Nondestructive testing method and system for dielectric constant of layered medium based on crcn

Abstract

The application discloses a CRCN-based layered medium dielectric coefficient nondestructive detection method and system. First, the CycleGAN network is trained to suppress noise in a data set and enhance image target features. Secondly, a ResCBAMNet detection network is proposed, which is based on ResNet-18 and introduces a CBAM module in a residual block, and improves an activation function in the network to LeakyReLU. The network can classify and detect the dielectric coefficient of the layered medium. The simulation experiment results show that the CRCN model can realize fast and nondestructive detection of the dielectric coefficient of each layer medium, so as to judge the structural health state of each medium layer, and is suitable for engineering application.

Description

Technical Field

[0001] This invention belongs to the field of deep learning, specifically relating to a method and system for non-destructive testing of the dielectric constant of layered media. Background Technology

[0002] Non-destructive testing (NDT) is a geophysical method that uses electromagnetic waves to detect the distribution of underground media. Its core principle lies in using the differences in the propagation characteristics of physical fields such as electromagnetic waves and sound waves to infer the spatial distribution of underground media. It is widely used in many fields, including construction engineering, archaeology, geological exploration, and environmental monitoring. Ground-penetrating radar (GPR) NDT technology can accurately locate the position and spatial distribution of underground targets, quickly acquiring relevant information without damaging the object being tested. Traditional NDT methods include ultrasonic, X-ray, and magnetic particle methods, which require manual analysis and judgment of the acquired information.

[0003] To reduce human intervention, many studies in recent years have introduced deep learning networks to automatically learn and extract key features from detection data, thereby improving the ability to identify and judge the condition of underground media. CRCN (CycleGAN-ResCBAMNet) is a deep learning architecture that integrates recurrent neural networks (RNNs), generative adversarial networks (GANs), and convolutional neural networks (CNNs). This architecture aims to demonstrate good performance in image-to-image conversion tasks by using RNNs to process sequence information, CNNs to extract spatial features, and leveraging the adversarial training mechanism of GANs. This network structure is commonly used in image style transfer, image super-resolution, and image inpainting. Therefore, a solution is needed to propose that uses CRCN for non-destructive detection of media while improving the ability to focus image features and extract spatial information. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a non-destructive testing method for the dielectric constant of layered media based on CRCN, which detects the corrosivity of the medium based on the curvature change of the echo image. This CRCN non-destructive testing network can realize rapid detection of the health of the medium structure and is suitable for engineering applications.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] This invention proposes a non-destructive testing method for the dielectric constant of layered dielectrics based on CRCN, comprising the following steps:

[0007] S1. Generate CycleGAN training set: Construct the first training set trainA containing noise and attenuation of the simulated echo image, and the second training set trainB containing no noise and no amplitude attenuation of the simulated echo image.

[0008] S2, CycleGAN Network Training: PyTorch is used as the deep learning framework to train CycleGAN, and the feature-enhanced echo image dataset is obtained using the training set of S1.

[0009] S3. Construct an improved CRCN network: Based on the ResNet-18 architecture, embed the convolutional block attention module CBAM into the residual block to obtain the deep learning model ResCBAMNet, and replace the activation function with the Leaky ReLU nonlinear unit.

[0010] S4. CRCN Network Training and Classification: Divide the dataset generated in step S2 into a training set and a test set in a 4:6 ratio. The training set is then divided into training data and validation data in a 6:4 ratio. Use the Adam optimizer to train the deep learning model from step S3 to obtain the detection model.

[0011] S5. Input the feature-enhanced echo image obtained in S2 into the detection model obtained in step S4, and realize rapid corrosion classification and detection of the medium based on the curvature change of the echo image.

[0012] Furthermore, in step S1 of the method proposed in this invention, the generation of the training set for the CycleGAN network specifically involves:

[0013] The dataset was generated using the Matlab simulation platform and employs a SIMO ground-penetrating radar single-transmitter multiple-receiver system. The radar transmits a Gaussian pulse signal, and multiple array elements simultaneously receive the signal, i.e., a transmit-receive bistatic mode. The noisy single-cycle Gaussian pulse transmitted signal g(t) is as follows:

[0014]

[0015] Where σ is the standard deviation of the Gaussian pulse, f0 is the radar operating center frequency, and N(t) represents clutter and Gaussian noise. The clutter in each layer is generated by superimposing clutter from different point targets with Gaussian noise. If the target is in the nth layer of the medium, in SIMO mode, the echo signal S(t) after carrier frequency removal from each array element... n M) is

[0016]

[0017] Where t = [t1, t2, ... t L ] T L is the number of range signal sampling points, S(t) n Let A(M) be an L×M B-scan two-dimensional signal matrix, where M is the number of array elements. i ε is the amplitude of the signal received by array element i. n R is the dielectric constant of the nth layer. inτ is the propagation path distance from the target to the i-th element in the n-th layer of the medium. in Let C be the propagation delay from the target to the i-th element in the n-layer medium, and C be 3 × 10⁻⁶. 8 m / s, where λ is the signal wavelength. The simulation uses a two-layer medium as an example; the echo image exhibits a hyperbolic characteristic. The dielectric constant of the medium is assumed to range from [6, 155]. The dielectric constant of the uncorroded portion ranges from [6, 10].

[0018] The training set of the CycleGAN network consists of two parts: trainA and trainB. TrainA is a Matlab-simulated echo image containing noise and attenuation. TrainB is a simulated echo image without noise and without amplitude attenuation.

[0019] To further facilitate subsequent processing and classification operations, the images are first processed into grayscale. Before processing, the dataset images are normalized to a size of 64×64, making it easier for the model to learn the curvature features in the images. At the same time, the normalized data helps to avoid problems such as gradient vanishing or exploding.

[0020] Furthermore, the CycleGAN network training in step S2 of the method proposed in this invention is specifically as follows:

[0021] The network training experimental platform is a Windows 10 system, equipped with one CPU processor (13th Gen Intel(R) Core(TM) i7-13700F) and one GPU (NVIDIA RTX 4090) with 16GB of video memory. PyTorch version 1.8.1 is used as the deep learning framework, with Python 3.8 and CUDA version 10.2. The best performance obtained on the test set is used to evaluate the algorithm's performance. The dataset generated after training features hyperbolic enhancement, and noise suppression reaches 6dB under the simulation parameters.

[0022] Furthermore, in step S3 of the method proposed in this invention, the improved Cyclic Adversarial Generative-Residual Network (CRCN) specifically includes:

[0023] The non-destructive testing classification model proposed in this invention adopts an improved ResNet-18 architecture. This network implements the following innovative optimizations on the basis of the classic ResNet-18 framework: (1) Introducing a Convolutional Block Attention Module (CBAM) to enhance the model's feature representation ability and accurately focus on key features; (2) The activation function is improved to a Leaky ReLU nonlinear unit, which solves the problem of zero gradient in the case of negative input and avoids the problem of neuron death. This model gives full play to the feature extraction and classification capabilities of ResNet-18 and the focusing and enhancement effect of CBAM on key features, thereby significantly improving the performance of corrosion classification and discrimination.

[0024] We introduce CBAM into the ResNet-18 residual blocks. CBAM (Convolutional Block Attention Module) is a lightweight hybrid attention mechanism that achieves a balance between feature selection and region focusing with extremely low computational cost. It improves the accuracy of feature map focusing and model generalization ability by sequentially deploying the Channel Attention Module (CAM) and the Spatial Attention Module (SAM).

[0025] Activation functions are a core component of deep learning neural networks. Their main principle is to inject nonlinear transformation capabilities into neurons, enabling the network to learn complex patterns and relationships. The ReLU function outputs 0 when the input is negative, and y = x when the input is greater than 0. This shows that it is not differentiable over the entire range, thus solving the vanishing gradient problem. Furthermore, it only needs to check if the input is greater than 0, making its computational efficiency relatively high. It also exhibits network sparsity and mitigates overfitting to some extent. However, it suffers from the Daed ReLU Problem, where some neurons become permanently inactive, preventing updates to the corresponding connection weights and potentially leading to gradient explosion. The Leaky ReLU function, on the other hand, outputs y = x when x is greater than or equal to 0, and introduces a coefficient α in the negative interval (y = α * x). Addressing the Daed ReLU Problem, Leaky ReLU applies a small slope to negative inputs. This non-zero gradient effectively mitigates the permanent inactivation of neurons in deep networks caused by improper weight initialization or excessively high learning rates.

[0026] Furthermore, in step S4 of the method proposed in this invention, the CRCN network training specifically includes:

[0027] The training dataset for this non-destructive testing corrosion classification network consists of noise-free echo images generated by CycleGAN training, whose curvature features are clearer compared to before noise suppression. Due to the significant curvature features after data augmentation, to avoid overfitting, the dataset is randomly divided into training and testing sets in a 4:6 ratio, with the training set randomly divided into training and validation data in a 6:4 ratio. Taking a two-layer medium as an example, the dataset is labeled into four categories: 1 represents no corrosion in the first layer and corrosion in the second layer (hereinafter referred to as "top no corrosion, bottom corrosion"); 2 represents no corrosion in both layers (hereinafter referred to as "top no corrosion, bottom no corrosion"); 3 represents corrosion in the first layer and no corrosion in the second layer (hereinafter referred to as "top corrosion, bottom no corrosion"); and 4 represents corrosion in both layers (hereinafter referred to as "top corrosion, bottom corrosion"). If there are n layers, the dataset is labeled as 2. n kind.

[0028] ResCBAMNet was trained over 20 epochs with a batch size of 64, using the Adam optimizer, a learning rate of 0.0001, and the cross-entropy loss function. The trained model can achieve rapid classification and detection of media corrosion.

[0029] Meanwhile, this invention also proposes a non-destructive testing device for the dielectric constant of layered dielectrics based on CRCN, comprising:

[0030] The signal generation module is used to generate a dataset of noisy echo signals from the SIMO ground-penetrating radar, and to construct a first training set trainA containing clutter and attenuation of the simulated echo images, and a second training set trainB containing no clutter and no amplitude attenuation of the simulated echo images.

[0031] The noise reduction and enhancement module uses the CycleGAN network to suppress noise and enhance features in the echo images, resulting in a hyperbolic feature-enhanced echo image dataset.

[0032] The CRCN network improvement module, based on the ResNet-18 architecture, embeds the convolutional block attention module CBAM into the residual block to obtain the deep learning model ResCBAMNet, while replacing the activation function with the Leaky ReLU nonlinear unit;

[0033] CRCN Network Training and Classification Module: The generated dataset is divided into training and test sets in a 4:6 ratio, and the training set is further divided into training data and validation data in a 6:4 ratio; the deep learning model ResCBAMNet is trained using the Adam optimizer to obtain the trained CRCN network model;

[0034] The classification and detection module inputs real-time data into the ResCBAMNet network and outputs the classification results of the medium corrosion status.

[0035] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0036] Finally, the present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, performs the method steps proposed in this invention.

[0037] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0038] 1. The CycleGAN network was introduced to suppress noise and enhance features in the dataset, which facilitates subsequent classification and recognition.

[0039] 2. The ResCBAMNet network model is proposed. Based on ResNet-18, the CBAM attention mechanism is introduced and the activation function in the network is updated to improve the ability to focus on image features and extract spatial information. At the same time, the network degradation problem of gradient explosion is avoided, making the model more generalizable and enabling fast detection of media corrosion, which is beneficial to engineering implementation.

[0040] In summary, the method proposed in this invention has a stronger advantage in curvature feature extraction and better recognition performance in non-destructive testing. Attached Figure Description

[0041] Figure 1 This is the training set for CycleGAN.

[0042] Figure 2 The loss function curve for training CycleGAN.

[0043] Figure 3 This describes the principle of the B-Scan image enhancement algorithm based on CycleGAN.

[0044] Figure 4 It uses the ResNet-18 network architecture.

[0045] Figure 5 This refers to the residual block after the introduction of CBAM.

[0046] Figure 6 This is a Convolutional Block Attention Module (CBAM).

[0047] Figure 7 This is a channel attention mechanism.

[0048] Figure 8 This is a spatial attention mechanism.

[0049] Figure 9 A comparison chart of the ReLU activation function and the Leaky ReLU activation function.

[0050] Figure 10 The data are for each category after noise suppression.

[0051] Figure 11 The loss function and accuracy curves for training ResCBAMNet.

[0052] Figure 12 For confusion matrix comparison.

[0053] Figure 13 To evaluate the indicator curve.

[0054] Figure 14 This is a flowchart of the method of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0056] This invention proposes a non-destructive testing method for the dielectric coefficient of layered media based on CRCN. First, it introduces a noise-suppressed image generation algorithm based on CycleGAN, including CycleGAN dataset creation and network training. Second, it proposes a method based on ResNet-18, introducing a CBAM module into the residual block and updating the traditional activation function to LeakyReLU. Through CycleGAN-ResCBAMNet, non-destructive testing of the medium can be achieved, thereby determining whether it has suffered geological damage. This invention mainly proposes a noise-suppressed image generation algorithm based on CycleGAN and an improved ResCBAMNet non-destructive testing network. (References) Figure 14 As shown, the main steps are as follows:

[0057] Step 1: Generating the training set for the CycleGAN network, specifically as follows:

[0058] The dataset was generated using the Matlab simulation platform and employs a SIMO ground-penetrating radar single-transmitter multiple-receiver system. The radar transmits a Gaussian pulse signal, and multiple array elements simultaneously receive the signal, i.e., a transmit-receive bistatic mode. The noisy single-cycle Gaussian pulse transmitted signal g(t) is as follows:

[0059]

[0060] Where σ is the standard deviation of the Gaussian pulse, f0 is the radar operating center frequency, and N(t) represents clutter and Gaussian noise. The clutter in each layer is generated by superimposing clutter from different point targets with Gaussian noise. If the target is in the nth layer of the medium, in SIMO mode, the echo signal S(t) after carrier frequency removal from each array element... n M) is

[0061]

[0062] Where t = [t1, t2, ... t L ] T L is the number of range signal sampling points, S(t) n Let A(M) be an L×M B-scan two-dimensional signal matrix, where M is the number of array elements. i ε is the amplitude of the signal received by array element i. n R is the dielectric constant of the nth layer. in τ is the propagation path distance from the target to the i-th element in the n-th layer of the medium. in Let C be the propagation delay from the target to the i-th element in the n-layer medium, and C be 3 × 10⁻⁶. 8 m / s, where λ is the signal wavelength. The simulation uses a two-layer medium as an example; the echo image exhibits a hyperbolic characteristic. The dielectric constant of the medium is assumed to range from [6, 155]. The dielectric constant of the uncorroded portion ranges from [6, 10].

[0063] The training set of the CycleGAN network consists of two parts: trainA and trainB. TrainA is a Matlab-simulated echo image containing noise and attenuation. TrainB is a simulated echo image without noise and without amplitude attenuation.

[0064] To facilitate subsequent processing and classification, the images are first processed into grayscale. Before processing, the dataset images are normalized to a size of 64×64, making it easier for the model to learn the curvature features in the images. Normalized data also helps avoid problems such as vanishing or exploding gradients. The training data for trainA and trainB are as follows: Figure 1 As shown.

[0065] Step 2, CycleGAN network training, specifically:

[0066] The network training experimental platform is a Windows 10 system equipped with a CPU processor (13th Gen Intel(R) Core(TM) i7-13700F) and a GPU (NVIDIA RTX4090) with 16GB of video memory. PyTorch version 1.8.1 is used as the deep learning framework, and the software environment is Python 3.8 with CUDA version 10.2. The best performance obtained on the test set is used to evaluate the performance of the algorithm.

[0067] Figure 2 The training loss function curve for CycleGAN is shown, where D_A and D_B are the loss functions of the discriminator, G_A and G_B are the loss functions of the generator, and cycle_A and cycle_B are terms in the cycle consistency loss. The training function gradually converges and remains stable after 20 epochs. Figure 3 Based on the CycleGAN-based B-Scan image enhancement algorithm, the generated dataset features hyperbolic enhancement. This embodiment uses a two-layer medium, where the feature curves correspond to the number of medium layers; the same applies to other medium layers. Furthermore, noise suppression can reach 6dB under simulation parameters. The CycleGAN training parameters are shown in Table 1.

[0068] Table 1 CycleGAN Network Parameters

[0069] Initial learning rate Training times 20 Learning rate decay training times 20 Optimizer Adam Initial learning rate 0.0001 Batchsize 1

[0070] Step 3: The improved Cyclic Adversarial Generative-Residual Network (CRCN) is as follows:

[0071] The non-destructive testing classification model proposed in this invention adopts an improved ResNet-18 architecture. This network implements the following innovative optimizations on the basis of the classic ResNet-18 framework: (1) Introducing a Convolutional Block Attention Module (CBAM) to enhance the model's feature representation ability and accurately focus on key features; (2) The activation function is improved to a Leaky ReLU nonlinear unit, which solves the problem of zero gradient in the case of negative input and avoids the problem of neuron death. This model gives full play to the feature extraction and classification capabilities of ResNet-18 and the focusing and enhancement effect of CBAM on key features, thereby significantly improving the performance of corrosion classification and discrimination.

[0072] Figure 4 For the ResNet-18 network architecture, the residual blocks after introducing CBAM are as follows: Figure 5 As shown. Figure 7 For channel attention mechanism, Figure 8As a spatial attention mechanism, CBAM (Convolutional Block Attention Module) is a lightweight hybrid attention mechanism, such as... Figure 6 As shown, it can achieve a balance between feature selection and region focusing with extremely low computational cost. This is achieved through the serial deployment of a Channel Attention Module (CAM, see...) Figure 7 ) and Spatial Attention Module (SAM, see Figure 8 This improves the accuracy of feature map focusing and the model's generalization ability.

[0073] Activation functions are a core component of deep learning neural networks. Their main principle is to enable neural networks to learn complex patterns and relationships by injecting nonlinear transformation capabilities into neurons. Figure 9 This image compares the ReLU and LeakyReLU activation functions. The ReLU function outputs 0 when the input is negative, and y = x when the input is greater than 0. This shows that ReLU is not differentiable over the entire range, thus solving the vanishing gradient problem. Since it only needs to check if the input is greater than 0, its computational efficiency is relatively high, and it also exhibits network sparsity, mitigating overfitting to some extent. However, ReLU suffers from the Daed ReLU Problem, where some neurons become permanently inactive, preventing updates to the corresponding connection weights and potentially leading to gradient explosion. The Leaky ReLU function, on the other hand, outputs y = x when x is greater than or equal to 0, and introduces a coefficient α in the negative region (y = α * x). Addressing the Daed ReLU Problem, Leaky ReLU applies a small slope to negative inputs. This non-zero gradient effectively mitigates the permanent inactivation of neurons caused by improper weight initialization or excessively high learning rates in deep networks.

[0074] Step 4: CRCN network training, specifically as follows:

[0075] The training dataset for this non-destructive testing corrosion classification network consists of 3410 noise-free echo images generated by CycleGAN training, whose curvature features are clearer than before noise suppression. The test set contains 3435 images. Due to the significant curvature features after data augmentation, to avoid overfitting, the dataset is randomly divided into training and test sets in a 4:6 ratio, with the training set randomly divided into training and validation data in a 6:4 ratio. Taking a two-layer medium as an example, the dataset is labeled into four categories: 1 represents no corrosion in the first layer and corrosion in the second layer (hereinafter referred to as "top no corrosion, bottom corrosion"); 2 represents no corrosion in both layers (hereinafter referred to as "top no corrosion, bottom no corrosion"); 3 represents corrosion in the first layer and no corrosion in the second layer (hereinafter referred to as "top corrosion, bottom no corrosion"); 4 represents corrosion in both layers (hereinafter referred to as "top corrosion, bottom corrosion"). If there are n layers, the dataset is labeled as 2. n kind. Figure 10 The data are for each category after noise suppression.

[0076] ResCBAMNet uses 20 epochs during training, with a batch size of 64, the Adam optimizer, a learning rate of 0.0001, and the cross-entropy function as the loss function. Figure 11 The loss function and accuracy curves for ResCBAMNet are shown, where train represents the training set and val represents the validation set. It can be seen that after noise suppression, the loss function stabilizes around the 10th epoch, while the accuracy reaches its maximum and remains stable. The validation set curves demonstrate that the model has achieved excellent generalization ability.

[0077] To verify the classification performance of the ReaCBAMNet model proposed in this invention, Figure 12The recognition performance of different network architectures was compared and analyzed using confusion matrix visualization. Figures (a), (b), (c), (d), and (e) show the classification decision distributions of VGG16, AlexNet, ResNet-18, ResNet18-CBAM (with CBAM), and the improved model ResCBAMNet proposed in this invention, respectively, under noise-free and attenuation-free conditions. Figure (f) shows the confusion matrix of ResCBAMNet on a training set with noise and attenuation. Experimental data show that the classification accuracies of each model on the benchmark set are 99.03%, 99.08%, 99.13%, 99.37%, 99.51%, and 98.79%, respectively. The classic architectures VGG16 and AlexNet have relatively poor classification performance due to limited feature representation capabilities. However, after training with CycleGAN image feature enhancement, ResCBAMNet's accuracy in determining media corrosion improved by 0.72 percentage points, and this model has a higher accuracy advantage and stronger feature extraction capability compared to other models. Image features augmented by CycleGAN achieve high recognition accuracy in classification networks, facilitating the determination of media corrosivity in engineering applications. To quantitatively analyze the experimental results and further validate the effectiveness of the proposed method, four evaluation metrics were used: accuracy, precision, recall, and F1-measure. The evaluation results for each model are as follows: Figure 13 As shown.

[0078] Experimental results show that, in terms of accuracy, the proposed ReaCBAMNet model, through synergistic optimization of residual connections and attention mechanisms, achieves a high level in all classifications and demonstrates relatively strong learning ability for various features. Specifically, the ResNet18-CBAM model improves the overall accuracy of all test samples by 0.24% compared to ResNet-18, while ReaCBAMNet further improves it by 0.14%. In terms of precision, the improved model performs excellently in classifications 3 and 4, and outperforms other models in recall across all four categories, effectively capturing samples from each category and avoiding missed detections. On the F1-Measure, the proposed model outperforms other models in classifications 3 and 4, achieving higher values ​​for each category. This indicates that the proposed method has a stronger advantage in curvature feature extraction and better recognition performance in non-destructive testing.

[0079] Example 2: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the method described in this invention.

[0080] Example 3: This example proposes a computer-readable storage medium storing computer instructions, which are used to cause the computer to perform the steps of the method described in this invention.

[0081] Example 4: This example proposes a non-destructive testing device for the dielectric constant of layered dielectrics based on CRCN, comprising:

[0082] The signal generation module is used to generate a dataset of noisy echo signals from the SIMO ground-penetrating radar, and to construct a first training set trainA containing clutter and attenuation of the simulated echo images, and a second training set trainB containing no clutter and no amplitude attenuation of the simulated echo images.

[0083] The noise reduction and enhancement module uses the CycleGAN network to suppress noise and enhance features in the echo images, resulting in a hyperbolic feature-enhanced echo image dataset.

[0084] The CRCN network improvement module, based on the ResNet-18 architecture, embeds the convolutional block attention module CBAM into the residual block to obtain the deep learning model ResCBAMNet, while replacing the activation function with the Leaky ReLU nonlinear unit;

[0085] CRCN Network Training and Classification Module: The generated dataset is divided into training and test sets in a 4:6 ratio, and the training set is further divided into training data and validation data in a 6:4 ratio; the deep learning model ResCBAMNet is trained using the Adam optimizer to obtain the trained CRCN network model;

[0086] The classification and detection module inputs real-time data into the ResCBAMNet network and outputs the classification results of the medium corrosion status.

[0087] It should be noted that the processing flows of Embodiments 2 to 4 correspond to the specific steps of the method provided in the embodiments of the present invention, and have the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0088] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A non-destructive testing method for the dielectric constant of layered dielectrics based on CRCN, characterized in that, Includes the following steps: S1. Generate CycleGAN training set: Construct the first training set trainA containing noise and attenuation of the simulated echo image, and the second training set trainB containing no noise and no amplitude attenuation of the simulated echo image. S2, CycleGAN Network Training: PyTorch is used as the deep learning framework to train CycleGAN, and the feature-enhanced echo image dataset is obtained using the training set of S1. S3. Construct an improved CRCN network: Based on the ResNet-18 architecture, embed the convolutional block attention module CBAM into the residual block to obtain the deep learning model ResCBAMNet, and replace the activation function with the Leaky ReLU nonlinear unit. S4. CRCN Network Training and Classification: Divide the dataset generated in step S2 into a training set and a test set in a 4:6 ratio. The training set is then divided into training data and validation data in a 6:4 ratio. Use the Adam optimizer to train the deep learning model from step S3 to obtain the detection model. S5. Input the feature-enhanced echo image obtained in S2 into the detection model obtained in step S4, and realize rapid corrosion classification and detection of the medium based on the curvature change of the echo image. In step S1, a dataset is generated using the Matlab simulation platform. The SIMO ground-penetrating radar employs a transmit-receive dual-base mode, where the radar transmits Gaussian pulse signals, and multiple array elements simultaneously receive the signals. Specifically, the radar transmits a single-cycle Gaussian pulse signal with added clutter. as follows: ， in, It is the standard deviation of the Gaussian pulse. The radar's operating center frequency, The clutter and Gaussian noise are generated by superimposing different point target clutter and Gaussian noise in each layer; In step S1, if the target is in the nth layer of medium, in SIMO mode, the echo signal after carrier frequency removal of each array element... for ， ； in, L is the number of range signal sampling points. Let M be an L×M B-scan two-dimensional signal matrix, where M is the number of array elements. Let i be the amplitude of the signal received by array element i. Let n be the dielectric constant of the nth layer. Let the distance be the propagation path from the target to the i-th array element in the n-th layer of medium. Let C be the propagation delay from the target to the i-th array element in the n-layer medium. , The wavelength is the signal wavelength.

2. The method according to claim 1, characterized in that, When the layered medium has n layers, the training dataset in step S4 is labeled in a total of 100 layers. Class, i.e., output Classification of corrosion states of different media.

3. The method according to claim 1 or 2, characterized in that, When the layered medium is bilayered, the training dataset in step S4 is labeled into four categories, resulting in four categories of medium corrosion state classifications. The classification is based on the hyperbolic curvature feature changes of the echo image, where: Category 1: First layer without corrosion, second layer with corrosion; Category 2: No corrosion in either layer; Category 3: First layer corroded, second layer uncorroded; Category 4: Both layers are corroded.

4. The method according to claim 1, characterized in that, ResCBAMNet uses 20 epochs during model training, with a batch size of 64, the Adam optimizer, a learning rate of 0.0001, and the cross-entropy function as the loss function.

5. The method according to claim 1, characterized in that, In step S3, the CBAM module achieves feature focusing by serially connecting the channel attention module CAM and the spatial attention module SAM.

6. A non-destructive testing device for the dielectric constant of a layered dielectric based on the method of any one of claims 1-5, characterized in that, include: The signal generation module is used to generate a dataset of noisy echo signals from the SIMO ground-penetrating radar, and to construct a first training set trainA containing clutter and attenuation of the simulated echo images, and a second training set trainB containing no clutter and no amplitude attenuation of the simulated echo images. The noise reduction and enhancement module uses the CycleGAN network to suppress noise and enhance features in the echo images, resulting in a hyperbolic feature-enhanced echo image dataset. The CRCN network improvement module, based on the ResNet-18 architecture, embeds the convolutional block attention module CBAM into the residual block to obtain the deep learning model ResCBAMNet, while replacing the activation function with the Leaky ReLU nonlinear unit; CRCN network training and classification module: The generated dataset is divided into training set and test set in a 4:6 ratio, and the training set is further divided into training data and validation data in a 6:4 ratio; The deep learning model ResCBAMNet is trained using the Adam optimizer to obtain the trained CRCN network model. The classification and detection module inputs real-time data into the ResCBAMNet network and outputs the classification results of the medium corrosion status.

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

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by a processor to perform the steps of the method according to any one of claims 1 to 5.

Images