Damage contour prediction method for composite materials based on multi-layer convolutional sparse coding and U-net

Abstract

The present invention provides a composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net, which includes: constructing a data set for predicting the damage profile of a composite laminated structure, exciting a sensor to generate an excitation signal and obtaining a full-array waveguide array signal, using the full-array waveguide array signal as input, using a delayed sum (DAS) algorithm to generate a damage image for each damage sample, selecting 256×256 pixels containing the damage profile in the damage image with the center of gravity of the simulated damage profile as the center, constructing a damage profile prediction network model (MLCSC-Unet) based on multi-layer convolutional sparse coding and U-net network, obtaining a trained MLCSC-Unet through training, inputting a test set of damage samples to obtain a corresponding damage profile prediction map, and quantitatively evaluating the performance of the damage profile prediction network model using the maximum intersection-over-union (IoU) ratio. The present invention combines multi-layer convolutional sparse coding and U-net to achieve highly accurate quantitative prediction of the damage profile of composite laminated structures with a relatively small number of parameters.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-destructive testing, in particular to a composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net. Background Art

[0002] In recent years, composite materials have been widely used in aviation, aerospace, energy, transportation, and other fields due to their lightweight and excellent fatigue and corrosion resistance. However, during the manufacturing and service life of composite materials, they may be subject to damage such as delamination, debonding, inclusions, fiber fracture, and matrix fracture. With the increasing application of composite laminated structures, the demand for composite damage detection has shifted from qualitative determination of damage to quantitative prediction, namely, the need to accurately predict the location, geometry, and spatial dimensions of damage within the structure. With the continuous development of Lamb wave detection technology, accurate, high-resolution estimation of damage location and quantitative prediction of damage information through signal processing and feature extraction of Lamb wave response signals have become important research areas in Lamb wave detection. Damage such as delamination and debonding in composite laminated structures is often of non-negligible size, and damage size is often considered an important indicator of the degree of failure in composite structures. Therefore, damage detection in composite structures requires not only accurate damage location but also quantitative prediction of damage parameters. In active Lamb wave detection, sensor arrays are often deployed to obtain structural information for qualitative damage determination and location detection. Among damage localization algorithms tested using sparse arrays, ellipse localization and its extensions typically extract time delay information from reflected signals as damage signatures, which are then used for damage localization imaging, achieving high detection accuracy. However, these algorithms assume point sources, making it difficult to accurately predict the contours of damages of a certain size.

[0003] In recent years, deep learning (DL) methods have been introduced into the field of structural health monitoring (SHM). DL technology can convert high-level and abstract features into simple and low-level learning features. Therefore, this enables DL technology to handle the complex problem of composite material damage identification by splitting it into a large number of simple modules. U-Net was proposed in the 2015 ISBI competition. It consists of shrinking and symmetrically expanding subnets to form a U-shaped architecture. Since it only requires a small number of training samples to achieve good semantic segmentation effects, it has considerable potential in damage identification of composite materials. However, compared with other damage identification methods in the damage detection field, the U-net network model is complex, often requires a large number of parameters, lacks interpretability, and the network working mechanism is difficult to correspond to actual physical indicators. Therefore, in order to solve the above problems, it is very urgent and important to seek a composite material damage contour prediction method based on multi-layer convolutional sparse coding and U-net to achieve accurate prediction of the delamination damage contour of composite laminated structures. Summary of the Invention

[0004] In response to the defects in the above-mentioned prior art, the present invention proposes a composite material damage contour prediction method based on multi-layer convolutional sparse coding and U-net. The method includes constructing a data set for predicting the damage contour of a composite material laminate structure, exciting the sensor to generate an excitation signal and obtaining a full-array waveguide array signal, using the full-array waveguide array signal as input, and using the delayed sum DAS algorithm to generate a damage imaging map of each damage sample. With the center of gravity of the simulated damage contour as the center, 256×256 pixels containing the damage contour in the damage imaging map are selected to construct a damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network. The trained MLCSC-Unet is obtained through training, and the damage sample of the test set is input to obtain the corresponding damage contour prediction map. The performance of the damage contour prediction network model is quantitatively evaluated by the maximum intersection-over-union ratio IoU. The present invention combines multi-layer convolutional sparse coding and U-net to achieve a high-accuracy quantitative prediction of the damage contour of a composite material laminate structure with a small number of parameters.

[0005] The present invention provides a composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net, which comprises the following steps:

[0006] S1. Construct a dataset for damage profile prediction of composite laminate structures;

[0007] S2. Excite the sensor to generate an excitation signal and obtain a full-array guided wave array signal: For each damaged sample, each sensor is individually excited in turn to generate an incident ultrasonic excitation signal, while all remaining sensors receive and obtain the corresponding full-array guided wave array signal;

[0008] S3, using the full array guided wave array signal as input, after normalization and dispersion compensation preprocessing, the delayed sum DAS algorithm is used to generate the damage imaging map of each damaged sample;

[0009] S4, taking the center of gravity of the simulated damage contour as the center, selecting a local imaging image containing the damage contour in the damage imaging image;

[0010] S5. Constructing a damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network, training the training set damage samples generated in step S1 to obtain a trained damage contour prediction network model; the input of the damage contour prediction network model MLCSC-Unet is a damage imaging image based on the DAS algorithm, and its output is a damage contour prediction image corresponding to the damage imaging image;

[0011] S6. Input the damage sample of the test set in step S1 into the damage contour prediction network model trained in step S5 to obtain the damage contour prediction map of the sample. The performance of the damage contour prediction network model is quantitatively evaluated by the maximum intersection over union (IoU). The maximum intersection over union (IoU) is:

[0012]

[0013] Among them, Y i Indicates the label of the i-th pixel in the label image; P i N represents the predicted probability that the i-th pixel of the predicted output image is defective; P Represents the total number of pixels in the label and prediction images.

[0014] Furthermore, the construction of the damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network in step S5 specifically includes the following steps:

[0015] S51. Construct a damage contour prediction network model based on a U-shaped convolutional network and including an encoding layer and a decoding layer. The encoding layer path is responsible for capturing detailed information of the input image, and the decoding layer path is responsible for achieving precise positioning. In order to maintain the size and content information of the high-dimensional feature vector generated by the decoding layer, a skip connection is added between each downsampling block of the encoding layer and each upsampling block of the decoding layer.

[0016] S52. A multi-layer convolutional sparse coding (ML-CSC) module is used to replace the convolutional layer in the U-shaped convolutional network. The damage image is used as the input image y of the damage contour prediction network model, which is expressed as:

[0017]

[0018] Among them, Γi represents the convolution result of the i-th layer, i = 1, 2, ..., L, L represents the number of layers; D i Represents a special dictionary; each D i Represents the convolution matrix W i The transpose of , that is:

[0019] D i =W i T (6);

[0020] S53. Based on the basis pursuit algorithm, an alternating layered pursuit algorithm is constructed, that is, solving the following equation:

[0021]

[0022] in, represents the input of the alternating layered tracking algorithm and has represents the sparse solution of the i-th layer of the alternating layered tracking algorithm; represents the square of the L2 norm; ||||1 represents the sparse regularization constraint; λ i Denotes control Γ i Regularization parameter for sparsity;

[0023] S54, the sparse solution is approximately represented in the form of a forward propagation of a convolutional neural network (CNN), and the approximate solver of the alternating layered tracking algorithm uses a multi-layer iterative soft threshold algorithm to solve the sparse solution:

[0024]

[0025] in, Represents the convolution result of the i-th layer; μ i and θ i Represent the first and second parameters of training respectively; W i Represents the convolution operation; W i T represents the deconvolution operation; k represents the expansion number; ReLU represents the linear activation function.

[0026] Preferably, the step S2 specifically includes the following steps:

[0027] S21. The sensor is excited to generate an incident ultrasonic excitation signal, and the response signal r(t) collected by the other sensors is:

[0028] r(t)=∫S(ω)H(ω)exp(-iωt)dω (1)

[0029] Where S(ω) represents the frequency domain expression of the excitation signal; H(ω) represents the transfer function of the system under test; ω represents frequency; t represents time; i represents an imaginary number;

[0030] S22. Assuming that the sensor only excites a single guided wave mode and ignoring the energy attenuation of the guided wave during propagation, the transfer function H(ω) of the system to be measured is expressed as:

[0031] H(ω)=A(ω)exp[ig(ω)x] (2)

[0032] Where x represents the propagation distance of the guided wave from excitation to reception; g represents the wave number of the guided wave mode; A(ω) represents the amplitude-frequency curve, which represents the relationship between the amplitude of the ultrasonic guided wave and the frequency in the detection system, and depends on the plate thickness, mode type, and sensor size.

[0033] S23. Based on formula (2), for each damaged sample, each sensor is excited separately and in turn to generate an incident ultrasonic excitation signal, and at the same time all other sensors receive the corresponding response ultrasonic signals to construct corresponding groups of full-array guided wave array signals.

[0034] Preferably, the step S3 specifically includes the following steps:

[0035] S31. When the full array of ultrasonic guided wave data is collected at the detection area (x, y) position of each damage sample, the corresponding damage imaging pixel value I (x, y) is expressed as:

[0036]

[0037] Among them, (x i ,y i ) and (x j ,y j ) represent the coordinates of the sensor for excitation and the sensor for reception respectively; C g represents the group velocity of the incident guided wave mode; K represents the total number of sensors;

[0038] S32. Normalize the pixel value of each point to obtain a damage image of each damaged sample:

[0039] I(x,y)=I(x,y) / max(I(x,y)) (4).

[0040] Preferably, the step S1 specifically includes the following steps:

[0041] S11. Set a sparse sensor array consisting of N sensors, set the area covered by the sensor array as the detection area, and divide the detection area into 512×512 grids;

[0042] S12, set three types of damage samples, namely square, rhombus and parallelogram;

[0043] S13. The damage is set to appear randomly at any position in the area to be detected, and its side length varies randomly within 20 to 40 grid lengths and corresponds to 10.9375 to 21.875 mm.

[0044] Preferably, in step S2, the reciprocity of ultrasonic propagation is taken into consideration, and a reciprocity is obtained for each damaged sample. A time series of full-array guided wave array signals.

[0045] Preferably, the IoU in step S6 quantitatively reflects the degree of overlap between the damage contours in the output image and the label image, and its value is between 0 and 1. An IoU value close to 1 indicates accurate damage contour prediction performance, while an IoU value close to 0 indicates inaccurate contour prediction.

[0046] Preferably, the side length of the square in step S12 is a random size in the range of 20 to 40 grid lengths; the rhombus randomly selects two relative vertices to be fixed on the basis of the square, and the other two vertices change within a circular ring with an original position as the center and a radius of 10 to 20 grids; the side lengths of two adjacent sides of the parallelogram are each randomly selected in the range of 20 to 40 grid lengths, and the directions include four directions of 0°, 90°, 180°, and 270°.

[0047] Preferably, the sensor in step S2 uses a circular piezoelectric chip and only excites the A0 mode to propagate in a quasi-isotropic carbon fiber reinforced plastic CFRP laminate, and the excitation signal is a 5-cycle tonebrust signal with a center frequency of 50 kHz.

[0048] Compared with the prior art, the technical effects of the present invention are:

[0049] 1. The present invention designs a composite material damage contour prediction method based on multi-layer convolutional sparse coding and U-net. Combining multi-layer convolutional sparse coding and U-net networks, it can extract damage contour information from traditional sparse sensor array signals and display it; at the same time, it can achieve quantitative prediction of damage contours of composite material laminated structures with high accuracy, and the number of parameters is less than that of neural networks of the same specifications.

[0050] 2. The present invention designs a composite material damage contour prediction method based on multi-layer convolutional sparse coding and U-net. On the basis of the basis pursuit algorithm, an alternating layered pursuit algorithm is proposed, which is beneficial to the forward propagation of the neural network, and can better capture the semantic and appearance information of the image, improve the segmentation performance, and increase the interpretability; in addition, with the help of the U-net network under deep learning, the quantitative prediction of the delamination damage contour of the composite sandwich material is realized using Lamb signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Other features, objects and advantages of the present application will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings.

[0052] Figure 1 This is a flow chart of the composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net of the present invention;

[0053] Figure 2 is a schematic diagram of a sparse sensor array in an embodiment of the present invention;

[0054] Figure 3a is a schematic diagram of a square damaged sample in an embodiment of the present invention;

[0055] Figure 3b Schematic diagram of a diamond-shaped damage sample in an embodiment of the present invention;

[0056] Figure 3c is a schematic diagram of a parallelogram sample in an embodiment of the present invention;

[0057] Figure 4 The damage imaging image obtained by the DAS time-delay superposition algorithm in the embodiment of the present invention and the MLCSC-Unet network input image intercepted with the damage contour center as the center;

[0058] Figure 5 is the model structure of the MLCSC-Unet network in an embodiment of the present invention;

[0059] Figure 6 is a schematic diagram of the ML-CSC module structure in an embodiment of the present invention;

[0060] Figure 7 These are the test results corresponding to three different samples in the embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the relevant invention are shown in the accompanying drawings.

[0062] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0063] Figure 1The present invention shows a composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net, which includes the following steps:

[0064] S1. Construct a dataset for damage profile prediction of composite laminated structures. The dataset for damage profile prediction of composite laminated structures includes n×3 damage samples as a training set and m×3 damage samples as a test set. In one embodiment, n is 1000 and m is 40.

[0065] S11. Set a sparse sensor array consisting of N sensors, set the area covered by the sensor array as the detection area, and divide the detection area into 512×512 grids; the sensor uses a circular piezoelectric chip.

[0066] In a specific embodiment, N is 8, and the sensor is an ultrasonic transducer, which is a circular piezoelectric chip with a diameter of 8 mm. The sparse sensor waveguide array is composed of 8 ultrasonic transducers, forming a square detection area of ​​280 mm × 280 mm. Figure 2 shown.

[0067] S12. Set three types of damage samples, namely square, rhombus and parallelogram, as shown in Figure 3a-3c As shown in the figure, the side length of the square is a random size in the range of 20 to 40 grid lengths; the rhombus has two opposite vertices randomly selected and fixed on the basis of the square, and the other two vertices are changed within a circle with a radius of 10 to 20 grids and the original position as the center; the side lengths of two adjacent sides of the parallelogram are randomly selected in the range of 20 to 40 grid lengths, and the directions include 0°, 90°, 180°, and 270°.

[0068] S13. The damage is set to appear randomly at any position in the area to be detected, and its side length varies randomly within 20 to 40 grid lengths and corresponds to 10.9375 to 21.875 mm.

[0069] S2. Excite the sensor to generate an excitation signal and obtain the full-array guided wave array signal: For each damaged sample, each sensor is excited separately and in turn to generate an excitation signal of the incident ultrasonic wave, while all other sensors receive and obtain the corresponding full-array guided wave array signal.

[0070] S21. The entire detection system includes the test instrument, sensor, and the structure under test, which can be approximated as a linear system. The sensor generates an incident ultrasonic excitation signal, and the response signal r(t) collected by the other sensors is:

[0071] r(t)=∫S(ω)H(ω)exp(-iωt)dω (1)

[0072] Where S(ω) represents the frequency domain expression of the excitation signal; H(ω) represents the transfer function of the system to be measured; ω represents frequency; t represents time; and i represents an imaginary number.

[0073] S22. Assuming that the sensor only excites a single guided wave mode and ignoring the energy attenuation of the guided wave during propagation, the transfer function H(ω) of the system to be measured is expressed as:

[0074] H(ω)=A(ω)exp[ig(ω)x] (2)

[0075] Where x represents the propagation distance of the guided wave from excitation to reception; g represents the wave number of the guided wave mode; and A(ω) represents the amplitude-frequency curve, which represents the relationship between the amplitude of the ultrasonic guided wave and the frequency in the detection system. This curve depends on the plate thickness, mode type, and sensor size and can be obtained through theoretical derivation or experimental testing.

[0076] S23. Based on formula (2), for each damaged sample, each sensor is excited to generate an incident ultrasonic excitation signal in turn, and only the A0 mode is excited to propagate in the quasi-isotropic carbon fiber reinforced plastic CFRP laminate. The excitation signal is a 5-cycle tonebrust signal with a center frequency of 50kHz. At the same time, all other sensors receive the corresponding response ultrasonic signals and construct several corresponding groups of full-array guided wave array signals. Considering the reciprocity of ultrasonic propagation, a reciprocal waveguide array is obtained for each damaged sample. A time series of full-array guided wave array signals.

[0077] In one embodiment, the sensor excites only the A0 mode on a 2 mm thick quasi-isotropic structure [0 / +45 / -45 / 90 / 90 / -45 / +45 / 0] 2s The system propagates through 16 layers of carbon fiber reinforced plastic (CFRP) laminates, generating 28 corresponding guided wave array signals. The guided wave array signals contain damage reflection waves, which are reflected at the damage edge. Therefore, the guided wave array signals carry damage contour information.

[0078] S3. Taking the full-array guided wave array signal as input, after normalization and dispersion compensation preprocessing, the delayed sum (DAS) algorithm is used to generate the damage imaging map of each damaged sample. Figure 4 Schematic diagram showing the DAS algorithm principle of the arranged guided wave array.

[0079] S31. When the full array of ultrasonic guided wave data is collected at the detection area (x, y) position of each damage sample, the corresponding damage imaging pixel value I (x, y) is expressed as:

[0080]

[0081] Among them, (xi ,y i ) and (x j ,y j ) represent the coordinates of the sensor for excitation and the sensor for reception respectively; C g represents the group velocity of the incident guided wave mode; K represents the total number of sensors.

[0082] S32. Normalize the pixel value of each point to obtain a damage image of each damaged sample:

[0083] I(x,y)=I(x,y) / max(I(x,y)) (4).

[0084] S4, such as Figure 4 As shown in the figure, in order to remove a lot of redundant interference information and improve the accuracy of the damage contour prediction network, the center of gravity of the simulated damage contour (x D ,y D ) as the center, and select 256×256 pixels containing the damage outline in the damage imaging image.

[0085] S5. Construct a damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network, such as Figure 5 As shown, the training set damage samples generated in step S1 are trained to obtain a trained damage contour prediction network model; the input of the damage contour prediction network model MLCSC-Unet is a damage imaging image based on the DAS algorithm, and its output is a damage contour prediction image corresponding to the damage imaging image.

[0086] The construction of the damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network specifically includes the following steps:

[0087] S51. Construct a damage contour prediction network model based on a U-shaped convolutional network and containing an encoding layer and a decoding layer. The encoding layer path is responsible for capturing detailed information of the input image, and the decoding layer path is responsible for achieving precise positioning. Considering the need to maintain the size and content information of the high-dimensional feature vector generated by the decoding layer, a jump connection is added between each downsampling block of the encoding layer and each upsampling block of the decoding layer to ensure that they have the same size and the feature map learned during downsampling is used to reconstruct the image.

[0088] S52, such as Figure 6As shown in the figure, a multi-layer convolutional sparse coding ML-CSC module (ML-CSCBlock) is used to replace the convolutional layer in the U-shaped convolutional network, playing the role of extracting damage feature maps. Considering that the size of the damage is very small relative to the detection area and the damage is sparse and accidental, the damage image obtained by the DAS algorithm meets the conditions of multi-layer convolutional sparse coding. The damage image is used as the input image y of the damage contour prediction network model, which is expressed as:

[0089]

[0090] Among them, Γ i represents the convolution result of the i-th layer, i = 1, 2, ..., L, L represents the number of layers; D i Represents a special dictionary; each D i Represents the convolution matrix W i The transpose of , that is:

[0091] D i =W i T (6).

[0092] S53, the threshold algorithm is the simplest tracking algorithm for sparse signals. In order to obtain a more accurate sparse solution, an alternating layered tracking algorithm is constructed based on the basis tracking algorithm, that is, solving the following equation:

[0093]

[0094] in, represents the input of the alternating layered tracking algorithm and has represents the sparse solution of the i-th layer of the alternating layered tracking algorithm; represents the square of the L2 norm; ||||1 represents the sparse regularization constraint; λ i Denotes control Γ i Regularization parameter for sparsity.

[0095] S54. The sparse solution is approximated using a forward propagation form of a convolutional neural network (CNN). The approximate solver of the alternating layered tracking algorithm uses a multi-layer iterative soft thresholding algorithm to solve the sparse solution:

[0096]

[0097] in, Represents the convolution result of the i-th layer; μ i and θ i Represent the first and second parameters of training respectively; W i Represents the convolution operation; W i Trepresents the deconvolution operation; k represents the expansion number; ReLU represents the linear activation function.

[0098] In a specific embodiment, L=3 and k=1 correspond to the number of convolutional layers and the number of expansions in the traditional U-net model, respectively.

[0099] From the perspective of sparsity, since the multi-layer iterative soft threshold algorithm is better than the hierarchical threshold algorithm, the ML-CSC block can extract more accurate It is beneficial to the forward propagation of the neural network, and can better capture the semantic and appearance information of the image, improve the segmentation performance, and increase the interpretability.

[0100] S6. Input the damage samples of the test set in step S1 into the damage contour prediction network model trained in step S5 to obtain the corresponding damage contour prediction map. The performance of the damage contour prediction network model is quantitatively evaluated using the maximum intersection over union (IoU) to verify the accuracy and effectiveness of the model. The maximum intersection over union (IoU) is:

[0101]

[0102] Among them, Y i Indicates the label of the i-th pixel in the label image; P i N represents the predicted probability that the i-th pixel of the predicted output image is defective; P Represents the total number of pixels in the label and prediction images.

[0103] IoU quantitatively reflects the degree of overlap between the damage contours in the output image and the label image, and its value is between 0 and 1. An IoU value close to 1 indicates accurate damage contour prediction performance, while an IoU value close to 0 indicates inaccurate contour prediction.

[0104] In a specific embodiment, Figure 7 The test results corresponding to three different damage samples are shown. It can be seen that the damage contour prediction map based on multi-layer convolutional sparse coding and U-net network model basically coincides with the actual prefabricated damage contour, which proves the effectiveness of this method and its ability to quantitatively predict the damage contour of composite laminated structures. Further comparison with the traditional U-net network model, as shown in Table 1, shows that the network model proposed by this method has significantly higher accuracy, which proves the improvement ability of this method.

[0105]

[0106] Table 1

[0107] The present invention designs a composite material damage contour prediction method based on multi-layer convolutional sparse coding and U-net. Combined with multi-layer convolutional sparse coding and U-net network, it can extract damage contour information from traditional sparse sensor array signals and display it; at the same time, it can achieve quantitative prediction of damage contours of composite laminated structures with high accuracy, and the number of parameters is less than that of neural networks of the same specifications; based on the basis tracking algorithm, an alternating layered tracking algorithm is proposed, which is beneficial to the forward propagation of the neural network, and can better capture the semantic and appearance information of the image, improve the segmentation performance, and increase the interpretability; in addition, with the help of the U-net network under deep learning, the quantitative prediction of the delamination damage contour of the composite sandwich material using Lamb signals is realized.

[0108] Finally, it should be noted that the above embodiments are only intended to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net, characterized in that: It includes the following steps: S1. Construct a dataset for damage profile prediction of composite laminate structures; S2. Excite the sensor to generate an excitation signal and obtain a full-array guided wave array signal: For each damaged sample, each sensor is individually excited in turn to generate an incident ultrasonic excitation signal, while all remaining sensors receive and obtain the corresponding full-array guided wave array signal; S3, using the full array guided wave array signal as input, after normalization and dispersion compensation preprocessing, the delayed sum DAS algorithm is used to generate the damage imaging map of each damaged sample; S4, taking the center of gravity of the simulated damage contour as the center, selecting a local imaging image containing the damage contour in the damage imaging image; S5. Constructing a damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network, training the training set damage samples generated in step S1 to obtain a trained damage contour prediction network model; the input of the damage contour prediction network model MLCSC-Unet is a damage imaging image based on the DAS algorithm, and its output is a damage contour prediction image corresponding to the damage imaging image; S6. Input the damage sample of the test set in step S1 into the damage contour prediction network model trained in step S5 to obtain the damage contour prediction map of the sample. The performance of the damage contour prediction network model is quantitatively evaluated by the maximum intersection over union (IoU). The maximum intersection over union (IoU) is: Among them, Y i Indicates the label of the i-th pixel in the label image; P i N represents the predicted probability that the i-th pixel of the predicted output image is defective; P Represents the total number of pixels in the label and prediction images.

2. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1 is characterized in that: The construction of the damage contour prediction network model MLCSC-Unet based on multi-layer convolutional sparse coding and U-net network in step S5 specifically includes the following steps: S51. Construct a damage contour prediction network model based on a U-shaped convolutional network and including an encoding layer and a decoding layer. The encoding layer path is responsible for capturing detailed information of the input image, and the decoding layer path is responsible for achieving precise positioning. In order to maintain the size and content information of the high-dimensional feature vector generated by the decoding layer, a skip connection is added between each downsampling block of the encoding layer and each upsampling block of the decoding layer. S52. A multi-layer convolutional sparse coding (ML-CSC) module is used to replace the convolutional layer in the U-shaped convolutional network. The damage image is used as the input image y of the damage contour prediction network model, which is expressed as: Among them, Γ i represents the convolution result of the i-th layer, i = 1, 2, ..., L, L represents the number of layers; D i Represents a special dictionary; each D i Represents the convolution matrix W i The transpose of , that is: D i =W i T (6); S53. Based on the basis pursuit algorithm, an alternating layered pursuit algorithm is constructed, that is, solving the following equation: in, represents the input of the alternating layered tracking algorithm and has represents the sparse solution of the i-th layer of the alternating layered tracking algorithm; represents the square of the L2 norm; || ||1 represents the sparse regularization constraint; λ i Denotes control Γ i Regularization parameter for sparsity; S54, the sparse solution is approximately represented in the form of a forward propagation of a convolutional neural network (CNN), and the approximate solver of the alternating layered tracking algorithm uses a multi-layer iterative soft threshold algorithm to solve the sparse solution: in, Represents the convolution result of the i-th layer; μ i and θ i Represent the first and second parameters of training respectively; W i Represents the convolution operation; W i T represents the deconvolution operation; k represents the expansion number; ReLU represents the linear activation function.

3. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: The step S2 specifically The following steps are involved: S21. The sensor is excited to generate an incident ultrasonic excitation signal, and the response signal r(t) collected by the other sensors is: r(t)=∫S(ω)H(ω)exp(-iωt)dω (1) Where S(ω) represents the frequency domain expression of the excitation signal; H(ω) represents the transfer function of the system under test; ω represents frequency; t represents time; i represents an imaginary number; S22. Assuming that the sensor only excites a single guided wave mode and ignoring the energy attenuation of the guided wave during propagation, the transfer function H(ω) of the system to be measured is expressed as: H(ω)=A(ω)exp[ig(ω)x] (2) Where x represents the propagation distance of the guided wave from excitation to reception; g represents the wave number of the guided wave mode; A(ω) represents the amplitude-frequency curve, which represents the relationship between the amplitude of the ultrasonic guided wave and the frequency in the detection system, and depends on the plate thickness, mode type, and sensor size. S23. Based on formula (2), for each damaged sample, each sensor is excited separately and in turn to generate an incident ultrasonic excitation signal, and at the same time all the remaining sensors receive the corresponding response ultrasonic signals to construct several corresponding groups of full-array guided wave array signals.

4. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: The step S3 specifically includes the following steps: S31. When the full array of ultrasonic guided wave data is collected at the detection area (x, y) position of each damage sample, the corresponding damage imaging pixel value I (x, y) is expressed as: Among them, (x i ,y i ) and (x j ,y j ) represent the coordinates of the sensor for excitation and the sensor for reception respectively; C g represents the group velocity of the incident guided wave mode; K represents the total number of sensors; S32. Normalize the pixel value of each point to obtain a damage image of each damaged sample: I(x,y)=I(x,y) / max(I(x,y)) (4).

5. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11. Set a sparse sensor array consisting of N sensors, set the area covered by the sensor array as the detection area, and divide the detection area into 512×512 grids; S12, set three types of damage samples, namely square, rhombus and parallelogram; S13. The damage is set to appear randomly at any position in the area to be detected, and its side length varies randomly within 20 to 40 grid lengths and corresponds to 10.9375 to 21.875 mm.

6. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: In step S2, the reciprocity of ultrasonic propagation is taken into consideration, and the reciprocity is obtained for each damaged sample. A time series of full-array guided wave array signals.

7. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: The IoU in step S6 quantitatively reflects the degree of overlap between the damage contours in the output image and the label image, and its value is between 0 and 1. An IoU value close to 1 indicates accurate damage contour prediction performance, while an IoU value close to 0 indicates inaccurate contour prediction.

8. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 5, characterized in that: The side length of the square in step S12 is a random size in the range of 20 to 40 grid lengths; the rhombus randomly selects two relative vertices based on the square and fixes them, and the other two vertices change within a circular ring with a radius of 10 to 20 grids and the original position as the center; the side lengths of two adjacent sides of the parallelogram are each randomly selected in the range of 20 to 40 grid lengths, and the directions include four directions: 0°, 90°, 180° and 270°.

9. The composite material damage profile prediction method based on multi-layer convolutional sparse coding and U-net according to claim 1, characterized in that: The sensor in step S2 uses a circular piezoelectric chip and only excites the A0 mode to propagate in a quasi-isotropic carbon fiber reinforced plastic (CFRP) laminate. The excitation signal is a 5-cycle tonebrust signal with a center frequency of 50 kHz.

Images