A method for predicting the properties of carbon-carbon composites

Abstract

The present application relates to a kind of carbon carbon composite material performance prediction method, by inputting three-dimensional needle carbon carbon composite material process parameters to the back propagation neural network model trained to processing, realize the fast prediction of three-dimensional needle carbon carbon composite material mechanical properties.The present application will part of data set as training data set when establishing database, train back propagation neural network model, learn different features in various three-dimensional needle carbon carbon composite material needle process, another part of data set is used as test data set to judge whether the back propagation neural network model after training meets the set prediction qualified rate, also divide part of data set as preparation data set, using data rotation and data set re-grouping mode is completed to the training of neural network model, both reduce training time, and increase the training number of different data combination, ensure that each group of data in database is fully utilized, improve the prediction speed of model.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional needle-punched carbon-carbon composite materials, and more specifically to a method for predicting the performance of carbon-carbon composite materials. Background Technology

[0002] Three-dimensional needle-punched carbon-carbon composites, as an important type of fiber-reinforced composite material, have the advantages of simpler preparation methods and lower costs compared to three-dimensional braided reinforced composites. The needle-punching process parameters, such as the arrangement of needles in the needle plate, the shape of the needles, and the stacking method of fiber layers in the fabric laminate, are significant factors affecting the compatibility of three-dimensional needle-punched carbon-carbon composites.

[0003] To predict material properties, traditional simulation calculations of three-dimensional needle-punched carbon-carbon composites are often used for performance analysis. This method is suitable for specific composite material models. Patent CN109241650A discloses a method for predicting the mechanical properties of carbon fiber reinforced composites based on cross-scale simulation. It uses the micro-finite element method to establish a unit cell prediction model of the carbon fiber composite at the micro-level and predict the mechanical properties of the carbon fiber reinforced composite. However, if different material parameters are given, it is necessary to remodel the composite material when predicting its properties, which is a cumbersome process and cannot be performed quickly. Summary of the Invention

[0004] This invention discloses a method for predicting the performance of carbon-carbon composite materials. By inputting the process parameters of three-dimensional needle-punched carbon-carbon composite materials into a trained backpropagation neural network model, the mechanical properties of three-dimensional needle-punched carbon-carbon composite materials can be rapidly predicted.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the properties of carbon-carbon composite materials, comprising: Step 1: Establish a database containing the needle-punching process parameters of three-dimensional needle-punched carbon-carbon composites and the corresponding performance values ​​of the three-dimensional needle-punched carbon-carbon composites; Step 2: Construct the backpropagation neural network model; Step 3: Divide the data in the database into three groups: M, N, and H. Group M is used as the training dataset, group N as the test dataset, and group H as the preliminary dataset. The datasets in the database are regrouped and rotated according to the rotation rate P. The number of rotation groups in each dataset = P × (M + N + H). The rotation order of the datasets is: Preliminary dataset → Test dataset → Training dataset → Preliminary dataset. Step 4: Train the backpropagation neural network model. Use M training datasets to train the backpropagation neural network model to obtain a pre-trained model. The output parameters of the backpropagation neural network model are:y The predicted output parameters are To describe the extent to which the neural network performs poorly in each training session, a cost function J is defined as the mean of the training error for each sample in the training dataset, where L is the training error. , , Step 5: The backpropagation neural network model after initial training is tested with N sets of test datasets to obtain the prediction results for each set of test data; Step 6: Calculate the prediction results Y of each set of test data obtained in Step 5. p The performance values ​​Y in the corresponding N test datasets e Calculate the mean squared error (MSE) and mean absolute error percentage (MAEP) to determine whether each set of test data is qualified. , , Step 7: Set the standard value of the mean square error of the test data to M1, and the standard value of the mean absolute error percentage to M2. If the following conditions are met... If the initial training of the backpropagation neural network model is successful, then the backpropagation neural network model after initial training will be used as the backpropagation neural network model after training is complete; otherwise, proceed to steps 8 and 9. Step 8: Specify the standard value for relative error in the test. The relative errors of the detection error and the mean absolute error are respectively

[0006]

[0007] Step 9: Adjust the network structure and parameters of the backpropagation neural network model after initial training, regroup the data in the database, return to step 3 to retrain the model until a trained backpropagation neural network model is obtained.

[0008] Furthermore, in step 1, the needle punching process parameters include the arrangement of needles in the needle plate (ARB) and the diameter of the needles (R), the stacking angle (α) of the fiber layers in the fabric stack, and the stacking order (STO).

[0009] Furthermore, the fabric stack consists of a layer of chopped fibers and a stack of unidirectional fiber nonwoven fabric.

[0010] Furthermore, the unidirectional fiber nonwoven fabric can be stacked as 0° nonwoven fabric and 90° nonwoven fabric.

[0011] Furthermore, the stacking method is carried out in the following order: 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer, and so on in a cyclical manner.

[0012] Furthermore, the stacking method is carried out in the following order: 0° nonwoven fabric - chopped fiber layer - 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric, and so on in a cyclical manner.

[0013] Furthermore, in step 1, the performance values ​​of the three-dimensional needled carbon-carbon composite material include the transverse tensile strength (TTS), the longitudinal tensile strength (LTS), and the flexural strength (BS) of the three-dimensional needled carbon-carbon composite material.

[0014] Furthermore, in step 2, the constructed backpropagation neural network model includes an input layer, hidden layers, and an output layer, and the activation function used is... Sigmoid The function is trained 10,000 times with a learning rate of [missing information]. η =0.3.

[0015] Furthermore, in step 3, the ratio of the three datasets in the database is M:N:H = 7:2:1.

[0016] This invention predicts the performance of needle-punched carbon-carbon composite materials based on a backpropagation neural network model. When establishing a numerical database of needle-punching process parameters and performance of three-dimensional needle-punched carbon-carbon composite materials, a portion of the dataset is used as a training dataset to train the backpropagation neural network model, learning different characteristics in various needle-punching processes of three-dimensional needle-punched carbon-carbon composite materials. Another portion of the dataset is used as a test dataset to judge whether the trained backpropagation neural network model meets the set prediction pass rate. A portion of the dataset is also set as a reserve dataset. The training of the neural network model is completed by data rotation and dataset regrouping. This reduces training time and increases the number of training times for different data combinations, ensuring that each group of data in the database is fully utilized and improving the prediction speed of the model. Attached Figure Description

[0017] Figure 1 This is a flowchart of the carbon-carbon composite material performance prediction method in an embodiment of the present invention; Figure 2 This refers to the arrangement of needles in the needle plate mentioned in the embodiment; Figure 3 for Figure 2 A schematic diagram of the needle's diameter within the dashed square. Figure 4 This is a schematic diagram of the stacking method of the fiber layers in the fabric laminate mentioned in the embodiment; Figure 5 This is a schematic diagram of the second method of stacking fiber layers in the fabric layering mentioned in the embodiment. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] This embodiment discloses a method for predicting the properties of three-dimensional needle-punched carbon-carbon composite materials, such as... Figure 1 As shown, this method mainly includes the following: Step 1: Obtain the needle-punching process parameters of the 3D needle-punched carbon-carbon composite material. Use these parameters as input features of a pre-defined neural network, and use the performance values ​​of the 3D needle-punched carbon-carbon composite material as output features. Divide the obtained sets of needle-punching process parameters and corresponding performance values ​​of the 3D needle-punched carbon-carbon composite material to establish a database of needle-punching process parameters and performance values ​​for the 3D needle-punched carbon-carbon composite material.

[0020] Specifically, the needle-punching process parameters include the arrangement of needles in the needle plate (ARB) and the needle diameter (R), the stacking angle (α) of the fiber layers in the fabric laminate, and the stacking order (STO). The performance values ​​of the three-dimensional needle-punched carbon-carbon composite include the transverse tensile strength (TTS), the longitudinal tensile strength (LTS), and the flexural strength (BS) of the three-dimensional needle-punched carbon-carbon composite.

[0021] The needle arrangement in the needle plate mentioned above is for reference. Figure 2 As shown, Figure 2 This is a schematic diagram of the macroscopic shape of the needle plate. The smallest periodic unit of the needle arrangement in the needle plate is... Figure 3 As shown, Figure 3 The diagram shows the needle diameter, x is the shortest distance between needles in a row, y1 is the shortest row spacing between two rows, y2 is the shortest row spacing between the first and third rows, and R is the needle diameter.

[0022] In addition, the stacking angle and stacking order requirements of the fiber layers in the fabric stack mentioned above are as follows: the fabric stack consists of chopped fiber layers and unidirectional fiber nonwoven fabric stacked together. The unidirectional fiber nonwoven fabric can be placed as 0° nonwoven fabric and 90° nonwoven fabric during stacking. This embodiment provides... Figure 4 and Figure 5 Two stacking methods, among which Figure 4 The stacking method is carried out in the following order: 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer. Each stacking unit is formed by following the above order. Multiple stacking units are stacked in this way. Figure 5 The stacking process is carried out in the following order: 0° nonwoven fabric - chopped fiber layer - 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric. Each stacking unit is formed by following the above order. Multiple stacking units are stacked in this manner.

[0023] Step 2: Build a backpropagation neural network model using the Python language and Python framework.

[0024] Specifically, the backpropagation neural network model constructed in this embodiment includes an input layer, a hidden layer, and an output layer, and the activation function used is... Sigmoid The function processes more continuous and smooth data, is easy to differentiate, and the model is trained 10,000 times with a learning rate of [missing information]. η =0.3.

[0025] Step 3: Divide the data in the database into three datasets: M, N, and H. The M dataset is used as the training dataset, the N dataset as the test dataset, and the H dataset as the preparatory dataset. The ratio of the three datasets is M:N:H = 7:2:1.

[0026] To prevent errors in the data itself from affecting model training, this embodiment requires that during training, a new round of data grouping be performed each time, and the datasets be rotated according to a rotation rate P. The number of rotation groups in each dataset = P × (M + N + H). The rotation order of the datasets is: preparatory dataset → test dataset → training dataset → preparatory dataset. Specifically, in this embodiment, the rotation rate P does not exceed 5%. Assuming P is 5%, the preparatory dataset has 10 groups, the test dataset has 20 groups, and the training dataset has 70 groups. Then, the number of rotation groups in each dataset is 5 groups. This avoids the situation of insufficient or excessive datasets during the rotation process. That is, the 5 groups of data from the previous round of preparatory dataset are rotated to the test dataset, the 5 groups of data from the previous round of test dataset are rotated to the training dataset, the 5 groups of data from the previous round of training dataset are rotated to the preparatory dataset, and so on.

[0027] Traditional models typically divide their databases into training and test sets. The training set contains a large amount of data, resulting in lengthy model training times, and the division between training and test sets is fixed. This invention, however, reduces the size of the training set by fine-tuning the data grouping ratio. This reduces training time without affecting model training and increases the number of training iterations with different data combinations. Furthermore, a rotation rate is used in data processing. After each model iteration, the dataset is rotated, and the rotation rate is updated in real-time based on the error magnitude to ensure the usability of each data set.

[0028] Step 4: Train the backpropagation neural network model. Use M training datasets to train the backpropagation neural network model to obtain a pre-trained model. The output parameters of the backpropagation neural network model are: y The predicted output parameters are To describe the extent to which the neural network performs poorly in each training session, a cost function J is defined as the mean of the training error for each sample in the training dataset, where L is the training error.

[0029]

[0030]

[0031] Step 5: The backpropagation neural network model after initial training is tested with N sets of test datasets to obtain the prediction results for each set of test data.

[0032] Step 6: Calculate the prediction results Y of each set of test data obtained in Step 5. p The performance values ​​Y in the corresponding N test datasets e Calculate the mean squared error (MSE) and mean absolute error percentage (MAEP) to determine whether each set of test data is qualified.

[0033]

[0034]

[0035] Step 7: Set the standard value of the mean square error of the test data to M1, and the standard value of the mean absolute error percentage to M2. If the following conditions are met... If the initial training of the backpropagation neural network model is successful, then the pre-trained backpropagation neural network model will be used as the completed backpropagation neural network model; otherwise, proceed to steps 8 and 9.

[0036] Step 8: Specify the standard value for relative error in the test. The relative errors of the detection error and the mean absolute error are respectively

[0037]

[0038] Step 9: Adjust the network structure and parameters of the backpropagation neural network model after initial training, regroup the data in the database, return to step 3 to retrain the model until a trained backpropagation neural network model is obtained.

[0039] Compared with traditional finite element simulation calculation methods, this invention greatly improves the efficiency of predicting the mechanical properties of three-dimensional needle-punched carbon composite materials and can significantly reduce costs.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the performance of carbon-carbon composite materials, characterized in that, include: Step 1: Establish a database containing the needle-punching process parameters of three-dimensional needle-punched carbon-carbon composites and the corresponding performance values ​​of the three-dimensional needle-punched carbon-carbon composites; Step 2: Construct the backpropagation neural network model; Step 3: Divide the data in the database into three groups: M, N, and H. Group M is used as the training dataset, group N is used as the test dataset, and group H is used as the preparatory dataset. The datasets in the database are regrouped and rotated according to the rotation rate P each time. The number of rotation groups in each dataset = P × (M + N + H). The rotation order of the datasets is: preparatory dataset → test dataset → training dataset → preparatory dataset. Step 4: Train the backpropagation neural network model. Use M training datasets to train the backpropagation neural network model to obtain a pre-trained model. The output parameters of the backpropagation neural network model are: y The predicted output parameters are To describe the extent to which the neural network performs poorly in each training session, a cost function J is defined as the mean of the training error for each sample in the training dataset, where L is the training error. Step 5: The backpropagation neural network model after initial training is tested with N sets of test datasets to obtain the prediction results for each set of test data; Step 6: Calculate the prediction results Y of each set of test data obtained in Step 5. p The performance values ​​Y in the corresponding N test datasets e Calculate the mean squared error (MSE) and mean absolute error percentage (MAEP) to determine whether each set of test data is qualified. Step 7: Set the standard value of the mean square error of the test data to M1, and the standard value of the mean absolute error percentage to M2. If the following conditions are met... Then the backpropagation neural network model after initial training will be used as the backpropagation neural network model after training is completed. Otherwise, proceed to steps 8 and 9; Step 8: Specify the standard value for relative error in the test. The relative errors of the detection error and the mean absolute error are respectively when hour: ; when hour: ; Step 9: Adjust the network structure and parameters of the backpropagation neural network model after initial training, regroup the data in the database, return to step 3 to retrain the model until a trained backpropagation neural network model is obtained.

2. The method for predicting the performance of carbon-carbon composite materials according to claim 1, characterized in that: In step 1, the needle punching process parameters include the arrangement of needles in the needle plate (ARB) and the diameter of the needles (R), the stacking angle (α) of the fiber layers in the fabric stack, and the stacking order (STO).

3. The method for predicting the performance of carbon-carbon composite materials according to claim 2, characterized in that: The fabric stack consists of a layer of chopped fibers and a stack of unidirectional fiber nonwoven fabric.

4. The method for predicting the performance of carbon-carbon composite materials according to claim 3, characterized in that: The unidirectional fiber nonwoven fabric can be stacked as 0° nonwoven fabric and 90° nonwoven fabric.

5. The method for predicting the performance of carbon-carbon composite materials according to claim 4, characterized in that: The unidirectional fiber nonwoven fabric is stacked in the following order: 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer, and so on in a cyclical manner.

6. The method for predicting the performance of carbon-carbon composite materials according to claim 4, characterized in that: The unidirectional fiber nonwoven fabric is stacked in the following order: 0° nonwoven fabric - chopped fiber layer - 0° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric - chopped fiber layer - 90° nonwoven fabric, and so on in a cyclical manner.

7. The method for predicting the performance of carbon-carbon composite materials according to claim 1, characterized in that: In step 1, the performance values ​​of the three-dimensional needled carbon-carbon composite material include the transverse tensile strength (TTS), the longitudinal tensile strength (LTS), and the flexural strength (BS) of the three-dimensional needled carbon-carbon composite material.

8. The method for predicting the performance of carbon-carbon composite materials according to claim 1, characterized in that: In step 2, the constructed backpropagation neural network model includes an input layer, a hidden layer, and an output layer, and the activation function used is... Sigmoid The function is trained 10,000 times with a learning rate of [missing information]. η =0.

3.

9. The method for predicting the performance of carbon-carbon composite materials according to claim 1, characterized in that: In step 3, the ratio of the three datasets in the database is M:N:H=7:2:1.

Images