A Guided Wave Full-Wavefield Inversion Method Based on a Physics-Informed Fusion Generative Adversarial Neural Network

Abstract

The present invention proposes a guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial network. The method includes damage wavefield feature extraction, network architecture design of the generative adversarial network, and a method for reconstructing the guided wavefield and damage geometry image, etc. In the method of the present invention, the output signal of the guided wavefield generator of the generative adversarial network after training is matched with the actually measured signal, and the guided wavefield and damage image are reconstructed through the random sequence corresponding to the matching result. This method can be flexibly applied to working conditions with various measuring point layout schemes, and has a high reconstruction efficiency and certain robustness to noise.

Description

Technical Field

[0001] The present invention belongs to the technical fields of machine learning and civil engineering, and particularly relates to a full-wavefield inversion method of guided waves based on a physics-informed generative adversarial network. The method is applicable to restoring the wavefield signal of ultrasonic guided waves and the geometric information of structural defects. Background Art

[0002] Nowadays, civil engineering construction has gradually become saturated, and the inspection and maintenance of existing engineering structures have received increasing attention. Ultrasonic non-destructive testing technology can effectively detect hidden defects in civil engineering structures, locate the defects and estimate the severity of the defects, so as to evaluate the safety status of existing engineering structures and take corresponding maintenance measures in a timely manner. Ultrasonic guided wave non-destructive testing has the advantages of large detection range and high efficiency compared with other ultrasonic testing methods, thus greatly reducing the labor and time costs. Therefore, it is an urgent research work to develop and utilize new ultrasonic guided wave non-destructive testing technologies to efficiently detect practical problems in civil engineering structures.

[0003] Guided waves have natural characteristics such as dispersion, multimodality, and mode conversion at discontinuities during propagation. Usually, there are a large number of overlapping wave packets in the guided wave signals collected from a small number of measurement points, making it very difficult to effectively interpret the signals, which severely limits the practicality of ultrasonic guided wave non-destructive testing technology in the detection of actual engineering structures. If the wavefield signal of the structure to be detected over the entire domain can be obtained, the information such as the direct waves and defect scattering waves of each mode can be directly observed from the wavefield image, greatly reducing the difficulty of interpreting the guided wave signals. Therefore, studying how to obtain the wavefield signal over the entire domain from the signals collected at a small number of measurement points is a problem with great practical value. In addition, a large amount of defect-related information is often contained in the overlapping wave packets of the guided wave signals collected at a small number of measurement points. How to make full use of this information to obtain the location information and geometric information of the defects is also a problem worthy of research.

[0004] As a deep generative model, the generative adversarial network realizes the training of the neural network through an adversarial game process, can learn the inherent characteristics of the wavefield in the guided wave signal dataset, and establish a mapping relationship between the guided wave field, defect images and vectors in the low-dimensional space, so as to generate diverse guided wave field signals and defect image data through randomly sampled low-dimensional vectors. In the present invention, a generative adversarial network for integrating guided wave physical information is trained using a guided wave signal dataset, and the low-dimensional vector is inverted by minimizing the gap between the wavefield signal output by the generator and the measured guided wave signal, so as to realize the reconstruction of the guided wave field signal and the excavation of the defect location and geometric information. Summary of the Invention

[0005] The object of the present invention is to solve the problems in the prior art to meet the actual situation requirements, and a guided wave full-wavefield inversion method based on a physics-informed generative adversarial neural network is proposed.

[0006] The present invention is achieved by the following technical solutions. The present invention proposes a guided wave full-wavefield inversion method based on a physics-informed generative adversarial neural network, and the method includes the following steps:

[0007] Step 1: Establish an original dataset containing structural defects and with diversity, and extract guided wave field features from the original dataset to form a new dataset for training the generative adversarial network;

[0008] Step 2: Design the network architectures of the generator G θ,I and G θ,II and the discriminator D θ in the generative adversarial neural network; set the initial values of the network training hyperparameters, train the generative adversarial neural network, and obtain the generators G θ,I and G θ,II that can accurately generate guided wave field features and defect images and have diversity through hyperparameter tuning;

[0009] Step 3: Measure the guided wave signal of the target structure and record it as the actual measurement signal Y;

[0010] Step 4: The process of full-wavefield inversion is achieved by optimizing the low-dimensional vector z and the defect position parameter x d ; generate the initial reconstructed full-wavefield feature G θ,I (z), convert the reconstructed full-wavefield feature G θ,I (z) into the reconstructed full-wavefield signal u[x d , G θ,I (z)]; based on the actual measurement signal Y, the measurement matrix Φ of the target structure, and the reconstructed full-wavefield signal u[x d , G θ,I (z)], obtain the optimal low-dimensional vector d and the defect position according to the objective function for optimizing the low-dimensional vector and the defect position ; substitute the low-dimensional vector and the defect position θ,I into the generators G θ,II to reconstruct the full-wavefield signal and the corresponding reconstructed defect geometry image

[0011] Furthermore, the specific content of Step 1 is as follows:

[0012] Step 1.1: Use the physical numerical calculation method to calculate the guided wave measurement signal Y of the structure under different defect conditions FEM , according to the physical equation of guided wave propagation, excite the guided wave signal v0[t] at the structure excitation point s. The guided wave signal scatters at the defect point r, and the wave field characteristics in the measurement signal Y FEM can be extracted as follows:

[0013]

[0014] where the propagation operator can be calculated by the following formula:

[0015]

[0016] where a and b take the corresponding r, q or s, q or s, r respectively; the operators and are the Fourier transform and its inverse transform respectively; d ref is the reference distance, which can be taken as the mean of the distances from the excitation point to each receiving point; c p (f) is the phase velocity of the guided wave at frequency f;

[0017] Step 1.2: Combine the wave field characteristics extracted in Step 1.1 with the corresponding defect image D d to form a new data set where N is the number of samples.

[0018] Furthermore, in Step 2, for the generation tasks of wave field characteristics and defect images, convolutional neural networks that can effectively extract image features are respectively selected to construct the generator G θ,I and G θ,II as well as the discriminator D θ network architectures; among them, the generators G θ,I and G θ,II input the low-dimensional vector z randomly sampled from the standard normal distribution i , and the low-dimensional vector z i is projected by the generator G θ,I and G θ,II networks into the generated wave field characteristics and defect images G θ,I (z i ) and G θ,II (z i ); the discriminator D θ inputs the samples in the real data set or the output of the generator (G θ,I (z i ), G θ,II (z i )) and passes through the discriminator Dθ Output a numerical value; the closer the numerical value corresponding to the overall output of the generator is to the numerical value corresponding to the overall real dataset sample, the closer the data distribution output by the generator is to the sample distribution of the real dataset.

[0019] Furthermore, in step two, an adversarial training method is used to train the generative adversarial network. Among them, the discriminator network aims to determine whether the input comes from the real dataset, and its training error function is:

[0020]

[0021] Among them, and x i respectively represent the output of the generator and the real dataset sample, that is, (G θ,I (z i ), G θ,II (z i )) and is the interpolation value between the output of the generator and the real dataset sample, that is and ∈ i follows the uniform distribution U[0,1] from 0 to 1; m is the number of samples in each minbatch; λ is the gradient penalty sparsity.

[0022] Furthermore, in step two, the generator network aims to generate as realistic wavefield features as possible to confuse the discriminator's judgment of the defect image, and its training error function is:

[0023]

[0024] Among them, E d and E ud are the total energies of the ultrasonic guided wave at the defect / intact structure at time t1, which can be calculated by the expression ; ρ μ is the ratio of the guided wave energy to the signal energy in the μ mode of the guided wave, which can be obtained through finite element calculation; is the amplitude of the guided wave at the polar coordinate (r,θ) position in the defect / intact structure.

[0025] Furthermore, in step two, the parameters of the discriminator network and the generator network are alternately updated, and through a dynamic game process, the network's learning degree of the real dataset features is continuously improved.

[0026] Furthermore, in step two, by calculating the quality and diversity indexes of the generated wavefield features and defect images, the generator G θ,I and G θ,IIThe learning degree of the real data set is used to optimize the network architecture and training hyperparameters of the generative adversarial network, and the generator G that effectively represents the wave field characteristics and defect geometric images is obtained. θ,I and G θ,II .

[0027] Furthermore, the specific content of step four is as follows:

[0028] Step 4.1: Establish a physical model of the guided wave energy conservation discriminant condition and the converted wave field signal u[x d , G θ,I (z)], and use the following physical model-based expression to convert the reconstructed full wave field feature G θ,I (z) into the full wave field signal u[x d , G θ,I (z)]:

[0029]

[0030] Step 4.2: Based on the full wave field signal and the actual measurement signal obtained in step 4.1, iteratively optimize x d and z to minimize the value of the following objective function:

[0031]

[0032] Denote the obtained optimal low-dimensional vector and defect position as and

[0033] Step 4.3: Based on the defect position and the low-dimensional vector obtained in step 4.2, substitute them into the following formula to reconstruct the full wave field signal:

[0034]

[0035] Substitute into the following formula to reconstruct the defect image:

[0036]

[0037] The present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the guided wave full wave field inversion method based on the physical information fusion generative adversarial neural network are implemented.

[0038] The present invention also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the guided wave full wave field inversion method based on the physical information fusion generative adversarial neural network are implemented.

[0039] The beneficial effects of the present invention:

[0040] 1. The method described in the present invention uses a wave field feature generator as a constraint to learn the mapping relationship between low-dimensional vectors and (high-dimensional) guided wave fields. By optimizing the low-dimensional vectors, the reconstruction of the guided wave field and the defect image is realized, significantly reducing the ill-posedness of the inverse problem of reconstructing the guided wave field from the signals of a small number of measurement points, thereby reducing the difficulty of reconstructing the guided wave field and the defect image.

[0041] 2. The method described in the present invention realizes the fusion of physical information into the generative adversarial neural network by adding a guided wave energy conservation constraint to the generator training error function and connecting a physical model transformed from wave field features after the generator, making the output of the generator more in line with the physical characteristics of guided waves, thereby improving the accuracy of guided wave field reconstruction.

[0042] 3. The method described in the present invention generates the guided wave field through the generator and the physical model, and the time required for this process is less than that consumed by traditional physical calculation models such as the finite element method and the finite difference method, which can increase the speed of reconstructing the guided wave field and the defect image.

[0043] 4. Compared with the method of using a neural network model to replace the entire full-wave field inversion, the method described in the present invention can be applied to different measurement point arrangement schemes. Compared with the method of inputting signals and outputting defect images based on a neural network model, it can output the guided wave field and explain the sources of each wave packet in the measurement signal, having interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without creative efforts.

[0045] Figure 1 Schematic diagram of the guided wave full-wave field inversion method based on the generative adversarial neural network described in the present invention;

[0046] Figure 2 Schematic diagram of the architecture and training of the physical information fusion generative adversarial neural network described in the present invention;

[0047] Figure 3 Schematic diagram of the measurement point arrangement and defect position in the embodiment of the present invention;

[0048] Figure 4Schematic diagrams of the guided wave field, defect location, and defect image obtained by using the method of the present invention in the embodiments of the present invention. Among them, (a) is the schematic diagram of the guided wave field; (b) is the schematic diagram of the defect location; (c) is the schematic diagram of the defect image. Detailed implementation manners

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] The purpose of the present invention is to propose a guided wave full-wavefield inversion method based on a generative adversarial neural network with physical information fusion to recover the guided wave field using a small number of measurement point signals and obtain the defect location and geometry. First, a generative adversarial neural network with physical information fusion is designed and trained using a guided wave field dataset to learn the mapping relationship between the low-dimensional vector and the (high-dimensional) guided wave field. Then, by minimizing the difference between the generated guided wave field signal obtained by the generator and the actual measured signal, the low-dimensional vector and the defect location parameters are inverted. Finally, based on the inverted low-dimensional vector and defect location parameters, the guided wave full-wavefield signal and the defect image are reconstructed.

[0051] Combined with Figure 1 , the present invention proposes a guided wave full-wavefield inversion method based on a generative adversarial neural network with physical information fusion, which specifically includes the following steps:

[0052] Step 1: Establish an original dataset containing structural defects and with diversity, and extract the guided wave field features from the original dataset to form a new dataset for training the generative adversarial network;

[0053] Step 2: Design the network architectures of the generator G θ,I and G θ,II and the discriminator D θ in the generative adversarial neural network; set the initial values of the network training hyperparameters, train the generative adversarial neural network, and through tuning the hyperparameters, obtain the generators G θ,I and G θ,II that can accurately generate the guided wave field features and defect images and have diversity;

[0054] Step 3: Measure the guided wave signal of the target structure and record it as the actual measured signal Y;

[0055] Step 4: The full-wavefield inversion process optimizes the low-dimensional vector z and the defect location parameter x dImplementation. Generate the initial reconstructed full-wavefield feature G θ,I (z), and transform the reconstructed full-wavefield feature G θ,I (z) into the reconstructed full-wavefield signal u|x d ,G θ,I (z)]. According to the actually measured signal Y, the measurement matrix Φ of the target structure, and the reconstructed full-wavefield signal u[x d ,G θ,I (z)], based on the objective function of optimizing the low-dimensional vector z and the defect position x d Finally, obtain the optimal low-dimensional vector and the defect position Substitute the low-dimensional vector and the defect position into the generator G θ,I and G θ,II to reconstruct the full-wavefield signal and the corresponding reconstructed defect image

[0056] The specific content of Step 1 is as follows:

[0057] Step 1.1: Use the physical numerical calculation method to calculate the guided-wave measurement signal Y of the structure under different defect conditions FEM . According to the physical equation of guided-wave propagation, excite the guided-wave signal v0[t] at the excitation point s of the structure. The guided-wave signal scatters at the defect point r, and the wavefield feature FEM in the measurement signal Y can be extracted by the following formula:

[0058]

[0059] where the propagation operator can be calculated by the following formula:

[0060]

[0061] where a and b take the corresponding r, q or s, q or s, r respectively; the operators and are the Fourier transform and its inverse transform respectively. d ref is the reference distance, which can be taken as the mean value of the distances from the excitation point to each receiving point. c p (f) is the phase velocity of the guided wave at frequency f.

[0062] Step 1.2: Combine the wavefield feature extracted in Step 1.1 with the corresponding defect image D d to form a new data set where N is the number of samples. ​

[0063] Step 2 is specifically as follows:

[0064] Step 2.1: For the generation tasks of wave field features and defect images, convolutional neural networks that can effectively extract image features are respectively selected to construct the generator G θ,I and G θ,II as well as the discriminator D θ network architectures; among them, the generator G θ,I and G θ,II input the low-dimensional vector z randomly sampled from the standard normal distribution i , and the low-dimensional vector z i is projected by the generator G θ,I and G θ,II networks respectively to generate the wave field feature and defect image G θ,I (z i ) and G θ,II (z i ); the discriminator D θ inputs the samples in the real dataset or the output of the generator (G θ,I (z i ), G θ,II (z i )) and outputs a numerical value through the discriminator D θ . The closer the numerical value corresponding to the overall generator output is to the numerical value corresponding to the overall samples in the real dataset, the closer the data distribution of the generator output is to the sample distribution of the real dataset;

[0065] Step 2.2: Use the method of adversarial training to train the generative adversarial network. Among them, the discriminator network aims to determine whether the input comes from the real dataset, and its training error function is:

[0066]

[0067] where and x i respectively represent the generator output and the real dataset sample, that is, (G θ,I (z i ), G θ,II (z i )) and is the interpolation value between the generator output and the real dataset sample, that is and ∈ i obeys the uniform distribution U[0,1] from 0 to 1; m is the number of samples in each min batch; λ is the gradient penalty sparsity, and the default value is taken as 10.

[0068] The generator network aims to generate as realistic wavefield features as possible to confuse the discrimination of the discriminator, and its training error function is:

[0069]

[0070] Where, E d and E ud are the total energies of the ultrasonic guided wave at time t1 on the defect / intact structure, which can be calculated by the expression ; ρ μ is the ratio of the guided wave energy to the signal energy in the μ mode of the guided wave, which can be obtained through finite element calculation; is the amplitude of the guided wave at the polar coordinate (r,θ) position in the defect / intact structure.

[0071] Alternately update the parameters of the discriminator network and the generator network, and continuously improve the network's learning degree of the features of the real dataset through a dynamic game process;

[0072] Step 2.3. Evaluate the learning degree of the generator G θ,I and G θ,II for the real dataset by calculating the quality and diversity indicators of the generated wavefield features and defect images, and then optimize the network architecture and training hyperparameters of the generative adversarial network to obtain the generators G θ,I and G θ,II that effectively represent the wavefield features and defect images.

[0073] The specific content of the fourth step is as follows:

[0074] Step 4.1. Establish the physical model of the guided wave energy conservation discrimination condition and the transformed wavefield signal u[x d ,G θ,I (z)], and use the following expression based on the physical model to transform the reconstructed full wavefield feature G θ,I (z) into the full wavefield signal u[x d ,G θ,I (z)]:

[0075]

[0076] Step 4.2. Based on the full wavefield signal and the actual measurement signal obtained in Step 4.1, iteratively optimize x d and z to minimize the value of the following objective function:

[0077]

[0078] Denote the obtained optimal low-dimensional vector and defect position as and

[0079] Step 4.3. Based on the defect positions obtained in Step 4.2 and the low-dimensional vectors substitute them into the following formula to reconstruct the full-wavefield signal:

[0080]

[0081] Substitute them into the following formula to reconstruct the defect image:

[0082]

[0083] Embodiment

[0084] This embodiment applies the present invention to the defect detection of aluminum plate components Figure 1 and gives the flowchart of the method described in the present invention Figure 2 and gives the schematic diagram of the training process of the neural network model Figure 3 and gives the schematic diagram of the arrangement of measurement points and the actual defect positions. The excitation signal is a sine signal with a Hanning window narrowband five-peak, and the center frequency is taken as 100 kHz. The plate material is aluminum, and the thickness is 6 mm.

[0085] The specific content of Step 1 is as follows: Use the finite element method to calculate the guided wave measurement signal Y of the aluminum plate under different defect conditions FEM , and extract the corresponding wavefield features The extracted wavefield features and the corresponding defect image D d are combined to form a new dataset

[0086] The specific content of Step 2 is as follows: After obtaining the above dataset, refer to the deep convolutional generative adversarial network to construct the generator G θ,I and G θ,II and the discriminator D θ . Both generators G θ,I and G θ,II are composed of four layers of transposed convolutional layers, batch normalization layers, and activation function layers. Among them, the low-dimensional vector z input by the input layer is randomly sampled from the standard normal distribution, its dimension is 100, and the resolution of the output image is 32 pixels × 32 pixels; the discriminator is composed of four layers of convolutional layers, batch normalization layers, and activation function layers. Among them, the resolution of the input image of the input layer is 32 pixels × 32 pixels, and the output is a value between 0 and 1. Use the Adam optimization algorithm to initialize various training parameters (batch size, number of training epochs, learning rate, etc.), use the adversarial training method to train the generative adversarial network, and optimize the training parameters to obtain the generators G θ,I and G θ,II that have effectively learned the wavefield features and defect images, and fix their network parameters θ without change.

[0087] Step 3 is specifically as follows: In the numerical simulation example, a guided wave signal is excited at measurement point 1, and measurement signals are received at the remaining measurement points. By analogy, the guided wave signal is alternately excited at each measurement point, and measurement signals are received at the remaining measurement points. Finally, 30 measurement signals are obtained to form a measurement signal matrix Y.

[0088] Step 4 is specifically as follows: According to the generator G obtained in Step 2 θ,I and the measurement signal Y obtained in Step 3, and based on the physical model u[x d , G θ,I (z)], an optimization algorithm is used to minimize the objective function to obtain the optimal low-dimensional vector and defect position as and Substitute into:

[0089]

[0090] to obtain the reconstructed full-wavefield signal and the reconstructed defect image. The snapshots of the reconstructed full-wavefield signal, the positions of the defects, and the reconstructed defect images are respectively as shown in Figure 4 (a) of Figure 4 (b) of Figure 4 (c) of

[0091] The present invention proposes a guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial network. The method includes damage wavefield feature extraction, network architecture design of the generative adversarial network, and a method for reconstructing the guided wave field and damage geometry images, etc. The method of the present invention uses the output signal of the guided wave field generator of the generative adversarial network after training to match the actual measurement signal, and reconstructs the guided wave field and damage images through the random sequence corresponding to the matching result. This method can be flexibly applied to working conditions with various measurement point arrangement schemes, and has a high reconstruction efficiency and certain robustness to noise.

[0092] The present invention also proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial network are implemented.

[0093] The present invention also proposes a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps of the guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial network are implemented.

[0094] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DRRAM). It should be noted that the memory of the method described in the present invention is intended to include but not limited to these and any other suitable types of memory.

[0095] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a high-density digital video disc (DVD)), or a semiconductor medium (such as a solid state disc (SSD)), etc.

[0096] In the implementation process, the steps of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed by the hardware processor or completed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0097] It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above method embodiments can be completed by the integrated logic circuit in the hardware of the processor or instructions in the form of software. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by the combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0098] The above has introduced in detail a guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial neural network proposed by the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A guided wave full-wavefield inversion method based on a physics-informed fusion generative adversarial neural network, characterized in that, The method includes the following steps: Step 1: Establish an original dataset containing structural defects and with diversity, extract guided wave field features from the original dataset to form a new dataset for training a generative adversarial network; Step 2: Design the generators G θ,I and G θ,II and the discriminator D θ of the generative adversarial neural network respectively; set the initial values of the network training hyperparameters, train the generative adversarial neural network, and obtain the generators G θ,I and G θ,II that can accurately generate guided wave field features and defect images and have diversity through hyperparameter tuning; Step 3: Measure the guided wave signal of the target structure and record it as the actual measured signal Y; Step 4. The process of full-wavefield inversion is achieved by optimizing the low-dimensional vector z and the defect position parameter x d ; Generate the initial reconstructed full-wavefield feature G θ,I (z) through the generator, and transform the reconstructed full-wavefield feature G θ,I (z) into the reconstructed full-wavefield signal u[x d , G θ,I (z)]; Based on the actual measured signal Y, the measurement matrix Φ of the target structure, and the reconstructed full-wavefield signal u[x d , G θ,I (z)], obtain the optimal low-dimensional vector d and the defect position x according to the objective function for optimizing the low-dimensional vector z and the defect position Substitute the low-dimensional vector and the defect position into the generators G θ,I and G θ,II to reconstruct the full-wavefield signal and the corresponding reconstructed defect geometric image The specific content of Step 1 is as follows: Step 1.1: Use physical numerical calculation methods to calculate the guided wave measurement signal Y of the structure under different defect conditions FEM , according to the physical equation of guided wave propagation, excite the guided wave signal v0[t] at the excitation point s of the structure. The guided wave signal is scattered at the defect point r, and the wave field characteristics in the measurement signal Y FEM can be extracted as follows: can be extracted according to the following formula: Among them, the propagation operator can be calculated by the following formula: Among them, a and b respectively take the corresponding r, q or s, q or s, r; the operator and are respectively the Fourier transform and its inverse transform; d ref is the reference distance, which can be taken as the mean value of the distances from the excitation point to each receiving point; c p (f) is the phase velocity of the guided wave at frequency f; Step 1.

2. Combine the wave field features extracted in Step 1.1 with the corresponding defect image D d to form a new data set where N is the number of samples; In Step 2, for the generation tasks of wave field features and defect images, convolutional neural networks that can effectively extract image features are respectively selected to construct the generators G θ,I and G θ,II as well as the discriminator D θ 's network architectures; among them, the generators G θ,I and G θ,II input the low-dimensional vector z randomly sampled from the standard normal distribution i , and the low-dimensional vector z i is projected by the networks of the generators G θ,I and G θ,II into the generated wave field features and defect images G θ,I (z i ) and G θ,II (z i ); the discriminator D θ inputs the samples in the real dataset or the output of the generator (G θ,I (z i ), G θ,II (z i )) and outputs a numerical value through the discriminator D θ ; the closer the numerical value corresponding to the overall generator output is to the numerical value corresponding to the overall samples in the real dataset, the closer the data distribution of the generator output is to the sample distribution of the real dataset; In Step 2, the generative adversarial network is trained by using an adversarial training method. Among them, the discriminator network aims to judge whether the input comes from the real dataset, and its training error function is: Among them, and x i respectively represent the generator output and the real dataset sample, that is, (G θ,I (z i ), G θ,II (z i )) and is the interpolation value between the generator output and the real dataset sample, that is and ∈ i follows the uniform distribution U[0, 1] from 0 to 1; m is the number of samples in each min batch; λ is the gradient penalty sparsity.

2. The method according to claim 1, wherein In Step 2, the generator network aims to generate as realistic wave field features as possible to confuse the judgment of the discriminator, and its training error function is: Among them, E d and E ud are the total energies of the ultrasonic guided wave at time t1 on the defect / intact structure, which can be calculated by the expression ρ μ is the ratio of the guided wave energy to the signal energy in the μ mode of the guided wave, which can be obtained through finite element calculation; is the amplitude of the guided wave at the polar coordinate (r, θ) position in the defect / intact structure.

3. The method according to claim 2, characterized in that In Step 2, alternately update the parameters of the discriminator network and the generator network, and continuously improve the network's learning degree of the real dataset features through a dynamic game process.

4. The method according to claim 3, characterized in that, In Step 2, evaluate the generator G by calculating the indicators of the quality and diversity of the generated wavefield features and defect images, and then optimize the network architecture and training hyperparameters of the generative adversarial network to obtain the generator G that effectively represents the wavefield features and defect geometry images θ,I and G θ,II 's learning degree of the real dataset, thereby optimizing the network architecture and training hyperparameters of the generative adversarial network to obtain the generator G that effectively represents the wavefield features and defect geometry images θ,I and G θ,II .

5. The method according to claim 4, wherein The specific content of Step 4 is as follows: Step 4.

1. Establish a physical model of the guided wave energy conservation discrimination condition and the converted wavefield signal u[x d ,G θ,I (z)], and use the following expression based on the physical model to convert the reconstructed full-wavefield feature G θ,I (z) into the full-wavefield signal u[x d ,G θ,I (z)]: Step 4.

2. Based on the full-wavefield signal and the actually measured signal obtained in Step 4.1, iteratively optimize x d and z to minimize the value of the following objective function: Remember that the optimal low-dimensional vector and the defect position are and Step 4.3: Based on the defect positions obtained in Step 4.2 and the low-dimensional vectors substitute them into the following formula to reconstruct the full-wavefield signal: Substitute into the following formula to reconstruct the defect image:

6. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-5.

7. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, it implements the steps of the method according to any one of claims 1-5.

Images