Composite material structure damage ultrasonic detection method based on deep learning and storage medium

Abstract

The invention relates to a composite material structure damage ultrasonic detection method and a storage medium, in particular to a composite material structure damage ultrasonic detection method based on deep learning and a storage medium. The invention aims to solve the problems that the traditional nondestructive testing method depends on professional knowledge of human beings, is time-consuming, expensive and easy to make mistakes. The method comprises the following steps of: 1, establishing a composite material laminated plate model with a layering defect and a composite material laminated plate model without a defect in the same size through simulation software; 2, setting a probe; 3, generating a time displacement curve; obtaining a group of time-frequency diagrams; 4, changing the size and the position of the defect and the position of the probe, and repeatedly executing for three to four N times to obtain N groups of time-frequency diagrams; 5, constructing a convolutional neural network model; 6, obtaining a trained convolutional neural network model; and 7, based on the trained convolutional neural network model, carrying out classification detection on a to-be-detected composite material laminated plate model. The method is applied to the field of composite material structure damage ultrasonic detection.

Description

Technical Field

[0001] The invention relates to a composite material structure damage ultrasonic detection method and a storage medium. Background Art

[0002] Composite materials have good mechanical properties and are widely used in aerospace, shipbuilding and other fields. They have advantages in many fields, such as specific strength, specific stiffness, fatigue resistance, safety and shock absorption. Because of the low density of composite materials, they are also very cost-effective for use in the industrial field. The process of manufacturing composite materials is relatively simple compared to other materials. However, during the preparation and working process, because the preparation environment may be under high pressure and high temperature, different types of damage will inevitably be caused inside the composite specimen. Including the formation of voids, defects between layers are the main forms of defects. Non-destructive testing is a key process in various industries such as aerospace, automobiles and manufacturing. It involves inspecting materials and components without causing any damage to them. Traditional non-destructive testing methods rely on human expertise, which is time-consuming, expensive and error-prone. Summary of the invention

[0003] The purpose of the present invention is to solve the problem that traditional non-destructive testing methods rely on human expertise, are time-consuming, expensive and error-prone, and to propose an ultrasonic detection method and storage medium for composite material structure damage based on deep learning.

[0004] The specific process of the ultrasonic detection method for composite material structure damage based on deep learning is as follows:

[0005] Step 1: Use simulation software to establish a composite laminate model with the same size, one with delamination defects and one without defects;

[0006] Step 2: Setting probes in composite laminate models with and without delamination defects by using simulation software;

[0007] Step 3, simulating in the simulation software that the laser irradiates the surface of the composite laminate model, the surface of the composite laminate model generates a heat source, and the probe in the composite laminate model absorbs the heat source to generate a time displacement curve;

[0008] The time displacement curves generated by the probe are processed to obtain a set of time-frequency diagrams;

[0009] Step 4: using simulation software to change the defect size and position in the composite laminate model with delamination defects, and change the probe position in the composite laminate model with delamination defects and without defects, and repeating steps 3 to 4 N times to obtain N groups of time-frequency diagrams;

[0010] Step 5: Build a convolutional neural network model;

[0011] Step 6: N groups of time-frequency graphs are used as inputs of the convolutional neural network model, and whether the composite laminate model has delamination defects is used as outputs of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model.

[0012] Step 7: Classify and detect the composite laminate model to be tested based on the trained convolutional neural network model.

[0013] A computer storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the deep learning-based ultrasonic detection method for composite material structure damage.

[0014] The beneficial effects of the present invention are:

[0015] Laser ultrasonic testing has several advantages over traditional ultrasonic testing methods, including higher accuracy, higher sensitivity, and the ability to detect defects in complex geometries. It is particularly useful for inspecting materials that are difficult to access or have irregular shapes. Convolutional neural networks can preprocess defect signals, extract features, and classify them. This reduces the difficulty of operation and shortens the time spent on non-destructive testing. Compared with general traditional signal processing methods, the use of convolutional neural networks can greatly reduce the time and knowledge and experience required by relevant personnel to engage in this work. Improve detection accuracy.

[0016] Considering the existing ultrasonic detection technology and combining the advantages of convolutional neural networks in image recognition, the present invention proposes a non-destructive testing method based on deep learning. Based on the theory and excitation mechanism of laser ultrasound, COMSOL finite element software is used to simulate the process of laser ultrasonic detection of delamination defects in composite laminates, and the displacement time curve of the set probe point is used to simulate the process of ultrasonic wave propagation in the structure, and a number of data are obtained by changing the size and position of the set defects. The time-frequency diagram obtained by the obtained time domain response after continuous wavelet transformation is processed by size and then brought into the neural network as a data set of the convolutional neural network for training, and the trained model is referenced by programming methods, so as to achieve the purpose of identifying delamination defects in composite laminates.

[0017] Laser ultrasonic technology has many advantages over other ultrasonic testing methods due to its characteristics in the field of non-destructive testing. The use of lasers can achieve non-contact testing, and can also detect components in complex environments. For example, when the components are in high temperature, high pressure and other environments, the use of laser detection is a more appropriate method. In addition, lasers also have the characteristics of high temporal and spatial resolution. The combination of ultrasonics and lasers can solve the problems of traditional ultrasonic technology. Convolutional neural networks can pre-process defect signals, extract features and classify them. It reduces the difficulty of operation and shortens the time spent on non-destructive testing. Compared with general traditional signal processing methods, the use of convolutional neural networks can greatly reduce the time and knowledge and experience of relevant personnel in this work. Improve detection accuracy; solve the problem that traditional non-destructive testing methods rely on human expertise, are time-consuming, expensive and prone to errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a simulation flow chart;

[0019] Figure 2 is the time-space distribution diagram of the set boundary heat source, a is the time-space distribution diagram of the set boundary heat source, and b is the time distribution diagram of the set boundary heat source;

[0020] Figure 3 Modeling of composite laminates, heat source and probe positions, and cloud maps of ultrasonic wave propagation in the structure, a is modeling of defective composite laminates, heat source and probe positions, and cloud maps of ultrasonic wave propagation in the structure, b is modeling of defective composite laminates, heat source and probe positions, and cloud maps of ultrasonic wave propagation in the structure, a and b are in chronological order, a is before and b is after; c is modeling of non-defective composite laminates, heat source and probe positions, and cloud maps of ultrasonic wave propagation in the structure, d is modeling of non-defective composite laminates, heat source and probe positions, and cloud maps of ultrasonic wave propagation in the structure, c and d are in chronological order, c is before and d is after;

[0021] Figure 4 is the probe point displacement time curve;

[0022] Figure 5 for Figure 4 Time-frequency response diagram obtained after continuous wavelet transform. DETAILED DESCRIPTION

[0023] Specific implementation method 1: The specific process of the composite material structure damage ultrasonic detection method based on deep learning in this implementation method is as follows:

[0024] Step 1: Use simulation software to establish a composite laminate model with the same size, one with delamination defects and one without defects;

[0025] Step 2: Setting probes in composite laminate models with and without delamination defects by using simulation software;

[0026] Step 3, simulating in the simulation software that the laser irradiates the surface of the composite laminate model, the surface of the composite laminate model generates a heat source, and the probe in the composite laminate model absorbs the heat source to generate a time displacement curve;

[0027] The time displacement curves generated by the probe are processed to obtain a set of time-frequency diagrams;

[0028] Step 4: using simulation software to change the defect size and position in the composite laminate model with delamination defects, and change the probe position in the composite laminate model with delamination defects and without defects, and repeating steps 3 to 4 N times to obtain N groups of time-frequency diagrams;

[0029] Step 5: Build a convolutional neural network model;

[0030] Step 6: N groups of time-frequency graphs are used as inputs of the convolutional neural network model, and whether the composite laminate model has delamination defects is used as outputs of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model.

[0031] Step 7: Classify and detect the composite laminate model to be tested based on the trained convolutional neural network model.

[0032] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the simulation software is COMSOL finite element simulation software.

[0033] The other steps and parameters are the same as those in the first embodiment.

[0034] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that in step three, the time displacement curve generated by the probe is processed to obtain a set of time-frequency diagrams; the specific process is:

[0035] The time displacement curves generated by the probe are processed using continuous wavelet transform to obtain a set of time-frequency diagrams.

[0036] The other steps and parameters are the same as those in the first or second embodiment.

[0037] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that a convolutional neural network model is constructed in step 5; the specific process is as follows:

[0038] The convolutional neural network model was built using the pytorch package.

[0039] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0040] Specific implementation mode five: This implementation mode is different from any of specific implementation modes one to four in that the convolutional neural network model constructed in step five sequentially includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fully connected layer, and an output layer.

[0041] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0042] Specific implementation method six: This implementation method is different from any one of specific implementation methods one to five in that the convolution kernel sizes of the first convolution layer, the second convolution layer, and the third convolution layer are 5×5.

[0043] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0044] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the N groups of time-frequency graphs in step 6 are used as inputs of the convolutional neural network model. The specific process is as follows:

[0045] The N groups of time-frequency graphs are processed into images with a resolution of 300×300 as the input of the convolutional neural network model.

[0046] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0047] Specific implementation eight: This implementation is different from any one of specific implementations one to seven in that in step six, the N groups of time-frequency graphs are used as the input of the convolutional neural network model, and the presence or absence of delamination defects in the composite laminate model is used as the output of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model; the specific process is:

[0048] Set the maximum number of iterations to M;

[0049] The number of samples used in each iterative training is set to L, L<N;

[0050] Set the learning rate to 0.01;

[0051] N groups of time-frequency graphs are used as the input of the convolutional neural network model, and whether the composite laminate model has delamination defects is used as the output of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model.

[0052] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0053] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that, in step 7, the composite material laminate model to be tested is classified and tested based on the trained convolutional neural network model; the specific process is:

[0054] Based on the trained convolutional neural network model, the composite laminate model to be tested is classified and tested, and the classification refers to whether the composite laminate model to be tested has delamination defects.

[0055] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0056] Specific implementation method ten: This implementation method is a computer storage medium, in which at least one instruction is stored. The at least one instruction is loaded and executed by a processor to implement the ultrasonic detection method for composite material structure damage based on deep learning.

[0057] It should be understood that the instructions include computer program products, software or computerized methods corresponding to any method described in the present invention; the instructions can be used to program a computer system, or other electronic device. Computer storage media may include readable media on which instructions are stored, which may include but are not limited to magnetic storage media, optical storage media; magneto-optical storage media include read-only memory ROM, random access memory RAM, erasable programmable memory (e.g., EPROM and EEPROM) and flash memory layers, or other types of media suitable for storing electronic instructions.

[0058] The COMSOL finite element simulation software was used to establish a composite laminate model with the same size, one with delamination defects and one without defects. The heating effect generated when the laser irradiates the material surface was simulated using a boundary heat source, and the multi-physics field was set to thermal expansion to simulate the process of laser irradiation generating ultrasound. The time displacement curve of the set probe point was obtained by transient research to show the propagation characteristics of the wave in the model.

[0059] In order to extract the frequency components contained in the longitudinal wave time domain signal and obtain the input image of the convolutional neural network, the longitudinal wave time domain signal is analyzed in the time and frequency domain. Since wavelet analysis can simultaneously display the time domain and frequency domain characteristics of the defect echo signal and the characteristics of the longitudinal wave signal, continuous wavelet transform is used to extract the characteristics of the defect signal, and the obtained time-frequency diagram is used as the input of the artificial convolutional neural network.

[0060] Use the pytorch software package to build a convolutional neural network for model training. Construct three convolutional layers with a convolution kernel of 5×5 to extract features. Construct three pooling layers with maximum pooling to reduce feature dimensions. Construct a fully connected layer to convert the extracted feature values ​​into a one-dimensional vector.

[0061] By changing the set defect size and position, changing the probe position and other operations, multiple sets of simulations are performed to obtain a number of data, and the time-frequency response diagram is obtained through continuous wavelet transform, which is used as the data set of the convolutional neural network after batch processing. The learning rate is set to 0.01, substituted into the convolutional neural network for training, and the trained model is referenced through programming to achieve the purpose of hierarchical defect recognition.

[0062] The process of obtaining the finite element form of the dynamic equation in S1 in the step:

[0063] The heat conduction equation is expressed in finite element form:

[0064]

[0065] Where C represents the specific heat capacity matrix, and the matrix of node temperature derivatives with respect to time is given by Indicates that K represents the heat conduction matrix, T is the node heat matrix, and P represents the heat load matrix.

[0066] Laser radiation on the solid surface is equivalent to a boundary heat load acting on the boundary S. The boundary condition is written as:

[0067]

[0068] Among them, n represents the unit normal vector (on the boundary), and q0 represents the density of the equivalent heat flow energy.

[0069] Express the dynamic equations in finite element form:

[0070]

[0071] In formula (3), the mass matrix, stiffness matrix and particle load matrix of the system are represented by M, K and F respectively. U represents the particle displacement matrix, It represents the second-order derivative matrix of node displacement with respect to time.

[0072] The boundary conditions for the traction and displacement on the free boundary are:

[0073] σ·n=0 (4)

[0074]

[0075] For the wave problem, when damping is ignored, the time domain displacement is obtained using the finite element format (3) and the Newmark time integration method. Assuming a time step of △t, the displacement and its primary differential with respect to time are:

[0076]

[0077] Different schemes of numerical integration in the Newmark integration method, that is, α and δ are important parameters for determining the stability and progress of the integration. Substituting (6) and (7) into the dynamic equations, we can obtain:

[0078]

[0079] This equation can be used to obtain the displacement value at each time point through iterative integration, and the instantaneous displacement U at time t+Δt in the time-displacement curve (8) can be obtained. t+Δt It can also be solved by the equation of motion. That is:

[0080]

[0081] Once the displacements are solved, the stress and strain components can be found using the geometric and physical equations:

[0082] ε=BU (10)

[0083] σ=DB(ε-ε0)=S(ε-ε0) (11) In the above equation, B represents the relevant strain matrix, D represents the elastic matrix, which is related to the material, S is the relevant stress matrix, and the initial strain is represented as ε0.

[0084] The calculation formula for heat source is:

[0085] q=γ(1-R)I0e -γz f(r)g(t) (12)

[0086] In the formula, γ represents the light absorption coefficient of the material used, the medium reflectance is represented by R, I0 represents the maximum energy power density at the center of the laser spot, and the spatial distribution and time distribution are represented by f(r) and g(t) respectively. The expressions are as follows:

[0087]

[0088]

[0089] In the above equation, a represents the irradiated spot radius, and the laser pulse width is represented by t0. Because γ is very small in most cases, e -γz The term can be ignored and equal to 1. So the relevant formula can be simply written as:

[0090] q=γ(1-R)I0f(r)g(t) (15)

[0091] Theoretical process of continuous wavelet transform in S2:

[0092] (1) Mother wavelet

[0093] If Ψ(t)∈L 2(R), whose Fourier transform is Ψ(Ω), if it satisfies:

[0094]

[0095] Then Ψ(t) is called the mother wavelet.

[0096] (2) Wavelet basis function

[0097] After the mother wavelet is selected, Ψ(t) can be translated or scaled. Different copies can be obtained, namely:

[0098]

[0099] Among them Ψ a,τ (t) is called the wavelet basis function. The scale parameter and translation parameter are represented by a and τ respectively, where a is a positive number.

[0100] (3) Wavelet transform

[0101] For x(t) = L 2 (R), the continuous wavelet transform formula of function x(t) is:

[0102]

[0103] The calculation formula of each layer in step S3 is:

[0104] The mathematical expression of the convolution operation is as follows:

[0105]

[0106] Among them, the output of the lth layer is z (l) , the weight matrix of the lth layer is the filter w (l) , a (l) is the output of the l-1th layer, b l is the bias term in layer 1.

[0107] The calculation method of the maximum pooling layer used is:

[0108]

[0109] In the formula, x i Indicates area The activation value of each neuron in .

[0110] The Softmax function of the fully connected layer is commonly used to evaluate the distribution differences of different samples. It can obtain the probability distribution of the current sample belonging to different categories. The calculation process is:

[0111]

[0112] In the formula, θ(i) (1≤i≤K) represents the parameters of each layer of the network.

[0113] Example:

[0114] The present invention is further described below in conjunction with the accompanying drawings and embodiments. A method for ultrasonic detection of composite material structures based on deep learning is characterized in that: delamination defects in composite laminates can be detected. Compared with manual identification of defects only through ultrasonic detection signals, it does not require relevant personnel to have high skills and experience, and has higher robustness and accuracy. The structural modeling of the present invention includes the following steps:

[0115] S1: Based on the relevant principles and excitation mechanisms of laser ultrasound, and according to the relevant elastic mechanics theory and thermodynamic coupling theory, the dynamic equations in finite element form are obtained. A composite laminate model with the same size, one with delamination defects and one without defects is established through COMSOL finite element simulation software. The heating effect generated when the laser irradiates the material surface is simulated using a boundary heat source, and the multi-physics field is set to thermal expansion to simulate the process of laser irradiation generating ultrasound. The time displacement curve of the set probe point is obtained by transient research to show the propagation characteristics of the wave in the model.

[0116] The heat conduction equation is expressed in finite element form:

[0117]

[0118] Where C represents the specific heat capacity matrix, and the matrix of node temperature derivatives with respect to time is given by Indicates that K represents the heat conduction matrix, T is the node heat matrix, and P represents the heat load matrix.

[0119] Laser radiation on the solid surface is equivalent to a boundary heat load acting on the boundary S. The boundary condition is written as:

[0120]

[0121] Among them, n represents the unit normal vector (on the boundary), and q0 represents the density of the equivalent heat flow energy.

[0122] Express the dynamic equations in finite element form:

[0123]

[0124] In formula (3), the mass matrix, stiffness matrix and particle load matrix of the system are represented by M, K and F respectively. U represents the particle displacement matrix, It represents the second-order derivative matrix of node displacement with respect to time.

[0125] The boundary conditions for the traction and displacement on the free boundary are:

[0126] σ·n=0 (4)

[0127]

[0128] For the wave problem, when damping is ignored, the time domain displacement is obtained using the finite element format (3) and the Newmark time integration method. Assuming a time step of △t, the displacement and its primary differential with respect to time are:

[0129]

[0130] Different schemes of numerical integration in the Newmark integration method, that is, α and δ are important parameters for determining the stability and progress of the integration. Substituting (6) and (7) into the dynamic equations, we can obtain:

[0131]

[0132] This equation can be used to obtain the displacement value at each time point through iterative integration, and the instantaneous displacement U at time t+Δt in the time-displacement curve (8) can be obtained. t+Δt It can also be solved by the equation of motion. That is:

[0133]

[0134] Once the displacements are solved, the stress and strain components can be found using the geometric and physical equations:

[0135] ε=BU (10)

[0136] σ=DB(ε-ε0)=S(ε-ε0) (11)

[0137] In the above equation, B represents the relevant strain matrix, D represents the elastic matrix, which is related to the material, S is the relevant stress matrix, and the initial strain is represented by ε0.

[0138] The calculation formula for heat source is:

[0139] q=γ(1-R)I0e -γz f(r)g(t) (12)

[0140] In the formula, γ represents the light absorption coefficient of the material used, the medium reflectance is represented by R, I0 represents the maximum energy power density at the center of the laser spot, and the spatial distribution and time distribution are represented by f(r) and g(t) respectively. The expressions are as follows:

[0141]

[0142] In the above equation, a represents the irradiated spot radius, and the laser pulse width is represented by t0. Because γ is very small in most cases, e -γz The term can be ignored and equal to 1. So the relevant formula can be simply written as:

[0143] q=γ(1-R)I0f(r)g(t) (15)

[0144] S2: In order to extract the frequency components contained in the longitudinal wave time domain signal and obtain the input image of the convolutional neural network, the longitudinal wave time domain signal is analyzed in the time and frequency domain. Since wavelet analysis can simultaneously display the time domain and frequency domain characteristics of the defect echo signal and the characteristics of the longitudinal wave signal, continuous wavelet transform is used to extract the characteristics of the defect signal, and the obtained time-frequency diagram is used as the input of the artificial convolutional neural network.

[0145] Theoretical process of continuous wavelet transform:

[0146] (1) Mother wavelet

[0147] If Ψ(t)∈L 2 (R), whose Fourier transform is Ψ(Ω), if it satisfies:

[0148]

[0149] Then Ψ(t) is called the mother wavelet.

[0150] (2) Wavelet basis function

[0151] After the mother wavelet is selected, Ψ(t) can be translated or scaled. Different copies can be obtained, namely:

[0152]

[0153] Among them, a,τ (t) is called the wavelet basis function. The scale parameter and translation parameter are represented by a and τ respectively, where a is a positive number.

[0154] (3) Wavelet transform

[0155] For x(t) = L 2 (R), the continuous wavelet transform formula of function x(t) is:

[0156]

[0157] S3: Use the pytorch software package to build a convolutional neural network for model training. Construct three convolutional layers with a convolution kernel of 5×5 to extract features. Construct three pooling layers with maximum pooling to reduce feature dimensions. Construct a fully connected layer to convert the extracted feature values ​​into a one-dimensional vector.

[0158] The calculation formula for each layer is:

[0159] The mathematical expression of the convolution operation is as follows:

[0160]

[0161] Among them, the output of the lth layer is z (l) , the weight matrix of the lth layer is the filter w (l) , a (l) is the output of the l-1th layer, b l is the bias term in layer 1.

[0162] The calculation method of the maximum pooling layer used is:

[0163]

[0164] Where x i Indicates area The activation value of each neuron in .

[0165] The Softmax function of the fully connected layer is commonly used to evaluate the distribution differences of different samples. It can obtain the probability distribution of the current sample belonging to different categories. The calculation process is:

[0166]

[0167] In the above formula, θ (i) (1≤i≤K) represents the parameters of each layer of the network.

[0168] S4: By changing the set defect size and position, changing the probe position and other operations, a number of data are obtained, and the time-frequency response diagram is obtained through continuous wavelet transform, which is used as the data set of the convolutional neural network after batch processing. The training set and the validation set are divided into a ratio of 7:3, the learning rate is set to 0.01, and the convolutional neural network is substituted for training, and the trained model is referenced through programming to achieve the purpose of hierarchical defect recognition.

[0169] The following is a further example

[0170] COMSOL finite element software is used to simulate the laser ultrasonic detection process of composite laminate delamination defects. The simulation process is as follows: Figure 1 As shown, the basic physical parameters are as follows: Figure 3The composite laminate model shown in the figure is used to perform laser ultrasonic finite element simulation on a four-layer laminate of carbon fiber composite materials with defects. When the pulsed laser is irradiated vertically on the upper surface of the test model, the ultrasonic waves it excites continue to propagate in the material. In the simulation process, it is assumed that the laser source and the laminate model are infinitely extended in the width direction. In this way, only the propagation of ultrasonic waves in the length and height directions of the model need to be considered, so the three-dimensional model can be reduced to a two-dimensional model, such as Figure 3 As shown. The length of the two-dimensional simplified model is fixed at L = 10mm, and the height is H = 4mm, where the height of each layer is 1mm. The ambient temperature is set to 293.15K, and the surface of the material is insulated. At the same time, in order to reduce the interference of the reflected wave generated by the boundary on the received signal, the bottom and left and right boundaries of the model are set as low-reflection boundaries. The carbon fiber composite material used is anisotropic, and the structural parameters are shown in Table 1.

[0171] Table 1 Basic parameters of the structure

[0172]

[0173] The thermal elastic mechanism similar to laser excitation is adopted to set the boundary heat source. The heat source is distributed in space to meet the strict Gaussian distribution; the time course meets the Gaussian distribution, such as Figure 2 shown and in Figure 3 The probe is set at the position shown to obtain the displacement-time curve of the point. Figure 3 The propagation process of ultrasonic waves in the structure can be clearly seen. It can be seen that when ultrasonic waves propagate to the delamination defects, they are reflected, while this will not happen in structures without defects. This is the basis for identifying whether the structure has delamination defects.

[0174] like Figure 4 The figure shows the time domain response of the longitudinal displacement of the probe point. When the laser begins to irradiate, the displacement of the probe point in the y direction increases significantly due to the absorption of heat, reaching a peak at about 90ns. Based on the propagation speed of ultrasound in the material and considering the geometric relationship, it is concluded that the defect reflection wave reaches the upper boundary at about 590ns. Combined with the probe table graph when there is a defect, it fluctuates at about 590ns, which verifies that the simulation results are relatively accurate.

[0175] Since the ultrasonic echo signal changes with time, the width and other characteristics of the defect echo cannot be well reflected. In order to facilitate the training of the convolutional neural network, Figure 4 The time domain signal shown in the figure is transformed by continuous wavelet transform to obtain Figure 5The time-frequency diagram shown in the figure continuously changes the defect size and position, and then repeatedly obtains several time-frequency response diagrams through the above method. They are uniformly resized into images with a resolution of 300×300 as the data set of the convolutional neural network. Due to the large amount of data required for training, another 400 images are obtained. These images are shuffled and divided into training set and verification set in a ratio of 7:3. The time-frequency diagrams obtained from structures with layered defects and without defects are grouped and named flaw and noflaw.

[0176] Construct a convolutional neural network with three convolutional layers, three pooling layers, and one fully connected layer. The above-obtained dataset is trained through the convolutional neural network. The number of samples used in each iteration of training is set to 12, the learning rate is set to 0.01, the number of training rounds is set to 10, and the loss function value every 10 times and the total accuracy of each round are displayed by programming. The training results are shown below, and it can be seen that the loss function value is showing a downward trend. The accuracy of the first two rounds of training is very low, and the accuracy has increased significantly since the third round of training, and the accuracy has tended to be accurate since the seventh round. At this point, the model training is completed. Finally, the images that are not in the dataset are used to identify defects through the trained model, and the accuracy rate reaches more than 99.4%.

[0177] -----The first round of training begins---

[0178] Training times 10, loss 0.6937801837921143

[0179] The accuracy on the overall test set: 0.25531914830207825

[0180] -------The second round of training begins------

[0181] Training time: 2 days, loss: 2.0219740867614746

[0182] The accuracy on the overall test set: 0.25531914830207825

[0183] --------The third round of training begins------

[0184] Training times: 30, loss: 0.788518488407135

[0185] The accuracy on the overall test set: 0.7446808218955994

[0186] -------The 4th round of training begins------

[0187] Training times 40, loss 0.40501078963279724

[0188] The accuracy on the overall test set: 0.7446808218955994

[0189] -------The 5th round of training begins------

[0190] Training times: 5 days, loss: 0.46480903029441833

[0191] The accuracy on the overall test set: 0.7446808218955994

[0192] ----The 6th round of training begins------

[0193] Training times 60, loss 0.3422389030456543

[0194] The accuracy on the overall test set: 0.7446808218955994

[0195] --------The second round of training begins------

[0196] Training times: 70, loss: 0.26504096388816833

[0197] The accuracy on the overall test set: 1.0

[0198] ------8th round of training begins-----

[0199] Training times 80, loss 0.5292102694511414

[0200] The accuracy on the overall test set: 0.7446808218955994

[0201] --------The 9th round of training begins------

[0202] Training times 90, loss 0.1289680451154709

[0203] The accuracy on the overall test set: 1.0

[0204] -------The 10th round of training begins------

[0205] Training times 100, loss 0.03122044913470745

[0206] The correct bowl rate on the overall test set: 1.0

[0207] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A composite material structure damage ultrasonic detection method based on deep learning, characterized by: The specific process of the method is: Step 1: Use simulation software to establish a composite laminate model with the same size, one with delamination defects and one without defects; Step 2: Setting probes in composite laminate models with and without delamination defects by using simulation software; Step 3, simulating in the simulation software that the laser irradiates the surface of the composite laminate model, the surface of the composite laminate model generates a heat source, and the probe in the composite laminate model absorbs the heat source to generate a time displacement curve; The time displacement curves generated by the probe are processed to obtain a set of time-frequency diagrams; Step 4: using simulation software to change the defect size and position in the composite laminate model with delamination defects, and change the probe position in the composite laminate model with delamination defects and without defects, and repeating steps 3 to 4 N times to obtain N groups of time-frequency diagrams; Step 5: Build a convolutional neural network model; Step 6: N groups of time-frequency graphs are used as inputs of the convolutional neural network model, and whether the composite laminate model has delamination defects is used as outputs of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model. Step 7: Classify and detect the composite laminate model to be tested based on the trained convolutional neural network model.

2. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 1, characterized in that: The simulation software is COMSOL finite element simulation software.

3. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 2, characterized in that: In step 3, the time displacement curve generated by the probe is processed to obtain a set of time-frequency diagrams; the specific process is: The time displacement curves generated by the probe are processed using continuous wavelet transform to obtain a set of time-frequency diagrams.

4. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 3 is characterized in that: In step 5, a convolutional neural network model is constructed; the specific process is as follows: The convolutional neural network model was built using the pytorch package.

5. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 4, characterized in that: The convolutional neural network model constructed in the step 5 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fully connected layer, and an output layer in sequence.

6. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 5, characterized in that: The convolution kernel sizes of the first convolution layer, the second convolution layer, and the third convolution layer are 5×5.

7. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 6, characterized in that: In step 6, the N groups of time-frequency graphs are used as the input of the convolutional neural network model. The specific process is as follows: The N groups of time-frequency graphs are processed into images with a resolution of 300×300 as the input of the convolutional neural network model.

8. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 7, characterized in that: In step 6, the N groups of time-frequency graphs are used as the input of the convolutional neural network model, and the presence or absence of delamination defects in the composite laminate model is used as the output of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model; the specific process is: Set the maximum number of iterations to M; The number of samples used in each iterative training is set to L, L<N; Set the learning rate to 0.01; N groups of time-frequency graphs are used as the input of the convolutional neural network model, and whether the composite laminate model has delamination defects is used as the output of the convolutional neural network model. The convolutional neural network model is trained to obtain a trained convolutional neural network model.

9. The method for ultrasonic detection of composite material structure damage based on deep learning according to claim 8, characterized in that: In step seven, classification detection is performed on the composite laminate model to be tested based on the trained convolutional neural network model; the specific process is: Based on the trained convolutional neural network model, the composite laminate model to be tested is classified and tested, and the classification refers to whether the composite laminate model to be tested has delamination defects.

10. A computer storage medium, characterized in that: The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the deep learning-based ultrasonic detection method for composite material structure damage as described in any one of claims 1 to 9.

Images