Layered defect linear frequency modulation Lamb wave and full convolutional network imaging method and system

Abstract

The invention discloses a linear frequency modulation Lamb wave and full convolutional network imaging method and system for layering defects. The method belongs to the technical field of neural network imaging methods. The imaging system comprises a plurality of piezoelectric transducers distributed on the surface of the carbon fiber plate to be detected, a channel switcher electrically connected with the piezoelectric transducers, a signal generator which is electrically connected with the channel switcher and can generate multi-frequency lamb waves, and an acquisition card; wherein the signal generator carries out Lamb wave excitation by adopting a linear frequency modulation mode, the imaging method can receive a linear frequency modulation Lamb wave signal through the piezoelectric transducer, and through time-frequency domain feature extraction of the linear frequency modulation signal, the Lamb wave signal is obtained. And multi-frequency information in the linear frequency modulation Lamb waves is fused through a full convolutional neural network, and a defect imaging result is directly output.

Description

Technical Field

[0001] The present invention relates to the field of imaging technology based on neural network, and in particular to a linear frequency modulated Lamb wave and full convolution network imaging method of layered defects. Background Art

[0002] During the service of an aircraft, the carbon fiber panels used in it may be subjected to low-speed impacts such as bird and foreign object impacts, resulting in delamination defects caused by debonding between carbon fiber panels. Once delamination defects occur, the strength of the material will drop sharply, causing great safety hazards. Therefore, real-time monitoring and quantitative evaluation of delamination defects in carbon fiber panels are of great significance.

[0003] Lamb wave detection is one of the commonly used methods for detecting delamination defects in carbon fiber plates. It usually uses an excitation Lamb wave to propagate in the carbon fiber plate, and then extracts signal features from the scattering that occurs when it encounters a defect, thereby achieving imaging, positioning and identification of the defect. As shown in Chinese patent applications CN201910379806.9, CN202111082997.6, CN202210901402.3, CN202210412454.4, etc.

[0004] However, the Lamb wave imaging technology with single frequency excitation adopted in the above existing technologies can only detect defects with a specific size that can scatter Lamb waves of a fixed wavelength, and cannot comprehensively detect defects of different sizes.

[0005] To solve the above problems, a crack evaluation method based on linear frequency modulation Lamb waves is also disclosed in the prior art. As shown in Chinese patent application CN202111134614.5. This method uses linear frequency modulation excitation to obtain Lamb waves of multiple frequencies, which is beneficial for the detection of defects of different sizes. However, the signal processing uses the Crazy-Climber algorithm, which can only obtain single linear defect information such as length or width, and cannot obtain accurate defect imaging images. It is not suitable for layered defect types that require information on area, morphology, etc. Summary of the invention

[0006] In view of the defects of the prior art, the purpose of the present invention is to propose a linear frequency modulated Lamb wave and full convolutional network imaging method and system for delamination defects. The imaging method can detect defects through linear frequency modulated Lamb waves, and then fuse the multi-frequency information in the linear frequency modulated Lamb waves through a fully convolutional neural network and directly output the defect imaging results, rather than a simple size value, so as to achieve accurate imaging, identification and evaluation of carbon fiber delamination defects.

[0007] The technical solution of the present invention is as follows:

[0008] A linear frequency modulated Lamb wave and fully convolutional network imaging system for delamination defects, comprising:

[0009] A plurality of piezoelectric transducers distributed on the surface of the carbon fiber plate to be detected, a channel switcher electrically connected to the piezoelectric transducers, a signal generator capable of generating multi-frequency Lamb waves and an acquisition card capable of acquiring Lamb wave signals electrically connected to the channel switcher, and a computer for storing and processing the acquired signals of the acquisition card; wherein the signal generator adopts a linear frequency modulation method for Lamb wave excitation.

[0010] The present invention further provides an imaging method using the above-mentioned imaging system, which comprises: using the signal generator to make each piezoelectric transducer perform Lamb wave excitation on the carbon fiber plate to be inspected in turn, while each piezoelectric transducer performs Lamb wave excitation, respectively collecting the Lamb wave signals received by the remaining piezoelectric transducers that have not been excited by Lamb waves through the channel switcher and the acquisition card, and completing the acquisition of Lamb wave signals after all piezoelectric transducers have been excited by Lamb waves, storing and imaging all the collected Lamb wave signals through a computer, and obtaining the internal delamination defects of the carbon fiber plate to be inspected.

[0011] According to some preferred embodiments of the present invention, the imaging method further comprises:

[0012] S1 simulates the delamination defect of the carbon fiber plate through a mass block to obtain a binary image of the simulated image, collects the corresponding carbon fiber plate delamination defect simulation data through the linear frequency modulation Lamb wave of the delamination defect and the full convolution network imaging system, and constructs a pre-training set for pre-training the imaging model according to the collected carbon fiber plate delamination defect simulation data;

[0013] S2 builds an imaging model for layered defect imaging based on a fully convolutional neural network;

[0014] S3 pre-trains the imaging model using the binary image and the pre-training set to obtain a pre-trained imaging model;

[0015] S4: performing actual detection on the carbon fiber plate containing delamination defects by using the linear frequency modulated Lamb wave of the delamination defect and the fully convolutional network imaging system, and constructing a verification set for parameter adjustment of the pre-trained imaging model according to the delamination defect data of the carbon fiber plate obtained by the detection, i.e., the measured data;

[0016] S5 adjusts the parameters of the pre-trained imaging model using the verification set to obtain a verified imaging model;

[0017] S6 identifies and images the delamination defect in the carbon fiber plate to be inspected by using the verified imaging model;

[0018] Wherein, obtaining the binary image of the simulated image includes:

[0019] The carbon fiber plate is divided into a plurality of grid units of equal size, and mass blocks of different sizes are pasted on a plurality of positions of the carbon fiber plate by adhesive, and an image of the carbon fiber plate containing the mass blocks is collected to obtain a simulated image. In the simulated image, the pixel value of the grid unit where the mass block is located is set to 1, and the pixel values ​​of other grid units are set to 0 to obtain the binary image.

[0020] According to some preferred embodiments of the present invention, obtaining the delamination defect simulation data includes: sticking mass blocks of different sizes at multiple positions of the carbon fiber plate by adhesive to simulate delamination defects of different positions and sizes; after the mass blocks are pasted, collecting Lamb wave signals through the linear frequency modulation Lamb wave of the delamination defect and the fully convolutional network imaging system to obtain the delamination defect simulation data of the corresponding position and size.

[0021] According to some preferred embodiments of the present invention, in obtaining the delamination defect simulation data, a mass block with a diameter of 5-50 mm is used, and the diameter of the mass block used in each simulation is changed at a gradient of 5 mm / time, and the setting position of the mass block varies between the center and the edge of the carbon fiber plate, and the step length of each change is 40 mm.

[0022] According to some preferred embodiments of the present invention, the construction of the pre-training set includes: pre-processing the carbon fiber plate delamination defect simulation data, and forming the pre-training set with the data obtained after the pre-processing, wherein the pre-processing includes:

[0023] S121 performs wavelet transformation on the obtained simulation data of each type of delamination defect, i.e., the received signal under the excitation of the linear frequency modulation Lamb wave distributed in the time domain, to obtain a defect image corresponding to the delamination defect type;

[0024] S122 crops the defective image to remove invalid areas in the image, that is, areas with too small pixel values, to obtain a cropped defective image.

[0025] According to some preferred embodiments of the present invention, the cropping uses straight line cropping, and the invalid area is a pixel area whose pixel value is less than 5% of the maximum pixel value of the defective image.

[0026] According to some preferred embodiments of the present invention, the imaging model includes: an input layer for inputting an original image, a first convolution layer containing an activation function for performing a first convolution and activation process on the input original image, a first pooling layer for performing a first maximum pooling process on a feature image obtained by the first convolution layer, i.e., a first feature image, a second convolution layer containing an activation function for performing a second convolution and activation process on a feature image obtained by the first pooling layer, i.e., a second feature image, a second pooling layer for performing a second maximum pooling process on a feature image obtained by the second convolution layer, i.e., a third feature image, and a feature image obtained by the second pooling layer, i.e., a fourth feature image. The third convolution layer containing the activation function performs the third convolution and activation processing, the first deconvolution layer performs the first transposed convolution on the feature image obtained by the third convolution layer, i.e., the fifth feature image, the fourth convolution layer containing the activation function performs the fourth convolution and activation processing on the feature image obtained by the first deconvolution layer, i.e., the sixth feature image, the second deconvolution layer performs the second transposed convolution on the feature image obtained by the fourth convolution layer, i.e., the seventh feature image, the fifth convolution layer containing the activation function performs the fifth convolution and activation processing on the feature image obtained by the second deconvolution layer, and the output layer outputs the feature image obtained by the five convolution layers, i.e., the output image.

[0027] According to some preferred embodiments of the present invention, the number of channels of the input layer is D1, the number of channels of the output layer is D2, and the size of the output image is L2×W2, wherein D1 is consistent with the number of Lamb wave signals received by the piezoelectric transducer, L2 and W2 are the numbers of grid units set in the length and width directions for gridding and discretizing the detection area, respectively, and D2 is set to 1 or the number of layers of the carbon fiber plate to be detected.

[0028] According to some preferred embodiments of the present invention, the parameter adjustment includes:

[0029] The time-frequency domain feature image sequence of NS simulated data in the pre-training set, i.e., the simulated feature sequence, and the time-frequency domain feature image sequence of NE measured data in the validation set, i.e., the measured feature sequence, are respectively obtained through the pre-trained imaging model;

[0030] Calculating the pixel mean and variance of all feature images in the simulated feature sequence, and simultaneously calculating the pixel mean and variance of all feature images in the measured feature sequence;

[0031] According to the obtained pixel mean and variance, normalizing any feature image in the measured feature sequence relative to the simulated feature sequence to obtain normalized feature data;

[0032] The normalized feature data is used to fine-tune the pre-trained imaging model.

[0033] The present invention has the following beneficial effects:

[0034] The imaging method of the present invention adopts linear frequency modulation excitation of Lamb waves and performs wavelet transformation on the collected multi-frequency Lamb wave signals, thereby obtaining richer Lamb wave time-frequency domain signal characteristics and more accurate multi-size defect imaging.

[0035] The present invention extracts the time-frequency domain signal characteristics of Lamb waves after wavelet transformation through an imaging model constructed based on a fully convolutional neural network, and can directly obtain accurate surface defect imaging results.

[0036] The present invention adopts a mass block simulation method to obtain defect simulation data, avoids processing a large number of defective samples for experiments, and reduces the model training cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the composition structure of the depth imaging system in a specific implementation manner.

[0038] Figure 2 FIG. 4 is a schematic diagram of linear frequency modulation in a specific implementation manner.

[0039] Figure 3 Schematic diagram of the process of depth imaging method in a specific implementation manner.

[0040] Figure 4 It is a schematic diagram of the process of preprocessing simulation data in a specific implementation manner.

[0041] Figure 5 It is a structural schematic diagram of an imaging model constructed based on FCN in a specific implementation manner.

[0042] Figure 6 Schematic diagram of the process of obtaining defect imaging images in Example 1.

[0043] Figure 7 This is a time domain signal diagram obtained in Example 1.

[0044] Figure 8 This is the time-frequency domain feature image obtained in Example 1.

[0045] Fig. 9 This is the final output image obtained in Example 1. DETAILED DESCRIPTION

[0046] The technical scheme of the present invention will be further described below in conjunction with the embodiments and drawings of the present invention. The embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0047] See attached Figure 1 According to the technical solution of the present invention, in some specific embodiments, the linear frequency modulated Lamb wave and full convolution network imaging system of the delamination defect includes:

[0048] A plurality of piezoelectric transducers 1 distributed on the surface of a carbon fiber plate 2 to be detected, a channel switcher 3 electrically connected to the piezoelectric transducer 1, a signal generator 4 capable of generating multi-frequency Lamb waves and an acquisition card 5 capable of acquiring Lamb wave signals, each electrically connected to the channel switcher 3, and a computer 6 for storing and processing the acquired signals of the acquisition card 5.

[0049] The number and distribution positions of the piezoelectric transducers 1 can be set according to actual needs, preferably, 8 or 4.

[0050] In some specific implementations, the piezoelectric transducer 1 may be arranged along the boundary of the area to be detected of the carbon fiber plate and surround the area to be detected in a square shape, and may also be adjusted according to actual needs.

[0051] In some specific embodiments, the piezoelectric transducer 1 can be attached to the surface of the to-be-detected carbon fiber plate 2 by means of an adhesive medium such as epoxy resin adhesive.

[0052] The working mode of the above system is: each piezoelectric transducer 1 is used in turn by the signal generator 4 to perform Lamb wave excitation on the surface of the carbon fiber plate 2 to be detected. While each piezoelectric transducer 1 is performing Lamb wave excitation, the channel switcher 3 and the acquisition card 5 respectively collect the Lamb wave signals received by the remaining piezoelectric transducers that have not been excited by Lamb waves. After all piezoelectric transducers 1 have been excited by Lamb waves, the collection of Lamb wave signals is completed. All the collected Lamb wave signals are stored and imaged by the computer 6 to obtain the internal delamination defects of the carbon fiber plate 2 to be detected.

[0053] More specifically, when there are eight piezoelectric transducers 1, the signal generator 4 is used to make one of the piezoelectric transducers 1 perform Lamb wave excitation in turn, and at the same time, the channel switch 3 and the acquisition card 5 respectively collect the Lamb wave signals received by the remaining piezoelectric transducers 1 that have not been excited by the Lamb wave, and a total of In some specific implementations, the eight sensors can be simplified to four sensors, and a total of Lamb wave signal.

[0054] The signal generator uses Figure 2 The linear frequency modulation method shown is used to perform Lamb wave excitation to generate multi-frequency Lamb wave signals.

[0055] For further information, see Attachment Figure 3In some embodiments, the linear frequency modulated Lamb wave and fully convolutional network imaging method of delamination defects according to the above imaging system includes the following steps:

[0056] S1 simulates the delamination defect of the carbon fiber plate through a mass block to obtain a binary image of the simulated image, and collects the corresponding carbon fiber plate delamination defect simulation data through the linear frequency modulation Lamb wave of the delamination defect and the fully convolutional network imaging system, and constructs a pre-training set for pre-training the imaging model according to the collected carbon fiber plate delamination defect simulation data.

[0057] In some more specific embodiments, obtaining the binary image of the simulated image includes:

[0058] The carbon fiber plate is divided into a plurality of grid units of equal size, mass blocks of different sizes are pasted on a plurality of positions of the carbon fiber plate by adhesive, an image of the carbon fiber plate containing the mass blocks is collected, and a simulated image is obtained. In the simulated image, the pixel value of the grid unit where the mass block is located is set to 1, and the pixel values ​​of other grid units are set to 0, so as to obtain the binary image.

[0059] The acquisition of the delamination defect simulation data includes:

[0060] Mass blocks of different sizes are attached to multiple locations of the carbon fiber plate by adhesive to simulate delamination defects of different locations and sizes;

[0061] After the mass block is pasted, Lamb wave signals are collected through the linear frequency modulation Lamb wave of the delamination defect and the full convolution network imaging system to obtain the delamination defect simulation data of the corresponding position and size.

[0062] Among them, mass blocks of different sizes can simulate delamination defects of different sizes. In some specific implementations, 5-10 mass blocks of different sizes can be selected to cover the size range of defects to be detected.

[0063] In some specific embodiments, the material of the mass block is stainless steel, which is adhered to the carbon fiber plate by epoxy resin adhesive, thereby forming a change in material properties of a local area of ​​the carbon fiber plate.

[0064] In some specific implementations, during the interlayer defect simulation, each mass block is placed multiple times, and is placed at 10-20 different positions in total.

[0065] In some more specific embodiments, refer to the attached Figure 4 The construction of the pre-training set includes: pre-processing the carbon fiber plate delamination defect simulation data, and forming the pre-training set with the data obtained after the pre-processing, wherein the pre-processing includes:

[0066] S121 performs wavelet transformation on the obtained simulation data of each type of delamination defect, i.e., the received signal under the excitation of the linear frequency modulation Lamb wave distributed in the time domain, to obtain a defect image corresponding to the delamination defect type;

[0067] Wherein, the wavelet transform is realized by the following calculation formula:

[0068]

[0069] Among them, wt(s,τ) represents the transformed image data, x(t) represents the original simulation data, s represents the scale of controlling the expansion and contraction of the wavelet function, and t express time, τ represents the time shift, ψ * (·) represents the conjugate of the following wavelet function:

[0070]

[0071] Wherein, ψ represents a wavelet basis function, preferably a Morlet wavelet basis function, and R represents a real number set.

[0072] S122 crops the defective image to remove invalid areas in the image, that is, areas with too small pixel values, to obtain a cropped defective image.

[0073] In some specific implementations, the cropping may be performed with 5% of the maximum pixel value as a threshold, that is, pixel regions that are less than 5% of the maximum pixel value are cropped.

[0074] Preferably, straight-line cutting is maintained during cutting so that the cut image is a rectangle or a square.

[0075] S2 builds an imaging model for layered defect imaging based on Fully Convolutional Neural Networks (FCN).

[0076] In some more specific embodiments, refer to the attached Figure 5The imaging model includes: an input layer for inputting an original image, a first convolution layer containing an activation function for performing a first convolution and activation process on the input original image, a first pooling layer for performing a first maximum pooling process on a feature image obtained by the first convolution layer, i.e., a first feature image, a second convolution layer containing an activation function for performing a second convolution and activation process on a feature image obtained by the first pooling layer, i.e., a second feature image, a second pooling layer for performing a second maximum pooling process on a feature image obtained by the second convolution layer, i.e., a third feature image, a third convolution layer for performing a third convolution process on a feature image obtained by the second pooling layer, i.e., a fourth feature image a third convolutional layer containing an activation function for performing an activation process, a first deconvolutional layer for performing a first transposed convolution on the feature image obtained by the third convolutional layer, i.e., the fifth feature image, a fourth convolutional layer containing an activation function for performing a fourth convolution and activation process on the feature image obtained by the first deconvolutional layer, i.e., the sixth feature image, a second deconvolutional layer for performing a second transposed convolution on the feature image obtained by the fourth convolutional layer, i.e., the seventh feature image, a fifth convolutional layer containing an activation function for performing a fifth convolution and activation process on the feature image obtained by the second deconvolutional layer, and an output layer for outputting the feature image obtained by the five convolutional layers, i.e., the output image;

[0077] Among them, the number of channels of the input layer is D1, the size of the input original image is L1×W1, the number of channels of the output layer is D2, and the size of the output image is L2×W2. Among them, L1 and W1 are the length and width of the cropped image, and the number of channels D1 is consistent with the number of Lamb wave signals received by the piezoelectric transducer. For example, if 8 sensors are used, it is 28, and if 4 sensors are used, it is 6. L2 and W2 are the number of grid units set in the length and width directions to discretize the detection area. For example, if the detection area is 200mm×200mm and it is discretized into small squares of 1mm×1mm, the values ​​of L2 and W2 are both 200. The number of output channels D2 is generally set to 1, which corresponds to the area of ​​the output image and the detection area. It can also be set to the number of interlayers of the carbon fiber plate to be detected to output the defects of all interlayers.

[0078] S3 pre-trains the imaging model using the binary image and the pre-training set to obtain a pre-trained imaging model.

[0079] In some embodiments, the pre-training includes:

[0080] Inputting the cropped defect image into the imaging model, extracting its image features through the molding model, performing dimensionality reduction processing, and generating a predicted defect image through transposed convolution;

[0081] The intersection-over-union ratio between the predicted defect image and the binary image is used as a loss function, and the forming model is iteratively adjusted to obtain the pre-trained forming model.

[0082] S4 performs actual detection of the carbon fiber plate containing delamination defects through the linear frequency modulated Lamb wave of the delamination defect and the fully convolutional network imaging system, and constructs a verification set for parameter adjustment of the pre-trained imaging model based on the delamination defect data of the carbon fiber plate obtained by the detection, that is, the measured data.

[0083] In some specific implementations, the construction of the validation set includes: processing the carbon fiber plate delamination defect data in accordance with the preprocessing, and forming the validation set with the processed data.

[0084] S5 adjusts the parameters of the pre-trained imaging model using the verification set to obtain a verified imaging model.

[0085] In some specific embodiments, when eight piezoelectric transducers are provided in the linear frequency modulated Lamb wave and full convolution network imaging system of the delamination defect, the parameter adjustment includes:

[0086] The N in the pre-training set are obtained respectively by using the pre-training imaging model S A time-frequency domain feature image sequence of simulated data That is, the simulated feature sequence and the N E The time-frequency domain feature image sequence of the measured data That is, the measured feature sequence, where represents the signal characteristic diagram of the mass block obtained by the j-th group of piezoelectric transducers at the i-th layer; Characteristic diagram representing the corresponding measured delamination defect;

[0087] Calculate the simulated feature sequence The pixel mean of all feature images in and variance At the same time, the measured characteristic sequence is calculated The pixel mean of all feature images in and variance

[0088] According to the obtained pixel mean and variance, the measured feature sequence Any feature image in Perform normalization processing relative to the simulated feature sequence to obtain normalized feature data Among them, the normalization processing calculation model is as follows:

[0089]

[0090] In some specific implementations, during fine-tuning, a smaller learning rate may be set, such as 10% of the pre-training learning rate, and the results of different learning rates may be tested and compared to find the optimal value.

[0091] The normalized feature data is used to fine-tune the pre-trained imaging model. During the fine-tuning process, the network is trained according to the time-frequency feature map of the actual defect, and the network parameters are updated to make the network suitable for layered defect data.

[0092] S6 identifies and images the delamination defects in the carbon fiber plate to be inspected by using the verified imaging model.

[0093] Example 1

[0094] See attached Figure 6 The piezoelectric transducer in the linear frequency modulated Lamb wave and full convolutional network imaging system of the delamination defect uses 4 lead zirconate titanate piezoelectric ceramics (PZT), which are pasted on a T300 carbon fiber plate for testing. The frequency of the linear frequency modulated Lamb wave signal used is 400kHz to 25kHz.

[0095] In the acquisition of simulation data, a mass block with a diameter of 5-50mm is used, and the mass block diameter used in each simulation is changed at a gradient of 5mm / time. The mass block setting position changes between the coordinates (0mm, 0mm) and (200mm, 200mm), and the step length of each change is 40mm, for a total of 36 positions. The linear frequency modulated Lamb wave signal of each mass block is collected at each position, and based on the setting of its 4 PZTs, 6 sets of detection signals can be obtained each time.

[0096] In obtaining the measured data, five carbon fiber plates containing real defects were processed for testing.

[0097] After the model verification is completed, another carbon fiber plate with delamination defects is tested. During the test, its time domain signal is collected as follows: Figure 7 As shown, the defect image is obtained as Figure 8 The time-frequency domain feature image is shown in the figure, and the final defect image is shown in the figure. Fig. 9 shown.

[0098] It should be noted that the above is only a preferred embodiment of the present invention, which should not limit the protection scope of the technical solution of the present invention. Any modification of the technical solution recorded in the above embodiments and equivalent replacement of technical features by ordinary technicians in the field within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A linear frequency modulated Lamb wave and fully convolutional network imaging system for delamination defects, characterized in that: It includes: A plurality of piezoelectric transducers distributed on the surface of the carbon fiber plate to be detected, a channel switcher electrically connected to the piezoelectric transducers, a signal generator capable of generating multi-frequency Lamb waves and an acquisition card capable of acquiring Lamb wave signals electrically connected to the channel switcher, and a computer for storing and processing the acquired signals of the acquisition card; wherein the signal generator adopts a linear frequency modulation method for Lamb wave excitation.

2. The imaging method using the linear frequency modulated Lamb wave and full convolution network imaging system of the layered defect according to claim 1 is characterized in that: It includes: The signal generator is used to make each piezoelectric transducer perform Lamb wave excitation on the carbon fiber plate to be inspected in turn. While each piezoelectric transducer is performing Lamb wave excitation, the channel switcher and the acquisition card respectively acquire Lamb wave signals received by the remaining piezoelectric transducers that have not been excited by Lamb waves. After all piezoelectric transducers have been excited by Lamb waves, the acquisition of Lamb wave signals is completed. All the acquired Lamb wave signals are stored and imaged by a computer to obtain the internal delamination defects of the carbon fiber plate to be inspected.

3. The imaging method according to claim 2, characterized in that: It also includes: S1 simulates the delamination defect of the carbon fiber plate through a mass block to obtain a binary image of the simulated image, collects the corresponding carbon fiber plate delamination defect simulation data through the linear frequency modulation Lamb wave of the delamination defect and the full convolution network imaging system, and constructs a pre-training set for pre-training the imaging model according to the collected carbon fiber plate delamination defect simulation data; S2 builds an imaging model for layered defect imaging based on a fully convolutional neural network; S3 pre-trains the imaging model using the binary image and the pre-training set to obtain a pre-trained imaging model; S4: performing actual detection on the carbon fiber plate containing delamination defects by using the linear frequency modulated Lamb wave of the delamination defect and the fully convolutional network imaging system, and constructing a verification set for parameter adjustment of the pre-trained imaging model according to the delamination defect data of the carbon fiber plate obtained by the detection, i.e., the measured data; S5 adjusts the parameters of the pre-trained imaging model using the verification set to obtain a verified imaging model; S6 identifies and images the delamination defect in the carbon fiber plate to be inspected by using the verified imaging model; Wherein, obtaining the binary image of the simulated image includes: The carbon fiber plate is divided into a plurality of grid units of equal size, and mass blocks of different sizes are pasted on a plurality of positions of the carbon fiber plate by adhesive, and an image of the carbon fiber plate containing the mass blocks is collected to obtain a simulated image. In the simulated image, the pixel value of the grid unit where the mass block is located is set to 1, and the pixel values ​​of other grid units are set to 0 to obtain the binary image.

4. The imaging method according to claim 3, characterized in that: The acquisition of the delamination defect simulation data includes: sticking mass blocks of different sizes at multiple positions of the carbon fiber plate by adhesive to simulate delamination defects of different positions and sizes; after the mass blocks are stuck, the Lamb wave signal is collected by the linear frequency modulation Lamb wave of the delamination defect and the full convolution network imaging system to obtain the delamination defect simulation data of the corresponding position and size.

5. The imaging method according to claim 4, characterized in that: In obtaining the delamination defect simulation data, a mass block with a diameter of 5-50 mm is used, and the diameter of the mass block used in each simulation is changed at a gradient of 5 mm / time. The setting position of the mass block varies between the center and the edge of the carbon fiber plate, and the step length of each change is 40 mm.

6. The imaging method according to claim 3, characterized in that: The construction of the pre-training set includes: pre-processing the carbon fiber plate delamination defect simulation data, and forming the pre-training set with the data obtained after the pre-processing, wherein the pre-processing includes: S121 performs wavelet transformation on the obtained simulation data of each type of delamination defect, i.e., the received signal under the excitation of the linear frequency modulation Lamb wave distributed in the time domain, to obtain a defect image corresponding to the delamination defect type; S122 crops the defective image to remove invalid areas in the image, that is, areas with too small pixel values, to obtain a cropped defective image.

7. The imaging method according to claim 6, characterized in that: The cropping uses straight line cropping, and the invalid area is a pixel area whose pixel value is less than 5% of the maximum pixel value of the defective image.

8. The imaging method according to claim 3, characterized in that: The imaging model includes: an input layer for inputting an original image, a first convolution layer containing an activation function for performing a first convolution and activation process on the input original image, a first pooling layer for performing a first maximum pooling process on a feature image obtained by the first convolution layer, i.e., a first feature image, a second convolution layer containing an activation function for performing a second convolution and activation process on a feature image obtained by the first pooling layer, i.e., a second feature image, a second pooling layer for performing a second maximum pooling process on a feature image obtained by the second convolution layer, i.e., a third feature image, a third convolution layer for performing a third maximum pooling process on a feature image obtained by the second pooling layer, i.e., a fourth feature image and activation processing on the feature image obtained by the third convolution layer, i.e., the fifth feature image; the first deconvolution layer performing the first transposed convolution on the feature image obtained by the third convolution layer, i.e., the fifth feature image; the fourth convolution layer performing the fourth convolution and activation processing on the feature image obtained by the first deconvolution layer, i.e., the sixth feature image; the second deconvolution layer performing the second transposed convolution on the feature image obtained by the fourth convolution layer, i.e., the seventh feature image; the fifth convolution layer performing the fifth convolution and activation processing on the feature image obtained by the second deconvolution layer; and the output layer outputting the feature image obtained by the five convolution layers, i.e., the output image.

9. The imaging method according to claim 8, characterized in that: The number of channels of the input layer is D1, the number of channels of the output layer is D2, and the size of the output image is L2×W2, wherein D1 is consistent with the number of Lamb wave signals received by the piezoelectric transducer, L2 and W2 are respectively the number of grid units set in the length and width directions for gridding and discretizing the detection area, and D2 is set to 1 or the number of layers of the carbon fiber plate to be detected.

10. The imaging method according to claim 3, characterized in that: The parameter adjustment includes: The N in the pre-training set are obtained respectively by using the pre-training imaging model S The time-frequency domain feature image sequence of the simulated data is the simulated feature sequence and the N E The time-frequency domain feature image sequence of the measured data is the measured feature sequence; Calculating the pixel mean and variance of all feature images in the simulated feature sequence, and simultaneously calculating the pixel mean and variance of all feature images in the measured feature sequence; According to the obtained pixel mean and variance, normalizing any feature image in the measured feature sequence relative to the simulated feature sequence to obtain normalized feature data; The normalized feature data is used to fine-tune the pre-trained imaging model.

Images