A terahertz thermal barrier coating defect image reconstruction method based on transfer learning

Abstract

The application provides a terahertz thermal barrier coating defect image reconstruction method based on transfer learning. A simulation image is obtained through model driving to train a convolutional neural network and obtain a pre-training network. A terahertz detection signal is obtained by detecting a thermal barrier coating debonding defect through a terahertz time-domain spectroscopy system, a terahertz image is obtained through feature imaging, and a supplementary data set is constructed. The supplementary data set is used for retraining the pre-training network through transfer learning, the original network is fine-tuned, a channel attention mechanism module is introduced to further optimize the edge effect of image reconstruction, to solve the problem of blurred edges of the terahertz image, and to improve the generalization and efficiency of the network. Through transfer learning of the pre-training network, the network can reconstruct a real terahertz image with clear edges without a large amount of actual sample data, and provides an efficient and excellent reconstruction method for the terahertz image.

Description

Technical Field

[0001] This invention belongs to the field of terahertz nondestructive testing and image processing, specifically relating to a method for reconstructing images of defects in terahertz thermal barrier coatings based on transfer learning. Background Technology

[0002] In the aerospace field, gas turbine engines are considered the most critical components of spacecraft. However, alloys are insufficient to withstand the high temperatures generated by fuel combustion in the engine compartment. Therefore, thermal barrier coatings (TBCs) are applied to the alloy substrate to improve the high-temperature resistance of the blades and extend their service life. TBCs consist of a ceramic top coat (TC) and a metallic bond coat (BC). However, during manufacturing and service, internal and external factors can cause internal damage and failure of the material, reducing its high-temperature resistance and service life, and creating potential safety hazards. Therefore, researchers hope to use non-destructive testing (NDT) techniques to detect and quantify the types and extent of internal damage to materials, and to analyze and study the results.

[0003] Common non-destructive testing (NDT) techniques for thermal barrier coatings include eddy current testing, thermal infrared (TIR) ​​testing, ultrasonic testing, and penetrant testing. However, these methods have limitations for different types of thermal barrier coating materials. Compared to other NDT methods, terahertz (THz) NDT technology shows broad application prospects in this field due to its superior safety and material penetration. However, when using a terahertz time-domain spectroscopy (THz-TDS) system for NDT, the probe spot diameter must be larger than the defect edge width due to hardware limitations, resulting in blurred images and unclear edges. Furthermore, the acquisition efficiency of terahertz images is very low due to the THz-TDS testing process.

[0004] The mainstream approach to improving the quality and imaging efficiency of terahertz images is to employ super-resolution (SR) technology. This technology uses software algorithms to restore a given low-resolution (LR) image to a high-resolution (HR) image without requiring hardware updates. This technique can effectively improve the resolution of existing images and sharpen blurred images, thereby overcoming the resolution limitations of optical imaging and detection instruments and significantly improving imaging efficiency.

[0005] The existing method (disclosed in patent CN116824088A) utilizes a model-driven approach to establish a terahertz simulation model using the finite-difference time-domain (FDTD) method. It then obtains terahertz simulation signals through model simulation, images these signals to acquire a dataset of simulated terahertz debonding defects, and imports this dataset into a newly established neural network for training. However, due to dataset limitations, the network trained by this method only achieves good reconstruction results for simulation images of the same type, and its reconstruction performance is unsatisfactory when facing different defect types, structures, and real terahertz detection images. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to overcome the deficiencies of existing technologies and provide a terahertz thermal barrier coating defect image reconstruction method based on transfer learning. This method can improve the resolution of terahertz-detected defect images while exhibiting good versatility for different types of defects, thus enhancing the reconstruction quality of terahertz defect images.

[0007] The technical solution of the present invention is as follows: A method for image reconstruction of terahertz thermal barrier coating defects based on transfer learning, comprising the following steps:

[0008] (1) A terahertz simulation model of thermal barrier coating is established using the finite-difference time-domain method. The terahertz simulation time-domain signal is obtained through simulation. The simulated terahertz image is obtained after signal processing. The simulation model dataset is constructed as a high-resolution image dataset.

[0009] (2) The image is downsampled by bicubic interpolation to obtain a low-resolution image dataset, which is then paired with the high-resolution image to form a sample pair, and divided into a training set and a test set in a 7:3 ratio.

[0010] (3) Input the low-resolution image into the pre-trained convolutional neural network to train the network. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and magnification are achieved through the sub-pixel convolution module, thereby completing image reconstruction.

[0011] (4) The thermal barrier coating was detected by the terahertz time-domain spectroscopy system to obtain the real terahertz signal. After feature imaging, various feature images were obtained to supplement the real terahertz dataset. The dataset was divided into training set and test set in a ratio of 7:3.

[0012] (5) The neural network structure is adjusted by using the supplementary real terahertz dataset and an attention mechanism is introduced to make the network focus on the reconstruction of the edges and defective parts of the image, so as to achieve the network's recognition of real terahertz images.

[0013] (6) The neural network structure adjusted in step (5) is applied to the field of real terahertz thermal barrier coating debonding defect images through transfer learning to enhance and reconstruct the thermal barrier coating images.

[0014] Furthermore, the pre-trained convolutional neural network in step (3) consists of three residual module structures. Each residual module consists of three convolutional layers and three PReLU activation function layers. The convolutional layers and activation function layers are arranged alternately in the residual module. The residual module is divided into two paths: identity mapping and residual mapping, which are connected in the form of skip connections. The output features of the upper layer are used as the input features of the lower layer. At the end of the residual structure, the identity mapping and residual mapping are summed and output, so that the network can pass shallow information to deep layers.

[0015] Specifically, the expression for the PReLU activation function is:

[0016]

[0017] Where x i It is the input of the activation function to the i-th channel, a i Control the slope of the negative half-axis.

[0018] Specifically, the expression for the output of the residual module is:

[0019] Out Res-block =F(X)+X,

[0020] Where X is the identity mapping of the residual module, and F(X) is the residual mapping of the image after feature extraction and activation by the activation function of the convolutional layer.

[0021] Furthermore, in step (5), the neural network structure is adjusted by adding a channel attention module to the final feature extraction layer of the network. The channel attention module consists of a pooling layer, a fully connected layer, and ReLU and Sigmoid activation function layers. The input features are obtained by global maximum pooling and average pooling to obtain global statistical information for each channel. The results obtained by the two different methods are summed, and the features are mapped through the fully connected layer. The output mapping is activated by the ReLU and Sigmoid functions. Different feature weights are assigned to each channel of the upper layer input features, and the feature image is multiplied with the weight of the corresponding channel to obtain the importance of each channel to the final feature extraction result. Based on the feature image feature extraction, the channel weights are adaptively assigned to ensure that the feature image with the larger feature weight affects its result, thereby guiding the reconstruction result to highlight the features and ultimately improving the network reconstruction capability.

[0022] Specifically, the expression for the Sigmoid activation function is:

[0023]

[0024] This invention also provides a terahertz thermal barrier coating defect image reconstruction device based on transfer learning, comprising the following steps:

[0025] The simulation model dataset module is constructed by using the finite-difference time-domain method to establish a terahertz simulation model of the thermal barrier coating, obtaining the terahertz simulation time-domain signal through simulation, processing the signal to obtain the simulated terahertz image, and constructing the simulation model dataset as a high-resolution image dataset.

[0026] Preprocessing high-resolution image dataset module: The image is downsampled using bicubic interpolation to obtain a low-resolution image dataset, which is then paired with the high-resolution image to form sample pairs, and divided into training and test sets in a 7:3 ratio;

[0027] Constructing a pre-trained network module for image reconstruction: Inputting low-resolution images into a pre-trained convolutional neural network for network training. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and magnification are achieved through a sub-pixel convolution module, thereby completing image reconstruction.

[0028] The module for constructing a real terahertz dataset is as follows: a terahertz time-domain spectroscopy system is used to detect thermal barrier coatings to obtain real terahertz signals. After feature imaging, various feature images are obtained to supplement the real terahertz dataset, which is divided into training and test sets in a 7:3 ratio.

[0029] Image reconstruction module: The neural network structure is adjusted using a supplementary real terahertz dataset, and an attention mechanism is introduced to make the network focus on the reconstruction of the edges and defective parts of the image, so as to achieve the network's recognition of real terahertz images.

[0030] Image enhancement module: The neural network structure adjusted in step (5) is applied to the field of real terahertz thermal barrier coating debonding defect images through transfer learning to enhance and reconstruct the thermal barrier coating images.

[0031] The present invention also provides an electronic device, comprising:

[0032] One or more processors;

[0033] Memory, used to store one or more programs;

[0034] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for reconstructing terahertz thermal barrier coating defect images based on transfer learning.

[0035] A computer-readable storage medium is provided that stores computer instructions which, when executed by a processor, implement the steps of the method for reconstructing images of terahertz thermal barrier coating defects based on transfer learning.

[0036] Compared with existing detection methods, the present invention has the following advantages:

[0037] (1) Based on the propagation mechanism of terahertz in thermal barrier coating, the finite-difference time-domain method is used to numerically simulate the terahertz signal, establish a terahertz simulation model of the debonding defect of thermal barrier coating, obtain the terahertz simulation signal, and use it to train the convolutional neural network. This avoids the problem that the terahertz thermal barrier coating detection process is complicated, time-consuming, and costly, making it difficult to obtain the real terahertz detection signal and reducing the training cost of the neural network.

[0038] (2) For the pre-trained network that uses simulated images to build the dataset, it can only achieve good reconstruction results for the image types and forms given in the dataset. However, for the problem that the reconstruction results of different defect types, structures and real terahertz images are not ideal, the network model is fine-tuned by transfer learning, and the network is retrained by introducing real terahertz image datasets. This achieves the goal of saving training time and learning real image features, thereby improving the versatility of the network and the reconstruction effect of the images.

[0039] Therefore, this invention provides a high-quality technical solution for terahertz image reconstruction of thermal barrier coating debonding defects. It utilizes numerical simulation to establish a model of thermal barrier coating debonding defects, obtains terahertz simulation images to prepare a dataset, and trains the network to learn general features. A small dataset of real terahertz images is prepared, and the pre-trained network is fine-tuned through transfer learning. An attention mechanism is introduced to improve the network's reconstruction effect on defect features. This avoids the time-consuming and costly problems associated with acquiring a large number of detection images using terahertz time-domain spectroscopy systems, and effectively improves the network's generalization ability, significantly reducing the economic and time costs of acquiring terahertz images. Attached Figure Description

[0040] Figure 1 This is the simulation model of the terahertz thermal barrier coating constructed in this invention;

[0041] Figure 2 The simulated terahertz image obtained in this invention;

[0042] Figure 3 Here is a flowchart of the neural transfer learning neural network constructed in this invention;

[0043] Figure 4 This refers to the attention mechanism model introduced in this invention;

[0044] Figure 5 This is a structural diagram of the electronic device of the present invention. Detailed Implementation

[0045] Based on the above-described technical solutions, the present invention will be described in detail below through specific examples and accompanying drawings. The examples listed herein are only used to verify and illustrate the feasibility and effectiveness of the technical solutions of the present invention, and are not intended to limit the present invention.

[0046] The solution to the technical problem of this invention is: a method for reconstructing images of terahertz thermal barrier coatings based on transfer learning, comprising the following steps:

[0047] (1) Constructing a simulation model dataset: Driven by a simulation model, a terahertz simulation model of the thermal barrier coating is established using the finite-difference time-domain method. Terahertz simulation time-domain signals are obtained through simulation, and simulated terahertz images are obtained after signal processing. A dataset is constructed as a high-resolution image dataset. The constructed terahertz thermal barrier coating simulation model is as follows: Figure 1 As shown; simulated terahertz image as Figure 2 As shown;

[0048] (2) Preprocessing high-resolution dataset: The images are downsampled by bicubic interpolation to obtain low-resolution images, which are then paired with high-resolution images to form sample pairs. The samples are divided into training and test sets in a 7:3 ratio.

[0049] (3) Constructing an image reconstruction pre-training network: The low-resolution image is input into the constructed convolutional neural network for network training. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and amplification are achieved through a sub-pixel convolution module, thereby completing image reconstruction. The neural network flowchart is as follows: Figure 3 As shown;

[0050] (4) Construct a supplementary real terahertz dataset: The thermal barrier coating is detected by the terahertz time-domain spectroscopy system to obtain real terahertz signals. After feature imaging, various feature images are obtained to establish a supplementary real terahertz dataset, which is divided into training set and test set in a ratio of 7:3.

[0051] (5) Image Reconstruction: The network structure is adjusted using a supplementary real terahertz dataset, and an attention mechanism is introduced to focus the network on the reconstruction of image edges and defective parts, thereby enabling the network to recognize real terahertz images. The attention mechanism model is as follows: Figure 4 As shown;

[0052] (6) Image enhancement: The neural network model adjusted in step (5) is applied to the field of real terahertz thermal coating debonding defect images through transfer learning to enhance and reconstruct the images obtained by the terahertz time-domain spectroscopy system.

[0053] The pre-trained convolutional neural network for the image in step (3) consists of three residual block structures. Each residual block consists of three convolutional layers and three PReLU activation function layers. The expression for the PReLU activation function is:

[0054]

[0055] Where x i It is the input of the activation function to the i-th channel, a i The slope of the negative half-axis is controlled; convolutional layers and activation function layers are alternately arranged in the residual block, which is divided into two paths: identity mapping and residual mapping, connected in a skip connection manner. The output features of the upper layer are used as the input features of the lower layer, and the identity mapping and residual mapping are summed at the end of the residual structure, so that the network can pass information from shallow layers to deep layers. The residual module outputs Out. Res-block The expression is:

[0056] Out Res-block =F(X)+X,

[0057] Where X is the identity mapping of the residual module, and F(X) is the residual mapping of the image after feature extraction and activation by the activation function after convolutional layer; the residual structure can alleviate the gradient vanishing problem caused by network deepening, improve the network's expressive power, and enhance the model's robustness.

[0058] In step (5), the network structure is adjusted by adding a channel attention module to the final feature extraction layer. This channel attention module consists of a pooling layer, a fully connected layer, and ReLU and Sigmoid activation function layers. The input features are processed using both global max pooling and average pooling to obtain global statistical information for each channel. The results from these two methods are summed, and the features are mapped through a fully connected layer. The ReLU and Sigmoid functions are then used to activate the output mapping. The Sigmoid activation function expression is:

[0059]

[0060] By assigning different feature weights to each channel of the upper-layer input features, the feature image is multiplied with the weight of the corresponding channel to obtain the importance of each channel to the final feature extraction result. Based on the feature extraction of the feature image, the channel weights are adaptively assigned to ensure that the feature image with the larger feature weight affects the result, thereby guiding the reconstruction result to highlight the features and ultimately improving the network reconstruction capability.

[0061] The present invention also provides a terahertz thermal barrier coating defect image reconstruction device based on transfer learning, comprising the following steps:

[0062] The simulation model dataset module is constructed by using the finite-difference time-domain method to establish a terahertz simulation model of the thermal barrier coating, obtaining the terahertz simulation time-domain signal through simulation, processing the signal to obtain the simulated terahertz image, and constructing the simulation model dataset as a high-resolution image dataset.

[0063] Preprocessing high-resolution image dataset module: The image is downsampled using bicubic interpolation to obtain a low-resolution image dataset, which is then paired with the high-resolution image to form sample pairs, and divided into training and test sets in a 7:3 ratio;

[0064] Constructing a pre-trained network module for image reconstruction: Inputting low-resolution images into a pre-trained convolutional neural network for network training. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and magnification are achieved through a sub-pixel convolution module, thereby completing image reconstruction.

[0065] The module for constructing a real terahertz dataset is as follows: a terahertz time-domain spectroscopy system is used to detect thermal barrier coatings to obtain real terahertz signals. After feature imaging, various feature images are obtained to supplement the real terahertz dataset, which is divided into training and test sets in a 7:3 ratio.

[0066] Image reconstruction module: The neural network structure is adjusted using a supplementary real terahertz dataset, and an attention mechanism is introduced to make the network focus on the reconstruction of the edges and defective parts of the image, so as to achieve the network's recognition of real terahertz images.

[0067] Image enhancement module: The neural network structure adjusted in step (5) is applied to the field of real terahertz thermal barrier coating debonding defect images through transfer learning to enhance and reconstruct the thermal barrier coating images.

[0068] like Figure 5 As shown, the present invention also provides an electronic device, comprising:

[0069] One or more processors;

[0070] Memory, used to store one or more programs;

[0071] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for reconstructing terahertz thermal barrier coating defect images based on transfer learning.

[0072] The present invention also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the described method for reconstructing terahertz thermal barrier coating defect images based on transfer learning.

[0073] The above is a detailed description of the method of the present invention in conjunction with the accompanying drawings. The specific embodiments described herein are merely illustrative examples and are not intended to limit the applicability of the present invention. For those skilled in the art, any modifications, equivalent substitutions, and improvements made based on the concept of the present invention within the scope of the specific embodiments and applications should be included within the protection scope of the present invention.

Claims

1. A method for image reconstruction of terahertz thermal barrier coating defects based on transfer learning, characterized in that, Includes the following steps: (1) A terahertz simulation model of thermal barrier coating is established using the finite-difference time-domain method. The terahertz simulation time-domain signal is obtained through simulation. The simulated terahertz image is obtained after signal processing. The simulation model dataset is constructed as a high-resolution image dataset. (2) The image is downsampled by bicubic interpolation to obtain a low-resolution image dataset, which is then paired with the high-resolution image to form a sample pair, and divided into a training set and a test set in a 7:3 ratio. (3) Input the low-resolution image into the pre-trained convolutional neural network to train the network. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and magnification are achieved through the sub-pixel convolution module, thereby completing image reconstruction. (4) The thermal barrier coating was detected by the terahertz time-domain spectroscopy system to obtain the real terahertz signal. After feature imaging, various feature images were obtained to supplement the real terahertz dataset. The dataset was divided into training set and test set in a ratio of 7:

3. (5) The neural network structure is adjusted by using the supplementary real terahertz dataset and an attention mechanism is introduced to make the network focus on the reconstruction of the edges and defective parts of the image, so as to achieve the network's recognition of real terahertz images. (6) The neural network structure adjusted in step (5) is applied to the field of real terahertz thermal barrier coating debonding defect images through transfer learning to enhance and reconstruct the thermal barrier coating images.

2. The method for image reconstruction of terahertz thermal barrier coating defects based on transfer learning according to claim 1, characterized in that, The pre-trained convolutional neural network in step (3) consists of three residual module structures. Each residual module consists of three convolutional layers and three PReLU activation function layers. The convolutional layers and activation function layers are arranged alternately in the residual module. The residual module is divided into two paths: identity mapping and residual mapping, which are connected in the form of skip connections. The output features of the upper layer are used as the input features of the lower layer. At the end of the residual structure, the identity mapping and residual mapping are summed and output, so that the network can pass shallow information to deep layers.

3. The method for reconstructing terahertz thermal barrier coating defect images based on transfer learning according to claim 2, characterized in that, The expression for the PReLU activation function is: Where x i It is the input of the activation function to the i-th channel, a i Control the slope of the negative half-axis.

4. The method for reconstructing images of terahertz thermal barrier coating defects based on transfer learning according to claim 2, characterized in that, The expression for the output of the residual module is: Out Res-block =F(X)+X, Where X is the identity mapping of the residual module, and F(X) is the residual mapping of the image after feature extraction and activation by the activation function of the convolutional layer.

5. The method for image reconstruction of terahertz thermal barrier coating defects based on transfer learning according to claim 1, characterized in that, In step (5), the neural network structure is adjusted by adding a channel attention module to the final feature extraction layer of the network. The channel attention module consists of a pooling layer, a fully connected layer, and ReLU and Sigmoid activation function layers. The input features are obtained by global maximum pooling and average pooling to obtain global statistical information for each channel. The results obtained by the two different methods are summed and mapped through the fully connected layer. The output mapping is activated by the ReLU and Sigmoid functions. Different feature weights are assigned to each channel of the upper layer input features. The feature image is multiplied with the weight of the corresponding channel to obtain the importance of each channel to the final feature extraction result. Based on the feature image feature extraction, the channel weights are adaptively assigned to ensure that the feature image with the larger feature weight affects its result, thereby guiding the reconstruction result to highlight the features and ultimately improving the network reconstruction capability.

6. The method for image reconstruction of terahertz thermal barrier coating defects based on transfer learning according to claim 5, characterized in that, The expression for the Sigmoid activation function is:

7. A device for reconstructing images of terahertz thermal barrier coating defects based on transfer learning, characterized in that, Includes the following steps: The simulation model dataset module is constructed by using the finite-difference time-domain method to establish a terahertz simulation model of the thermal barrier coating, obtaining the terahertz simulation time-domain signal through simulation, processing the signal to obtain the simulated terahertz image, and constructing the simulation model dataset as a high-resolution image dataset. Preprocessing high-resolution image dataset module: The image is downsampled using bicubic interpolation to obtain a low-resolution image dataset, which is then paired with the high-resolution image to form sample pairs, and divided into training and test sets in a 7:3 ratio; Constructing a pre-trained network module for image reconstruction: Inputting low-resolution images into a pre-trained convolutional neural network for network training. The network consists of a shallow feature extraction layer and a deep feature extraction layer constructed based on the residual module. Upsampling and magnification are achieved through a sub-pixel convolution module, thereby completing image reconstruction. The module for constructing a real terahertz dataset is as follows: a terahertz time-domain spectroscopy system is used to detect thermal barrier coatings to obtain real terahertz signals. After feature imaging, various feature images are obtained to supplement the real terahertz dataset, which is divided into training and test sets in a 7:3 ratio. Image reconstruction module: The neural network structure is adjusted using a supplementary real terahertz dataset, and an attention mechanism is introduced to make the network focus on the reconstruction of the edges and defective parts of the image, so as to achieve the network's recognition of real terahertz images. Image enhancement module: The neural network structure adjusted in step (5) is applied to the field of real terahertz thermal barrier coating debonding defect images through transfer learning to enhance and reconstruct the thermal barrier coating images.

8. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-6.

Images