Strain measurement method based on GCN neural network

Abstract

The invention provides a strain measurement method based on a GCN neural network, and the method comprises the steps: collecting a speckle image and a speckle displacement field of a tested piece according to the actual working condition demands, and forming a data set for the subsequent training and testing of a measurement model; building a speckle displacement measurement model according to the GCN neural network; inputting the data set into a speckle displacement measurement model for training to obtain optimal measurement model configuration, and performing displacement measurement on deformation of the speckle image to obtain a prediction result of speckle displacement; and calculating the strain of all points in the speckle image based on the obtained speckle displacement prediction result, and gathering to form a surface strain field of the measured object. According to the method provided by the invention, the deformation stress of the high-temperature material component can be accurately measured by building the speckle displacement measurement model based on the GCN neural network, and high-precision and real-time strain measurement in a high-temperature environment is realized.

Description

Technical Field

[0001] The present invention belongs to the fields of deep learning and experimental mechanics technology, and particularly relates to a strain measurement method based on a GCN neural network. Background Technique

[0002] With the rapid development of fields such as aerospace and composite materials, the demand for the performance of composite materials in high-temperature environments is increasing. Therefore, studying the mechanical properties of these materials in high-temperature environments is of great significance for material selection and structural design. Strain is one of the main characteristics of material mechanical properties. At present, there are mainly two types of methods for measuring the strain of materials: contact type and non-contact type.

[0003] The contact type method mainly realizes the measurement of strain through strain gauges and extensometers. However, this method can only obtain displacement strain data at certain specific points and cannot perform full-field strain measurement on the surface of the specimen. Secondly, connecting the high-temperature strain gauge to the surface of the specimen will have a local strengthening effect on the material, affecting the measurement accuracy.

[0004] The non-contact measurement method is mainly based on optical measurement methods, which have the characteristics of non-contact and full-field measurement. The most widely used optical measurement methods are the Digital Image Correlation (DIC) method and the stress analysis method based on the U-net convolutional neural network.

[0005] Currently, there are mainly two problems faced in strain measurement using optical methods: (1) The setting of some parameters of the DIC algorithm (such as subset size, step size, window function, etc.) under extreme working conditions such as high temperature will cause great differences in the strain measurement results; (2) The calculation time of the DIC measurement method for obtaining the strain measurement results is too long. At present, the environment for strain measurement of high-temperature materials is generally above 800°C. Poor-quality high-temperature-resistant speckles will seriously affect the measurement results of traditional DIC. Secondly, the high-temperature measurement environment is complex, and the measurement results of DIC are affected by multiple parameters such as subset size, step size, window function, etc. If these parameters cannot be set well, it may lead to the failure of the measurement results. At the same time, in the aerospace field, a large number of digital images are collected, and the calculation amount is large. Due to conditions, high-performance computers cannot be widely used for the measurement environment. Therefore, the DIC method is difficult to apply in on-site high-temperature real-time measurement. The stress analysis method based on the U-net convolutional neural network has problems such as insufficient extraction of local features of images and low accuracy of strain measurement. Especially in high-temperature environments, the stress analysis method based on the U-net convolutional neural network is no longer applicable. The Graph Convolutional Networks (GCN) can effectively capture the relationship between pixels or regions in the image and has high accuracy and robustness in complex scenarios. Summary of the Invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention proposes a strain measurement method based on a GCN neural network to achieve high-precision and real-time strain measurement in a high-temperature environment.

[0007] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0008] A strain measurement method based on a GCN neural network specifically includes the following steps:

[0009] S1. Collect the speckle images and speckle displacement fields of the test piece according to the actual working condition requirements and form them into a data set for subsequent training and testing of the measurement model;

[0010] S2. Build a speckle displacement measurement model according to the GCN neural network;

[0011] S3. Input the data set into the speckle displacement measurement model for training to obtain the optimal measurement model configuration, measure the deformation of the speckle image to obtain the prediction result of the speckle displacement;

[0012] S4. Based on the prediction result of the speckle displacement obtained in step S3, calculate the strain of all points in the speckle image, and the set constitutes the surface strain field of the measured object.

[0013] Further, step S1 specifically includes:

[0014] S11. For test pieces of different materials, combine the surface grinding method and the spraying method to prepare high-temperature-resistant speckles on the surface of the test piece;

[0015] S12. Use the digital image correlation method to measure the processed test piece, extract the speckle images at each time, and measure the displacement field to establish a data set based on the speckle images and speckle displacement fields.

[0016] More specifically, step S12 includes:

[0017] S121. Use a material tensile testing machine to conduct a material tensile experiment on the processed test piece to extract the displacement fields of the test piece at different times under this working condition and establish a data set;

[0018] S122. Turn on the calibrated CCD binocular camera and the light source, align them with the center of the specimen, adjust the level to make the specimen surface parallel to the CCD binocular camera, use a calibration plate for calibration to eliminate systematic errors; start the measurement program according to the main deformation area of the tensile specimen;

[0019] S123. Conduct a test. Start the material tensile testing machine, set the experimental tensile rate. After the material tensile testing machine starts working, turn on the CCD binocular camera to take pictures.

[0020] S124. Obtain the tensile force and displacement through the material tensile testing machine, and obtain the original speckle image and the deformed speckle image after applying the load through the CCD binocular camera.

[0021] S125. Divide the speckle image at a certain moment into multiple sub-regions, match the original speckle image and the deformed speckle image for each sub-region to obtain the displacement value of the speckle sub-region under the applied load; calculate the displacement values of all subset regions to obtain the speckle displacement field at a certain moment.

[0022] Further, step S2 specifically includes:

[0023] S21. Design and improve the non-local mean filtering algorithm to denoise the speckle image.

[0024] S22. For the denoised speckle image, select the interested rectangular region ROI, represent the speckle feature points as a graph structure to represent the correlation between speckle feature points.

[0025] S23. Obtain the coordinates of the speckle feature points and the adjacency matrix of the speckle feature points from the graph structure, and introduce the multi-scale graph attention module MGAM to integrate the speckle feature point information at different levels.

[0026] S24. Build a deep residual graph neural network model. Based on the GCN neural network, combine the initial speckle image residual and the high-order neighborhood speckle image residual, and improve the learning effect of the measurement model through a multi-input residual structure.

[0027] More specifically, step S21 specifically includes:

[0028] S211. Adopt a sliding window method to slide a search box of a fixed size on the global speckle image, and perform non-local mean filtering processing on the speckle image in each search box; first fit a low-order polynomial by the least squares method to smooth the speckle image in the window, use the polynomial value at the center coordinate in the window as the new gray value of the center pixel of the smoothed speckle image, compare the similarity of the polynomial values at the center points of different windows, and determine the weight of the search box in the global image reconstruction.

[0029] S212. For each search box, let \(I = \{I(x, y)|x, y\in S\}\) be the speckle image containing noise within the search box, where \(S\) represents the set of pixel point indices, \(x\) is the central pixel point, i.e., the target pixel point to be denoised, and \(y\) is the pixel point in the search box, i.e., any pixel point other than the central pixel point. The denoised image is obtained by weighted averaging all the pixels in \(I\), and the formula is expressed as:

[0030] O(x)=\sum y∈I \(\mu\cdot\omega(x,y)I(y)\);

[0031] Among them, \(\mu\) is the denoising coefficient; \(\omega(x,y)\) is the similarity weight coefficient; \(I(y)\) represents the image block centered on the search box pixel point \(y\); \(O(x)\) represents the filtered image block centered on the central pixel point \(x\). The denoising coefficient \(\mu\) is dynamically adjusted according to the local density \(\rho\) of the speckle distribution to achieve an adaptive balance between the denoising intensity and the preservation of details. The local density \(\rho\) of the speckle is defined as the proportion of speckle pixels in the similarity box, and the calculation formula is as follows:

[0032]

[0033] Among them, \(N\) is the number of speckle pixels in the search box; \(W\) is the side length of the similarity box. The similarity box adopts a multi-scale sliding window strategy. For large-scale speckle structures, a large side length is selected to maintain coherence. For small-scale details, the side length is switched to retain fine features.

[0034] S213. Calculate the similarity weight coefficient \(\omega(x,y)\); the similarity degree between image blocks is quantified based on the Euclidean distance, and the gradient information of the image block is additionally introduced to improve the Euclidean distance. The gradient feature vectors of the image blocks \(V(x)\) and \(V(y)\) are defined as and Then the improved Euclidean distance \(d(x,y)\) between the search box pixel point \(y\) and the central pixel point \(x\) is:

[0035]

[0036] Among them, \(V(x)\) and \(V(y)\) are the gray values of the image blocks centered on \(x\) and \(y\) respectively, and are the gradient feature vectors of the image block; \(\alpha\) is the gradient weight coefficient;

[0037] The gradient weight coefficient \(\alpha\) is adaptively adjusted according to the complexity of the local structure of the speckle, and the formula is expressed as:

[0038]

[0039] Among them, is the average value of the gradient amplitudes of all pixels within the current similarity box, represents the maximum value of the pixel gradient amplitude in the search box, and γ is an empirical constant; when the local structure is complex, i.e., the gradient amplitude is high, α is increased to strengthen the weight of the gradient information, and vice versa;

[0040] Finally, the calculation formula for the similarity weight coefficient ω(x,y) is:

[0041]

[0042] where h is the filtering coefficient and M(x) is the normalization coefficient.

[0043] More specifically, step S22 specifically includes:

[0044] S221. Select the region of interest rectangle ROI, perform feature detection on the ROI region to form an initial set of interest feature points with coordinate information; then sequentially perform full-image feature extraction on the speckle images at each moment to obtain the full-image feature point sets at each moment, and then integrate the full-image feature point sets at each moment to obtain a sequence feature point set;

[0045] S222. Match the initial set of interest feature points with the sequence feature point set; obtain the coordinates of the speckle feature points in the ROI region on the images at different moments, and obtain the coordinates and displacements of the speckle feature points on the image through speckle feature point extraction and feature matching and displacement Thus, the speckle features are represented as a graph structure, where the nodes represent different speckle feature points and the edges represent the correlation relationships between different speckle feature points;

[0046] S223. For the correlation relationships between different speckle feature points, calculate the correlation between speckle feature points through the Pearson correlation coefficient, and construct an adjacency matrix A(r i ,r j );

[0047] S224. For the obtained adjacency matrix, only retain the upper triangular values of each adjacency matrix, and vectorize the adjacency matrix into a feature vector required for GCN network training to generate a flattened feature vector as the primary feature representation.

[0048] More specifically, step S23 specifically includes:

[0049] S231. Design a multi-scale graph attention module MGAM. Based on the graph structure of speckle feature points, firstly, the features of all nodes in the graph are averaged according to the channel dimension through global pooling guided by channel attention. Then, the importance of different feature channels is analyzed through a learnable channel attention mechanism to obtain global features. Then, the local features of each speckle node are input into the graph attention module GAM to generate local speckle features with the same dimension as the global features. Then, the global features are fused with the local speckle features.

[0050] S232, while applying the attention mechanism on each node through GAM to learn the relationship between nodes and aggregate information from neighboring nodes;

[0051] S233. In order to make the attention coefficients between different nodes easier to compare, the obtained attention coefficients are normalized to generate a new feature representation for each speckle node.

[0052] More specifically, step S24 specifically includes:

[0053] S241, the input feature matrix A (r i ,r j ) is transformed to obtain H( 0 ) as the initial feature;

[0054] S242. Add high-order neighborhood residuals to the basic GCN. Each layer inherits the convolution output of the previous layer to reduce the propagation speed of node features. At the same time, add initial residuals to ensure that each speckle node can eventually retain a part of the initial features. And by introducing hyperparameters to adjust the ratio between the initial residual and the high-order neighborhood residual, the processing performance of the displacement measurement model when facing different speckle images is improved.

[0055] Furthermore, step S3 specifically includes:

[0056] S31, acquiring sample data for a speckle displacement measurement model, the sample data being an original speckle image captured by a CCD industrial camera, a deformed speckle image after a load is applied, and a speckle displacement field calculated in S1;

[0057] S32. Randomly shuffle the data and divide the obtained sample data into a training set, a validation set and a test set in a ratio of 6:2:2; the training set is used to train the speckle displacement measurement model and optimize the weight coefficient by back propagation; the validation set is used to monitor the training process, adjust the hyperparameters and select the optimal model configuration; the test set is only used after the model training and parameter adjustment are completed, and is input into the final optimal model to evaluate its generalization ability and displacement measurement accuracy;

[0058] S33. Input the speckle images before and after deformation, as well as the calculated speckle displacement field, into the speckle displacement measurement model to obtain the predicted results of the speckle displacement.

[0059] Further, step S4 specifically includes:

[0060] S41. Calculate the displacements of all displacement points based on the predicted results of the speckle displacement obtained in step S3.

[0061] S42. Based on the displacements of all obtained displacement points, calculate the Green strain components in the case of finite deformation. Through the Green strain components, the strain distribution of the material in each direction can be obtained, thereby constructing the surface strain field of the test piece.

[0062] Based on the above technical solutions, the present invention has the following beneficial effects:

[0063] 1. A strain measurement method based on the GCN neural network constructed by the present invention solves the problem of the influence of DIC parameter settings on the strain measurement results in high-temperature environments. At the same time, the use of the GCN neural network makes the calculation speed decrease exponentially, thus making it possible to output the strain measurement results in real time.

[0064] 2. The present invention designs an improved non-local mean filtering algorithm. While suppressing noise, it maximally retains the true displacement information of the speckle features to achieve dynamic weight allocation, and better retains the local details and structural integrity of the speckle image while denoising, providing a high-fidelity data basis for the subsequent accurate prediction of the speckle displacement.

[0065] 3. By introducing a multi-scale graph attention module, the present invention realizes the adaptive learning of local information in the speckle image and captures the characteristics of the object at different times; it helps to improve the matching effect of speckle feature points, especially in the case of poor quality of speckle images under extreme working conditions such as high temperature.

[0066] 4. The present invention performs multi-layer graph convolution operations through a multi-input residual structure, allowing the model to better utilize different speckle image input information to improve performance when in the deep layer, and being more capable of obtaining comprehensive speckle node feature information, which is more effective when processing high-temperature speckle displacement image datasets.

[0067] In summary, the speckle displacement measurement model based on the graph neural network proposed by the present invention has good spatio-temporal feature extraction and processing capabilities, can significantly improve the efficiency of strain measurement, and output the strain measurement results in real time. Description of the Drawings

[0068] The accompanying 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:

[0069] Figure 1 is the overall flowchart of a strain measurement method based on a GCN neural network proposed by the present invention;

[0070] Figure 2 is the structural diagram of the speckle displacement measurement network based on the GCN neural network of the present invention;

[0071] Figure 3 is the schematic diagram of the speckle pattern structure expressed by the speckle adjacency matrix of the present invention. Detailed implementation manners

[0072] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0073] Although the steps in the present invention are numbered, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It can be understood that the term "and / or" used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0074] As Figure 1 shown, the present invention proposes a strain measurement method based on a GCN neural network, which specifically includes the following steps:

[0075] S1. Collect the speckle images and speckle displacement fields of the test piece according to the actual working condition requirements and form them into a data set for subsequent training and testing of the measurement model;

[0076] As a preferred implementation manner, step S1 specifically includes:

[0077] S11. For test pieces made of different materials, prepare high-temperature-resistant speckles on the surface of the test piece by combining the surface grinding method and the spraying method;

[0078] In this embodiment, the speckle material has a relatively high melting point (>2000 degrees Celsius) and ultra-high temperature stability, and maintains chemical stability with the measured material under high-temperature conditions; therefore, the surface of the object to be measured is ground step by step to remove the oxide film and stains, and a high-pressure spray gun is used to spray the speckle material mixture onto the surface of the test piece to make randomly distributed speckles.

[0079] S12. Use the digital image correlation method to measure the processed test piece, extract the speckle images at each time, measure the displacement field, and establish a data set based on the speckle images and the speckle displacement field.

[0080] More specifically, step S12 includes:

[0081] S121. Use a material tensile testing machine to conduct a material tensile experiment on the processed test piece to extract the displacement field of the test piece at different times under this working condition and establish a data set; in this embodiment, specifically: set the experimental method, experimental material, specimen number, specimen shape, gauge length, specimen width, specimen thickness, experimental temperature, and tensile speed in the material tensile testing machine, and zero the displacement and load in the system;

[0082] S122. Turn on the calibrated CCD binocular camera and the light source, align them with the center of the specimen, adjust the level to make the specimen surface parallel to the CCD binocular camera, use a calibration plate for calibration to eliminate systematic errors; start the measurement program according to the main deformation area of the tensile specimen; in this embodiment, the light source used is an ordinary white light lamp, adjust the light intensity and the light convergence degree to form a uniform light field on the specimen surface, check the number of speckles in the selected surface component, and adjust the camera frequency to 25 Hz;

[0083] S123. Conduct the experiment, start the material tensile testing machine, set the experimental tensile rate for slow loading. In this embodiment, the experimental tensile rate is set to 0.5 μm / s. After the material tensile testing machine starts working, when the load-displacement curve appears, turn on the CCD binocular camera to take pictures;

[0084] S124. When the load-displacement curve in the material tensile testing machine stops, end the CCD binocular camera shooting; obtain the tensile force and displacement through the material tensile testing machine, and obtain the original speckle image and the deformed speckle image after applying the load through the CCD binocular camera;

[0085] S125. Divide the speckle image at a certain moment into multiple sub-regions, match the original speckle image and the deformed speckle image for each sub-region to obtain the displacement value of the speckle sub-region under the applied load; calculate the displacement values of all sub-set regions to obtain the speckle displacement field at a certain moment.

[0086] For example, select subset A in the reference picture (i.e., the original speckle image), and in the deformed picture (i.e., the deformed speckle reference image), the corresponding subset A of subset A is matched warp , so that the displacement value of subset A can be obtained. After calculating the displacements of all sub-set regions, the speckle displacement field at a certain moment can be obtained. S2. Build a speckle displacement measurement model according to the GCN neural network;

[0087] As a preferred implementation method, such asFigure 2 As shown, step S2 specifically includes:

[0088] S21. Design an improved non - local mean filtering algorithm to perform noise reduction processing on the speckle image; the improved non - local mean filtering algorithm designed in the present invention can ensure that image details are not lost, and at the same time, the denoised image also has relatively high clarity; the specific filtering process is as follows:

[0089] S211. Adopt a sliding window method to slide a search box of a fixed size on the global speckle image, and perform non - local mean filtering processing on the speckle image within each search box; first, fit a low - order polynomial by the least - squares method to smooth the speckle image within the window, use the polynomial value at the center coordinate within the window as the new gray value of the central pixel of the smoothed speckle image, compare the similarity of the polynomial values at the center points of different windows, and determine the weight value of the search box in the global image reconstruction;

[0090] S212. For each search box, let I = {I(x, y)|x, y ∈ S} be the speckle image containing noise within the search box, S represents the set of pixel point indices, x is the central pixel point, that is, the target pixel point to be denoised, and y is the pixel point within the search box, that is, any pixel point other than the central pixel point; the denoised image is obtained by weighted averaging all the pixels in I, and the formula is expressed as:

[0091] O(x) = Σ y∈I μ·ω(x, y)I(y);

[0092] Among them, μ is the noise reduction coefficient; ω(x, y) is the similarity weight coefficient; I(y) represents the image block centered on the search box pixel point y; O(x) represents the filtered image block centered on the central pixel point x; the noise reduction coefficient μ is dynamically adjusted according to the local density ρ of the speckle distribution to achieve an adaptive balance between the denoising intensity and detail preservation; in this embodiment, the dynamic adjustment of the noise reduction coefficient is specifically as follows:

[0093]

[0094] The local density ρ of the speckle is defined as the proportion of speckle pixels within the similarity box, and the calculation formula is as follows:

[0095]

[0096] Among them, N is the number of speckle pixels within the search box; W is the side length of the similarity box; the similarity box adopts a multi - scale sliding window strategy; for large - scale speckle structures, a large side length is selected to maintain coherence; for small - scale details, the side length is switched to retain fine features; for example, in this embodiment, the dynamic adjustment formula of the side length W of the similarity box is:

[0097]

[0098] Such an adjustment strategy significantly improves the flexibility and accuracy of similarity evaluation by adaptively matching the speckle scale characteristics;

[0099] S213. Calculate the similarity weight coefficient ω(x,y); quantify the similarity degree between image patches based on the Euclidean distance, and additionally introduce the gradient information of the image patches to improve the Euclidean distance; define the gradient feature vectors of image patches V(x) and V(y) as and Then the Euclidean distance d(x,y) between the pixel point y in the search box and the central pixel point x after improvement is:

[0100]

[0101] where V(x) and V(y) are the gray values of the image patches centered at x and y respectively, and are the gradient feature vectors of the image patches; α is the gradient weight coefficient; by fusing the gradient features, the similarity coefficient ω(x,y) can more accurately reflect the local geometric correlation between speckle feature points and avoid the detail loss caused by ignoring the structural information in the traditional method;

[0102] The gradient weight coefficient α is adaptively adjusted according to the complexity of the local structure of the speckle, and the formula is expressed as:

[0103]

[0104] where is the average value of the gradient amplitudes of all pixels within the current similarity box, represents the maximum value of the gradient amplitudes of the pixels within the search box, γ is an empirical constant, which is defaulted to 1.2 in this embodiment; when the local structure is complex, that is, the gradient amplitude is high, α is increased to strengthen the weight of the gradient information, and vice versa.

[0105] Finally, the calculation formula of the similarity weight coefficient ω(x,y) is:

[0106]

[0107] where h is the filtering coefficient and M(x) is the normalization coefficient.

[0108] Compared with the traditional non - local mean filtering algorithm that only relies on gray - scale similarity and ignores local structural information, which may lead to detail loss or noise residue, the improved method proposed in this application maximally retains the true displacement information of speckle features while suppressing noise and realizes dynamic weight allocation; by introducing gradient information combined with gray - scale information to act simultaneously, it better retains the local details and structural integrity of the speckle image during denoising, providing a high - fidelity data basis for subsequent displacement field reconstruction.

[0109] S22. For the speckle image after noise reduction processing, select the rectangular region of interest ROI, represent the speckle feature points as a graph structure to represent the association between speckle feature points;

[0110] As a preferred implementation manner, step S22 specifically includes:

[0111] S221. Select the rectangular region of interest ROI, perform feature detection on the ROI region to form an initial - moment interest feature point set with coordinate information; then sequentially perform full - image feature extraction on the speckle images at each moment (t k , k = 1, 2, 3…n) to obtain the full - image feature point sets at each moment; the feature extraction process of the speckle images at each moment will cover the entire image, not just the ROI region, thus forming the full - image feature point sets. This set contains all the detected feature points in the image, providing complete information about the global features of the image; then integrate the full - image feature point sets at each moment to obtain the sequence feature point set;

[0112] S222. Match the interest feature point set at the initial moment (t 0 time) with the sequence feature point set; obtain the coordinates of the speckle feature points in the ROI region on the images at different moments. Through speckle feature point extraction and feature matching, obtain the coordinates and displacement In the displacement calculation process, the GCN network can effectively capture the dynamic relationships and interdependencies between speckles by utilizing the complex topological structure and interactions of speckle points. The calculation formula for the displacement of speckle feature points:

[0113]

[0114] Where: is the horizontal displacement of the speckle feature point numbered i at the k - th moment is the vertical displacement of the speckle feature point numbered i at the k - th moment; is the x - direction coordinate of the speckle feature point numbered i in the k - th moment image; is the y - direction coordinate of the speckle feature point numbered i in the k - th moment image;

[0115] The speckle features are thus represented as a graph structure, where nodes represent different speckle feature points and edges represent the correlation relationships between different speckle feature points;

[0116] S223. For the correlation relationships between different speckle feature points, the Pearson correlation coefficient is used to calculate the correlation between speckle feature points, and an adjacency matrix A(r i ,r j ) of speckle feature points is constructed; The formula is expressed as:

[0117]

[0118] where r i and r j respectively represent the time series extracted from the i-th speckle feature point and the j-th speckle feature point, and E(·) represents the mathematical expectation of the sequence;

[0119] S224. For the obtained adjacency matrix, in order to avoid feature redundancy, only the upper triangular values of each adjacency matrix are retained, and the adjacency matrix is vectorized into a feature vector required for GCN network training, and the flattened feature vector is generated as the primary feature representation.

[0120] The schematic diagram of the adjacency matrix A(r i ,r j ) of speckle feature points is as shown in Figure 3 . This connection relationship indicates that there is a certain correlation between speckle feature points. Such a speckle graph structure can help understand the displacement relationship between speckle feature points and provide important feature information for the subsequent speckle displacement measurement model based on the GCN neural network.

[0121] S23. Obtain the coordinates of speckle feature points and the adjacency matrix of speckle feature points from the graph structure, and introduce a multi-scale graph attention module MGAM to integrate speckle feature point information at different levels;

[0122] As a preferred implementation manner, step S23 specifically includes:

[0123] S231. Design a multi-scale graph attention module MGAM. Based on the graph structure of speckle feature points, first, through global pooling guided by channel attention, the features of all nodes in the graph are averaged according to the channel dimension, and the importance of different feature channels is analyzed through a learnable channel attention mechanism to obtain global statistical features; Then, the local features of each speckle node are input into the graph attention module (GAM) to generate local features with the same dimension as the global features; Then, the global features are fused with the local speckle features;

[0124] S232. At the same time, apply the attention mechanism on each node through GAM to learn the relationships between nodes and aggregate information from adjacent nodes;

[0125] For the graph attention module, the input of the layer is a set of node features V = {v 1 , v 2 , …, v N , v i ∈ R D}, where N represents the number of speckle nodes and D represents the dimension of each speckle node feature. The output of the layer is a new set of speckle node features V' = {v 1 ′, v 2 ′, …, ′, v i ′ ∈ R D ′}. Similarly, N represents the number of nodes after update, and D' represents the dimension of the speckle node features after update. Generally, the dimension of the speckle node features does not change.

[0126] During the update process of the speckle node features, the calculation of the attention coefficients is involved, and the calculation process is as follows:

[0127]

[0128] where e ij represents the importance of the feature point of node j to node i, and W q and W k are learnable parameters.

[0129] S233. In the speckle displacement calculation model, the model allows each speckle node to participate in the update of other speckle nodes, that is, all speckle nodes are connected. To make the attention coefficients easy to compare between different nodes, the obtained attention coefficients need to be normalized to generate a new feature representation for each speckle node. In this embodiment, the softmax function is used, and the formula is as follows:

[0130] a ij = softmax(e ij );

[0131]

[0132] where N i represents the neighbor speckle nodes of the i-th speckle node, that is, other speckle nodes except itself. Therefore, once the normalization process of the attention coefficients is completed, these coefficients can be used to calculate the linear combination of the corresponding speckle features, so as to generate a new feature representation for each speckle node; through step S23, the MGAM module can adaptively learn the local information in the speckle image and capture the characteristics of the object at different times. This will help improve the matching effect of the speckle feature points, especially in the case of poor quality of the speckle images under extreme working conditions such as high temperature.

[0133] In addition, the adjacency matrix obtained in step S22 only provides structural information, while the new feature representation generated by the MGAM module in step S23 contains optimized semantic information (such as displacement gradient and spatial topological constraints), thereby obtaining speckle feature point information that has both structural information and semantic information.

[0134] S24. Construct a deep residual graph neural network model based on the GCN neural network, combining the initial speckle image residual and the high-order neighborhood speckle image residual, and improve the learning effect of the measurement model through a multi-input residual structure;

[0135] As a preferred implementation, step S24 specifically includes:

[0136] S241, the input feature matrix A (r i ,r j ) is transformed to obtain H( 0 ) as the initial feature; the formula is expressed as:

[0137] H( 0 )=σ(AW

[0138] Among them, A∈R i×j is the adjacency matrix obtained above, σ represents the nonlinear activation function, and W is the learnable parameter;

[0139] S242, the initial graph convolution layer is the key part of GCN, which is used to propagate and aggregate node features on the graph structure. It is proposed to add high-order neighborhood residuals, that is, to inherit the convolution output of the previous layer, to reduce the propagation speed of node features. However, since the initial features of some speckle nodes will be lost as the convolution depth increases, and the displacement calculation is based on the initial speckle image, in order to ensure that the final representation of each speckle node retains part of the initial features, the initial residual H( 0 ), the formula is as follows:

[0140]

[0141] Among them, H (k) represents the k-th layer speckle node feature matrix, A represents the adjacency matrix, W (k) represents the weight matrix of the kth layer, D represents the diagonal matrix, σ is the activation function; this patent uses the commonly used ReLU activation function; the introduced hyperparameter γ k , used to adjust the ratio between the initial speckle image residual and the high-order speckle neighborhood residual;

[0142] As described above, this patent uses a multi-input residual structure to perform multi-layer graph convolution operations, allowing the model to better utilize different speckle image input information to improve performance when in the deep layer, and to obtain more comprehensive speckle node feature information, which is more effective in processing high-temperature speckle displacement image datasets.

[0143] S3. Input the speckle images and the speckle displacement field before and after deformation into the speckle displacement measurement model for training to obtain the optimal measurement model configuration, and perform displacement measurement on the deformation of the speckle image to obtain the prediction result of the speckle displacement. It should be noted here that the speckle displacement field obtained in step S1 is used for the training of the displacement measurement model, and the prediction result of the speckle displacement obtained here is the final result based on the speckle images before and after deformation, and it is essentially still the speckle displacement field.

[0144] As a preferred embodiment, step S3 specifically includes:

[0145] S31. Obtain the sample data for the speckle displacement measurement model. The sample data is the original speckle image captured by a CCD industrial camera and the deformed speckle image after applying a load, and the speckle displacement field calculated in S1.

[0146] S32. Randomly shuffle the data and divide the obtained sample data into a training set, a validation set, and a test set according to the ratio of 6:2:2. The training set is used to train the speckle displacement measurement model to optimize the weight coefficients through backpropagation. The validation set is used to monitor the training process, adjust hyperparameters (such as the learning rate and network depth), and select the optimal model configuration. The test set is only used after the model training and parameter tuning are completed, and is input into the finally determined optimal model to evaluate its generalization ability and displacement measurement accuracy. In this embodiment, the trained speckle displacement measurement model should have high accuracy, good generalization ability, robust performance, and reasonable confidence evaluation.

[0147] S33. Input the speckle images before and after deformation, and the calculated full-field speckle displacement field of the speckle image into the speckle displacement measurement model to obtain the displacement field established based on the deformed image.

[0148] S4. Based on the prediction result of the speckle displacement obtained in step S3, calculate the strain of all points in the speckle image, and the set constitutes the surface strain field of the object to be measured.

[0149] As a preferred embodiment, step S4 specifically includes:

[0150] S41. Calculate the displacements of all displacement points based on the predicted results of the speckle displacements obtained in step S3. To better filter out the noise in the discrete displacement data and obtain a better smoothing effect, in this embodiment, the fitting function method is considered for denoising, and only the two-dimensional first-order polynomial is considered as the fitting function in this embodiment. The specific approach is to fit the u and v field displacements of the discrete data in the local sub-domain of the speckle displacement field established in step S3 with a two-dimensional first-order polynomial, which is expressed by the formula:

[0151]

[0152] where x and y represent the coordinates of each pixel point in the image, a 0 , a 1 , a 2 , b 0 , b 1 , b 2 are the coefficients of the fitting polynomial to be determined; u(x, y) and v(x, y) are the discrete displacement data, referring to the horizontal displacement and vertical displacement amounts of this point respectively;

[0153] Assume that the horizontal displacement field u consists of (2M + 1)×(2M + 1) data points, then arrange the two-dimensional discrete displacement data matrix composed of these data points into a one-dimensional column vector:

[0154]

[0155] where n = (2M + 1)×(2M + 1) - 1, and further the horizontal displacement field can be rewritten in matrix form:

[0156] u = Xa;

[0157] Thus, the least squares method can be used to solve the vector of undetermined coefficients:

[0158]

[0159] In the formula, (X T X) -1 X T is the pseudo-inverse matrix of X, and for the vertical displacement field v, it can be obtained in the same way:

[0160]

[0161] Furthermore, the coefficients of the fitting polynomial are obtained. The purpose of these coefficients is to smooth the discrete displacement data, filter out the noise, and provide continuous displacement field data for the subsequent strain calculation by fitting a two-dimensional first-order polynomial. The fitted polynomial can more accurately describe the distribution of the displacement field and provide a basis for calculating the Green strain components;

[0162] S42. Based on the displacements of all the obtained displacement points, calculate the Green strain components under finite deformation conditions. Through the Green strain components, the strain distribution of the material in each direction can be obtained, thereby constructing the surface strain field of the test piece.

[0163] The Green strain components under finite deformation conditions are as follows:

[0164]

[0165] E xx represents the normal strain in the x direction, that is, the elongation of the material per unit length in the x direction; this strain component is caused by the displacement gradient in the x direction;

[0166] E yy represents the normal strain in the y direction, that is, the elongation of the material per unit length in the y direction; this strain component is caused by the displacement gradient in the y direction;

[0167] E xy represents the shear strain, which describes the angular change between two perpendicular directions (x and y) in the x - y plane of the material due to shear deformation; the shear strain is due to the combined action of a displacement gradient in the y direction in the x direction and a displacement gradient in the x direction in the y direction.

[0168] The purpose of calculating the Green strain components is to describe the deformation of the material after being stressed. By calculating these strain components, the strain distribution of the material in each direction can be obtained, thereby constructing the surface strain field of the object under test. The Green strain components can reflect the local deformation characteristics of the material, providing important data support for subsequent material mechanics performance analysis and structural design.

[0169] So far, when measuring the strain of high - temperature materials in the future, only the original speckle image and the speckle image after deformation under the applied load need to be input into the deep neural network measurement model in the figure, and the surface deformation measurement results of the material component to be measured can be obtained. To sum up, the method proposed by the present invention accurately measures the deformation stress of high - temperature material components through the deep neural network measurement model based on the GCN neural network, realizing high - precision and real - time strain measurement under high - temperature environments.

[0170] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

[0171] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A strain measurement method based on GCN neural network, characterized in that: The specific steps include: S1. According to the actual working condition requirements, the speckle image and speckle displacement field of the test piece are collected and formed into a data set for subsequent training and testing of the measurement model; S2. Build a speckle displacement measurement model based on the GCN neural network; S3, inputting the data set into the speckle displacement measurement model for training, obtaining the optimal measurement model configuration, performing displacement measurement on the deformation of the speckle image, and obtaining the prediction result of the speckle displacement; S4. Based on the prediction result of the speckle displacement obtained in step S3, the strains of all points in the speckle image are calculated, and the strains are collectively constituted into the surface strain field of the measured object.

2. The strain measurement method based on GCN neural network according to claim 1 is characterized in that: Step S1 specifically includes: S11. For the test pieces made of different materials, high temperature resistant speckles are prepared on the surface of the test pieces by combining surface grinding method and spraying method; S12. Use a digital image correlation method to measure the processed test piece, extract the speckle image at each time, and measure the displacement field to establish a data set based on the speckle image and the speckle displacement field.

3. The strain measurement method based on GCN neural network according to claim 2 is characterized in that: Step S12 specifically includes: S121, using a material tensile testing machine to perform a material tensile test on the treated test piece, so as to extract the displacement field of the test piece at different times under the working condition to establish a data set; S122, start the calibration CCD binocular camera and light source, align them with the center of the sample, adjust the level so that the sample surface is parallel to the CCD binocular camera, use the calibration plate for calibration, and eliminate the system error; start the measurement program according to the main deformation area of ​​the tensile sample; S123, conduct an experiment, start the material tensile testing machine, set the experimental tensile rate, and after the material tensile testing machine starts working, start the CCD binocular camera to take pictures; S124, obtaining the tensile force and displacement through a material tensile testing machine, and obtaining the original speckle image and the deformed speckle image after the load is applied through a CCD binocular camera; S125, dividing the speckle image at a certain moment into multiple sub-regions, matching the original speckle image with the deformed speckle image in each sub-region, and obtaining the displacement value of the speckle sub-region under the applied load; calculating the displacement values ​​of all subset regions, and obtaining the speckle displacement field at a certain moment.

4. The strain measurement method based on GCN neural network according to claim 1 is characterized in that: Step S2 specifically includes: S21. Design and improve the non-local mean filtering algorithm to reduce the noise of the speckle image; S22, for the speckle image after noise reduction processing, select a rectangular region of interest ROI, and represent the speckle feature points as a graph structure to represent the association between the speckle feature points; S23, obtaining the coordinates of the speckle feature points and the adjacency matrix of the speckle feature points from the graph structure, and introducing a multi-scale graph attention module MGAM to integrate the speckle feature point information at different levels; S24. Construct a deep residual graph neural network model based on the GCN neural network, combine the initial speckle image residual and the high-order neighborhood speckle image residual, and improve the learning effect of the measurement model through a multi-input residual structure.

5. The strain measurement method based on GCN neural network according to claim 4 is characterized in that: Step S21 specifically includes: S211, sliding a search box of a fixed size on the global speckle image by a sliding window method, performing non-local mean filtering on the speckle image in each search box; firstly fitting a low-order polynomial by the least square method to smooth the speckle image in the window, taking the polynomial value at the center coordinate in the window as the new gray value of the center pixel of the smoothed speckle image, comparing the similarity of the polynomial values ​​at the center points of different windows, and determining the weight of the search box in the global image reconstruction; S212, for each search box, let I = {I(x,y)|x,y∈S} be the speckle image containing noise in the search box, S represents the pixel index set, x is the center pixel, i.e., the target pixel to be denoised, y is the search box pixel, i.e., any pixel except the center pixel; the denoised image is obtained by weighted averaging of all pixels in I, and the formula is expressed as: O(x)=∑ y∈I μ·ω(x,y)I(y); Among them, μ is the noise reduction coefficient; ω(x, y) is the similarity weight coefficient; I(y) represents the image block centered on the search box pixel y; O(x) represents the filtered image block centered on the central pixel x; the noise reduction coefficient μ is dynamically adjusted according to the local density ρ of the speckle distribution to achieve an adaptive balance between denoising intensity and detail retention; the local density ρ of speckle is defined as the proportion of speckle pixels in the similarity box, and the calculation formula is as follows: Where N is the number of speckle pixels in the search box; W is the side length of the similarity box; the similarity box adopts a multi-scale sliding window strategy; for large-scale speckle structures, a large side length is selected to maintain coherence; for small-scale details, a small side length is switched to retain fine features; S213, calculate the similarity weight coefficient ω(x, y); quantify the similarity between image blocks based on the Euclidean distance, and additionally introduce the gradient information of the image blocks to improve the Euclidean distance; define the gradient feature vectors of the image blocks V(x) and V(y) as and Then the Euclidean distance d(x,y) between the improved search box pixel y and the center pixel x is: Among them, V(x) and V(y) are the grayscale values ​​of the image blocks centered on x and y respectively. and is the gradient feature vector of the image block; α is the gradient weight coefficient; The gradient weight coefficient α is adaptively adjusted according to the complexity of the local structure of the speckle, and the formula is expressed as: in, is the average value of the gradient amplitude of all pixels in the current similarity box, It represents the maximum value of the pixel gradient amplitude in the search box, and γ is an empirical constant. When the local structure is complex, that is, the gradient amplitude is high, α is increased to strengthen the weight of the gradient information, and vice versa. The final calculation formula of similarity weight coefficient ω(x,y) is: Among them, h is the filter coefficient and M(x) is the normalization coefficient.

6. The strain measurement method based on GCN neural network according to claim 4 is characterized in that: Step S22 specifically includes: S221, selecting a rectangular region of interest ROI, performing feature detection on the ROI region, and forming a set of feature points of interest at the initial moment with coordinate information; then performing full-image feature extraction on the speckle image at each moment in turn, obtaining a full-image feature point set at each moment, and then integrating the full-image feature point sets at each moment to obtain a sequence feature point set; S222, matching the feature point set of interest at the initial moment with the sequence feature point set; obtaining the coordinates of the speckle feature points in the ROI region on the image at different moments, and obtaining the coordinates of the speckle feature points on the image through speckle feature point extraction and feature matching. and displacement Therefore, the speckle feature is represented as a graph structure, where nodes represent different speckle feature points and edges represent the correlation between different speckle feature points; S223, for the correlation between different speckle feature points, the correlation between the speckle feature points is calculated by the Pearson correlation coefficient, and the adjacency matrix A (r i ,r j ); S224. For the obtained adjacency matrix, only the upper triangular value of each adjacency matrix is ​​retained, and the adjacency matrix is ​​vectorized into the feature vector required for GCN network training, and the flattened feature vector is generated as the primary feature representation.

7. The strain measurement method based on GCN neural network according to claim 5 is characterized in that: Step S23 specifically includes: S231. Design a multi-scale graph attention module MGAM. Based on the graph structure of speckle feature points, firstly, the features of all nodes in the graph are averaged according to the channel dimension through global pooling guided by channel attention. Then, the importance of different feature channels is analyzed through a learnable channel attention mechanism to obtain global features. Then, the local features of each speckle node are input into the graph attention module GAM to generate local speckle features with the same dimension as the global features. Then, the global features are fused with the local speckle features. S232, while applying the attention mechanism on each node through GAM to learn the relationship between nodes and aggregate information from neighboring nodes; S233. In order to make the attention coefficients between different nodes easier to compare, the obtained attention coefficients are normalized to generate a new feature representation for each speckle node.

8. The strain measurement method based on GCN neural network according to claim 5 is characterized in that: Step S24 specifically includes: S241, the input feature matrix A (r i ,r j ) is transformed to obtain H( 0 ) as the initial feature; S242. Add high-order neighborhood residuals to the basic GCN. Each layer inherits the convolution output of the previous layer to reduce the propagation speed of node features. At the same time, add initial residuals to ensure that each speckle node can eventually retain a part of the initial features. And by introducing hyperparameters to adjust the ratio between the initial residual and the high-order neighborhood residual, the processing performance of the displacement measurement model when facing different speckle images is improved.

9. The strain measurement method based on GCN neural network according to claim 1, characterized in that: Step S3 specifically includes: S31, acquiring sample data for a speckle displacement measurement model, the sample data being an original speckle image captured by a CCD industrial camera, a deformed speckle image after a load is applied, and a speckle displacement field calculated in S1; S32. Randomly shuffle the data and divide the obtained sample data into a training set, a validation set and a test set in a ratio of 6:2:2; the training set is used to train the speckle displacement measurement model and optimize the weight coefficient by back propagation; the validation set is used to monitor the training process, adjust the hyperparameters and select the optimal model configuration; the test set is only used after the model training and parameter adjustment are completed, and is input into the final optimal model to evaluate its generalization ability and displacement measurement accuracy; S33, inputting the speckle images before and after the deformation and the calculated speckle displacement field into a speckle displacement measurement model to obtain a prediction result of the speckle displacement.

10. The strain measurement method based on GCN neural network according to claim 1, characterized in that: Step S4 specifically includes: S41, calculating the displacements of all displacement points based on the prediction results of the speckle displacement obtained in step S3; S42. Based on the displacements of all displacement points obtained, the Green strain component is calculated under finite deformation conditions as follows. Through the Green strain component, the strain distribution of the material in various directions can be obtained, thereby constructing the surface strain field of the test piece.

Images