An ultrasonic guided wave defect inversion imaging method, system, device and medium based on unsupervised deep learning

Abstract

The application discloses an ultrasonic guided wave defect inversion imaging method and system based on unsupervised deep learning, equipment and medium, mainly relates to the field of inversion imaging technology, in order to solve the problem of high dependence on sample and insufficient model generalization ability of the existing scheme. Including: collecting and preprocessing ultrasonic guided wave detection signal, obtaining standardized input signal data set; using unsupervised defect inversion network, the input signal data set is mapped to the defect velocity field distribution matrix; based on wave equation, the defect velocity field distribution matrix is forward, and the predicted ultrasonic guided wave detection signal is generated; the error of the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, the error is minimized by reverse propagation through chain rule to optimize the unsupervised inversion network parameters, and the reconstruction error is obtained; when the reconstruction error converges to below the threshold value or reaches the set iteration number, the defect velocity field is output; the defect velocity field is converted into the residual thickness field as the defect quantitative imaging result.

Description

Technical Field

[0001] This application relates to the field of inversion imaging technology, and in particular to an ultrasonic guided wave defect inversion imaging method, system, device and medium based on unsupervised deep learning. Background Technology

[0002] Ultrasonic guided wave testing, with its advantage of enabling large-area, long-distance scanning of waveguide structures through single-point excitation, has become an important technology for detecting defects in structures such as plates and pipes. To visually represent the defect state, various imaging methods have emerged. Among them, supervised deep learning imaging technology, by establishing a direct mapping relationship between the detection signal and the defect distribution map, has achieved rapid imaging of aluminum plate defects, demonstrating its application value in specific scenarios.

[0003] However, this type of supervised imaging method has obvious limitations. First, it is highly dependent on samples, requiring a large number of samples labeled with real defect distribution maps to train the model. However, the morphology of defects in real scenes is complex and diverse, and it is difficult to fully cover them with standard samples made through precision processing. Furthermore, simulated samples deviate from actual detection signals, resulting in a lack of reliable basis for model training. Second, the model has insufficient generalization ability. Models trained for defects with known shapes, sizes, and locations have poor adaptability to unknown defects and are difficult to meet the needs of complex on-site detection. Summary of the Invention

[0004] This application provides a method, system, device, and medium for ultrasonic guided wave defect inversion imaging based on unsupervised deep learning, in order to solve the problems of high sample dependence and insufficient model generalization ability of existing solutions.

[0005] In a first aspect, this application provides an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning, the method comprising: The ultrasonic guided wave detection signal was acquired and preprocessed to obtain a standardized input signal dataset. An unsupervised defect inversion network is used to map the input signal dataset into a defect velocity field distribution matrix; Based on the wave equation, the defect velocity field distribution matrix is ​​forward modeled to generate a predicted ultrasonic guided wave detection signal; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, and the error is minimized by backpropagation using the chain rule to optimize the parameters of the unsupervised inversion network and obtain the reconstruction error. When the reconstruction error converges to below the threshold or reaches the set number of iterations, the defect velocity field is output. The defect velocity field is converted into the residual thickness field as the result of quantitative defect imaging.

[0006] In one implementation of this application, the ultrasonic guided wave detection signal is acquired and preprocessed to obtain a standardized input signal dataset, specifically including: An array of transceiver sensors is arranged on the surface of the structure under test; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

[0007] In one implementation of this application, an unsupervised defect inversion network is used to map the input signal dataset into a defect velocity field distribution matrix, specifically including: The input layer of the unsupervised defect inversion network receives a standardized input signal dataset. By using the feature extraction layer and fully connected layer of an unsupervised defect inversion network, the nonlinear mapping relationship between the signal and the defect in the input signal dataset is predicted, and the velocity field distribution matrix of the predicted defect is obtained. The output layer of the unsupervised defect inversion network outputs the predicted defect velocity field distribution matrix.

[0008] In one implementation of this application, a predicted ultrasonic guided wave detection signal is generated by performing a forward modeling of the defect velocity field distribution matrix based on the wave equation, specifically including: Based on the two-dimensional frequency domain elastic wave equation: To obtain the predicted ultrasonic guided wave detection signal; Where i is the imaginary unit, Angular frequency, To predict ultrasonic guided wave detection signals, Assuming a pre-defined frequency domain point source, The mass matrix corresponding to the defect velocity field distribution matrix. This is the stiffness matrix corresponding to the defect velocity field distribution matrix. For spatial coordinates, This is the damping parameter.

[0009] In one implementation of this application, the error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, and the error is minimized through backpropagation using the chain rule to update the parameters of the unsupervised inversion network, thereby obtaining the reconstruction error. Specifically, this includes: Define the standardized input signal as The predicted ultrasonic guided wave detection signal is defined as follows: ;in, Angular frequency, Spatial coordinates; Through the loss function : Calculate the error between the current unsupervised inversion network parameters used to predict the ultrasonic guided wave detection signal and the standardized input signal; Where N is the total number of samples and n is the sample number; Through the formula: Calculate the gradient of the loss function with respect to the predicted ultrasonic guided wave detection signal; where, To retrieve the training parameters for the inverted network; Through the formula: Calculate and predict the physical gradient of the ultrasonic guided wave detection signal with respect to the velocity field; in, Let c be the finite-difference small perturbation step size, and let c represent the defect velocity field distribution matrix. This represents the predicted ultrasonic guided wave detection signal corresponding to an increase in the finite difference small perturbation step size. This represents the predicted ultrasonic guided wave detection signal corresponding to a reduced finite difference small perturbation step size. Through the formula: Calculate the gradient of the loss with respect to the defect velocity field. ; Through the formula: Calculate the gradient of the defect velocity field with respect to the network parameters; where, This represents the mapping relationship between the input signal and the defect velocity field. Calculate the total gradient using the gradient chain rule: , Updating unsupervised inversion network parameters using the Adam optimizer : ,in, The learning rate; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated using the updated unsupervised inversion network parameters as the reconstruction error.

[0010] In one implementation of this application, the defect velocity field is converted into a residual thickness field as the result of quantitative defect imaging, specifically including: The converged inversion velocity field is converted into the residual thickness field using a dispersion relation mapping operator: The remaining thickness field is output as the result of quantitative imaging of defects.

[0011] Secondly, this application provides an ultrasonic guided wave defect inversion imaging system based on unsupervised deep learning, the system comprising: The module is used to acquire and preprocess ultrasonic guided wave detection signals to obtain a standardized input signal dataset. The mapping module is used to map the input signal dataset into a defect velocity field distribution matrix using an unsupervised defect inversion network; The generation module is used to perform forward modeling of the defect velocity field distribution matrix based on the wave equation to generate a predicted ultrasonic guided wave detection signal. The module is used to calculate the error between the predicted ultrasonic guided wave detection signal and the standardized input signal. The error is minimized by backpropagation using the chain rule to optimize the parameters of the unsupervised inversion network and obtain the reconstruction error. The output module is used to output the defect velocity field when the reconstruction error converges to below the threshold or reaches the set number of iterations; the defect velocity field is converted into the remaining thickness field as the result of quantitative imaging of the defect.

[0012] In one implementation of this application, the obtaining module includes an obtaining unit for arranging a transceiver integrated sensor array on the surface of the structure to be tested; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

[0013] Thirdly, this application provides an ultrasonic guided wave defect inversion imaging device based on unsupervised deep learning, the device comprising: processor; And a memory containing executable code, which, when executed, causes the processor to execute an unsupervised deep learning-based ultrasonic guided wave defect inversion imaging method, as described above.

[0014] Fourthly, this application provides a non-volatile computer storage medium storing computer instructions, which, when executed, implement an unsupervised deep learning-based ultrasonic guided wave defect inversion imaging method as described above.

[0015] As can be seen from the above technical solutions, this application has the following advantages: This application employs an unsupervised deep learning architecture, eliminating the need for training samples labeled with real defect distribution maps. This fundamentally avoids the data scarcity problem caused by the uniformity of standard specimen morphology and the distortion of simulation signals in traditional methods. By directly utilizing the original ultrasonic guided wave detection signal as input, combined with an unsupervised defect inversion network and a wave equation forward modeling closed-loop mechanism, the system can adaptively learn and reconstruct defect features without labels. This design frees the model training basis from the limitations of artificially prepared specimens or numerical simulations, broadening the applicable defect types and working conditions, and providing a feasible training paradigm for the detection of complex, non-standard defects in industrial settings.

[0016] Since this application does not rely on prior training on specific defect shapes, sizes, or locations, its unsupervised inversion network dynamically learns the physical mapping relationship between defects and ultrasonic responses by minimizing the reconstruction error between predicted and measured signals. This mechanism enables the system to perform reasonable inversion based on the physical constraints of the wave equation when faced with defect morphologies that have not appeared in the training phase, outputting a defect velocity field distribution that conforms to physical laws. Compared to the overfitting risk of traditional supervised models for known defects, this scheme achieves a paradigm shift from "memory-based recognition" to "inference-based inversion," enhancing its robustness and adaptability in complex field environments.

[0017] By directly converting the defect velocity field output by the unsupervised network into the residual thickness field via a wave equation physical model, an end-to-end quantitative imaging chain is formed. This process does not rely on empirical calibration or manual parameter setting; all conversions are based on the physical laws of wave propagation, ensuring that the imaging results have clear physical meaning and interpretability. Compared to existing methods that require additional calibration curves or empirical formulas for thickness estimation, this scheme achieves a direct, continuous, and high-fidelity quantitative expression of defect depth and size without additional calibration, providing a reliable and traceable structured output for defect assessment. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of an unsupervised defect inversion network structure provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the overall process of ultrasonic guided wave defect inversion imaging based on unsupervised deep learning, provided in an embodiment of this application.

[0022] Figure 4 This is a schematic diagram of the internal structure of an ultrasonic guided wave defect inversion imaging system based on unsupervised deep learning, provided in an embodiment of this application.

[0023] Figure 5 This is a schematic diagram of the internal structure of an ultrasonic guided wave defect inversion imaging device based on unsupervised deep learning, provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0026] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0027] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] The embodiment provides an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning, such as... Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Acquire and preprocess the ultrasonic guided wave detection signal to obtain a standardized input signal dataset.

[0029] In some embodiments, this step may specifically include: An array of transceiver sensors is arranged on the surface of the structure under test; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

[0030] It should be further explained that the use of a transceiver integrated sensor array in this step can reduce the phase deviation and signal delay inconsistency caused by the installation of multiple probes, improve the spatial consistency of the detected signal, and reduce the feature extraction error caused by differences in sensor layout. By filtering and extracting effective signal segments through window functions, environmental noise, non-target mode interference, and boundary reflection artifacts can be effectively suppressed, ensuring that the signal segments involved in the analysis have clear physical correspondence. After extracting the frequency domain amplitude features through fast Fourier transform, the changes in energy distribution caused by defects can be objectively characterized, providing a quantitative basis directly related to the material state for subsequent pattern recognition. Normalizing the frequency domain amplitude and mapping it to the [0,1] interval can eliminate the amplitude offset caused by amplifier gain fluctuations, sensor sensitivity differences, and excitation voltage drift under different detection conditions, making signal data from different devices, different times, or different structural parts comparable, thereby improving the convergence stability and generalization ability of the machine learning model during the training process and avoiding model performance degradation caused by inconsistent input scales.

[0031] Step 120: Using an unsupervised defect inversion network, the input signal dataset is mapped to a defect velocity field distribution matrix.

[0032] In some embodiments, this step may specifically include: The input layer of the unsupervised defect inversion network receives a standardized input signal dataset. By using the feature extraction layer and fully connected layer of an unsupervised defect inversion network, the nonlinear mapping relationship between the signal and the defect in the input signal dataset is predicted, and the velocity field distribution matrix of the predicted defect is obtained. The output layer of the unsupervised defect inversion network outputs the predicted defect velocity field distribution matrix.

[0033] Specifically, the unsupervised defect inversion network has an encoder structure, including: an input layer, a feature extraction layer, a fully connected layer, and an output layer. More specifically, such as... Figure 2The diagram shows the unsupervised defect inversion network structure of this application, including: an input layer, a feature extraction layer, a fully connected layer, and an output layer; the input layer receives a standardized input signal dataset; the feature extraction layer and the fully connected layer are used to predict the nonlinear mapping relationship between the received signal and the defect; the output layer outputs the predicted defect velocity field distribution matrix; since there is no real defect velocity field distribution, it is impossible to directly calculate the error between the real defect velocity field distribution and the predicted defect velocity field distribution and perform backpropagation of the error; subsequent operations are required to make the predicted defect velocity field distribution matrix close to the real defect velocity field distribution matrix.

[0034] It should be noted that this step uses an unsupervised defect inversion network to map the standardized input signal dataset into a defect velocity field distribution matrix. This avoids reliance on manually labeled defect samples or precise physical forward models, reducing dependence on prior knowledge and enabling the model to directly learn complex nonlinear mapping relationships from measured signals. The encoder structure preserves the multi-scale frequency domain features of the signal through multi-layer feature extraction and fully connected layers, achieving end-to-end reconstruction from discrete sensor responses to continuous spatial velocity fields. The output results have spatial continuity and can be directly used for quantitative characterization of defect boundaries and morphology without post-processing interpolation or mesh interpolation. The defect velocity field distribution matrix output by this method directly corresponds to the spatial layout of the sensor array, forming a seamless data chain with subsequent defect size assessment and localization algorithms. This improves the adaptability and automation of the overall detection system in complex geometric structures and multimodal wave propagation environments. Furthermore, since the input is the standardized data generated in step 110, no additional calibration or normalization processing is required, ensuring the stability and reproducibility of the process.

[0035] Step 130: Perform forward modeling of the defect velocity field distribution matrix based on the wave equation to generate the predicted ultrasonic guided wave detection signal.

[0036] In some embodiments, this step may specifically include: Based on the two-dimensional frequency domain elastic wave equation: To obtain the predicted ultrasonic guided wave detection signal; Where i is the imaginary unit, Angular frequency, To predict ultrasonic guided wave detection signals, For a pre-defined known frequency domain point source (an ultrasonic excitation source at a pre-defined point in space within the frequency domain, which is a known external force term on the right-hand side of the wave equation, used to describe the position, frequency, and intensity of the probe, thereby solving for the frequency domain displacement wave field on the left-hand side of the equation)). The mass matrix corresponding to the defect velocity field distribution matrix. This is the stiffness matrix corresponding to the defect velocity field distribution matrix. For spatial coordinates, This is the damping parameter.

[0037] It should be noted that this step uses a two-dimensional frequency domain elastic wave equation to perform forward modeling on the defect velocity field distribution matrix. This can realistically reconstruct the propagation response of ultrasonic guided waves in a defective medium based on a physical model, ensuring that the generated initial signal is consistent with the actual physical mechanism and avoiding systematic deviations introduced by empirical models.

[0038] Step 140: Calculate the error between the predicted ultrasonic guided wave detection signal and the standardized input signal, minimize the error through backpropagation using the chain rule to optimize the parameters of the unsupervised inversion network, and obtain the reconstruction error.

[0039] In some embodiments, the error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, and the error is minimized through backpropagation using the chain rule to update the parameters of the unsupervised inversion network, thereby obtaining the reconstruction error. Specifically, this includes: The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated. This error is then minimized through backpropagation using the chain rule to update the parameters of the unsupervised inversion network, yielding the reconstruction error, which specifically includes: Define the standardized input signal as The predicted ultrasonic guided wave detection signal is defined as follows: ;in, Angular frequency, Spatial coordinates; Through the loss function : Calculate the error between the current unsupervised inversion network parameters used to predict the ultrasonic guided wave detection signal and the standardized input signal; Where N is the total number of samples and n is the sample number; Through the formula: Calculate the gradient of the loss function with respect to the predicted ultrasonic guided wave detection signal; where, To retrieve the training parameters for the inverted network; Through the formula: Calculate and predict the physical gradient of the ultrasonic guided wave detection signal with respect to the velocity field; in, Let c be the finite-difference small perturbation step size, and let c represent the defect velocity field distribution matrix. This represents the predicted ultrasonic guided wave detection signal corresponding to an increase in the finite difference small perturbation step size. This represents the predicted ultrasonic guided wave detection signal corresponding to a reduced finite difference small perturbation step size. Through the formula: Calculate the gradient of the loss with respect to the defect velocity field. ; Through the formula: Calculate the gradient of the defect velocity field with respect to the network parameters; where, This represents the mapping relationship between the input signal and the defect velocity field. Calculate the total gradient using the gradient chain rule: , Updating unsupervised inversion network parameters using the Adam optimizer : ,in, The learning rate; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated using the updated unsupervised inversion network parameters as the reconstruction error.

[0040] It should be further explained that this step uses a loss function to calculate the mean square error between the predicted ultrasonic guided wave detection signal and the standardized input signal. This quantifies the deviation between the unsupervised inversion network output and the measured signal, providing a measurable convergence target for parameter optimization. The physical gradient of the predicted signal with respect to the defect velocity field is calculated using the finite difference method, explicitly embedding the physical constraints of the wave equation into the backpropagation process. This ensures that the gradient update direction conforms to the physical laws of elastic wave propagation, avoiding non-physical interpretations caused by purely data-driven approaches. Finally, the chain rule is used to propagate the gradient of the loss with respect to the network parameters layer by layer, achieving end-to-end parameter optimization from the output layer to each layer of the encoder. The data update does not rely on external supervision labels or analytical derivatives of the forward model; the Adam optimizer adaptively adjusts the learning rate, which can alleviate gradient vanishing or oscillation problems, achieve stable convergence in the non-convex loss space, and improve parameter iteration efficiency; the final output reconstruction error serves as a direct indicator of the network training state, which can reflect the evolution trend of defect inversion accuracy in real time, providing a traceable diagnostic basis for system operation, and this error value is fully compatible with the standardized processing flow of steps 110 and 130, ensuring the consistency of data scale and processing logic in the entire detection-inversion-verification closed loop without the need for additional normalization or calibration operations.

[0041] Step 150: When the reconstruction error converges to below the threshold or reaches the set number of iterations, output the defect velocity field; convert the defect velocity field into the remaining thickness field as the quantitative imaging result of the defect.

[0042] Specifically, converting the defect velocity field into the residual thickness field as the quantitative imaging result of the defect can be achieved as follows: The converged inversion velocity field is converted into the residual thickness field using a dispersion relation mapping operator: The remaining thickness field is output as the result of quantitative imaging of defects.

[0043] It should be further explained that this step uses a dispersion relation mapping operator to convert the converged defect velocity field into a residual thickness field. This allows for the direct spatial quantification of material thickness loss based on the physical mechanism of elastic wave propagation, avoiding the coupling uncertainties and spatial sampling sparsity caused by relying on mechanical thickness probes or contact measurements. This conversion process does not introduce additional interpolation or mesh re-division, preserving the original spatial resolution and coordinate consistency of the inverted velocity field. This ensures that each pixel in the residual thickness field corresponds to the measured position of the sensor array, guaranteeing the spatial fidelity of the imaging results. Since the mapping operator is analytically constructed from the dispersion equation, its calculation process is entirely determined by material properties and wave mode parameters, requiring no manual calibration or empirical correction. This ensures the reproducibility and device independence of the quantitative results. The output residual thickness field is a continuous two-dimensional distribution, which can be directly used for the automated calculation of defect area, contour, and volume. It provides directly input geometric parameters for structural residual life assessment and is fully compatible with the standardized data streams in steps 110 to 140, requiring no additional coordinate transformation or scale adaptation. This forms a seamless, end-to-end processing flow from signal acquisition to quantitative imaging.

[0044] A dispersion relation mapping operator is used to convert the converged defect velocity field into a residual thickness field, mapping the defect velocity field at each spatial location to the corresponding residual thickness field without interpolation or mesh reconstruction. The mapping process is performed on the original discrete coordinate system of the sensor array, ensuring that the output residual thickness field is completely consistent with the input velocity field in terms of spatial resolution and coordinate system, preserving the geometric details of the defect edges. Since the dispersion relation is uniquely determined by the material constitutive parameters and wave mode physical properties, this conversion does not rely on experimental calibration or empirical fitting. It only requires the material constants and operating frequency range of the effective signal segment to achieve device-independent quantitative output. The generated residual thickness field is a two-dimensional real matrix, where each pixel value represents the residual thickness of the material at that point. It can be directly used for numerical integration calculations of defect area, profile length, and volume, providing quantifiable geometric input for structural residual life assessment. Furthermore, it is seamlessly integrated with the standardized signal processing flow of the preceding steps in terms of data format and coordinate system, without the need for additional spatial registration or scale conversion.

[0045] Based on the foregoing description, this application can be specifically as follows: Figure 3The diagram illustrates the overall workflow of this application. Ultrasonic guided wave detection signals are acquired and preprocessed to obtain a standardized input signal dataset. An unsupervised defect inversion network is constructed to map the input signals into a defect velocity field. The inverted defect velocity field is then forward-modeled based on the wave equation to generate a predicted ultrasonic guided wave detection signal. The loss between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, and the loss is minimized through backpropagation using the chain rule to optimize the unsupervised inversion network parameters. When the reconstruction error converges to below a threshold or reaches the set number of iterations, the defect velocity field is output. The velocity field is then converted into a thickness field, and the quantitative defect imaging results are output.

[0046] In addition, this application Figure 4 This application provides an embodiment of an ultrasonic guided wave defect inversion imaging system based on unsupervised deep learning. For example... Figure 4 As shown in the embodiments of this application, the system mainly includes: Module 210 is obtained, which is used to acquire and preprocess ultrasonic guided wave detection signals to obtain a standardized input signal dataset.

[0047] The module 210 includes a receiving unit. Used to deploy a transceiver integrated sensor array on the surface of the structure under test; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

[0048] The mapping module 220 is used to map the input signal dataset into a defect velocity field distribution matrix using an unsupervised defect inversion network.

[0049] The generation module 230 is used to perform forward modeling of the defect velocity field distribution matrix based on the wave equation to generate a predicted ultrasonic guided wave detection signal.

[0050] Module 240 is used to calculate the error between the predicted ultrasonic guided wave detection signal and the standardized input signal. The error is minimized by backpropagation using the chain rule to optimize the parameters of the unsupervised inversion network and obtain the reconstruction error.

[0051] The output module 250 is used to output the defect velocity field when the reconstruction error converges to below the threshold or reaches the set number of iterations; and to convert the defect velocity field into the remaining thickness field as the result of quantitative imaging of the defect.

[0052] The above are method embodiments of this application. Based on the same inventive concept, this application also provides an ultrasonic guided wave defect inversion imaging device based on unsupervised deep learning. Figure 5 As shown, the device includes: a processor; and a memory connected to the processor via a bus, on which executable code is stored. When the executable code is executed, the processor performs an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning as described in the above embodiment.

[0053] Specifically, the server collects and preprocesses ultrasonic guided wave detection signals to obtain a standardized input signal dataset; using an unsupervised defect inversion network, the input signal dataset is mapped to a defect velocity field distribution matrix; based on the wave equation, the defect velocity field distribution matrix is ​​forward modeled to generate a predicted ultrasonic guided wave detection signal; the error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated, and the error is minimized through backpropagation using the chain rule to optimize the parameters of the unsupervised inversion network and obtain the reconstruction error; when the reconstruction error converges to below a threshold or reaches a set number of iterations, the defect velocity field is output; the defect velocity field is converted into a residual thickness field as the quantitative imaging result of the defect.

[0054] In addition, this application embodiment also provides a non-volatile computer storage medium storing executable instructions, which, when executed, implement the ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning as described above.

[0055] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for inverting ultrasonic guided wave defects based on unsupervised deep learning, characterized in that, The method includes: The ultrasonic guided wave detection signal was acquired and preprocessed to obtain a standardized input signal dataset. An unsupervised defect inversion network is used to map the input signal dataset into a defect velocity field distribution matrix; Based on the wave equation, the defect velocity field distribution matrix is ​​forward modeled to generate a predicted ultrasonic guided wave detection signal; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated. This error is then minimized through backpropagation using the chain rule to optimize the unsupervised inversion network parameters, yielding the reconstruction error, which specifically includes: Define the standardized input signal as The predicted ultrasonic guided wave detection signal is defined as follows: ;in, Angular frequency, Spatial coordinates; Through the loss function : Calculate the error between the current unsupervised inversion network parameters used to predict the ultrasonic guided wave detection signal and the standardized input signal; Where N is the total number of samples and n is the sample number; Through the formula: Calculate the gradient of the loss function with respect to the predicted ultrasonic guided wave detection signal; where, To retrieve the training parameters for the inverted network; Through the formula: Calculate and predict the physical gradient of the ultrasonic guided wave detection signal with respect to the velocity field; in, Let c be the finite-difference small perturbation step size, and let c represent the defect velocity field distribution matrix. This represents the predicted ultrasonic guided wave detection signal corresponding to an increase in the finite difference small perturbation step size. This represents the predicted ultrasonic guided wave detection signal corresponding to a reduced finite difference small perturbation step size. Through the formula: Calculate the gradient of the loss with respect to the defect velocity field. ; Through the formula: Calculate the gradient of the defect velocity field with respect to the network parameters; where, This represents the mapping relationship between the input signal and the defect velocity field. Calculate the total gradient using the gradient chain rule: ， Updating unsupervised inversion network parameters using the Adam optimizer : ,in, The learning rate; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated using the updated unsupervised inversion network parameters as the reconstruction error. When the reconstruction error converges to below the threshold or reaches the set number of iterations, the defect velocity field is output; the defect velocity field is converted into the remaining thickness field as the result of quantitative defect imaging.

2. The ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning according to claim 1, characterized in that, The ultrasonic guided wave detection signal was acquired and preprocessed to obtain a standardized input signal dataset, which specifically includes: An array of transceiver sensors is arranged on the surface of the structure under test; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

3. The ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning according to claim 1, characterized in that, An unsupervised defect inversion network is used to map the input signal dataset into a defect velocity field distribution matrix, specifically including: The input layer of the unsupervised defect inversion network receives a standardized input signal dataset. By using the feature extraction layer and fully connected layer of an unsupervised defect inversion network, the nonlinear mapping relationship between the signal and the defect in the input signal dataset is predicted, and the velocity field distribution matrix of the predicted defect is obtained. The output layer of the unsupervised defect inversion network outputs the predicted defect velocity field distribution matrix.

4. The ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning according to claim 1, characterized in that, Based on the wave equation, a forward modeling of the defect velocity field distribution matrix is ​​performed to generate a predicted ultrasonic guided wave detection signal, specifically including: Based on the two-dimensional frequency domain elastic wave equation: To obtain the predicted ultrasonic guided wave detection signal; Where i is the imaginary unit, Angular frequency, To predict ultrasonic guided wave detection signals, Assuming a pre-defined frequency domain point source, The mass matrix corresponding to the defect velocity field distribution matrix. This is the stiffness matrix corresponding to the defect velocity field distribution matrix. For spatial coordinates, This is the damping parameter.

5. The ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning according to claim 1, characterized in that, The defect velocity field is converted into the residual thickness field as the result of quantitative defect imaging, specifically including: The converged inversion velocity field is converted into the residual thickness field using a dispersion relation mapping operator: The remaining thickness field is output as the result of quantitative imaging of defects.

6. An ultrasonic guided wave defect inversion imaging system based on unsupervised deep learning, characterized in that, The system includes: The module is used to acquire and preprocess ultrasonic guided wave detection signals to obtain a standardized input signal dataset. The mapping module is used to map the input signal dataset into a defect velocity field distribution matrix using an unsupervised defect inversion network; The generation module is used to perform forward modeling of the defect velocity field distribution matrix based on the wave equation to generate a predicted ultrasonic guided wave detection signal. The module obtains the error between the predicted ultrasonic guided wave detection signal and the standardized input signal. It minimizes this error through backpropagation using the chain rule to optimize the unsupervised inversion network parameters, thus obtaining the reconstruction error, which specifically includes: Define the standardized input signal as The predicted ultrasonic guided wave detection signal is defined as follows: ;in, Angular frequency, Spatial coordinates; Through the loss function : Calculate the error between the current unsupervised inversion network parameters used to predict the ultrasonic guided wave detection signal and the standardized input signal; Where N is the total number of samples and n is the sample number; Through the formula: Calculate the gradient of the loss function with respect to the predicted ultrasonic guided wave detection signal; where, To retrieve the training parameters for the inverted network; Through the formula: Calculate and predict the physical gradient of the ultrasonic guided wave detection signal with respect to the velocity field; in, Let c be the finite-difference small perturbation step size, and let c represent the defect velocity field distribution matrix. This represents the predicted ultrasonic guided wave detection signal corresponding to an increase in the finite difference small perturbation step size. This represents the predicted ultrasonic guided wave detection signal corresponding to a reduced finite difference small perturbation step size. Through the formula: Calculate the gradient of the loss with respect to the defect velocity field. ; Through the formula: Calculate the gradient of the defect velocity field with respect to the network parameters; where, This represents the mapping relationship between the input signal and the defect velocity field. Calculate the total gradient using the gradient chain rule: ， Updating unsupervised inversion network parameters using the Adam optimizer : ,in, The learning rate; The error between the predicted ultrasonic guided wave detection signal and the standardized input signal is calculated using the updated unsupervised inversion network parameters as the reconstruction error. The output module is used to output the defect velocity field when the reconstruction error converges to below the threshold or reaches the set number of iterations; the defect velocity field is converted into the remaining thickness field as the result of quantitative imaging of the defect.

7. The ultrasonic guided wave defect inversion imaging system based on unsupervised deep learning according to claim 6, characterized in that, The module includes the unit of obtaining. Used to deploy a transceiver integrated sensor array on the surface of the structure under test; The host computer generates an excitation signal, which is output through the acquisition card and amplifier. The excitation signal propagates within the structure under test and is received by the sensor array to form an ultrasonic guided wave detection signal. The ultrasonic guided wave detection signal is filtered and a window function is used to extract the effective signal segment. Perform a Fast Fourier Transform on the effective signal segment to extract the frequency domain amplitude features; The frequency domain signal is normalized and mapped to the [0,1] interval to obtain a standardized input signal dataset.

8. An ultrasonic guided wave defect inversion imaging device based on unsupervised deep learning, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning as described in any one of claims 1-5.

9. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement an ultrasonic guided wave defect inversion imaging method based on unsupervised deep learning as described in any one of claims 1-5.

Images