Composite armor performance prediction method based on BP neural network and particle swarm algorithm

Abstract

The present invention relates to a composite armor performance prediction method based on BP neural network and particle swarm algorithm, comprising step 1: using numerical simulation means to establish a composite armor numerical simulation model, and obtain a composite armor structural parameter-protection performance parameter database; step 2: randomly dividing the above database into training set and test set data. Using the training set data, a heuristic algorithm is used to perform global optimization on the number of hidden layer nodes and learning rate of the BP neural network; step 3: using the training set data, adopting the BP neural network architecture, taking the structural feature matrix of the composite armor as input and the protection feature matrix as output, training to obtain a composite armor protection performance prediction model; step 4: using the test set data, verifying the accuracy and generalization ability of the performance prediction model, and obtaining a composite armor protection performance prediction model based on the BP neural network. The present invention can realize the digital characterization of anti-penetration performance and provide guidance for the design of composite armor structures.

Description

Technical Field

[0001] The invention belongs to the field of armor protection of tank armored vehicles and relates to a composite armor performance prediction method based on BP neural network and particle swarm algorithm, which realizes rapid and high-precision characterization of the composite armor's anti-penetration performance. Background Art

[0002] Ceramic composite structure is one of the most important lightweight additional armor for armored vehicles and one of the effective technical approaches to reduce the weight of armor protection systems. Accurately characterizing the mapping relationship between composite armor structural parameters and performance parameters is the basis and prerequisite for structural design and material selection. The performance parameters of composite structures have a complex nonlinear relationship with structural parameters, material parameters and incoming threat parameters, and it is difficult to characterize their protective performance using theoretical analytical methods. Using experimental methods to predict the performance of composite armor requires a lot of manpower, material resources and time, and has a low cost-effectiveness ratio. Summary of the invention

[0003] In order to solve the above technical problems existing in the prior art, the present invention proposes a composite armor performance prediction method based on BP neural network and particle swarm algorithm, comprising the following steps:

[0004] Step 1: Use numerical simulation methods to establish a numerical simulation model of composite armor and obtain a database of composite armor structure parameters and protection performance parameters;

[0005] Step 2: randomly divide the database into training set data and test set data, use the training set data, and use a heuristic algorithm to globally optimize the number of hidden layer nodes P and the learning rate η of the BP neural network;

[0006] Step 3: Using the training set data, adopting the BP neural network architecture, taking the structural feature matrix of the composite armor as input and the protection feature matrix as output, training to obtain the composite armor protection performance prediction model;

[0007] Step 4: Input the test set data in step 2 into the composite armor protection performance prediction model trained in step 3 to obtain the protection feature matrix prediction value corresponding to the test set data.

[0008] Furthermore, in step 1, the composite armor structure parameters include projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ, the protection performance parameters include penetration depth h and aftereffect penetration depth m, and the projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ of the incoming threat are set to have a value range of [a1 a2], [b1 b2], [c1 c2] and [ρ1 ρ2 … ρ n], where n is the number of material types. After each numerical simulation, the penetration depth h of the projectile into the composite armor and the after-effect penetration m of the after-effect plate are extracted. The projectile velocity a, projectile weight b, and the thickness c of each unit layer are taken at intervals of Δa, Δb, and Δc, respectively. The material density ρ of each unit layer is in the range of [ρ1 ρ2 … ρ n ] are randomly selected, the structural parameters are randomly combined, and a composite armor structural parameter-protection performance parameter database is obtained.

[0009] Furthermore, in step 2, the projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ are taken as the structural characteristic matrix of the composite armor, and the penetration depth h and aftereffect penetration m are taken as the protection characteristic matrix of the composite armor. The database obtained in step 1 is randomly divided into training set data and test set data according to the principle of 4:1 of training set data volume: test set data volume.

[0010] Furthermore, in step 2, the heuristic algorithm includes a particle swarm algorithm, a genetic algorithm or a differential evolution algorithm.

[0011] Furthermore, in step 2, the search ranges of the number of hidden layer nodes P and the learning rate η are [P min , P max ] and [η min , η max ], the fitness function of the heuristic algorithm is the average value of the prediction error of the training data by N-fold cross validation, and the optimal number of hidden layer nodes P of the BP neural network is obtained by iteration. best and the learning rate η best .

[0012] Furthermore, in step 3, the BP neural network architecture includes an input layer, a hidden layer and an output layer, wherein the number of input layers and output layers is 1, the number of hidden layers is at least 1, the number of neurons in the input layer is equal to the data dimension of the composite armor structure characteristic matrix, the number of neurons in the output layer is equal to the dimension of the composite armor protection characteristic matrix, and the number of neurons in the first hidden layer is greater than the number of neurons in the input layer.

[0013] The present invention can realize the quantitative characterization of the structural parameters and protective performance parameters of composite armor, and can guide the structural design of composite armor. The present invention aims at the accurate prediction of the protective performance of composite armor, combines numerical simulation and machine learning means, combines machine learning algorithms, and aims at the problem of accurate characterization of the protective performance of composite armor. It gives full play to the advantages of machine learning algorithms in high-dimensional data processing, establishes a composite armor structural parameter-protective performance mapping model, realizes the digital characterization of the anti-penetration performance of composite armor, provides guidance for the structural design of composite armor, and can greatly improve the design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention has 5 drawings, among which:

[0015] Figure 1 It is a schematic diagram of a composite armor solid model;

[0016] Figure 2 It is a flow chart of the optimization process of the number of hidden layer nodes and learning rate of BP neural network using particle swarm algorithm;

[0017] Figure 3 It is a schematic diagram of the BP neural network architecture;

[0018] Figure 4 It is the BP neural network error-iteration number change curve;

[0019] Figure 5 It is a schematic diagram comparing the predicted and actual values ​​of composite armor protection characteristics. DETAILED DESCRIPTION

[0020] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0021] The machine learning algorithm can automatically learn the rules and patterns of the input data set, has strong nonlinear mapping capabilities, and can directly use the structure-performance data of the composite armor for iterative optimization to obtain a mapping model that characterizes the protective performance.

[0022] The technical solution of a specific embodiment of the present invention is as follows:

[0023] Step 1: Use Solidworks or Creo 3D modeling software to build a composite armor geometric model, and use LS-DYNA or Abaqus numerical simulation software to build a composite armor numerical simulation model by dividing the grid, selecting material type, applying constraints, and other steps.

[0024] Set the value ranges of incoming threat projectile velocity a, projectile weight b, thickness of each unit layer c, and material density ρ to [a1a2], [b1 b2], [c1 c2], [ρ1 ρ2 … ρ n ], where n is the number of material types. After each numerical simulation, the penetration depth h of the projectile into the composite armor and the after-effect penetration m of the after-effect plate are extracted. The projectile velocity, projectile weight, and thickness are taken at intervals of Δa, Δb, and Δc, respectively. The material density of each unit layer is in [ρ1 ρ2 … ρ n ] are randomly selected.

[0025] The above structural parameters are randomly combined to construct a database containing four types of structural parameters, namely projectile velocity, projectile weight, thickness and material density, and two types of protection parameters, namely penetration depth and aftereffect penetration depth.

[0026] Step 2: With bullet velocity, bullet weight, thickness of each unit layer of composite armor and material density as the structural characteristic matrix of composite armor, penetration depth and aftereffect penetration depth as the protection characteristic matrix of composite armor, the database obtained in step 1 is randomly divided into training set and test set data according to the principle of 4:1 training set: test set data volume. Use heuristic algorithms such as particle swarm algorithm or differential evolution algorithm to globally optimize the two parameters of the hidden layer node number P and learning rate η of the BP neural network. The search ranges of the hidden layer node number P and learning rate η are [P min , P max ]、[η min , η max ], the fitness function of the optimization algorithm is the average value of the prediction error of the training data by N-fold cross validation, and the number of hidden layer nodes P must be an integer during the parameter optimization process. The optimal number of hidden layer nodes P of the BP neural network is obtained by iteration best , learning rate η best .

[0027] Step 3: Use the training set data and the BP neural network architecture, take the structural feature matrix of the composite armor as input and the protection feature matrix as output, and train to obtain the composite armor protection performance prediction model by setting the network parameters such as the number of neural network layers, the number of hidden layer nodes, the activation functions of the hidden layer and the output layer, the training algorithm, the training accuracy and the learning rate.

[0028] Step 4: Use the test set data in step 2 and input it into the composite armor protection performance prediction model in step 3. Compare the actual penetration depth and after-effect penetration depth of the composite armor with the predicted penetration depth and after-effect penetration depth to verify the accuracy and generalization ability of the performance prediction model.

[0029] in:

[0030] (1) The BP neural network architecture includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the data dimension of the composite armor structure feature matrix, and the number of neurons in the output layer is equal to the dimension of the composite armor protection feature matrix.

[0031] (2) In step 2, the number of hidden layer nodes and learning rate of the BP neural network are globally optimized, and heuristic algorithms such as particle swarm algorithm, genetic algorithm, differential evolution algorithm, etc. can be used.

[0032] In a specific embodiment of the present invention,

[0033] Step 1: Use Solidworks to build a numerical solid model of the composite armor, where the composite armor consists of a panel and a back plate. The dimensions of the two unit layers are 100mm×100mm, the dimensions of the after-effect plate are 100mm×100mm×80mm, and the distance between the after-effect plate and the composite armor back plate is 60mm. The materials of the composite armor panel layer are aluminum oxide, silicon carbide, and boron carbide, with densities of 4.1g / cm 3 、3.1g / cm 3 and 2.52g / cm 3 The back plate materials are armor steel, titanium alloy and aluminum alloy, with densities of 7.85g / cm 3 4.45g / cm 3 and 2.75g / cm 3 , the after-effect plate material is 603 armor steel. The value ranges of the incoming threat projectile velocity a (unit: m / s), projectile weight b (unit: g), thickness of each unit layer c (unit: mm), and material density are set to [750850], [0.8 1.4], [5 20], [7.95 4.45 2.754.1 3.1 2.52], respectively. The projectile velocity a, projectile weight b, and thickness of each unit layer c are set at intervals of 10, 0.1, and 0.5, respectively. The structural parameters can be randomly combined to obtain a composite armor solid model with wide discrete interval structural parameters.

[0034] Attached Figure 1 It is a schematic diagram of the composite armor entity model, where 1 represents the projectile body, 2 represents the composite armor panel, 3 represents the composite armor back plate, and 4 represents the after-effect plate.

[0035] LS-DYNA was used to establish a numerical simulation model of composite armor, in which the material of the projectile was steel, the contact of each unit layer material was set to erosion contact, the projectile, composite armor, and aftereffect plate all used eight-node three-dimensional solid meshes with a mesh size of 1 mm, the constitutive model of metal materials used the Johnson Cook model, and the constitutive model of ceramic materials used the JohnsonHolmquist Ceramics model to characterize the damage failure process of composite armor under impact. After each calculation, the penetration depth h of the projectile into the composite armor and the aftereffect penetration depth m of the aftereffect plate were extracted.

[0036] Through step 1, a database of four types of structural parameters, namely projectile velocity, projectile weight, thickness and material density, and two types of protection parameters, namely penetration depth and aftereffect penetration depth, can be established.

[0037] In a specific embodiment of the present invention,

[0038] Step 2: The projectile velocity, projectile weight, thickness of each unit layer of the composite armor, and material density are used as the structural characteristic matrix of the composite armor, with a dimension of 1×6; the penetration depth and aftereffect penetration depth are used as the protection characteristic matrix of the composite armor, with a dimension of 1×2.

[0039] The 2176 sets of data obtained in step 1 are randomly divided into training set and test set data according to the principle of 4:1 training set: test set data volume, of which 1740 sets of training set data and 436 sets of test set data. The particle swarm algorithm (PSO algorithm) is used to globally optimize the two parameters of the number of hidden layer nodes P and learning rate η of the BP neural network, where the search ranges of the number of hidden layer nodes P and learning rate η are [5, 15] and [0.01, 0.1] respectively, where P is an integer. The number of iterations of the optimization algorithm is set to 100, and the fitness function is the average value (R) of the prediction error of the training data by 5-fold cross validation, and the calculation formula is shown in Formula 1.

[0040]

[0041] Among them, y i1 and i1 are the label value and predicted value of the first output feature of the i-th group of samples, y i2 and i2 are the label value and predicted value of the first output feature of the i-th group of samples, respectively, and k is the number of samples.

[0042] Attached Figure 2 This is a flowchart of the optimization process of the number of hidden layer nodes and learning rate of the BP neural network using the particle swarm algorithm. The training set data is used to calculate the individual extreme value and global maximum value of the particle swarm, and the optimal solution is determined by combining the fitness function value, thereby obtaining the optimal solution for the two parameters of the BP neural network. Finally, the optimal number of hidden layer nodes P is obtained through iteration. best =10, learning rate η best =0.02.

[0043] In a specific embodiment of the present invention,

[0044] Step 3: Using the 1740 sets of data obtained in step 2, a composite armor performance prediction model was established using the BP neural network architecture.

[0045] As attached Figure 3As shown in the figure, the BP neural network consists of 3 layers, including 1 input layer, 1 output layer and 1 hidden layer. The structural feature matrix of the composite armor is used as input, the protection feature matrix is ​​used as output, the number of neurons in the input layer is 6, the number of neurons in the hidden layer is 10, and the number of neurons in the output layer is 2. The activation function of the hidden layer is the tansig function, the activation function of the output layer is the logsig function, the weights and thresholds of each layer are randomly generated, and the learning rate is 0.02.

[0046] The iteration number of the BP neural network is set to 200, and the training algorithm adopts the Levenberg-Marquardt algorithm (LM algorithm). The loss function of the network is the root mean square error (RMSE) between the composite armor protection performance label value and the predicted value. The calculation formula is shown in Formula 2:

[0047]

[0048] Using 1740 sets of training data and the above neural network architecture, a composite armor protection performance prediction model was obtained. Figure 4 This is the neural network error-iteration number change curve. The horizontal axis is the iteration number, and the vertical axis is the root mean square error of the training set. When the network is trained to 168 times, the model has basically converged, and the root mean square error is 2.166.

[0049] In a specific embodiment of the present invention,

[0050] Step 4: Use the test set data from step 2 and input it into the composite armor protection performance prediction model trained in step 3 to obtain the protection feature matrix prediction value corresponding to the test set data.

[0051] The predicted values ​​of the composite armor protection characteristic matrix are compared with the actual values ​​to verify the accuracy and generalization ability of the performance prediction model. Figure 5 The figure is a comparison diagram of the predicted and actual values ​​of composite armor protection characteristics. The horizontal axis is the penetration depth value and the aftereffect penetration depth value, and the vertical axis is the penetration depth (mm). The yellow represents the actual value of the protection characteristics, and the blue represents the predicted value. Figure 5 It can be seen that the errors between the actual values ​​and the predicted values ​​of the two output features are 2.75% and 5.88% respectively, indicating that the constructed composite armor performance prediction model based on BP neural network can be used to characterize the anti-penetration performance of armor.

[0052] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A composite armor performance prediction method based on BP neural network and particle swarm algorithm, characterized in that: The steps include: Step 1: Use numerical simulation methods to establish a numerical simulation model of composite armor and obtain a database of composite armor structure parameters and protection performance parameters; Step 2: randomly divide the database into training set data and test set data, use the training set data, and use a heuristic algorithm to globally optimize the number of hidden layer nodes P and the learning rate η of the BP neural network; Step 3: Using the training set data, adopting the BP neural network architecture, taking the structural feature matrix of the composite armor as input and the protection feature matrix as output, training to obtain the composite armor protection performance prediction model; Step 4: Input the test set data in step 2 into the composite armor protection performance prediction model trained in step 3 to obtain the protection feature matrix prediction value corresponding to the test set data.

2. The composite armor performance prediction method based on BP neural network and particle swarm algorithm according to claim 1 is characterized in that: In step 1, the composite armor structure parameters include projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ, and the protection performance parameters include penetration depth h and aftereffect penetration depth m. The projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ of the incoming threat are set to have a value range of [a1 a2], [b1 b2], [c1 c2] and [ρ1 ρ2 … ρ n ], where n is the number of material types. After each numerical simulation, the penetration depth h of the projectile into the composite armor and the after-effect penetration m of the after-effect plate are extracted. The projectile velocity a, projectile weight b, and the thickness c of each unit layer are taken at intervals of Δa, Δb, and Δc, respectively. The material density ρ of each unit layer is in the range of [ρ1 ρ2 … ρ n ] are randomly selected, the structural parameters are randomly combined, and a composite armor structural parameter-protection performance parameter database is obtained.

3. The composite armor performance prediction method based on BP neural network and particle swarm algorithm according to claim 2 is characterized in that: In step 2, projectile velocity a, projectile weight b, thickness of each unit layer c and material density ρ are taken as the structural characteristic matrix of the composite armor, penetration depth h and aftereffect penetration depth m are taken as the protection characteristic matrix of the composite armor, and the database obtained in step 1 is randomly divided into training set data and test set data according to the principle of 4:1 of training set data volume: test set data volume.

4. The composite armor performance prediction method based on BP neural network and particle swarm algorithm according to claim 3 is characterized in that: In step 2, the heuristic algorithm includes a particle swarm algorithm, a genetic algorithm or a differential evolution algorithm.

5. The composite armor performance prediction method based on BP neural network and particle swarm algorithm according to claim 4 is characterized in that: In step 2, the search ranges of the number of hidden layer nodes P and the learning rate η are [P min , P max ] and [η min , η max ], the fitness function of the heuristic algorithm is the average value of the prediction error of the training data by N-fold cross validation, and the optimal number of hidden layer nodes P of the BP neural network is obtained by iteration. best and the learning rate η best .

6. The composite armor performance prediction method based on BP neural network and particle swarm algorithm according to claim 1 is characterized in that: In step 3, the BP neural network architecture includes an input layer, a hidden layer and an output layer, wherein the number of input layers and output layers are both 1, the number of hidden layers is at least 1, the number of neurons in the input layer is equal to the data dimension of the composite armor structure characteristic matrix, the number of neurons in the output layer is equal to the dimension of the composite armor protection characteristic matrix, and the number of neurons in the first hidden layer is greater than the number of neurons in the input layer.

Images