lamb wave technology particularly important in preventive maintenance and long-term health monitoring.

[0004] For larger-sized damage such as circular holes and open cracks, the current mainstream damage detection methods are mainly based on linear Lamb wave technology. How to improve the detection precision and accuracy of linear damage defects in plates is a key research issue. Traditional Lamb wave detection technology mainly focuses on the changes in parameters such as the amplitude of the Lamb wave response signal to evaluate the damage condition. However, when exploring microcrack damage in the early stage when the damage size is much smaller than the wavelength, traditional linear analysis methods show obvious limitations and it is difficult to accurately capture and quantify the existence and evolution of microcracks. Nonlinear Lamb technology can overcome the shortcomings of linear technology in detecting microcracks.

[0005] In recent years, the defect detection method combining Lamb wave technology with data-driven deep learning methods has shown certain advantages in thin plate damage detection, but the existing methods have not achieved the prediction of quantitative parameters such as microcrack length, direction and specific position. The method of the present invention is based on nonlinear Lamb wave analysis theory, comprehensively considers the composite influence of multiple factors such as position, direction and length on the damage detection results, and uses deep learning technology to detect and evaluate plate microcracks. Summary of the invention

[0006] Technical problem to be solved by the invention

[0007] In view of the problem that the existing Lamb wave technology combined with the deep learning method cannot realize the quantitative parameter detection of microcracks, the present invention proposes a microcrack detection method based on superimposed spectrum diagram and neural network.

[0008] Technical Solution

[0009] In order to achieve the above object, the technical solution provided by the present invention is:

[0010] A microcrack detection method based on superimposed spectrum graph and neural network includes the following steps:

[0011] Step 1: Set up several acquisition sensors in the microcrack damage area of ​​the plate to be detected, place excitation sensors at the specified position of the plate, and symmetrically emit S0 mode Lamb wave excitation signals on the upper and lower surfaces of the plate at the excitation sensors.

[0012] Step 2: Process and discretize the signals received by the acquisition sensor to generate a spectrum data matrix and a superimposed spectrum diagram;

[0013] Step 3: Generate the spectrum data matrix and superimposed spectrum graph based on step 2 to train the neural network model

[0014] Step 4: Detect the coordinates, angles, and lengths of the microcracks based on the neural network model trained in step 3.

[0015] Furthermore, the method for generating the superimposed spectrum in step 2 is:

[0016] Step 1: Determine the sampling frequency fs based on the signal received by the acquisition sensor, the method is:

[0017]

[0018] where x start is the sampling start time, x end is the sampling end time, and N is the number of data points to be collected.

[0019] Step 2: Apply Hamming window processing to the received signal of the acquisition sensor.

[0020] Step 3, perform FFT transformation on the signal processed in step 2 to generate a spectrum diagram.

[0021] Step 4: In the spectrum obtained in step 3, calculate the discretized frequency point f in the range from 0 to the sampling frequency fs. p , f p The calculation method is:

[0022] f p =(0:N 2 -1)*(fs / N 2 )

[0023] Where N 2 is the data length processed by FFT, according to the frequency point f p The spectrum amplitude of the corresponding signal in the frequency domain is used to obtain the discretized signal amplitude value.

[0024] Step 5: normalize the discretized amplitude value of step 4 by:

[0025]

[0026] Where mag is the original frequency amplitude, mag norm is the normalized frequency amplitude, and mag(mag) is the maximum amplitude.

[0027] Step 6: Overlay the normalized spectrum data of all collected sensors into a data matrix.

[0028] Step 7: Generate a superimposed spectrum diagram based on the data in step 6.

[0029] Furthermore, the neural network model trained in step 3 includes a feature extraction part and a prediction part, wherein the feature extraction part performs convolution operations and correlation processing on the input data, and the prediction part performs feature processing through a multi-head attention mechanism and outputs microcrack coordinates, angles and length information.

[0030] Beneficial Effects

[0031] In terms of characteristics, the method of the present invention forms a superimposed spectrum diagram by superimposing the spectrum data of all sensors, which contains more information and has stronger visualization ability, and is helpful for predicting the characteristics of different microcracks;

[0032] In terms of deep learning, the method of the present invention uses the superimposed spectrum diagram obtained by simulation as the input of the CNN network to train the neural network. The results show that the method can generate characteristic information of the position coordinates, direction and length of microcracks at the same time, thereby improving the efficiency and accuracy of microcrack detection in plate structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the detection method of the present invention;

[0034] Figure 2 It is the Lamb aluminum plate dispersion curve diagram of the detection method of the present invention;

[0035] Figure 3 The excitation signal time domain waveform and normalized spectrum diagram of the detection method of the present invention;

[0036] Figure 4 The neural network model structure used in the detection method of the present invention;

[0037] Figure 5 It is a schematic diagram of a simulation experiment model of the detection method of the present invention;

[0038] Figure 6 It is a microcrack prediction flow chart of the detection method of the present invention;

[0039] Figure 7 It is a superimposed spectrum diagram of microcracks of different lengths of the detection method of the present invention;

[0040] Figure 8 It is the superimposed spectrum diagram of micro cracks at different angles of the detection method of the present invention;

[0041] Fig. 9 This is a superimposed spectrum diagram of microcracks at different positions of the detection method of the present invention. DETAILED DESCRIPTION

[0042] Figure 1 This is a flow chart of the microcrack detection method of the present invention. The following describes each step in detail.

[0043] Excitation signal selection

[0044] The characteristics of Lamb waves include dispersion and multimodality, which means that the modes excited by Lamb waves will be different under different frequency and plate thickness conditions, and the propagation speeds of different modes will also be different under the same frequency and thickness. Figure 2 The dispersion curve of Lamb wave propagation in 2mm thick aluminum plate is shown. It can be found that under the condition of constant plate thickness, there are at least two Lamb wave modes at a specific frequency. Since the dispersion phenomenon of A0 mode is more significant than that of S0 mode in the low frequency range, the Lamb wave of S0 mode is selected for excitation.

[0045] In this embodiment, a 5-cycle 200KHz Hanning windowed sine wave signal is used as the excitation source, and its mathematical expression is:

[0046]

[0047] Where H(t) is a step function, f c =200kHz is the center frequency of the excitation signal, φ is the phase of the signal, and N=5 is the number of peaks of the modulated sinusoidal signal. Figure 3 The time domain waveform and normalized spectrum of the excitation signal are shown.

[0048] When sound waves encounter closed micro-cracks or other types of contact damage during propagation, a series of unique nonlinear phenomena will be triggered, such as the generation of high-order harmonics and changes in the resonance frequency of sound waves, which is called contact acoustic nonlinearity. The mechanism behind this phenomenon is that when longitudinal sound waves act on cracks, they cause continuous vibrations on the crack surface. In the compression phase of the sound wave, the contact stiffness of the crack surface is high, causing the crack to close and allowing the sound wave to continue to transmit; in the tension phase of the sound wave, the crack will open, thereby hindering the further propagation of the sound wave. The present invention uses the second harmonic signal in the nonlinear spectrum to introduce a superimposed spectrum diagram, extract key damage information from the spectrum data, and use it as input data for training the neural network.

[0049] Overlay spectrum generation

[0050] The present invention uses the superimposed spectrum graph as a feature for microcrack detection, and the steps for generating the superimposed spectrum graph are:

[0051] Step 1), after reading the data, calculate the sampling frequency fs, the method is:

[0052]

[0053] where x start is the first value of the time data, x endis the last value of the time data, and N is the length of the time data. Specifically, the sampling frequency is determined by calculating the average interval between two adjacent points in the time series and taking its reciprocal.

[0054] Step 2), applying a Hamming window to the data to reduce spectrum leakage. The method used in this embodiment is:

[0055]

[0056] Where n is the index of the sample point and N is the length of the Hamming window.

[0057] Step 3), performing fast Fourier transform (FFT) on the signal processed in step 2), the method is:

[0058]

[0059] Among them, X(f) is the representation of the time domain signal x(t) in the frequency domain, and f is the frequency.

[0060] Step 4), calculate the frequency point f in the range from 0 to the sampling frequency fs p , according to the frequency point f p The spectrum amplitude of the corresponding signal in the frequency domain obtains the discretized amplitude value of the signal in the frequency range of 0 to 500kHz.

[0061] In this embodiment, the observation target range is defined as 0 to 500kHz. After performing FFT, f p The calculation method is as follows:

[0062] f p =(0:N 2 -1)*(fs / N 2 )

[0063] Where N 2 is the length of data processed by FFT.

[0064] Step 5), normalize the discretized amplitude value of step 4), the method is:

[0065]

[0066] Among them, mag is the original frequency amplitude, mag norm is the normalized frequency amplitude, and mag(mag) is the maximum amplitude.

[0067] Step 6), superimpose the normalized spectrum data of the eight sensors in step 5) to form a data matrix with 8 rows and 501 columns, which represents the normalized spectrum results of the eight sensors in the frequency range of 0 to 500 kHz.

[0068] Step 7), generate a normalized superimposed spectrum diagram, use logarithmic scale color mapping to enhance the visual effect, the horizontal axis of the superimposed spectrum diagram represents the frequency point (from 0 to 500kHz), and the vertical axis represents the sensor number (from 1 to 8). Through visualization, the spectral characteristics of the signals received by each sensor can be more clearly displayed.

[0069] The superimposed spectrum will serve as the input of the subsequent neural network model for model training, verification and testing to better understand and predict the characteristics of microcracks.

[0070] Neural network model training and microcrack prediction

[0071] The network model designed in this embodiment combines convolutional neural network (CNN) and multi-head attention mechanism to achieve damage prediction tasks. Figure 4 The basic structure of the network is explained. The network model consists of two parts: feature extraction part and prediction part.

[0072] The feature extraction part includes the feature extraction part of the convolutional layer and the multi-head attention structure, which obtains the context information from the input feature vector.

[0073] First, the input data is resized into a matrix of size 1×501×8, and then the convolution operation is performed through the first convolution layer, whose mathematical expression is x 1 =W 1 *x+b 1 , where W 1 is the convolution kernel, * indicates the convolution operation, b 1 is a bias term. In this embodiment, the number of input channels of the first convolutional layer is 1, the number of output channels is 16, the convolution kernel size is 3×3, the step size is 1, and the padding is 1.

[0074] The convolutional output is then normalized through a batch normalization layer using Where μ is the mean of the batch data, σ 2 is the variance and ε is a small constant.

[0075] Then apply the ReLU activation function for nonlinear transformation, the method is: 3 =max(0,x 2 ).

[0076] Then down-sampling is performed through the maximum pooling layer. In this embodiment, the pooling kernel size is 2×2, the step size is 2, and the formula is x 4 =max pool_size＝2×2 (x 3 ).

[0077] Then, the convolution operation is performed through the second convolutional layer, the formula is x 5=W 2 *x 4 +b 2 , where W 2 is the convolution kernel, b 2 is the bias term. In this embodiment, the number of input channels of the second convolutional layer is 16, the number of output channels is 32, the convolution kernel size is 3×3, the step size is 1, and the padding is 1.

[0078] Perform normalization and ReLU function operation again, and the output vector is x 6 After two convolutional layers, the size of the feature map becomes (32, 125, 2).

[0079] The prediction part of the network model is a multi-head attention layer, which rearranges the output of the previous model into an appropriate shape and applies a multi-head attention mechanism to capture the dependencies between features. The multi-head attention layer uses multiple attention heads, each of which focuses on a different subspace of the input features, so that it can capture the diverse dependencies between features. The output shape is adjusted to a form suitable for the input of the multi-head attention mechanism, and the features are processed by the multi-head attention mechanism.

[0080] In this embodiment, the feature map is first rearranged into a sequence of (125, 32) and sent to the attention module. Through the attention mechanism, each feature vector incorporates the context information of other positions. Finally, the output of the attention module is rearranged into a feature map of (32, 125, 2), and the output vector is x 7 .

[0081] Flatten the features into a one-dimensional vector, randomly discard some neurons through the Dropout layer to prevent overfitting, and the output vector is x 8 In this embodiment, dropout=0.5, which means that the outputs of half of the neurons are randomly set to zero with a probability of 50%.

[0082] Finally, the classification or regression operation is performed through the fully connected layer, and the formula is y = W 3 ·x 8 +b 3 , where W 3 is the weight matrix of the fully connected layer, b 3 is the bias term. The input dimension of the fully connected layer is 32×125×2, and the output dimension is 1.

[0083] The final output is three values, representing the coordinates, angle, and length of the microcracks.

[0084] Next, the trained network model is tested with the prepared test set of 192 data sets. The mean absolute error (MAE) is used as the loss function to evaluate the performance of the network model for position and length. The results show that the MAE values ​​of the test set for position and length are 0.31 and 0.11 respectively, that is, the average absolute error between the predicted microcrack position and the actual microcrack position is 0.31 mm, and the average absolute error between the predicted microcrack length and the actual microcrack length is 0.11 mm. The cosine similarity loss function is used to evaluate the performance of the microcrack angle prediction because cosine similarity can more effectively handle the periodicity and boundary problems of angles than using the mean absolute error (MAE) as the loss function. In angle prediction, MAE does not consider the periodicity of angles, which may lead to misjudgment of values ​​close to the boundaries (such as 0 degrees and 180 degrees). In contrast, cosine similarity loss considers the periodicity of angles, making the angle measurement more natural and accurate, especially when the predicted value is close to the periodic boundary.

[0085] The formula and steps of the cosine similarity loss function are as follows:

[0086] First convert the angle from degrees to radians:

[0087]

[0088] Among them, target is the actual angle value, and output is the predicted angle value. Then for each angle, calculate the cosine and sine values, which will be used for subsequent similarity calculations.

[0089] cos_output=cos(output_rad)

[0090] cos_target = cos(target_rad)

[0091] sin_output=sin(output_rad)

[0092] sin_target = sin(target_rad)

[0093] The similarity between the two angles is then evaluated by computing the dot product of their cosine and sine vectors:

[0094] Cosine Similarity=cos_output·cos_target+sin_output·sin_target

[0095] The cosine similarity loss is 1 minus this similarity value, ensuring that exactly the same angle results in zero loss:

[0096] Cosine Loss=1-Cosine Similarity

[0097] The Cosine Loss value of the test set for angle is 0.01, and finally MAE is used to evaluate the accuracy of the model in angle prediction. After calculation, the MAE value of the angle is 2.49, that is, the average absolute error between the predicted microcrack angle and the actual microcrack angle is 2.49 degrees.

[0098] Experimental Results

[0099] In this embodiment, the finite element analysis software ABAQUS is used to establish a three-dimensional aluminum plate simulation model, and the display dynamic analysis is used to solve and analyze the propagation process of Lamb waves when there is damage in the three-dimensional aluminum plate. The parameters of the aluminum plate model created in the simulation are shown in Table 1.

[0100] Table 1 Aluminum plate model parameters

[0101]

[0102] Figure 5 The schematic diagram of the aluminum plate model used in the simulation is used to establish a coordinate system with the center of the aluminum plate as the origin. In order to prevent the boundary reflection signal from interfering with the experiment, an infinite boundary with a width of 100mm is arranged around the aluminum plate. With the origin as the center, a square damage layout area with a side length of 30mm is set. Then, outside the damage layout area, eight sensors are set to collect data, and the sensor numbers are 1-8 in sequence, and the position coordinates are (-40, 0) mm, (-28.28, 28.28) mm, (0, 40) mm, (28.28, 28.28) mm, (40, 0) mm, (28.28, -28.28) mm, (0, -40) mm, (-28.28, -28.28) mm. Then, an anti-symmetrical Lamb wave excitation signal is set on the upper and lower surfaces of the aluminum plate at (-70, 0) mm to excite the S0 mode Lamb wave required by the present invention. Finally, a through-going microcrack is set in the set damage layout area, and a neural network training data set is constructed through numerical simulation. Among them, the damage coordinates are defined as the coordinates of the microcrack damage center, the length is defined as the damage length, and the direction is defined as the angle between the microcrack (counterclockwise) and the positive direction of the y-axis.

[0103] In order to ensure the accuracy and stability of the calculation process while reducing the complexity of the calculation, the grid size and time step need to be set to appropriate values, according to the formula:

[0104]

[0105] Where, L max is the maximum mesh size, L minis the minimum grid size, λ min is the minimum wavelength, Δt is the time increment, c is the speed of Lamb wave propagation in the aluminum plate, and in this embodiment, L max =1mm, L min =0.5mm, Δt=9.5ns.

[0106] In order to more deeply analyze and reveal the relationship between the length, angle and position of the microcracks and the spectrum of the signals received by the eight sensors, the spectrum results of the data collected by the eight sensors are logarithmized and normalized at the whole frequency point from 0 to 500kHz, and then superimposed to form an 8×501 data matrix. This data matrix is ​​visualized to show the superimposed spectrum diagram, and this spectrum diagram is used as the input of the subsequent neural network model to carry out the training, verification and testing of the network model used. Figure 6 A flowchart for predicting microcrack length using a CNN network is given, which includes four parts: signal acquisition, data processing, model training, and microcrack detection.

[0107] 1) Microcrack length prediction

[0108] Firstly, the influence of different microcrack lengths on the nonlinear effect of the superimposed spectrum was verified. Three microcracks with lengths of 4 mm, 6 mm, and 8 mm, positions (0,0) mm, and angles (45 degrees) were selected. Figure 7 a-7c are the superimposed spectrum diagrams corresponding to 4 mm, 6 mm and 8 mm respectively. It can be seen that with the increase of length, the brightness of all sensors in the figure near the second frequency (400kHz) gradually increases.

[0109] By continuously changing the length of the microcracks, starting from 2 mm, with a step size of 0.01 mm, and gradually increasing to 8 mm, a total of 601 data sets were constructed and divided into training set, test set, and validation set in a ratio of 8:1:1.

[0110] The mean absolute error (MAE) is used as the loss function of the CNN network model to evaluate the model performance. The mean absolute error (MAE) is a commonly used loss function used to quantify the difference between the predicted value and the true value. The calculation formula of MAE is as follows:

[0111] where y i True value, is the predicted value, and n is the number of samples. The characteristic of MAE is that it gives the same weight to all error sizes, which makes it more robust when dealing with outliers. Compared with the mean square error (MSE), MAE can better reflect the actual effect of prediction in some cases.

[0112] In complex prediction models, it is usually necessary to predict multiple related output values, such as length, angle, and position (x and y). In order to accurately evaluate the performance of the model in these different dimensions, the present invention uses MAE to independently calculate the three types of losses. The length loss calculation formula for predicting the length of microcracks is as follows:

[0113]

[0114] Among them, outputs[i] is the length value of the i-th sample predicted by the model, and labels[i] is the actual length value of the i-th sample. After training, the test set is used for testing, and the value of LengthLoss is 0.03, that is, the average absolute error between the predicted microcrack length and the actual microcrack length is 0.03mm.

[0115] Table 2 shows the comparison between the actual length of microcracks and the length predicted by the model. It can be seen that the trained network model achieves a relatively accurate prediction of the length of microcracks.

[0116] Table 2 Microcrack length prediction results

[0117]

[0118] 2) Microcrack Angle Prediction

[0119] To verify the influence of different microcrack angles on the nonlinear effect of the superimposed spectrum, five different angles, 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees, were selected for analysis while keeping the position (0,0) mm and length (6 mm) unchanged. The figure shows the superimposed spectrum corresponding to the five angles. Similar to the superimposed spectrum of the microcrack length change, the display brightness of the sensors with different numbers at the double frequency in the spectrum is different. Figure 8 a shows the superimposed spectrum diagram when the microcrack angle is 30 degrees, and local bright spots appear near sensor No. 3. When the angle increases to 60 degrees, local bright spots appear in sensors No. 3, No. 4, No. 7, and No. 8, as shown in Figure 8 As shown in b. When the angle continues to increase to 90 degrees, Figure 8 There are local bright spots on sensors 2, 4, 6, and 8 in c. Figure 8 In d, a local bright spot appears near sensor No. 7, when the angle is 120 degrees. When the angle is 150 degrees, local bright spots appear in sensors No. 2, No. 3, No. 6, and No. 7. It is not difficult to find from the above simulation results that the change in the direction of the microcrack will lead to the relative change of the spectral relationship of these eight sensors, and the local bright spot generally appears near the area perpendicular to the microcrack.

[0120] Microcracks at five locations were selected, with coordinates of (0,0), (5,5), (-5,5), (-5,-5), and (5,-5). The angle of the microcrack at each location was continuously changed, with one data point taken for every 1 degree. 180 sets of data were obtained for each location, and a total of 900 sets of data were obtained, which were divided into training set, test set, and validation set in a ratio of 8:1:1.

[0121] The angle loss calculation formula for predicting the microcrack angle is as follows:

[0122]

[0123] Where outputs[i] is the length value of the i-th sample predicted by the model, and labels[i] is the actual length value of the i-th sample. After the model is trained, it is tested on the test set, and the AngleLoss value obtained is 3.84, that is, the average absolute value error between the predicted microcrack angle and the actual microcrack angle is 3.84 degrees.

[0124] Table 3 gives some comparisons between the actual angles of microcracks and the angles predicted by the model, from which it can be seen that the trained network model can effectively predict the microcrack angles.

[0125] Table 3 Microcrack angle prediction results

[0126]

[0127] 3) Microcrack location prediction

[0128] To verify the influence of different microcrack positions on the nonlinear effect of the superimposed spectrum, four coordinate positions were selected, namely (0, 0) mm, (13, 13) mm, (-13, 13) mm, and (13, -13) mm, keeping the crack angle (45 degrees) and length (6 mm) unchanged. Fig. 9 a is the superimposed spectrum diagram of the coordinate (0, 0) mm, from which it can be seen that local bright spots appear near sensors 3, 4, and 8. Fig. 9 b is the superimposed spectrum diagram of the coordinate (13,13) mm, and local bright spots appear near sensors 3 and 4. Fig. 9 c, it can be seen that a local bright spot appears near sensor 4. Finally, Fig. 9 The superimposed spectrum of d shows that local bright spots appear near sensors 3 and 8, and the coordinates of the microcracks are (13,13) mm. Based on these analyses, it can be found that local bright spots appear near the position perpendicular to the microcracks, which is similar to the effect of the microcrack direction factor on the superimposed spectrum.

[0129] The superimposed spectrum map is used as input to train the CNN network model, and the trained network is used to predict the specific locations of different microcracks. That is, when the eight sensors obtain new data, they are input into the trained network model to obtain the accurate location information of the microcracks. The method continuously changes the location of the microcracks. First, (-15, 15) mm is used as the starting position. The microcracks are set in sequence with a center distance of 1 mm along the x-axis until the 31st microcrack is set at the coordinate (15, 15) mm. Subsequently, the 32nd microcrack layout is moved down 1 mm in the y-axis direction to (-15, 14) mm, and the microcracks are continued to be set at the same x-axis interval. This sequence continues until the last microcrack (15, -15) mm is arranged. A total of 961 sets of data are obtained, and the data are divided into training set, test set and validation set according to the ratio of 8:1:1.

[0130] The calculation method for the Loss value of position prediction is:

[0131]

[0132] Among them, outputs[i,0] and outputs[i,1] are the x and y coordinates of the i-th sample predicted by the model. labels[i,0] and labels[i,1] are the x and y coordinates of the i-th sample predicted by the model. After training, the test set is used for testing. The value of LocationLoss is 0.14, which means that the average absolute error between the predicted microcrack location and the actual microcrack location is 0.14mm.

[0133] Table 4 shows the comparison between the actual coordinates of some microcracks and the coordinates predicted by the model, from which it can be seen that the trained network model can effectively predict the location of microcracks.

[0134] Table 4 Microcrack location prediction results

[0135]

[0136] 4) Comprehensive forecast

[0137] The above respectively verifies the influence of single variables of length, angle and position on the nonlinear effect of microcracks. However, in practical applications, these three variables (length, angle and position) are randomly distributed, and the number of combinations is large, the situation is complicated, and it is more difficult to predict the three variables at the same time than to predict the single variables mentioned in the previous three sections. Therefore, in the process of simulation modeling, the position is evenly distributed in the damage layout area, and the angle and length are randomly generated. In order to make the generated microcracks controllable, the coordinate positions are set as (0,0) mm, (13,13) mm, (-13,13) mm, (13,-13) mm in the damage layout area, and the angles are randomly distributed in the range of 0 to 180 degrees (the angles are accurate to two decimal places), and the lengths are randomly distributed in the range of 2 mm and 8 mm (the lengths are accurate to two decimal places). In order to increase the data set, the positions are rearranged according to the previous rules, and the angles and lengths are also randomly distributed as described above. A total of 1922 sets of microcrack data with length, angle and position information marked at the same time are obtained, and these data are divided into training set, test set and validation set in a ratio of 8:1:1. In the model training part, in order to improve the prediction ability of the model, the CNN network and multi-head attention mechanism are used to train the model to simultaneously predict the length, angle and position of the microcracks.

[0138] Table 5 shows some comparisons between actual parameters of microcracks and parameters predicted by the model. As can be seen from the table, the trained network model can accurately predict the location, angle and length information of microcracks.

[0139] Table 5 Microcrack detection test results

[0140]

[0141] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A microcrack detection method based on superimposed spectrum and neural network, characterized in that: The following steps are involved: Step S1, setting a number of acquisition sensors in the microcrack damage area of ​​the plate to be detected, placing an excitation sensor at a specified position of the plate, and symmetrically transmitting S0 mode Lamb wave excitation signals on the upper and lower surfaces of the plate at the excitation sensor; Step S2, processing and discretizing the signals received by the acquisition sensor to generate a spectrum data matrix and a superimposed spectrum diagram; Step S3, training a neural network model based on the spectrum data matrix and the superimposed spectrum graph generated in step S2; Step S4, detecting the coordinates, angles and lengths of the microcracks based on the neural network model trained in step S3.

2. A microcrack detection method based on superimposed spectrum and neural network as claimed in claim 1, characterized in that: The method for generating the superimposed spectrum diagram in step S2 is: Step S2-1, determining the sampling frequency fs based on the signal received by the acquisition sensor, the method is: where x start is the sampling start time, x end is the sampling end time, N is the number of data points to be collected; Step S2-2, applying Hamming window processing to the signal received by the acquisition sensor; Step S2-3, performing FFT transformation on the signal processed in step 2-2 to generate a spectrum diagram; Step S2-4, in the spectrum obtained in step 2-3, calculate the discretized frequency points f in the range from 0 to the sampling frequency fs p , f p The calculation method is: f p =(0:N2-1)*(fs / N2) Where N2 is the data length processed by FFT, according to the frequency point f p The spectrum amplitude of the corresponding signal in the frequency domain is used to obtain the discretized signal amplitude value; Step 2-5, normalize the discretized amplitude value of step 2-4, the method is: Where mag is the original frequency amplitude, mag norm is the normalized frequency amplitude, mag(mag) is the maximum amplitude; Step 2-6, superimpose the normalized spectrum data of all collected sensors into a data matrix; Step 2-7, generating a superimposed spectrum diagram based on the data in step 2-6.

3. A microcrack detection method based on superimposed spectrum and neural network as claimed in claim 1, characterized in that: The neural network model trained in step S3 includes a feature extraction part and a prediction part, wherein the feature extraction part performs convolution operation and correlation processing on the input data, and the prediction part performs feature processing through a multi-head attention mechanism and outputs microcrack coordinates, angles and length information.

Images