Method for dynamically measuring composite material under ultrahigh-temperature working condition

Abstract

The invention relates to a method for dynamically measuring a composite material under an ultra-high-temperature working condition, which comprises the following steps of: constructing a composite material high-temperature strain dynamic detection platform, and respectively acquiring speckle images of a tested piece under a cold state condition and an ultra-high-temperature working condition as data sets; constructing a supervised network model for eliminating speckle image thermal radiation and thermal disturbance influence, wherein the supervised network model comprises a multi-scale feature extraction module, a dual convolutional neural network and a polarization self-attention mechanism; constructing a joint loss function L for the supervised image restoration network model; training the model to obtain an optimal image restoration network model; an ultra-high-temperature speckle image serves as input, a speckle image restoration result is output through the image restoration network model, and a displacement field is obtained through DIC calculation; according to the invention, the influence of thermal radiation, thermal disturbance and haze phenomena in an ultra-high temperature environment on image quality is effectively eliminated, and the dynamic strain measurement precision of the composite material under an ultra-high temperature working condition is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of strain detection, and particularly to a method for dynamically measuring composite materials under ultra-high temperature conditions. Background Art

[0002] With the rapid development of fields such as aerospace, metallurgy, petrochemical industry, etc., composite materials in these fields need to work stably at high temperatures. Therefore, it is crucial to measure the mechanical properties of composite materials at high temperatures. Digital Image Correlation (DIC) is a full-field optical measurement method used to measure the displacement and deformation of an object. With its non-contact measurement advantage, it has been widely used in the field of high-temperature measurement.

[0003] However, there are still many problems in the practical application of DIC technology under ultra-high temperature conditions. First of all, in a high-temperature non-vacuum environment, due to the influence of heat sources such as a constant temperature furnace, the gas density distribution is uneven, which will form a temperature difference between the sample to be measured and the DIC lens, causing a change in the air refractive index, resulting in the drift and jitter of speckle feature points, and reducing the measurement accuracy.

[0004] Secondly, the sublimation of the carbon fiber matrix of the composite material in an ultra-high temperature environment produces a phenomenon similar to haze, and the vaporized particulate matter suspends around the object to be measured, affecting the image quality; at the same time, it causes a light scattering effect, the image is blurred, the sample features are not prominent enough, and the overall contrast decreases. Such imaging conditions will lead to the loss of image details and cause matching failures in the subsequent image matching calculation process. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method for dynamically measuring composites under ultra-high temperature conditions, which solves the problem that the prior art cannot achieve high-precision dynamic measurement of composites under ultra-high temperature conditions. This method extracts ultra-high temperature speckle image information through sparse representation with the help of a multi-scale feature extraction module in a deep neural network supervision model. At the same time, by virtue of the advantage of the multi-scale mechanism to broaden the feature coverage range, it effectively alleviates the low restoration accuracy caused by insufficient shallow feature extraction. Secondly, by designing a double convolutional neural network structure, the upper feature extraction path extracts global context features from the feature tensor, and the lower feature extraction path extracts local detail features from the feature tensor. Before each layer of convolutional neural network in the upper and lower feature extraction paths, a polarization self-attention mechanism is innovatively introduced to extract deep features. The core lies in constructing an orthogonal dimension feature screening system: while compressing features in a specific direction, maintaining high-resolution representation in the vertical dimension, enhancing the restoration effect of the model on ultra-high temperature speckle images. Therefore, through the deep neural network model ThermoClearNet, the influence of thermal radiation, thermal disturbance, and haze phenomena on the image quality in the ultra-high temperature environment is effectively eliminated, and the dynamic strain measurement accuracy of composites under ultra-high temperature conditions is significantly improved.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A method for dynamically measuring composites under ultra-high temperature conditions, including the following steps:

[0007] S1. Build a high-temperature strain dynamic detection platform for composites, and collect the speckle images of the test piece under cold conditions and ultra-high temperature conditions respectively as a data set;

[0008] S2. Construct a supervised network ThermoClearNet model for eliminating the influence of thermal radiation and thermal disturbance on speckle images, including a multi-scale feature extraction module MFEM, a double convolutional neural network, and a polarization self-attention mechanism;

[0009] S3. Input the ultra-high temperature speckle image into the multi-scale feature extraction module MFEM to obtain the multi-scale fusion feature Output of the ultra-high temperature speckle image MFEM-B2 ;

[0010] S4. The double convolutional neural network includes an upper feature extraction path and a lower feature extraction path. Use Output MFEM-B2 as the input of the upper and lower paths to extract global context features and local detail features respectively;

[0011] S5. Before each layer of convolutional neural network in the upper and lower feature extraction paths, introduce a polarization self-attention mechanism PSA, and input the input feature map X of the current layer of the upper and lower paths into the polarization self-attention mechanism respectively to extract deep features;

[0012] S6. Concatenate the final output global context feature map containing deep features and the local detail feature map along the channel dimension to obtain a complete information feature that preserves the global and local noise distributions;

[0013] S7. Construct a joint loss function L for the supervised image restoration network ThermoClearNet model;

[0014] S8. Train the ThermoClearNet model to obtain an optimal image restoration network model;

[0015] S9. Collect the ultra-high temperature speckle images of the composite material specimen under ultra-high temperature conditions as input, output the restored result of the ultra-high temperature speckle image through the image restoration network model, and perform DIC calculation to obtain the dynamic measurement of the force deformation of the composite material under ultra-high temperature conditions.

[0016] Furthermore, the specific process of step S1 includes the following steps:

[0017] S11. Set up a high-temperature DIC experimental system, including a thermostatic chamber, an industrial camera, a lens, a blue light source, and a tensile testing machine;

[0018] S12. Fix the test piece on the tensile rod of the tensile testing machine, open the observation window, eliminate the influence of the refraction of the observation window glass on the speckle image, and adjust the camera after clamping in place to collect the cold-state speckle image of the test piece to be measured;

[0019] S13. Start the thermostatic chamber to ensure a stable heat flow field is formed in the experimental environment. Under this stable state, use the camera to collect the ultra-high temperature speckle image of the test piece to be measured;

[0020] S14. Divide the collected ultra-high temperature speckle images into a training set, a validation set, and a test set according to 6:2:2.

[0021] Furthermore, the multi-scale feature extraction module MFEM in step S3 includes two sub-networks, MFEM-B1 and MFEM-B2. Using dilated convolution and various sizes of convolutional kernels, it realizes the extraction and fusion of multi-scale features of the ultra-high temperature speckle image, Output MFEM-B2 , and its specific process includes the following steps:

[0022] S31. MFEM-B1 constructs a three-level cascaded feature extraction architecture using the concept of sparse representation. The core of this system consists of three convolutional units with increasing parameters. Each processing unit contains three modules: a dilated convolution operation layer, a normalization processing module, and a non-linear activation layer. Among them, the coefficient combination (1 / 2 / 3) of the dilated convolution operation layer forms a progressive dilation mechanism. Through the stacking strategy of progressive dilated convolutions, the coverage range of the perception field of the ThermoClearNet model is significantly expanded;

[0023] Among them, the MFEM-B1 expression is as follows:

[0024] Output MFEM-B1 = D3(D2(D1(N)))

[0025] Among them, N represents the ultra-high temperature speckle image obtained by shooting, and D3, D2, and D1 respectively represent convolutional units with coefficients of 3, 2, and 1. Output MFEM-B1 represents the sparse feature output by the MFEM-B1 module;

[0026] S32 and MFEM-B2 receive the sparse feature Output processed by MFEM-B1 MFEM-B1 as the input. This module constructs a three-channel heterogeneous feature extraction network. Each feature processing channel uses a 128-dimensional feature vector space as the operation basis. First, standard convolution preprocessing is performed on each channel, and the basic convolution operation is combined with the ReLU activation function with gradient stability characteristics to ensure that the feature representations of each channel are consistent in the initial processing stage;

[0027] S33. Through the configuration of heterogeneous convolution kernels (1×1, 3×3, 5×5), multi-scale feature convolution operations are implemented to obtain feature information of ultra-high temperature speckle images at different scales. Then, PReLU operations are performed on the features of the branches. Through multi-dimensional feature tensor splicing and fusion, the complementary features extracted from the four heterogeneous channels are deeply integrated. The specific expression of MFEM-B2 is as follows:

[0028] Output MFEM-B2 = Fusion(R1, R3, R5)

[0029] R1 = F1(Output MFEM-B1 )

[0030] R3 = F3(Output MFEM-B1 )

[0031] R5 = F5(Output MFEM-B1 )

[0032] Among them, F1, F3, and F5 respectively represent convolution operations with different convolution kernel sizes. Fusion represents the feature splicing function after the PReLU operation. Output MFEM-B2 represents the extraction and fusion features of the multi-scale features of the ultra-high temperature speckle image output by MFEM-B2.

[0033] Furthermore, in step S4, the specific process of step S4 includes the following steps:

[0034] S41. The upper - side feature extraction path constructs a global context information capture system through the alternating stacking of conventional convolution and dilated convolution, in conjunction with a cross - layer connection mechanism: receiving Output MFEM-B2 , extracting basic textures through 3×3 conventional convolution, then concatenating the first - layer output and the input features along the channel dimension, applying a 3×3 dilated convolution with dilation rate r = 2 to the mixed features, expanding the receptive field to a 7×7 pixel range to capture cross - regional speckle correlations. The alternating configuration strategy of the convolutional kernels ensures the ability to capture large - scale speckle features while effectively controlling the consumption of computing resources by introducing batch normalization operations after each convolutional layer. Repeating the above operations, a global context feature map is finally generated, which contains the overall deformation trend information of the speckles caused by thermal distortion;

[0035] S42. The lower - side feature extraction path adopts a progressive dilated convolution design to enhance the ability to capture local detail features by expanding the range of the perception field. It also receives Output MFEM-B2 , and sequentially adopts progressive dilated convolutions. The output of each layer is added to the previous - order features after batch normalization. For a two - dimensional input feature map I and a convolutional kernel K, the output O of the dilated convolution can be expressed as:

[0036]

[0037] where r is the set known dilation rate of the dilated convolution, m and n are the sizes of the convolutional kernel. Finally, a local detail feature map is generated, focusing on the speckle edge blurring, micro - fractures, and noise distribution caused by thermal perturbations. I represents the extraction and fusion feature Output of the multi - scale features of the ultra - high - temperature speckle image MFEM-B2 ;

[0038] S43. The global context feature map and the local detail feature map of the upper - side and lower - side feature extraction paths output by S41 and S42 are concatenated along the channel dimension to retain the complete information of the global deformation trend and local noise distribution. Subsequently, a 3×3 conventional convolution operation is applied to the mixed features to eliminate cross - channel redundant information through local receptive field modeling.

[0039] Furthermore, both the upper - side feature extraction path and the lower - side feature extraction path adopt a seven - layer convolutional structure, with 32 3×3 convolutional kernels configured in each layer, optimizing the computational efficiency while ensuring the network's processing ability.

[0040] Furthermore, the polarization self - attention mechanism in step S5 includes two main branches: the channel self - attention mechanism and the spatial self - attention mechanism, synchronously optimizing the feature distribution in the channel and spatial dimensions through a parallel two - stream integration mode. Its specific process includes the following steps:

[0041] S51. The channel self-attention mechanism focuses on the semantic associations between feature channels. The formula of the channel attention mechanism of the ThermoClearNet model is expressed as:

[0042] I ch (X) = F SG [W Z (σ1(W v (X)) × F SM (σ2(W q (X)))]

[0043] Where ch represents the Channel channel branch, W Z , W v , W q are standard 1×1 convolution operations, σ1 and σ2 are two tensor reconstruction operators, F SG is the Sigmoid activation function, F SM is the SoftMax normalization function. By dynamically adjusting the weights of each channel, the model can better capture the dependencies between channels. The channel dimension needs to be finely modeled, so three-weight enhanced semantic screening is adopted to improve the feature expression ability of the ThermoClearNet model;

[0044] S52. The spatial self-attention mechanism then enhances the understanding of the correlations between image features. The formula of spatial self-attention is expressed as:

[0045] I sp (X) = F SG [σ3(F GP (σ1(W v (X))) × σ2(W v (X))]

[0046] Where sp represents the Spatial spatial branch, σ3 is the tensor reconstruction operator, F SG is the Sigmoid activation function, F GP represents the global pooling function, directly using global average pooling GPool to replace the explicit query-key W Z , W q interaction, directly performing element-wise multiplication on the original features to avoid loss of spatial resolution and ensure the integrity of speckle edge details; at the same time, since the computational cost of the spatial dimension is much higher than that of the channel dimension, omitting W Z , W q can reduce the parameter calculation amount;

[0047] S53. Add the output I ch (X) of the channel branch and the output I sp (X) of the spatial branch element-wise to obtain the feature map X' optimized by PSA:

[0048] X' = I ch (X) + I sp (X)

[0049] This design enables each feature extraction unit of the double convolutional neural network to recalibrate the feature response in the channel and spatial dimensions through the PSA mechanism before processing, strengthen the transmission of effective features and suppress noise interference, and input the fused feature map X' into the conventional convolution or dilated convolution module of the current layer to extract deep features through the convolution kernel operation;

[0050] S54. For the dilated convolution layer of the upper path, after channel concatenation of X' with the cross-layer connected features, convolution is performed; for the progressive dilated convolution layer of the lower path, X' is added to the residual features and then input to the next layer.

[0051] Furthermore, in step S7, a joint loss function L for constructing the supervised image restoration network ThermoClearNet model is used, and the pixel-wise Charbonnier loss l char and edge-aware loss l edge and frequency reconstruction loss l fr are used to restore the ultra-high temperature speckle image to a level similar to the cold-state speckle image. The specific process includes the following steps:

[0052] S71. The Charbonnier loss function alleviates the influence of outliers by introducing a square root operation, enhancing the robustness of the model to noise. The specific formula is as follows:

[0053]

[0054] In the formula, I R represents the restored image output by the model, that is, the ideal speckle distribution after disturbance elimination, which needs to be aligned with the cold-state speckle image Y in the feature space, and ε is a non-zero constant term;

[0055] S72. The edge-aware loss function extracts the high-frequency information of the image using the Laplacian operator, and focuses on optimizing the accurate reconstruction of the speckle edge contour. The specific formula is as follows:

[0056] ]>

[0057] In the formula, Δ represents the Laplacian operator, ΔI R is the predicted edge feature after restoration, and ΔY represents the true edge feature, which is used as a supervision signal, requiring the edges generated by the model to be consistent with the physical true speckle deformation pattern;

[0058] S73. The frequency domain reconstruction loss function converts the image to the frequency domain space through Fourier transform, ensuring that the restoration result is consistent with the real image in terms of frequency domain features. The specific formula is as follows:

[0059] L fr = ||F(I R ) - F(Y)||

[0060] In the formula, F represents the fast Fourier transform;

[0061] S74. Calculate the above loss components respectively in each training stage, and continuously adjust the network parameters through gradient backpropagation. Its overall objective function can be expressed in the following weighted combination form:

[0062] L = L char (I R , Y) + τ1L edge (I R , Y) + τ2L fr (I R , Y)

[0063] τ1 and τ2 represent the weight coefficients corresponding to the losses.

[0064] Furthermore, the specific process of step S8 includes the following steps:

[0065] S81. Classify the dataset collected in step S14 into a training set, a validation set, and a test set. The training set is used for parameter learning, the validation set is used for hyperparameter tuning, and the test set is used to evaluate the generalization performance of the model. This partitioning method follows the standard practice in the field of machine learning;

[0066] S82. Execute the iterative optimization process of the model. Update the model parameters driven by the training set data, monitor the overfitting situation of the model using the validation set and adjust the hyperparameters, and finally evaluate the prediction accuracy of the model using the test set;

[0067] S83. After a complete training - validation - test process, an optimized speckle image restoration ThermoClearNet network model is obtained.

[0068] Furthermore, the specific process of step S9 includes the following steps:

[0069] S91. Input the captured ultra - high - temperature speckle image into the optimized ThermoClearNet model. The trained network layer inside the model will automatically extract the input features and calculate the corresponding multiple linear regression parameters, which define the mathematical mapping relationship between the input ultra - high - temperature speckle image and the target restored image;

[0070] S92. Finally, perform DIC digital image correlation calculation on the restored image output by the optimized ThermoClearNet model to obtain a high - precision dynamic measurement of the force deformation of the composite material under ultra - high - temperature conditions.

[0071] With the above technical solutions, the present invention provides a method for dynamically measuring composite materials under ultra-high temperature conditions, having at least the following beneficial effects:

[0072] 1. The deep neural network model ThermoClearNet effectively eliminates the influence of thermal radiation and thermal disturbance on image quality in the ultra-high temperature environment, significantly improving the dynamic strain measurement accuracy of composite materials under ultra-high temperature conditions.

[0073] 2. The present invention is specifically designed for ultra-high temperature conditions and can effectively cope with complex problems such as uneven gas density, change in air refractive index, and sublimation of carbon fiber matrix in high temperature environments, solving problems such as image quality degradation, feature point drift and jitter of traditional DIC technology in ultra-high temperature environments, and ensuring the reliability of measurement results.

[0074] 3. Through the automated processing of the deep neural network, the present invention reduces manual intervention and improves measurement efficiency. It is not only applicable to the dynamic measurement of composite materials under ultra-high temperature conditions in fields such as aerospace, metallurgy, and petrochemical engineering, but also provides an important reference basis for subsequent material process optimization and assembly design. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The drawings described herein are used to provide a further understanding of the present application and form a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0076] Figure 1 is a schematic diagram of a composite material high-temperature strain dynamic detection platform built for the present invention;

[0077] Figure 2 is a schematic diagram of the multi-scale feature extraction module (MFEM) proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments. Thereby, a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects can be obtained and implemented accordingly.

[0079] Those of ordinary skill in the art can understand that all or part of the steps in implementing the methods of the above embodiments can be completed by instructing relevant hardware through a program. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0080] Please refer to Figure 1 - Figure 2 , which shows a specific implementation manner of this embodiment. In this embodiment, the multi-scale feature extraction module in the deep neural network supervision model extracts the ultra-high temperature speckle image information through sparse representation, and at the same time, by virtue of the advantage of the multi-scale mechanism to broaden the feature coverage range, effectively alleviates the low restoration accuracy caused by insufficient shallow feature extraction; secondly, by designing a double convolutional neural network structure, the upper feature extraction path extracts global context features from the feature tensor, and the lower feature extraction path extracts local detail features from the feature tensor. Before each layer of the convolutional neural network in the upper feature extraction path and the lower feature extraction path, a polarization self-attention mechanism is innovatively introduced to extract deep features. The core lies in constructing an orthogonal dimension feature screening system: while compressing features in a specific direction, maintaining a high-resolution representation in the vertical dimension, enhancing the restoration effect of the model on the ultra-high temperature speckle image; therefore, through the deep neural network model ThermoClearNet, the influence of thermal radiation, thermal disturbance, and haze phenomena in the ultra-high temperature environment on the image quality is effectively eliminated, and the dynamic strain measurement accuracy of the composite material under ultra-high temperature working conditions is significantly improved.

[0081] Please refer to Figure 1 , this embodiment proposes a dynamic measurement method for composite materials under ultra-high temperature working conditions, and this method includes the following steps:

[0082] S1. Build a high-temperature strain dynamic detection platform for composite materials, and respectively collect the speckle images of the test piece under cold conditions and ultra-high temperature working conditions as a data set;

[0083] As a preferred implementation manner of step S1, the specific process of step S1 includes the following steps:

[0084] S11. Build a high-temperature DIC experimental system, including a thermostatic chamber, an industrial camera, a lens, a blue light source, and a tensile testing machine; as Figure 1 shown; among them, the imaging speed of the industrial camera: 30 frames per second, resolution: 1624×1236 pixels), and the focal length is 35 mm lens (EDMUND OPTICS);

[0085] S12. Fix the test piece on the tensile rod of the tensile testing machine, open the observation window, eliminate the influence of the refraction of the observation window glass on the speckle image, adjust the camera after clamping in place to collect the cold-state speckle image of the test piece to be measured, and then use it for the training of the ThermoClearNet model;

[0086] S13. Start the thermostat, heat the thermostat to 1000 °C and keep it at a constant temperature for 20 minutes for image collection to ensure that a stable heat flow field is formed in the experimental environment. In this stable state, use the camera to collect the ultra-high temperature speckle image of the test piece to be measured;

[0087] S14. Divide the collected ultra-high temperature speckle images into a training set, a validation set, and a test set according to 6:2:2.

[0088] S2. Construct a supervised network ThermoClearNet model to eliminate the influence of thermal radiation and thermal disturbance on the speckle image, including a multi-scale feature extraction module MFEM, a double convolutional neural network, and a polarization self-attention mechanism;

[0089] S3. Take the ultra-high temperature speckle image as the input of the multi-scale feature extraction module MFEM to obtain the multi-scale fusion feature Output of the ultra-high temperature speckle image MFEM-B2 ;

[0090] As a preferred implementation manner of step S3, in step S3, the multi-scale feature extraction module MFEM includes two sub-networks, MFEM-B1 and MFEM-B2. As Figure 2 shown, using dilated convolution and multiple sizes of convolutional kernels, the extraction and fusion feature Output of the multi-scale features of the ultra-high temperature speckle image are realized MFEM-B2 , and its specific process includes the following steps:

[0091] S31. MFEM-B1 constructs a three-level cascaded feature extraction architecture using the concept of sparse representation. The core of this system consists of three convolutional units with increasing parameters. Each processing unit contains three modules: a dilated convolution operation layer, a normalization processing module, and a non-linear activation layer. Among them, the coefficient combination (1 / 2 / 3) of the dilated convolution operation layer forms a progressive dilation mechanism. Through the stacking strategy of progressive dilated convolutions, the coverage range of the perception field of the ThermoClearNet model is significantly expanded;

[0092] Among them, the expression of MFEM-B1 is as follows:

[0093] Output MFEM-B1 = D3(D2(D1(N)))

[0094] Among them, N represents the ultra-high temperature speckle image obtained by shooting, D3, D2, and D1 respectively represent convolutional units with coefficients of 3, 2, and 1, and OutputMFEM-B1 Represents the sparse features output by the MFEM-B1 module; this multi-scale feature fusion mechanism not only enhances the parsing accuracy of the speckle image pixel-level features, but also improves the model's capture efficiency of key gray information through composite feature extraction.

[0095] S32. MFEM-B2 receives the sparse feature Output processed by MFEM-B1 MFEM-B1 As input, this module constructs a three-channel heterogeneous feature extraction network. Each feature processing channel uses a 128-dimensional feature vector space as the operation basis, and realizes multi-granularity feature capture through differential convolution kernel configuration; first, standard convolution preprocessing is performed on each channel, and the basic convolution operation is combined with the ReLU activation function with gradient stability characteristics to ensure that each channel maintains feature representation consistency in the initial processing stage;

[0096] S33. Through heterogeneous convolution kernel configuration (1×1, 3×3, 5×5), multi-scale feature convolution operations are performed to obtain feature information of ultra-high temperature speckle images at different scales. Then, PReLU operations are performed on the features of the branches, and through multi-dimensional feature tensor splicing and fusion, the complementary features extracted from the four heterogeneous channels are deeply integrated. The specific expression of MFEM-B2 is as follows:

[0097] Output MFEM-B2 = Fusion(R1, R3, R5)

[0098] R1 = F1(Output MFEM-B1 )

[0099] R3 = F3(Output MFEM-B1 )

[0100] R5 = F5(Output MFEM-B1 )

[0101] Where F1, F3, and F5 represent convolution operations with different convolution kernel sizes respectively, Fusion represents the feature splicing function after PReLU operation, and Output MFEM-B2 Represents the extraction and fusion features of the multi-scale features of the ultra-high temperature speckle image output by MFEM-B2.

[0102] In this embodiment, to fully expand the receptive field of the ThermoClearNet model and enhance the feature extraction ability, a multi-scale feature extraction module (MFEM) is proposed. By leveraging the advantages of sparse representation and multi-scale analysis, this module extracts information from ultra-high temperature speckle images through sparse representation, and at the same time broadens the feature coverage range with a multi-scale mechanism to ensure that the model obtains sufficient feature information in subsequent processing, effectively alleviating the low restoration accuracy caused by insufficient shallow feature extraction.

[0103] S4. The dual convolutional neural network includes an upper-side feature extraction path and a lower-side feature extraction path, and uses Output MFEM-B2 as the input of the upper and lower paths to extract global context features and local detail features respectively;

[0104] As a preferred implementation manner of step S4, the specific process of step S4 includes the following steps:

[0105] S41. The upper-side feature extraction path constructs a global context information capture system through the alternating stacking of conventional convolution and dilated convolution, in conjunction with a cross-layer connection mechanism: receiving the H×W×64 tensor feature Output MFEM-B2 extracted by MFEM-B2, extracting basic textures through 3×3 conventional convolution, then concatenating the first-layer output and the input features through channels, applying a 3×3 dilated convolution with a dilation rate r = 2 to the mixed features to expand the receptive field to a 7×7 pixel range, capturing cross-region speckle correlations. The alternating configuration strategy of the convolutional kernels ensures the ability to capture large-range speckle features while effectively controlling the consumption of computing resources by introducing batch normalization operations after each convolutional layer. Repeating the above operations, finally generating an H×W×128 global context feature map, which contains the overall deformation trend information of the speckles caused by thermal distortion;

[0106] S42. The lower-side feature extraction path adopts a progressive dilated convolution design to enhance the ability to capture local detail features by expanding the perception domain range, effectively avoiding grid artifacts in the feature extraction process and significantly improving the noise suppression ability of the network. Also receiving the H×W×64 tensor feature Output MFEM-B2 extracted by MFEM-B2, successively using progressive dilated convolution, and adding the output of each layer to the previous-order feature (residual connection) after batch normalization. For a two-dimensional input feature map I and a convolutional kernel K, the output O of the dilated convolution can be expressed as:

[0107]

[0108] where r is the set known dilation rate of the dilated convolution, m and n are the sizes of the convolution kernel, and finally a local detail feature map of H×W×128 is generated, focusing on the speckle edge blurring, micro fractures, and noise distribution caused by thermal perturbations. I is expressed as the extraction and fusion feature Output of the multi-scale features of the ultra-high temperature speckle image MFEM-B2 ;

[0109] S43. For the H×W×128 global context feature map and the H×W×128 local detail feature map of the upper and lower side feature extraction paths output by S41 and S42, they are concatenated along the channel dimension to form a 256-dimensional mixed feature tensor, retaining the complete information of the global deformation trend and local noise distribution. Subsequently, a 3×3 conventional convolution operation is applied to the mixed features to eliminate cross-channel redundant information through local receptive field modeling

[0110] More specifically, both the upper side feature extraction path and the lower side feature extraction path adopt a seven-layer convolution structure, with 32 3×3 convolution kernels configured in each layer, optimizing the computational efficiency while ensuring the network processing ability

[0111] In this embodiment, to fully improve the overall performance of the ThermoClearNet model, a dual convolutional neural network structure is designed to enhance the restoration effect of the model on the speckle image by increasing the network width. Its input is the H×W×64-channel feature tensor output by the multi-scale feature extraction module MFEM. In the model architecture, the upper side feature extraction path extracts global context features from the feature tensor, and the lower side feature extraction path extracts local detail features from the feature tensor. The upper and lower sub-networks work together to significantly improve the expression ability of the ThermoClearNet model

[0112] S5. Before each layer of the convolutional neural network in the upper side feature extraction path and the lower side feature extraction path, a polarized self-attention mechanism (polarized self-attention module, PSA) is introduced, and the input feature maps X of the current layer of the upper and lower paths are respectively input into the polarized self-attention mechanism to extract deep features. Its core lies in constructing an orthogonal dimension feature screening system: while implementing feature compression in a specific direction, maintaining a high-resolution representation in the vertical dimension

[0113] As a preferred implementation of step S5, the polarized self-attention mechanism in step S5 includes two main branches: the channel self-attention mechanism and the spatial self-attention mechanism, synchronously optimizing the feature distribution in the channel and spatial dimensions through a parallel two-stream integration mode. Its specific process includes the following steps

[0114] S51. The channel self-attention mechanism focuses on the semantic association between feature channels. The formula of the channel attention mechanism of the ThermoClearNet model is expressed as

[0115] I ch (X) = F SG [W Z (σ1(W v (X)) × F SM (σ2(W q (X)))]

[0116] Where ch represents the Channel channel branch, W Z , W v , W q are standard 1×1 convolution operations, σ1 and σ2 are two tensor reconstruction operators, F SG is the Sigmoid activation function, F SM is the SoftMax normalization function. By dynamically adjusting the weights of each channel, the model can better capture the dependencies between channels. The channel dimension needs to be finely modeled, so three-weight enhanced semantic screening is adopted to improve the feature expression ability of the ThermoClearNet model;

[0117] S52. The spatial self-attention mechanism then enhances the understanding of the correlation between image features. The formula for spatial self-attention is expressed as:

[0118] I sp (X) = F SG [σ3(F GP (σ1(W v (X))) × σ2(W v (X))]

[0119] Where sp represents the Spatial spatial branch, σ3 is the tensor reconstruction operator, F SG is the Sigmoid activation function, F GP represents the global pooling function, directly using global average pooling GPool to replace the explicit query-key W Z , W q interaction, directly performing element-wise multiplication on the original features to avoid loss of spatial resolution and ensure the integrity of speckle edge details; at the same time, since the computational cost of the spatial dimension is much higher than that of the channel dimension, omitting W Z , W q can reduce the amount of parameter calculation;

[0120] S53. Add the output I ch (X) of the channel branch and the output I sp (X) of the spatial branch element-wise to obtain the feature map X' optimized by PSA:

[0121] X' = I ch (X) + I sp (X)

[0122] This design enables each feature extraction unit of the dual convolutional neural network to recalibrate the feature response in the channel and spatial dimensions through the PSA mechanism before processing, strengthen the transmission of effective features and suppress noise interference, and input the fused feature map X' into the conventional convolution or dilated convolution module of the current layer to extract deep features through the convolution kernel operation;

[0123] S54. For the dilated convolution layer of the upper-side path, after channel concatenation of X' and the cross-layer connected features, convolution is performed; for the progressive dilated convolution layer of the lower-side path, X' is added to the residual features and then input into the next layer.

[0124] S6. Concatenate the final output global context feature map containing deep features and the local detail feature map along the channel dimension to obtain a complete information feature that retains the global and local noise distributions;

[0125] S7. Construct a joint loss function L for the supervised image restoration network ThermoClearNet model;

[0126] As a preferred implementation of step S7, in step S7, to construct the joint loss function L for the supervised image restoration network ThermoClearNet model, the per-pixel Charbonnier loss L char , edge-aware loss L edge , and frequency reconstruction loss L fr are used to restore the ultra-high temperature speckle image to a level similar to the cold-state speckle image. The specific process includes the following steps:

[0127] S71. The Charbonnier loss function alleviates the influence of outliers by introducing a square root operation, enhancing the robustness of the model to noise. The specific formula is as follows:

[0128]

[0129] In the formula, I R represents the restored image output by the model, that is, the ideal speckle distribution after disturbance elimination, which needs to be aligned with the cold-state speckle image Y in the feature space, and ε is a non-zero constant term;

[0130] S72. The edge-aware loss function extracts the high-frequency information of the image using the Laplace operator and focuses on optimizing the accurate reconstruction of the speckle edge contour. The specific formula is as follows:

[0131]

[0132] In the formula, Δ represents the Laplace operator, and ΔI RFor the predicted edge features after restoration, ΔY represents the true edge features, which are used as the supervision signal, and the edges generated by the model are required to be consistent with the physical true speckle deformation pattern;

[0133] S73. The frequency-domain reconstruction loss function transforms the image into the frequency-domain space through Fourier transform to ensure that the restored result is consistent with the real image in terms of frequency-domain features. The specific formula is as follows:

[0134] L fr =||F(I R ) - F(Y)||

[0135] In the formula, F represents the fast Fourier transform;

[0136] S74. Calculate the above loss components respectively in each training stage, and continuously adjust the network parameters through gradient backpropagation. The overall objective function can be expressed in the following weighted combination form:

[0137] L=L char (I R , Y)+τ1L edge (I R , Y)+τ2L fr (I R , Y)

[0138] τ1 and τ2 represent the weight coefficients corresponding to the losses.

[0139] S8. Train the ThermoClearNet model to obtain the optimal image restoration network model;

[0140] As the preferred implementation manner of step S8, the specific process of step S8 includes the following steps:

[0141] S81. Classify the dataset collected in step S14 into a training set, a validation set, and a test set. The training set is used for parameter learning, the validation set is used for hyperparameter tuning, and the test set is used to evaluate the generalization performance of the model. This division method follows the standard practice in the field of machine learning;

[0142] S82. Execute the iterative optimization process of the model, update the model parameters driven by the training set data, monitor the overfitting situation of the model using the validation set and adjust the hyperparameters, and finally evaluate the prediction accuracy of the model using the test set;

[0143] S83. After a complete training-validation-test process, an optimized ThermoClearNet network model for speckle image restoration is obtained.

[0144] S9. Take the ultra-high temperature speckle image of the composite material specimen under ultra-high temperature conditions as the input. After passing through the image restoration network model, the restored result of the ultra-high temperature speckle image is output, and DIC calculation is performed to obtain the dynamic measurement of the stress deformation of the composite material under ultra-high temperature conditions.

[0145] As a preferred implementation manner of step S9, the specific process of step S9 includes the following steps:

[0146] S91. Input the captured ultra-high temperature speckle image into the optimized ThermoClearNet model. The trained network layer inside the model will automatically extract the input features and calculate the corresponding multiple linear regression parameters, which define the mathematical mapping relationship between the input ultra-high temperature speckle image and the target restored image.

[0147] S92. Finally, perform DIC digital image correlation calculation on the restored image output by the optimized ThermoClearNet model to obtain the high-precision dynamic measurement of the stress deformation of the composite material under ultra-high temperature conditions, and it can also provide an important reference basis for subsequent material processes, assembly, etc.

[0148] The beneficial effects of the present invention are as follows:

[0149] 1. In the present invention, the multi-scale feature extraction module in the deep neural network supervision model extracts the ultra-high temperature speckle image information by means of sparse representation, and at the same time, by virtue of the advantage of the multi-scale mechanism to broaden the feature coverage range, it effectively alleviates the low restoration accuracy caused by insufficient extraction of shallow features. Secondly, by designing a double convolutional neural network structure, the upper feature extraction path extracts global context features from the feature tensor, and the lower feature extraction path extracts local detail features from the feature tensor. Before each layer of the convolutional neural network in the upper feature extraction path and the lower feature extraction path, a polarization self-attention mechanism is innovatively introduced to extract deep features. The core lies in constructing an orthogonal dimension feature screening system: while compressing features in a specific direction, maintaining a high-resolution representation in the vertical dimension, and enhancing the restoration effect of the model on the ultra-high temperature speckle image. Therefore, the deep neural network model ThermoClearNet effectively eliminates the influence of thermal radiation, thermal disturbance, and haze phenomenon on the image quality in the ultra-high temperature environment, and significantly improves the dynamic strain measurement accuracy of the composite material under ultra-high temperature conditions.

[0150] 2. The present invention is specifically designed for ultra-high temperature conditions, and can effectively cope with complex problems such as uneven gas density, change in air refractive index, and sublimation of carbon fiber matrix in high-temperature environments, solve problems such as image quality degradation, feature point drift and jitter of traditional DIC technology in ultra-high temperature environments, and ensure the reliability of measurement results.

[0151] 3. Through the automated processing of the deep neural network, the present invention reduces manual intervention and improves the measurement efficiency. It is not only applicable to the dynamic measurement of composite materials under ultra-high temperature conditions in the fields of aerospace, metallurgy, petrochemical, etc., but also provides an important reference basis for subsequent material process optimization, assembly design, etc.

[0152] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0153] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices.

[0154] The above embodiments have introduced the present invention in detail. Specific examples are used herein to elaborate on the principle and implementation of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A dynamic measurement method for composite materials under ultra-high temperature conditions, characterized in that It includes the following steps: S1. Build a dynamic detection platform for high-temperature strain of composite materials, and collect the speckle images of the test piece under cold conditions and ultra-high temperature working conditions respectively as the data set; S2. Construct a supervised network ThermoClearNet model to eliminate the influence of thermal radiation and thermal disturbance on speckle images, including a multi-scale feature extraction module MFEM, a double convolutional neural network, and a polarization self-attention mechanism; S3. Take the ultra-high temperature speckle image as the input of the multi-scale feature extraction module MFEM to obtain the multi-scale fusion feature Output of the ultra-high temperature speckle image MFEM-B2 ; S4. The dual convolutional neural network includes an upper feature extraction path and a lower feature extraction path, and uses Output MFEM-B2 as the input of the upper and lower paths to extract global context features and local detail features respectively; S5. Before each layer of the convolutional neural network in the upper-side feature extraction path and the lower-side feature extraction path, introduce a polarization self-attention mechanism PSA, and input the input feature map X of the current layer of the upper and lower paths into the polarization self-attention mechanism respectively to extract deep features; S6. Concatenate the final output global context feature map containing deep features and the local detail feature map along the channel dimension to obtain a complete information feature that retains the global and local noise distributions; S7. Construct a joint loss function L for the supervised image restoration network ThermoClearNet model; S8. Train the ThermoClearNet model to obtain the optimal image restoration network model; S9. Collect the ultra-high temperature speckle image of the composite material test piece under ultra-high temperature working conditions as the input, output the restored result of the ultra-high temperature speckle image through the image restoration network model, and perform DIC calculation to obtain the dynamic measurement of the force deformation of the composite material under ultra-high temperature working conditions.

2. The dynamic measurement method of a composite material under ultra-high temperature conditions according to claim 1, characterized in that: The specific process of step S1 includes the following steps: S11. Build a high-temperature DIC experimental system, including a constant temperature box, an industrial camera, a lens, a blue light source, and a tensile testing machine; S12. Fix the test piece on the tensile rod of the tensile testing machine, open the observation window, eliminate the influence of the refraction of the observation window glass on the speckle image, adjust the camera after clamping in place to collect the cold-state speckle image of the test piece to be measured, and then use it for the training of the ThermoClearNet model; S13. Start the constant temperature box, heat the constant temperature box to 1000 °C and keep it constant for 20 minutes for image collection to ensure that a stable heat flow field is formed in the experimental environment. Under this stable state, use the camera to collect the ultra-high temperature speckle image of the test piece to be measured; S14. Divide the collected ultra-high temperature speckle images into a training set, a validation set, and a test set according to 6:2:

2.

3. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 1, characterized in that: In step S3, the multi-scale feature extraction module MFEM includes two sub-networks, MFEM-B1 and MFEM-B2. Using dilated convolution and convolution kernels of various sizes, the extraction and fusion of multi-scale features of the ultra-high temperature speckle image are realized, and the feature Output is obtained. MFEM-B2 , and the specific process includes the following steps: S31. MFEM-B1 uses the concept of sparse representation to construct a three-level cascaded feature extraction architecture. The core of this system consists of three convolutional units with increasing parameters. Each processing unit contains three modules: a dilated convolutional operation layer, a normalization processing module, and a non-linear activation layer. Among them, the coefficient combination of the dilated convolutional operation layer forms a progressive dilation mechanism with 1 / 2 / 3. Through the stacking strategy of progressive dilated convolutions, the coverage range of the perception field of the ThermoClearNet model is significantly expanded; Among them, the expression of MFEM-B1 is as follows: Output MFEM-B1 = D3(D2(D1(N))) Among them, N represents the ultra-high temperature speckle image obtained by shooting, D3, D2, and D1 respectively represent the convolutional units with coefficients of 3, 2, and 1, and Output MFEM-B1 represents the sparse feature output by the MFEM-B1 module; S32. MFEM-B2 receives the sparse feature Output processed by MFEM-B1 as input, and this module constructs a three-channel heterogeneous feature extraction network. Each feature processing channel uses a 128-dimensional feature vector space as the operation basis. First, standard convolution preprocessing is performed on each channel, and the basic convolution operation is paired with the ReLU activation function with gradient stability characteristics to ensure the consistency of feature representation in the initial processing stage of each channel; MFEM-B1 As input, this module constructs a three-channel heterogeneous feature extraction network. Each feature processing channel uses a 128-dimensional feature vector space as the operation basis. First, standard convolution preprocessing is performed on each channel, and the basic convolution operation is paired with the ReLU activation function with gradient stability characteristics to ensure the consistency of feature representation in the initial processing stage of each channel; S33. Implement multi-scale feature convolution operations through heterogeneous convolution kernels configured as 1×1, 3×3, and 5×5 convolutions to obtain super high-temperature speckle image feature information at different scales. Then, perform PReLU operations on the features of the branches, and through multi-dimensional feature tensor splicing and fusion, deeply integrate the complementary features extracted from the four heterogeneous channels. The specific expression of MFEM-B2 is as follows: Output MFEM-B2 = Fusion(R1, R3, R5) R1 = F1(Output MFEM-B1 ) R3 = F3(Output MFEM-B1 ) R5 = F5(Output MFEM-B1 ) Among them, F1, F3, and F5 respectively represent convolution operations with different convolution kernel sizes, Fusion represents the feature concatenation function after the PReLU operation, and Output MFEM-B2 represents the extraction and fusion features of the multi-scale features of the ultra-high temperature speckle image output by MFEM-B2.

4. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 1, characterized in that: The specific process of step S4 includes the following steps: S41. The upper-side feature extraction path constructs a global context information capture system through the alternating stacking of conventional convolution and dilated convolution, in cooperation with the cross-layer connection mechanism: receiving Output MFEM-B2 , extracting the basic texture through 3×3 conventional convolution, then splicing the first-layer output and the input features through channels, applying 3×3 dilated convolution with a dilation rate r = 2 to the mixed features, expanding the receptive field to the 7×7 pixel range, capturing the cross-region speckle correlation. The alternating configuration strategy of the convolution kernels ensures the ability to capture large-range speckle features while effectively controlling the consumption of computing resources by introducing batch normalization operations after each convolution layer. Repeating the above operations, a global context feature map is finally generated, which contains the overall deformation trend information of the speckles caused by thermal distortion; S42. The lower-side feature extraction path adopts a progressive dilated convolution design to enhance the ability to capture local detailed features by expanding the receptive field range, and also receives Output MFEM-B2 , and progressive dilated convolution is adopted in sequence. The output of each layer is added to the previous-order feature after batch normalization. For the two-dimensional input feature map I and the convolution kernel K, the output O of the dilated convolution can be expressed as: Where r is the set known dilation rate of the dilated convolution, m and n are the sizes of the convolution kernel, and finally a local detail feature map is generated, focusing on the speckle edge blurring, micro-fracture, and noise distribution caused by thermal perturbations. I is expressed as the extraction and fusion feature Output of the multi-scale features of the ultra-high temperature speckle image MFEM-B2 ; S43. Concatenate the global context feature map and the local detail feature map of the upper and lower side feature extraction paths output by S41 and S42 along the channel dimension to retain the complete information of the global deformation trend and local noise distribution. Subsequently, apply a 3×3 conventional convolution operation to the mixed features to eliminate cross-channel redundant information through local receptive field modeling.

5. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 4, characterized in that: Both the upper side feature extraction path and the lower side feature extraction path adopt a seven-layer convolution structure, with 32 3×3 convolution kernels configured in each layer, optimizing the computational efficiency while ensuring the network's processing ability.

6. The dynamic measurement method of the composite material under ultra-high temperature conditions according to claim 1, wherein: The polarization self-attention mechanism in step S5 includes two main branches: the channel self-attention mechanism and the spatial self-attention mechanism. It synchronously optimizes the feature distribution in the channel and spatial dimensions through a parallel two-stream integration mode. The specific process includes the following steps: S51. The channel self-attention mechanism focuses on the semantic association between feature channels. The formula for the channel attention mechanism of the ThermoClearNet model is expressed as: I ch (X) = F SG |W Z (σ1(W v (X)) × F SM (σ2(W q (X)))] where ch represents the Channel channel branch, W Z , W v , W q is a standard 1×1 convolution operation, σ1 and σ2 are two tensor reconstruction operators, F SG is the Sigmoid activation function, F SM is the SoftMax normalization function. By dynamically adjusting the weights of each channel, the model can better capture the dependencies between channels. The channel dimension needs to be finely modeled, so three-weight enhanced semantic screening is adopted to improve the feature expression ability of the ThermoClearNet model; S52. The spatial self-attention mechanism then strengthens the understanding of the correlation between image features. The formula for spatial self-attention is expressed as: I sp (X) = F SG [σ3(F GP (σ1(W v (X))) × σ2(W v (X))] where sp represents the Spatial branch, σ3 is the tensor reconstruction operator, F SG is the Sigmoid activation function, F GP represents the global pooling function, directly using global average pooling GPool to replace the explicit query-key W Z , W q interaction, directly performing pointwise multiplication on the original features to avoid loss of spatial resolution and ensure the integrity of speckle edge details; at the same time, since the computational cost of the spatial dimension is much higher than that of the channel dimension, omitting W Z , W q can reduce the amount of parameter calculation; S53. Output the channel branch I ch (X) and the spatial branch output I sp (X) are added element by element to obtain the feature map X' optimized by PSA: X' = I ch (X) + I sp (X) This design enables each feature extraction unit of the double convolutional neural network to first recalibrate the feature response in the channel and spatial dimensions through the PSA mechanism before processing, strengthening the transmission of effective features and suppressing noise interference. Input the fused feature map X' into the conventional convolution or dilated convolution module of the current layer, and extract deep features through convolution kernel operations; S54. For the dilated convolution layer of the upper side path, X' is concatenated with the features of the cross-layer connection and then convolution is performed; for the progressive dilated convolution layer of the lower side path, X' is added to the residual features and then input to the next layer.

7. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 1, characterized in that: In step S7, a joint loss function L for constructing the supervised image restoration network ThermoClearNet model is used, and the pixel-by-pixel Charbonnier loss L char , edge-aware loss L edge , and frequency reconstruction loss L fr are used to restore the ultra-high temperature speckle image to a level similar to the cold-state speckle image. The specific process includes the following steps: S71. The Charbonnier loss function alleviates the influence of outliers by introducing a square root operation, enhancing the model's robustness to noise. The specific formula is as follows: where I R represents the restored image output by the model, that is, the ideal speckle distribution after disturbance elimination, which needs to be aligned with the cold-state speckle image Y in the feature space, and ε is a non-zero constant term; S72. The edge-aware loss function uses the Laplace operator to extract the high-frequency information of the image, focusing on optimizing the precise reconstruction of the speckle edge contour. The specific formula is as follows: where Δ represents the Laplace operator, and ΔI R is the predicted edge feature after restoration, and ΔY represents the true edge feature, which is used as a supervision signal. It is required that the edges generated by the model be consistent with the physical true speckle deformation pattern; S73. The frequency-domain reconstruction loss function converts the image to the frequency-domain space through Fourier transform to ensure that the restored result is consistent with the real image in terms of frequency-domain features. The specific formula is as follows: L fr = ||F(I R ) - F(Y)|| In the formula, F represents the fast Fourier transform; S74. Calculate the above loss components separately in each training stage, and continuously adjust the network parameters through gradient backpropagation. Its overall objective function can be expressed in the following weighted combination form: L = L char (I R , Y) + τ1L edge (I R , Y) + τ2L fr (I R , Y) τ1 and τ2 represent the weight coefficients corresponding to the losses.

8. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 2, characterized in that: The specific process of step S8 includes the following steps: S81. Classify the dataset collected in step S14 into a training set, a validation set, and a test set. The training set is used for parameter learning, the validation set is used for hyperparameter tuning, and the test set is used to evaluate the generalization performance of the model. This partitioning method follows the standard practice in the field of machine learning; S82. Execute the iterative optimization process of the model. Update the model parameters driven by the training set data, monitor the overfitting situation of the model using the validation set and adjust the hyperparameters, and finally use the test set to evaluate the prediction accuracy of the model; S83. After a complete training-validation-test process, an adequately optimized ThermoClearNet network model for speckle image restoration is obtained.

9. A dynamic measurement method for composite materials under ultra-high temperature conditions according to claim 1, characterized in that: The specific process of step S9 includes the following steps: S91. Input the captured ultra-high temperature speckle image into the optimized ThermoClearNet model. The trained network layer inside the model will automatically extract the input features and calculate the corresponding multiple linear regression parameters, which define the mathematical mapping relationship between the input ultra-high temperature speckle image and the target restored image; S92. Finally, perform DIC digital image correlation calculation on the restored image output by the optimized ThermoClearNet model to obtain a high-precision dynamic measurement of the force deformation of the composite material under ultra-high temperature conditions.

Images