A neural network-based method for predicting full-field strain inside solid propellant

Abstract

The application provides a solid propellant internal full-field strain prediction method based on a neural network, comprising the following steps: 1, measuring the internal microstructure feature change rule of the solid propellant; 2, acquiring the solid propellant internal full-field strain by adopting a digital volume correlation method; 3, constructing a multi-modal convolutional neural network model; 4, constructing a sample data set; 5, training the multi-modal convolutional neural network model by using the data set, and obtaining the multi-modal convolutional neural network model under the optimal condition. The solid propellant internal full-field strain can be predicted by inputting the initial microstructure inside the solid propellant and the external loading condition into the model, and the problem that the solid propellant internal full-field strain cannot be quickly and accurately predicted in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of solid propellants, and particularly relates to a solid propellant internal full-field strain prediction method based on a multi-modal convolutional neural network. BACKGROUND

[0002] Solid rocket engines have the characteristics of simple structure, high reliability, strong maneuverability, short ground preparation time, and easy long-term storage, and are widely used in missile weapons. A solid rocket engine mainly consists of a shell, a solid propellant, an insulating layer, and a lining, etc. components, among which the solid propellant provides the main power source and occupies a core position in the solid rocket engine. With the continuous rise of the performance requirements of solid rocket engines, the development of high-energy solid propellants has attracted much attention.

[0003] As a kind of energetic composite material with high solid particle filling ratio, the solid propellant is mechanically mixed by multiple phases and has complex microstructure characteristics. During transportation, storage, launching, ignition, flight, etc., under the action of external loads such as temperature, pressure, vibration, etc., phenomena such as matrix fracture, micro-crack generation, particle breakage, and interface dewetting occur, and the internal full-field strain of the solid propellant is complex and variable. It is extremely difficult to obtain and predict the full-field strain evolution process through traditional testing methods, which has become a bottleneck problem restricting the long-term service of solid propellants.

[0004] The existing solving methods mainly have the following problems: 1. The solid propellant constitutive model can only give the macroscopic stress-strain relationship, and cannot predict the internal full-field strain according to the different microstructure of the propellant; the finite element calculation cannot reflect the influence of the real microstructure of the propellant on the strain field according to the ideal model, and the calculation is time-consuming and costly. It cannot realize the rapid and accurate prediction of the internal full-field strain of the solid propellant with different microstructure, and cannot meet the related needs of performance improvement, strength and life, and reliability evaluation of the solid propellant. SUMMARY

[0005] The technical problem solved by the application is to overcome the shortcomings of the prior art and provide a solid propellant internal full-field strain prediction method based on a neural network, which realizes the rapid prediction of the internal full-field strain of the propellant by shooting the microstructure characteristics of the initial state of the solid propellant and giving the macroscopic strain state, and solves the problem that the internal full-field strain of the solid propellant cannot be rapidly and accurately predicted in the prior art.

[0006] The technical solution of the application is a solid propellant internal full-field strain prediction method based on a neural network, which comprises the following steps:

[0007] Step one, obtain the initial state and the internal microstructure image of the solid propellant under different loading conditions, and compare to obtain the internal microstructure feature variation law; record the experimental conditions and the solid propellant material properties to obtain the sample state characteristics s;

[0008] Step two, process the obtained internal full-field strain of the solid propellant by using the digital volume correlation method, and convert the internal full-field strain into a gray scale image to obtain the measured micro strain field image E under different loading conditions;

[0009] Step three, construct a multi-modal convolutional neural network model, the input variable of the model is the microstructure image D of the solid propellant in the initial state, the sample state characteristics s, and the output variable is the predicted micro strain field image G corresponding to the sample state characteristics s;

[0010] Step four, construct a sample data set according to the internal microstructure features of the solid propellant and the internal full-field strain results;

[0011] Step five, train the multi-modal convolutional neural network model using the sample data set in step four to obtain the multi-modal convolutional neural network model under the optimal condition, so as to predict the internal full-field strain of the solid propellant.

[0012] Further, the step one of obtaining the initial state and the internal microstructure image of the solid propellant under different loading conditions comprises the following steps:

[0013] S1, for the solid propellant, carry out in-situ uniaxial tensile test based on the synchrotron radiation source to measure the synchrotron radiation X-ray projection of the solid propellant in the initial state and under different loading conditions:

[0014] S2, reconstruct the synchrotron radiation X-ray projection in step S1 by using phase recovery technology to obtain a 16-bit three-dimensional digital image;

[0015] S3, convert the 16-bit three-dimensional digital image in step S2 into an 8-bit three-dimensional digital image, and the gray scale value of the 8-bit three-dimensional digital image is linearly scaled between 0 and 255, that is, the maximum gray scale value and the minimum gray scale value in the 16-bit three-dimensional digital image correspond to 255 and 0 in the pixel gray scale value of the 8-bit three-dimensional digital image.

[0016] Further, the way of step two to obtain the internal full-field strain of the solid propellant is: calculate the 8-bit three-dimensional digital image in step S3 by using digital volume correlation, and obtain the internal full-field strain of the solid propellant sample under different loading conditions through the correlation operation between the three-dimensional digital image of the initial state of the solid propellant and the three-dimensional digital image of the solid propellant at different stages of the test loading.

[0017] Further, the multi-modal convolutional neural network model in step three is a multi-modal convolutional neural network model comprising 6 convolutional layers, 3 pooling layers, 3 up-sampling layers and 1 fully connected layer, and each network layer is connected in the following order: first convolutional layer-first pooling layer-second convolutional layer-second pooling layer-third convolutional layer-third pooling layer-fully connected layer-first up-sampling layer-fourth convolutional layer-second up-sampling layer-fifth convolutional layer-third up-sampling layer-sixth convolutional layer.

[0018] The activation function of the 6 convolutional layers and the fully connected layer in the multi-modal convolutional neural network model is a ReLU linear rectifier function.

[0019] Further, the data size of the solid propellant microstructure image D in the initial state in step three is m x m x 2, wherein m x m represents the number of pixel points contained in the image D, and 2 represents that each pixel point contains two variable values: a gray value and an identifier for determining whether the current pixel point is a particle: 0 represents that the current pixel point is not a particle, and 1 represents that the current pixel point is a particle.

[0020] The data size of the sample state feature s is n x 1; the state feature includes one or more of the tensile strain size, the elastic modulus of the material, the yield stress, and the Poisson's ratio, and n represents that the selected state feature contains n kinds.

[0021] The data size of the predicted microstructure strain field image G corresponding to the sample state feature s is m x m x 1, wherein the number of pixel points contained in the image G is also m x m, and 1 represents that each pixel point contains one variable value, i.e., a strain value.

[0022] Further, the sample data set is constructed according to the internal microstructure characteristics of the solid propellant and the full-field strain results in step four in the following manner:

[0023] The solid propellant microstructure image D in the initial state is selected from the 8-bit three-dimensional digital image in step S3; the internal full-field strain in step two is converted into a gray-scale image to obtain the microstructure strain field image E of the solid propellant under different loading conditions; the experimental conditions and the material properties of the solid propellant are recorded to obtain the sample state feature s; and the microstructure image D, the microstructure strain field image E and the sample state feature s are cropped and preprocessed to obtain the sample data set.

[0024] Further, the cropping and preprocessing process is as follows:

[0025] The obtained microstructure image D and microstructure strain field image E are cropped to obtain a picture set with a uniform size of m x m;

[0026] Then, the variable values ​​contained in each pixel of the microstructure image D and the microstrain field image E are calculated according to formula z. i '=(z i –min(z i )) / (max(z i )–min(z i Map the values ​​sequentially to [0,1] and calculate the normalized sample value corresponding to each variable value of each pixel.

[0027] Among them, z i Let max(z) be the gray value of the current pixel i in the microstructure image D or the strain value of the current pixel i in the microstrain field image E. i ) represents the maximum gray value in the microstructure image D or the maximum strain value in the microstrain field image E, and min(z) represents the minimum gray value in the microstructure image D or the minimum strain value in the microstrain field image E; z i 'for z i The normalized value obtained through calculation;

[0028] The state characteristics s of the specimens under different loading conditions are mapped to [0,1] in the same way.

[0029] Furthermore, the training of the multimodal convolutional neural network model using the dataset described in step five is achieved as follows: the preprocessed sample dataset is divided into a training set and a validation set; the mean squared error between the output variable value and the target variable value is used as the loss function, and the mean absolute error between the output variable value and the target variable value is selected as the supervision standard for training the multimodal convolutional neural network; the multimodal convolutional neural network is trained using the data from the training set and the validation set, and the weights and bias terms are continuously iterated using backpropagation and stochastic gradient descent optimization algorithms during training. When the loss function reaches its minimum, the optimal multimodal convolutional neural network model is obtained; the output variable value is the predicted micro-strain field image G corresponding to the sample state characteristics s, and the target variable value is the preprocessed micro-strain field image E'.

[0030] Furthermore, the pixels of the microstructure image D' and the microstrain field image E' in the preprocessed sample dataset are randomly divided into training and validation sets in a 7:3 ratio.

[0031] Furthermore, based on different loading conditions, the values ​​of the preprocessed sample state features s' are input into the multimodal convolutional neural network model for training.

[0032] The advantages of this invention compared to the prior art are:

[0033] (1) The method comprehensively considers the microstructure characteristics and macroscopic property parameters of the solid propellant, establishes the feature relationship between different data forms by using a multi-modal convolutional neural network model, trains the model through a series of solid propellant microstructure images, internal full-field strain and material state characteristics, and after training, the microstructure image and material state characteristics of the solid propellant can be used to realize the prediction of the microstructure image.

[0034] (2) The method can reduce the calculation cost of solving the full-field strain of the solid propellant, and meet the analysis requirements of the engineering field.

[0035] (3) The method is easy to operate, and has no technical and professional background knowledge requirements for the operator, and can be widely applied.

[0036] (4) The method fully utilizes the multi-modal fusion characteristics of the neural network, can consider the microstructure characteristics, macroscopic deformation parameters and external mechanical load parameters of the propellant, and better describe the distribution and evolution of the full-field strain in the deformation process, and has high reliability.

[0037] (5) The method has good generalization ability, and can be applied to different deformation conditions of similar materials after sufficient training. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is an in-situ uniaxial tensile test schematic diagram based on a synchrotron radiation source for an embodiment of the application.

[0039] Figure 2 It is a multi-modal convolutional neural network structure schematic diagram for an embodiment of the application.

[0040] Figure 3 It is a solid propellant microstructure image (a, b) and a full-field strain schematic diagram (c) for an embodiment of the application.

[0041] Figure 4 It is a comparison diagram of the full-field strain prediction effect and the experimental results for an embodiment of the application. DETAILED DESCRIPTION

[0042] The application will be further described below with reference to the drawings.

[0043] A solid propellant full-field strain prediction method based on a neural network comprises the following steps:

[0044] Step one, obtain the initial state and the internal microstructure image of the solid propellant under different loading conditions, and compare to obtain the internal microstructure characteristic change rule; record the experimental conditions and the solid propellant material properties to obtain the sample state characteristics s.

[0045] (1) In this example, the in-situ uniaxial tensile test based on the synchrotron radiation source is carried out for the solid propellant, and the synchrotron radiation X-ray projections of the solid propellant under 0, 0.16, 0.32 and 0.48 tensile strains are measured, and the test process is as shown in Figure 1

[0046] (2) The phase recovery technology in the PITRE software (Phase-sensitive x-ray Image processing and Tomography REconstruction) is used to reconstruct the synchrotron radiation X-ray projections in step (1), and 16bit three-dimensional digital images are obtained.

[0047] (3) The 16bit three-dimensional digital images in step (2) are converted into 8bit three-dimensional digital images, and the gray value of the 8bit three-dimensional digital images is linearly scaled between 0 and 255, that is, the maximum gray value and the minimum gray value in the 16bit three-dimensional digital images correspond to 255 and 0 in the pixel gray value of the 8bit three-dimensional digital images, respectively.

[0048] Step two, the digital volume correlation method is used to process the internal full-field strain of the solid propellant, and the internal full-field strain is converted into a gray image, and the measured micro strain field image E under different loading conditions is obtained.

[0049] The 8bit three-dimensional digital images in step (3) are calculated by using the digital volume correlation, and the internal full-field strain of the solid propellant sample under different load conditions is obtained by the correlation operation between the three-dimensional digital images of the initial state (0 tensile strain) of the solid propellant and the three-dimensional digital images of the solid propellant at different stages (0.16, 0.32 and 0.48 tensile strain) of the test loading. The present application simultaneously uses two kinds of technologies to measure the three-dimensional displacement field and strain field in the full volume range inside the test piece: a traditional subset-based DVC method (Subset-Based DVC Approach) and a FE-based DVC method (FE-based DVC Approach) with better robustness. The traditional subset-based DVC method divides the reference image and the deformed image into smaller difference sub-regions, and then performs correlation operation, so as to realize high-precision operation.

[0050] Step three, a multi-modal convolutional neural network model is constructed.

[0051] As shown in Figure 2 The present embodiment constructs a multi-modal convolutional neural network model comprising 6 convolutional layers, 3 pooling layers, 3 up-sampling layers and 1 fully connected layer.

[0052] ​The network layers are connected in the following order: a first convolutional layer-a first pooling layer-a second convolutional layer-a second pooling layer-a third convolutional layer-a third pooling layer-a fully connected layer-a first upsampling layer-a fourth convolutional layer-a second upsampling layer-a fifth convolutional layer-a third upsampling layer-a sixth convolutional layer. The activation functions of the six convolutional layers and the fully connected layer in the multi-modal convolutional neural network model are all ReLU linear rectification functions.

[0053] Further, the input variables of the multi-modal convolutional neural network model include a solid propellant microstructure image D in an initial state with a data size of mxm x 2, wherein mxm represents the number of pixel points contained in the image D, and 2 represents that each pixel point contains two variable values, i.e., a gray value and an identifier for determining whether the current pixel point is a particle: 0 represents that the current pixel point is not a particle, and 1 represents that the current pixel point is a particle.

[0054] The input variables of the multi-modal convolutional neural network model also include a sample state feature s with a data size of nx 1; wherein the sample state feature s is obtained by recording experimental conditions and solid propellant material properties, and the state feature includes one or more of a tensile strain size, an elastic modulus of the material, a yield stress, and a Poisson's ratio, and n represents that the selected state feature contains n kinds.

[0055] The output variables of the multi-modal convolutional neural network model include a predicted microstructure strain field image G corresponding to the sample state feature s with a data size of mxm x 1, wherein the number of pixel points contained in the image G is also mxm, and 1 represents that each pixel point contains a variable value, i.e., a strain value.

[0056] Step four, constructing a required sample data set according to the internal microstructure characteristics of the solid propellant and the full-field strain results.

[0057] In this example, the solid propellant microstructure image D in the initial state is selected from the 8-bit three-dimensional digital image in step (3). Figure 3 (a) represents the solid propellant microstructure in this example, Figure 3 (b) represents the microstructure image after extracting the particles from the propellant microstructure image. The internal full-field strain in step two is converted into a gray image and screened to obtain 0.16, 0.32 and 0.48 tensile strain images E, as shown in Figure 3 (c); the microstructure images in this example are taken from the same sample, and the characteristics such as the elastic modulus, the yield stress and the Poisson's ratio of the material are all the same, so the sample state feature s only needs to include the tensile strain size.

[0058] The obtained microstructure image D and tensile strain image E are cut to obtain a picture set with a uniform size of mxm;

[0059] Then, the variable values ​​contained in each pixel in the microstructure image D and the tensile strain image E are calculated according to formula z. i '=(z i –min(z i )) / (max(z i )–min(z i Map the values ​​sequentially to [0,1] and obtain the sample normalization value corresponding to each variable value of each pixel.

[0060] Among them, z i Let max(z) be the grayscale value of the current pixel i in the microstructure image D or the strain value of the current pixel i in the tensile strain image E. i ) represents the maximum gray value in the microstructure image D or the maximum strain value in the tensile strain image E, and min(z) represents the minimum gray value in the microstructure image D or the minimum strain value in the tensile strain image E; z i 'for z i The normalized value is obtained through calculation.

[0061] The state characteristics s of the specimens under different loading conditions are mapped to [0,1] in the same way.

[0062] Step 5: Train the multimodal convolutional neural network model using the dataset to obtain the multimodal convolutional neural network model under optimal conditions, in order to predict the full-field strain inside the solid propellant.

[0063] The preprocessed sample dataset was randomly divided into a training set and a validation set in a 7:3 ratio. The mean squared error between the output variable value and the target variable value was used as the loss function, and the mean absolute error between the output variable value and the target variable value was selected as the supervision standard for training the dynamic neural network. The dynamic neural network model was trained using the data from the training set and the validation set. During training, the weights and bias terms were continuously iterated using backpropagation and stochastic gradient descent optimization algorithms. When the loss function reached its minimum, the optimal dynamic neural network model was obtained. This model was then used to predict the full-field strain inside the solid propellant.

[0064] The above output variable value is the predicted micro-strain field image G corresponding to different tensile strains, and the target variable value is the preprocessed tensile strain image.

[0065] In this example, a model is trained and predicted using a set of microstructure images from uniaxial tensile tests of solid propellant materials, such as... Figure 4 The figure shows the strain distribution across the entire test set.

[0066] from Figure 4It can be seen that the predicted strain field is very close to the experimental strain field, and the method can quickly predict and analyze the full-field strain distribution of the solid propellant.

[0067] Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the present application by using the disclosed methods and technical contents without departing from the spirit and scope of the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solutions of the present application shall fall within the protection scope of the technical solutions of the present application.

Claims

1. A neural network-based method for predicting a full-field strain inside a solid propellant, characterized in that, The method comprises the following steps: Step one, obtaining the microstructure images of the solid propellant under initial state and different loading conditions, recording the experimental conditions and the material properties of the solid propellant to obtain the sample state characteristics s; Step two, processing the obtained full-field strain of the solid propellant by using the digital volume correlation method, and converting the internal full-field strain into a gray image to obtain the measured micro-strain field image E under different loading conditions; Step three, constructing a multi-modal convolutional neural network model, the input variable of the model being the microstructure image D of the solid propellant under initial state and the sample state characteristics s, and the output variable being the predicted micro-strain field image G corresponding to the sample state characteristics s; Step four, cropping the microstructure image D and the micro-strain field image E, and pre-processing the microstructure image D, the micro-strain field image E and the sample state characteristics s to construct a sample data set; Step five, training the multi-modal convolutional neural network model by using the sample data set in step four to obtain the multi-modal convolutional neural network model under the optimal condition, so as to predict the full-field strain of the solid propellant.

2. The method of claim 1, wherein, The step one comprises the following steps: S1, measuring the synchrotron X-ray projection of the solid propellant under initial state and different loading conditions by using the in-situ uniaxial tensile test based on the synchrotron radiation source; S2, reconstructing the synchrotron X-ray projection in step S1 by using the phase recovery technology to obtain a 16-bit three-dimensional digital image; S3, converting the 16-bit three-dimensional digital image in step S2 into an 8-bit three-dimensional digital image, and linearly scaling the gray value of the 8-bit three-dimensional digital image between 0 and 255, that is, the maximum gray value and the minimum gray value in the 16-bit three-dimensional digital image correspond to 255 and 0 in the pixel gray value of the 8-bit three-dimensional digital image.

3. The neural network-based method for predicting the full-field strain of a solid propellant according to claim 2, wherein The step two obtains the full-field strain of the solid propellant by using the digital volume correlation to calculate the 8-bit three-dimensional digital image in step S3, and obtains the full-field strain of the solid propellant sample under different loading conditions through the correlation operation between the three-dimensional digital image of the initial state of the solid propellant and the three-dimensional digital image of the solid propellant under different loading conditions.

4. The method of claim 1, wherein, The multi-modal convolutional neural network model in step three is a multi-modal convolutional neural network model comprising six convolutional layers, three pooling layers, three up-sampling layers and one fully connected layer, and each network layer is connected in the following order: first convolutional layer-first pooling layer-second convolutional layer-second pooling layer-third convolutional layer-third pooling layer-fully connected layer-first up-sampling layer-fourth convolutional layer-second up-sampling layer-fifth convolutional layer-third up-sampling layer-sixth convolutional layer. The activation function of the six convolutional layers and the fully connected layer in the multi-modal convolutional neural network model is a ReLU linear rectifier function.

5. The method of claim 1, wherein, The data size of the initial state solid propellant microstructure image D in step three is m×m×2, wherein m×m represents the number of pixel points contained in the image D, and 2 represents that each pixel point contains two variable values, i.e., a gray value and an identifier for determining whether the current pixel point is a particle, wherein 0 represents that the current pixel point is not a particle, and 1 represents that the current pixel point is a particle; The data size of the sample state feature s is n×1; the state feature includes one or more of the following: tensile strain size, elastic modulus of the material, yield stress, and Poisson's ratio, and n represents that the selected state feature contains n kinds; The data size of the predicted microstructure strain field image G corresponding to the sample state feature s is m×m×1, wherein the number of pixel points contained in the image G is also m×m, and 1 represents that each pixel point contains one variable value, i.e., a strain value.

6. The neural network-based method for predicting the full-field strain inside a solid propellant according to claim 2, wherein In step four, the sample data set is constructed according to the internal microstructure characteristics and the full-field strain results of the solid propellant, and the method is as follows: The initial state solid propellant microstructure image D is selected from the 8-bit three-dimensional digital image in step S3; the internal full-field strain in step two is converted into a gray image to obtain the microstructure strain field image E of the solid propellant under different loading conditions; and the experimental conditions and the material properties of the solid propellant are recorded to obtain the sample state feature s; The microstructure image D and the microstructure strain field image E are cropped, and the microstructure image D, the microstructure strain field image E and the sample state feature s are preprocessed to obtain the sample data set.

7. The neural network-based method for predicting the full-field strain inside a solid propellant according to claim 6, wherein The cropping and preprocessing process is as follows: The obtained microstructure image D and the microstructure strain field image E are cropped to obtain a picture set with a uniform size of m×m; Then, the variable values ​​contained in each pixel of the microstructure image D and the microstrain field image E are calculated according to formula z. i '=(z i -min(z i )) / (max(z i )-min(z i Map the values ​​sequentially to [0,1] and calculate the normalized sample value corresponding to each variable value of each pixel. Among them, z i Let max(z) be the gray value of the current pixel i in the microstructure image D or the strain value of the current pixel i in the microstrain field image E. i ) represents the maximum gray value in the microstructure image D or the maximum strain value in the microstrain field image E, and min(z) represents the minimum gray value in the microstructure image D or the minimum strain value in the microstrain field image E; z i 'for z i The normalized value obtained through calculation; The sample state features s under different loading conditions are mapped to [0, 1] in the same way.

8. The neural network-based method for predicting the full-field strain inside a solid propellant according to claim 7, wherein, In step five, the data set is used to train the multi-modal convolutional neural network model, and the training is realized by the following method: the preprocessed sample data set is divided into a training set and a validation set; the mean square error between the output variable value and the target variable value is used as the loss function, and the mean absolute error between the output variable value and the target variable value is selected as the supervision standard for the multi-modal convolutional neural network during training; the data of the training set and the validation set are used to train the multi-modal convolutional neural network, and the weight and bias term are iterated by using the back propagation and the stochastic gradient descent optimization algorithm during training, and the optimal multi-modal convolutional neural network model is obtained when the loss function reaches the minimum value; The output variable value is the predicted microstructure strain field image G corresponding to the sample state feature s, and the target variable value is the preprocessed microstructure strain field image E'.

9. The neural network-based method for predicting the full-field strain inside a solid propellant according to claim 8, wherein, The pixels of the microstructure image D' and the pixels of the microstructure strain field image E' in the preprocessed sample data set are randomly divided into a training set and a validation set according to a ratio of 7:

3.

10. The neural network-based method for predicting the full-field strain inside a solid propellant according to claim 8, wherein, According to different loading conditions, the values of the preprocessed sample state feature s' are input into the multi-modal convolutional neural network model for training.

Images