A method for inverting strain rate effect of brittle materials based on data processing

Abstract

The present invention discloses a method for inverting the strain rate effect of brittle materials based on data processing, and belongs to the field of predicting material parameters based on electrical digital data processing. The method dynamically loads the brittle material through a Hopkinson bar test device, and measures the material surface strain field distribution in real time in combination with digital image correlation technology to generate a strain field cloud map sequence with time series information; at the same time, a finite element model is established, the parameters in the strain rate strengthening model are changed, and the strain field cloud map data under different parameter combinations are obtained; these cloud map data are trained through a 3D convolutional neural network, and a mapping relationship between the cloud map and the parameters is established, so that the strain rate strengthening model parameters under the test conditions can be inverted according to the experimental data. The inverted parameters are back-substituted into the finite element analysis, the macroscopic dynamic constitutive relationship of the specimen is calculated, and a closed-loop verification is performed with the experimental results. The method can quickly and accurately obtain the constitutive model parameters of brittle materials under complex strain rates.

Description

Technical Field

[0001] The invention relates to the field of electrical digital data processing, in particular to a method for inverting strain rate effect of brittle materials based on data processing. Background Art

[0002] Brittle materials (a general term for materials whose compressive strength is sensitive to hydrostatic pressure, such as concrete, rock, and polymers) are widely used in various defense and civil engineering projects, and the study of their mechanical properties has always been a hot topic in the engineering community. Under strong dynamic loads, the mechanical behavior of such materials is significantly different from that under quasi-static loading, which is mainly reflected in the influence of loading strain rate on the constitutive relationship and strength of the material. This behavior is called strain rate effect. Currently, the split Hopkinson bar technology is a new technology for the study of such materials under medium and high strain rates (10 2 -10 4 s -1 ) is the main means of dynamic mechanical properties testing under .

[0003] In the early days, the discrete results obtained from such experimental tests were directly used to fit mathematical formulas and then develop dynamic constitutive relations for such materials. However, as the research progressed, people gradually realized that the significant dynamic enhancement exhibited by such materials at high strain rates may represent a change in the stress state in the specimen. The apparent strain rate effect observed in the experiment cannot be entirely attributed to the true material properties. The inertial effects (also called structural effects) caused by material properties (such as density, Poisson's ratio, friction angle, expansion angle, etc.), size, end face friction, etc., also made significant contributions. If the experimental data is not corrected and taken into account in the strain rate effect model, the dynamic strength of such materials will be exaggerated, leading to dangerous structural design.

[0004] Therefore, how to identify the true strain rate effect in test data has become an important topic in this field. Some scholars choose to determine the true power amplification value by comparing the input and output power amplification coefficient values ​​and iteratively adjusting them, but the calculation efficiency is too low and the convergence is difficult, and sometimes it is impossible to obtain an accurate value.

[0005] At present, digital data processing technology has been applied in many aspects. For example, the Chinese invention patent with application publication number CN118536072A discloses a bridge crack detection and identification method and system based on unsupervised learning and big data, which uses a clustering algorithm to perform cluster analysis on the preprocessed structural data, and uses a convolutional neural network to locate and segment the crack area of ​​the bridge for the image data in the preprocessed structural data; and extracts the crack features of the segmented crack area to build a health assessment model through a machine learning algorithm, thereby realizing automatic detection of bridge cracks. Summary of the invention

[0006] The purpose of the present invention is to propose a method for inverting the strain rate effect of brittle materials based on data processing, which improves the efficiency and accuracy of determining the strain rate effect of brittle materials.

[0007] The technical solution adopted by the present invention is:

[0008] A method for inverting strain rate effect of brittle materials based on data processing comprises the following steps:

[0009] S1. Prepare several Hopkinson bar test specimens of brittle materials, and spray speckles around the specimens; arrange the experimental device, set up a binocular high-speed camera, and adjust its position so that the binocular high-speed camera can capture clear and complete speckle images around the specimens;

[0010] S2. Conduct dynamic SHPB test, by changing the bullet length, bullet diameter and incident speed in the SHPB test, conduct dynamic impact test of brittle material Hopkinson bar test specimens at different strain rates, and use a high-speed camera to shoot the speckle deformation process of the specimen under impact in the test;

[0011] S3, analyzing the photos taken by the high-speed camera in step S2, calculating the strain field cloud map of the speckle area of ​​the specimen; arranging the strain field cloud map of the speckle area of ​​the specimen in chronological order, extracting the strain field information, and assembling it into a cloud map sequence containing time information;

[0012] S4. Carry out finite element simulation under different strain rate strengthening model parameters, simulate the impact test under different bullet lengths, bullet diameters, and incident velocities in step S2, and obtain the yield surface of the brittle material after dynamic improvement; substitute the yield surface parameters of the brittle material into the finite element analysis, and change the parameters. α , λ, δ, γtr for numerical analysis; according to the frame rate of the high-speed camera, the circumferential strain field cloud map of the specimen with the same time dimension is output, and the S3 step processing method is adopted to assemble a three-dimensional cloud map sequence containing time information; the three-dimensional cloud map sequence is divided into a training set, a validation set, and a test set;

[0013] S5, normalize the pixel values ​​in the three-dimensional cloud image sequence in step S4, as shown in formula (3-1), to ensure that the data distribution is suitable for neural network training;

[0014]

[0015] Where x′ is the strain value of a pixel in the 3D cloud sequence, x′ min , x′ max are the minimum and maximum values ​​of strain in the three-dimensional cloud sequence, respectively;

[0016] S6, building a 3D CNN convolutional neural network, which includes several 3D convolutional layers and pooling layers, and sequentially processes the three-dimensional cloud image sequence processed in S5 using the 3D convolutional layers and pooling layers to obtain a high-dimensional feature map;

[0017] S7, establish a fully connected layer, after passing through the 3D convolution layer and the pooling layer, output a high-dimensional feature map, use the fully connected layer to flatten the 3D features into a one-dimensional vector, and then perform a linear transformation;

[0018] S8, input the 3D cloud image sequence in step S4 into the network model built for learning, and establish parameters α, λ, δ, γ tr The mapping relationship with the S4 step three-dimensional cloud image sequence is shown in formula (6-1):

[0019] [α,λ,δ,γ tr ] = f 3DCNN (D1,H1,W1,ε) (6-1)

[0020] Among them, α, λ, δ, γ tr is the parameter to be predicted, f 3DCNN The functional relationship established for 3D CNN, D1, H1, W1, and ε are the depth, height, width, and strain values ​​of the cloud image, respectively;

[0021] The input data is forward propagated through each layer of the network, and the output feature map of each layer is calculated until the final prediction result is obtained.

[0022] The beneficial technical effects of the present invention are as follows:

[0023] The present invention proposes a method for predicting the strain rate effect of brittle materials based on data processing. The method combines digital image correlation technology (DIC), finite element analysis (FEA) and 3D convolutional neural network (3DCNN) for parameter inversion of strain rate hardening model of brittle materials under dynamic loading conditions. The surface strain field is measured in real time by DIC, and the experimental data is combined with the cloud map sequence generated by finite element simulation. 3DCNN is used to perform parameter inversion on different strain fields to achieve automatic parameter estimation. Finally, the accuracy and efficiency of the model are improved through closed-loop verification of experimental and simulation results. This method integrates physical modeling and deep learning technology, providing a fast and accurate solution for the study of dynamic behavior of materials.

[0024] Specifically, the present invention has the following advantages:

[0025] (1) By establishing a finite element model of brittle materials and changing the relevant parameters in the strain rate hardening model, a large number of strain field cloud map sequences under different parameter combinations are generated; this provides rich simulation data for subsequent neural network training, and by continuously adjusting the parameters, the dynamic mechanical behavior of the real material can be better approximated.

[0026] (2) Using 3D CNN to train the cloud image sequence generated by finite element method, the corresponding strain rate hardening model parameters can be inverted based on the input strain field cloud image sequence. This deep learning-based inversion technology effectively improves the estimation ability of complex parameters and reduces human experience intervention.

[0027] (3) The framework combines finite element analysis with deep learning to form a complete process from experimental data acquisition, parameter inversion to model verification, which improves the efficiency and accuracy of constructing dynamic constitutive models of brittle materials.

[0028] The present invention not only improves the efficiency and accuracy of determining the strain rate effect of brittle materials, but also provides strong support for material science research and engineering applications. The method of the present invention has important application prospects in improving the level of material design and analysis and optimizing the performance of engineering structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0030] Figure 1 A flow chart of a method for inverting the strain rate effect of brittle materials based on data processing provided by an embodiment of the present invention;

[0031] Figure 2 Schematic diagram of a Hopkinson bar (SHPB) specimen and its speckle spraying involved in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of a Hopkinson bar (SHPB) experimental device involved in an embodiment of the present invention;

[0033] Figure 4 Schematic diagram of the parameters of a Hopkinson bar (SHPB) bullet involved in an embodiment of the present invention;

[0034] Figure 5 It is a schematic diagram of assembling a sequence of cloud images after DIC recognition of speckle photos in an embodiment of the present invention;

[0035] Figure 6 It is a schematic diagram of a finite element model of a Hopkinson bar (SHPB) finite element simulation involved in an embodiment of the present invention;

[0036] Figure 7 Schematic diagram of the strain cloud diagram of the test piece after impact in an embodiment of the present invention;

[0037] Figure 8 Schematic diagram of the convolution and other processing flow of the 3D CNN convolutional neural network model in an embodiment of the present invention;

[0038] Fig. 9 It is a schematic diagram of the dynamic amplification factor-strain rate curve in the embodiment of the present invention. Changing the parameters in the formula will obtain different curves to serve as the input of the constitutive strengthening model in the finite element method;

[0039] Fig.10 It is a schematic diagram comparing the dynamic constitutive results of numerical simulation and the dynamic constitutive results of experiments according to an embodiment of the present invention;

[0040] Fig.11 It is a schematic diagram of the workflow of inverting the strain rate effect of brittle materials based on data processing in an embodiment of the present invention.

[0041] In the figure: 1. SHPB bullet; 2. Signal acquisition instrument; 3. Incident rod; 4. First strain gauge; 5. Hopkinson bar specimen; 6. Binocular high-speed camera; 7. Signal line; 8. Second strain gauge; 9. Transmission rod. DETAILED DESCRIPTION

[0042] The present invention discloses a method for inverting the strain rate effect of brittle materials based on data processing. On the one hand, the method performs a dynamic compression test on the brittle material through a Hopkinson bar experiment, and performs real-time measurement based on an optical measurement approach; on the other hand, a 3D CNN convolutional neural network is established based on a numerical analysis method and machine learning means to identify the cloud map sequence of the weekly strain field of the test piece during the test and establish a mapping relationship between the cloud map sequence and the strain rate strengthening model parameters. Given a test cloud map sequence, the strain rate strengthening model parameters that can be directly input into a numerical simulation platform are obtained. The present invention comprehensively applies test measurement, mechanical analysis, machine learning and inversion theory, and can quickly and accurately predict the strain rate effect of brittle materials.

[0043] like Figure 1 As shown, a method for inverting the strain rate effect of brittle materials based on data processing specifically includes the following steps:

[0044] S1, such as Figure 2 As shown in FIG. 1 , several brittle material Hopkinson bar (SHPB) specimens are prepared, and black digital speckle patterns are sprayed around the specimens.

[0045] like Figure 3As shown, an experimental device is arranged, which includes an incident rod 3, a transmission rod 9 and a SHPB bullet 1, and the Hopkinson rod specimen is arranged between the incident rod 3 and the transmission rod 9. A first strain gauge 4 is arranged on the incident rod 3, and a second strain gauge 8 is arranged on the transmission rod 9. The first strain gauge 4 and the second strain gauge 8 are both connected to the signal acquisition instrument 2 through a signal line 7. A binocular high-speed camera is set up on one side of the incident rod 3 and the transmission rod 9, and lights are arranged and the position of the binocular high-speed camera is adjusted until the camera can capture a clear and complete speckle image of the specimen.

[0046] S2. Perform a dynamic SHPB test. Figure 4 As shown in the figure, by changing the SHPB test bullet length, bullet diameter and incident speed, dynamic impact tests of brittle materials at different strain rates are carried out. The high-speed camera captures the speckle deformation process of the specimen under the impact in the experiment. According to the strain data of the incident rod and the transmission rod recorded by the SHPB experimental system, the average strain rate of the specimen in this experiment is calculated by formulas (1-1)-(1-3) The mean strain ε(t) and the effective mean stress σ(t).

[0047]

[0048] Among them, ls and As are the length and cross-sectional area of ​​the specimen, respectively, and c, E, and A are the elastic wave velocity, elastic modulus, and cross-sectional area of ​​the Hopkinson bar, respectively. i , ε r , ε t are the incident wave, reflected wave, and transmitted wave measured respectively; v is the particle velocity, and the subscripts Ⅰ and Ⅱ represent the incident and transmission ends of the specimen respectively, that is, v Ⅰ represents the particle velocity at the incident end of the specimen, v Ⅱ Represents the particle velocity at the transmission end of the specimen.

[0049] S3. Import the photos taken by the high-speed camera in step S2 into the digital image correlation technology software (DIC) to calculate the strain field cloud map of the speckle area of ​​the specimen. Arrange the strain field cloud map in chronological order, extract the strain field information through programming software, and assemble it into a cloud map sequence containing time information, such as Figure 5 shown.

[0050] S4. Carry out Figure 6 The finite element simulation under different strain rate hardening model parameters is shown, simulating the impact test under the bullet parameters (bullet length, bullet diameter, and incident speed) in step S2. The yield surface Δσ of the brittle material after dynamic enhancement e , controlled by formula (2-1).

[0051]

[0052] Among them, Δσ e The yield surface after dynamic improvement of the material, is the strain rate, σ is the flow stress, dc is the intermediate quantity, β is the friction angle, fc is the static compressive strength. γ is the dynamic enhancement coefficient, which can be calculated by formula (2-2):

[0053]

[0054] In the formula, is the strain rate corresponding to quasi-static loading. According to common practice, is the conversion strain rate; α , λ, δ are the control parameters of the function graph. Considering the continuity of the two-segment function given by equation (2-2) at the conversion strain rate, the parameters α, λ can be calculated by the following formula:

[0055]

[0056] Among them, γ tr is the dynamic strength enhancement factor corresponding to the converted strain rate.

[0057] The above brittle material constitutive model (such as Fig. 9 The finite element analysis was performed by changing the parameters α, λ, δ, and γtr. According to the frame rate of the high-speed camera, the circumferential strain field cloud map of the specimen with the same time dimension was output, and the S3 step processing method was used to assemble it into a three-dimensional cloud map sequence containing time information. According to the needs of the model, the three-dimensional cloud map sequence was divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5.

[0058] S5. Normalize the pixel values ​​in the cloud image in step S4, as shown in formula (3-1), to ensure that the data distribution is suitable for neural network training.

[0059]

[0060] Where x′ is the strain value of a pixel in the cloud map, x′ min , x′ max are the minimum and maximum values ​​of strain in the cloud diagram, respectively.

[0061] The cloud map data is then formatted into a 3D tensor (size: depth × height × width × channel), and each cloud map is stored in pairs with the corresponding parameter label.

[0062] S6, such as Figure 8As shown, a 3D CNN convolutional neural network model is built. The neural network model includes a 3D convolution layer, a pooling layer, and a connection layer. Several 3D convolution layers and pooling layers are set. A three-dimensional convolution kernel is used to slide on the three-dimensional input to extract spatiotemporal features. The dimension of the convolution kernel is defined as (D, H, W), where D is the time dimension, H is the horizontal coordinate of a certain point in the cloud map, and W is the vertical coordinate of a certain point in the cloud map. The convolution kernel takes the form of formula (4-1).

[0063]

[0064] Where I(x,y,z) is the value of the input feature map at position (x,y,z), K(i,j,k) is the weight of the convolution kernel at position (i,j,k), d, h, and w are the depth, height, and width of the convolution kernel, respectively;

[0065] The 3D pooling layer performs spatial and temporal pooling on the input to reduce the dimension of the data. The pooling method uses maximum pooling. The maximum value is taken for each sub-region in the pooling window to retain the most significant features of the cloud image. The kernel size is (2×2×2), that is, 2×2×2 pooling is performed in the time, width and height dimensions. The maximum pooling layer output value O m (x,y,z) is determined by formula (4-2).

[0066]

[0067] Among them, I(x,y,z) is the value of the input feature map at the (x,y,z) position, and i,j,k is the coordinates of the convolution kernel at the (i,j,k) position.

[0068] Leaky ReLU is used as the activation function to introduce nonlinearity, as shown in formula (4-3).

[0069]

[0070] Here α is a very small positive number, and for this type of regression problem, α=0.01 is taken.

[0071] S7. Establish a fully connected layer. First, the input cloud map data is normalized. Next, the preliminary spatial and temporal features are extracted through the first 3D convolution to obtain the convolved feature map. Then, these feature maps are pooled to reduce the size and retain the main information to obtain the pooled feature map. Repeat several rounds of 3D convolution and pooling operations to further extract high-level features, and output the corresponding convolution and pooling results each time. Finally, after multiple 3D convolution and pooling processes, a high-dimensional feature map is obtained. These high-dimensional feature maps are flattened into one-dimensional vectors and linearly transformed through several fully connected layers. In the last layer, the fully connected layer uses the Softmax function to convert the output of the network into a probability distribution to complete the regression task. The fully connected process is controlled by formula (5-1), and the Softmax function is shown in formula (3-5).

[0072] O=W·X+b (5-1)

[0073]

[0074] Where W is the weight matrix, X is the flattened input feature vector, b is the bias vector, O is the output matrix, xi is the output of the i-th neuron in the output layer, and N is the total number of categories.

[0075] S8. Input the cloud map of the training set in step S4 into the constructed 3D CNN for learning, and establish the mapping relationship between the parameters α, λ, δ, γtr and the cloud map sequence in step S4, as shown in formula (6-1).

[0076] [α,λ,δ,γ tr ] = f 3DCNN (D1,H1,W1,ε) (6-1)

[0077] Among them, α, λ, δ, γtr are the parameters to be predicted, f 3DCNN The functional relationship established for 3D CNN, D1, H1, W1, and ε are the depth, height, width, and strain values ​​of the cloud image, respectively.

[0078] The input data is forward propagated through each layer of the network, and the output feature map of each layer is calculated until the final prediction result is obtained, such as Figure 8 The mean square error is used to evaluate the loss function to measure the difference between the model prediction value and the true value, as shown in formula (6-2), to evaluate the model performance.

[0079]

[0080] Among them, MSE is the mean square error, which represents the average value of the squared errors of all samples; N is the total number of samples; yi is the true value, which represents the true label or target value of the i-th sample; is the predicted value, which represents the predicted output of the model for the i-th sample.

[0081] Based on the results of MSE calculation, the network weights are updated through the back propagation algorithm to minimize the loss. The gradient descent method is used to adjust the network parameters.

[0082] S9. Repeat the forward propagation and back propagation process in the training set until the predetermined number of training rounds is reached or the model converges.

[0083] S10. Use the validation set and test set to evaluate the model, calculate the accuracy, recall and other indicators to ensure the generalization ability of the model. When the MSE calculated by the model is less than 1‰, the neural network training is completed.

[0084] S11. Input the cloud sequence obtained in step S3 into the machine trained in step S10. The machine recognizes the cloud sequence according to formula (6-1) and predicts the parameters corresponding to the cloud sequence. α , λ, δ, γtr.

[0085] S12, Substitute the parameters obtained in step S11 into the finite element software for numerical simulation, and output Fig.10 The macroscopic dynamic constitutive model of the specimen is shown and compared with the dynamic constitutive model obtained in the test. The mean absolute error (MAPE) is used to evaluate the accuracy of the predicted parameters.

[0086]

[0087] Where N is the total number of samples output by numerical simulation, is the ith value of the dynamic constitutive model obtained from the experiment, is the ith value of the dynamic constitutive model calculated by numerical simulation. When MAPE converges to 2‰, it is considered that the predicted parameters are sufficient to characterize the true dynamic enhancement coefficient. Otherwise, adjust the size of the convolution kernel and the pooling window size in step S6, and repeat steps S6-S11 until MAPE converges to within 2‰.

[0088] Parts not described in the above implementation manner can be implemented by adopting or drawing on existing technologies.

[0089] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above implementation modes. Any changes, modifications, additions or substitutions made by those skilled in the art within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for inverting strain rate effect of brittle materials based on data processing, characterized in that: The following steps are involved: S1. Prepare a number of brittle material Hopkinson bar test specimens, and measure the spraying speckle around the brittle material Hopkinson bar test specimens; Arrange the experimental device, set up the binocular high-speed camera, and adjust its position so that the binocular high-speed camera can capture clear and complete speckle images of the brittle material Hopkinson bar test specimen; S2. Conduct dynamic SHPB test, by changing the bullet length, bullet diameter and incident speed in the SHPB test, conduct dynamic impact test of brittle material Hopkinson bar test specimens at different strain rates, and use a high-speed camera to shoot the speckle deformation process of the specimen under impact in the test; S3, analyzing the photos taken by the high-speed camera in step S2, and calculating the strain field cloud map of the speckle area of ​​the specimen; Arrange the strain field cloud images of the speckle area of ​​the specimen in chronological order, extract the strain field information, and assemble them into a cloud image sequence containing time information; S4. Carry out finite element simulation under different strain rate strengthening model parameters, simulate the impact test under different bullet lengths, bullet diameters, and incident velocities in step S2, and obtain the yield surface of the brittle material after dynamic improvement; substitute into the finite element analysis, change the parameters α, λ, δ, γtr for numerical analysis; output the circumferential strain field cloud map of the specimen with the same time dimension according to the frame rate of the high-speed camera, and use the processing method in step S3 to assemble it into a three-dimensional cloud map sequence containing time information; divide the three-dimensional cloud map sequence into a training set, a validation set, and a test set; S5, normalize the pixel values ​​in the three-dimensional cloud image sequence in step S4, as shown in formula (3-1), to ensure that the data distribution is suitable for neural network training; Where x′ is the strain value of a pixel in the 3D cloud sequence, x′ min , x′ max are the minimum and maximum values ​​of strain in the three-dimensional cloud sequence respectively; S6. Building a 3D CNN convolutional neural network model, the 3D CNN convolutional neural network model includes several 3D convolutional layers and pooling layers, and sequentially processes the three-dimensional cloud image sequence processed in S5 using the 3D convolutional layers and pooling layers to obtain a high-dimensional feature map; S7, establish a fully connected layer, after passing through the 3D convolution layer and the pooling layer, output a high-dimensional feature map, use the fully connected layer to flatten the 3D features into a one-dimensional vector, and then perform a linear transformation; S8, input the 3D cloud image sequence in step S4 into the 3D CNN convolutional neural network model for learning, and establish parameters α, λ, δ, γ tr The mapping relationship with the S4 step three-dimensional cloud image sequence is shown in formula (6-1): [α,λ,δ,γ tr ] = f 3DCNN (D1,H1,W1,ε) (6-1)where, α, λ, δ, γ tr is the parameter to be predicted, f 3DCNN The functional relationship established for the 3D CNN convolutional neural network model, D1, H1, W1, and ε are the depth, height, width, and strain values ​​of the cloud image, respectively; The input data is forward propagated through each layer of the network, and the output feature map of each layer is calculated until the final prediction result; In step S4: The yield surface Δσ of the brittle material after dynamic enhancement e , controlled by formula (2-1): Among them, Δσ e Dynamically improve the yield surface of brittle materials, is the strain rate, σ is the flow stress, d c is the intermediate quantity, β is the friction angle, fc is the static compressive strength, and γ is the dynamic enhancement coefficient, which is calculated by formula (2-2): In the formula, is the strain rate corresponding to quasi-static loading; is the conversion strain rate; α , λ, δ are the control parameters of the function graph; parameters α, λ are calculated by the following formula: Among them, γ tr is the dynamic strength enhancement factor corresponding to the converted strain rate.

2. The method for inverting the strain rate effect of brittle materials based on data processing according to claim 1, characterized in that: In step S2: the average strain rate of the Hopkinson bar test specimen of brittle material is calculated by formula (1-1)-(1-3): Mean strain ε(t), effective mean stress σ(t); Where ls and As are the length and cross-sectional area of ​​the specimen, c, E, and A are the elastic wave velocity, elastic modulus, and cross-sectional area of ​​the Hopkinson bar, respectively; εi, ε r , εt are the measured incident wave, reflected wave and transmitted wave respectively; v is the particle velocity, and the subscripts Ⅰ and Ⅱ represent the incident and transmission ends of the specimen respectively.

3. The method for inverting the strain rate effect of brittle materials based on data processing according to claim 1, characterized in that: In step S6: Use a three-dimensional convolution kernel to slide on the three-dimensional input to extract spatiotemporal features. The dimension of the convolution kernel is defined as (D, H, W), where D is the time dimension, H is the horizontal coordinate of a certain point in the cloud map, and W is the vertical coordinate of a certain point in the cloud map; the convolution kernel takes the form of formula (4-1): Where I(x,y,z) is the value of the input feature map at position (x,y,z), K(i,j,k) is the weight of the convolution kernel at position (i,j,k), d, h, and w are the depth, height, and width of the convolution kernel, respectively; The pooling layer performs spatial and temporal pooling on the input. The pooling method uses maximum pooling, taking the maximum value of each sub-region in the pooling window to retain the most significant features of the cloud image. The kernel size is 2×2×2, that is, 2×2×2 pooling is performed in the time, width and height dimensions; the maximum pooling layer output value O m (x, y, z) is determined by formula (4-2): Among them, I(x,y,z) is the value of the input feature map at the position (x,y,z), and i,j,k are the coordinates of the convolution kernel at the position (i,j,k); Leaky ReLU is used as the activation function to introduce nonlinearity, as shown in formula (4-3):

4. The method for inverting strain rate effect of brittle materials based on data processing according to claim 3, characterized in that: In step S7: The last layer of the fully connected layer uses the Softmax function to convert the output of the network into a probability distribution; the fully connected process is controlled by formula (5-1), and the Softmax function is shown in formula (5-2): O=W e ·X+b (5-1) Where W e is the weight matrix, X is the flattened input feature vector, b is the bias vector, O is the output matrix, xi is the output of the i-th neuron in the output layer, and N is the total number of categories.

5. The method for inverting strain rate effect of brittle materials based on data processing according to claim 4, characterized in that: In step S8: The mean square error is used to evaluate the loss function to measure the difference between the predicted value and the true value of the 3D CNN convolutional neural network model, as shown in formula (6-2), to evaluate the model performance: Among them, MSE is the mean square error, which represents the average value of the squared errors of all samples; n is the total number of samples; yq is the true value, which represents the true label or target value of the qth sample; is the predicted value, which represents the predicted output of the model for the qth sample.

6. The method for inverting strain rate effect of brittle materials based on data processing according to claim 5, characterized in that: The following steps are also included: S9, repeating the forward propagation and back propagation process in the training set of the three-dimensional cloud image sequence until a predetermined number of training rounds is reached or the model converges; S10. Use the validation set and test set of the 3D cloud image sequence to evaluate the 3D CNN convolutional neural network model, calculate the accuracy and recall rate indicators, and ensure the generalization ability of the model; when the MSE calculated by the model is less than 1‰, the machine training is completed; S11, input the cloud sequence obtained in step S3 into the machine trained in step S10, and the machine recognizes the cloud sequence according to formula (6-1) and predicts the parameters α, λ, δ, γtr corresponding to the cloud sequence; S12, substituting the parameters obtained in step S11 into the finite element software for numerical simulation, outputting the macroscopic dynamic constitutive model of the specimen, comparing it with the dynamic constitutive model obtained in the test, and using the mean absolute error to evaluate the accuracy of the predicted parameters; Where R is the total number of samples output by numerical simulation, is the rth value of the dynamic constitutive model obtained from the experiment, is the rth value of the dynamic constitutive model calculated by numerical simulation; when MAPE converges to 2‰, it is considered that the predicted parameters are sufficient to characterize the true dynamic enhancement coefficient; otherwise, adjust the size of the convolution kernel and the pooling window size in step S6, and repeat steps S6-S11 until MAPE converges to within 2‰.

Images