Multilayer concrete penetration trajectory prediction method fusing physical information CNN
By constructing a CNN model that integrates physical information, the penetration process is decomposed into a single-layer target penetration stage and an inter-target flight stage. Combining a dual-branch prediction architecture driven by data and physics, the problem of low accuracy and efficiency in ballistic prediction in existing technologies is solved, and high-precision and fast ballistic trajectory prediction is achieved.
Patent Information
- Application Number
- CN202511842628.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-01-23
AI Technical Summary
Existing technologies struggle to quickly and accurately predict a projectile's attitude changes, trajectory, and velocity decay during penetration of a multi-layered concrete target. Traditional methods are either costly or computationally inefficient, making rapid trajectory prediction difficult.
A CNN model integrating physical information is constructed to decompose the penetration process into a single-layer target penetration stage and an inter-target flight stage. Combining a data-driven and physics-driven dual-branch prediction architecture, the accurate prediction of the projectile's motion state is achieved through a one-dimensional convolutional neural network and a physical constraint penalty term.
It achieves high-precision and fast ballistic trajectory prediction, with good generalization performance and computational efficiency, and reduces research costs.
Smart Images
Figure CN121389807A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of projectile penetration trajectory prediction, and particularly relates to a multi-layer concrete penetration trajectory prediction method fusing a CNN of physical information. BACKGROUND
[0002] In order to effectively attack a solid target such as a command building with a multi-layer spaced concrete structure, a penetration weapon is usually used to penetrate the multi-layer floor and then detonate at a predetermined position. However, due to the fact that the projectile incidence attitude is usually not vertical and the thickness of each layer of the target is relatively thin, the projectile attitude and trajectory are prone to deflection during penetration, which seriously affects the damage performance. Therefore, it is of important engineering application value to quickly and accurately predict the attitude change, motion trajectory and speed attenuation of the projectile during penetration for weapon damage efficiency evaluation and design optimization. Since the penetration process of the projectile penetrating through the multi-layer concrete target involves multi-variable coupling, combined with the non-uniformity of the concrete material and the complexity of the dynamic mechanical properties thereof, the determination of the projectile attitude and trajectory during penetration becomes a complex nonlinear problem. In the traditional penetration trajectory research method, the multi-layer concrete target penetration test is high in cost and the initial conditions are difficult to guarantee; the theoretical analysis method often needs to introduce simplified assumptions and involves complex iterative calculations, and has a small applicable range; although the numerical simulation method can accurately reflect the actual deflection behavior of the projectile, since the number of target layers in the multi-layer concrete target penetration model is large, the projectile occupies a large amount of computing resources during the flight between the target layers, resulting in low calculation efficiency and difficulty in realizing rapid prediction of the trajectory. With the rapid development of artificial intelligence technology, the data-driven scientific research paradigm provides a new technical approach to solve such problems. SUMMARY
[0003] The application aims to overcome the shortcomings of the existing penetration trajectory prediction method and provide a multi-layer concrete penetration trajectory prediction method fusing a CNN of physical information.
[0004] The technical solution for achieving the application is as follows: a multi-layer concrete penetration trajectory prediction method fusing a CNN of physical information, comprising the following steps:
[0005] Step (1): constructing a single-layer concrete target penetration dataset, containing multiple groups of input features and corresponding output parameters; the input features include projectile parameters, target parameters and projectile-target encounter conditions, and the output parameters are motion state parameters of the projectile after penetrating through the single-layer target;
[0006] Step (2): based on the single-layer concrete target penetration dataset, constructing and training a single-layer target prediction model of a deep learning network fusing a physical mechanism;
[0007] Step (3): the trained single-layer target prediction model is used for single-layer target penetration prediction to predict the motion state parameters of the projectile after penetrating the i-th layer of target; a free flight model between the projectile and the target is established based on rigid body dynamics;
[0008] Step (4): for a given multi-layer spaced concrete target configuration and projectile initial condition, the post-target motion state parameters of the i-th layer of target are input into the free flight model between the projectile and the target to calculate the projectile-target encounter condition parameters before the projectile hits the i+1-th layer of target;
[0009] Step (5): the calculated projectile-target encounter condition parameters and the projectile and target parameters are input into the single-layer target prediction model to predict the motion state parameters of the projectile after penetrating the i+1-th layer of target;
[0010] Step (6): steps (4) and (5) are executed in a loop until the final motion state of the projectile after penetrating all the preset target layers is predicted.
[0011] Further, in step (1), the projectile parameters include diameter d, length L, CRH, mass m and distance l between the projectile tip and the center of mass; the target parameters include strength f c and thickness H t ; the projectile-target encounter conditions include velocity v, angular velocity ω, attack angle α and incidence angle φ; the post-target projectile motion parameters include post-target horizontal velocity v z , vertical velocity v y , attitude angle β, post-target angular velocity ω' and vertical distance h y between the projectile tail after penetrating a single layer of target and the projectile tip before penetration.
[0012] Further, in step (1), the input and output parameters of the data set are normalized and preprocessed, and the values of each parameter are mapped to the interval [0, 1], and the normalization formula is: , where x' is the normalized value of the feature and label, x is the original feature and label value, x min and x max are the minimum and maximum values of the input and output values; and the training set and the test set are divided according to the proportions of 80% and 20%.
[0013] Further, in step (2), the single-layer target prediction model includes a data-driven unit, a physics-driven unit and a feature fusion layer.
[0014] Further, the data-driven unit is a pure deep learning path, which is a one-dimensional convolutional neural network 1D-CNN model with attention mechanism, including:
[0015] an input layer for receiving an input tensor with a shape of (N, 11, 1), where N is the batch size;
[0016] The first one-dimensional convolution layer uses 32 convolution kernels with a size of 2, a step of 1, and a two-end zero padding strategy, followed by a ReLU activation function and a 2x1 maximum pooling layer;
[0017] The second one-dimensional convolution layer uses 64 convolution kernels with a size of 2, a step of 1, and a two-end zero padding strategy, followed by a ReLU activation function and a 2x1 maximum pooling layer;
[0018] The flattening layer is used to convert the feature map into a feature vector.
[0019] The first fully connected layer contains 128 neurons and uses a ReLU activation function.
[0020] The output layer contains 5 neurons and uses a linear activation function to output the predicted 5 target post-movement state parameters.
[0021] Further, the loss function used by the data-driven unit to train the one-dimensional convolutional neural network model is a combined loss function, and its expression is:
[0022]
[0023] wherein, is the mean square error between the predicted value and the true value; is a physical constraint penalty term for penalizing prediction results that violate physical laws; is a weight coefficient, with a value range of 0.05 to 0.1;
[0024] Physical constraint penalty term At least includes the following cases: the horizontal velocity v of the projectile out of the target z is greater than the initial velocity v into the target, the absolute value of the angular velocity ω' of the projectile out of the target is greater than a preset threshold, or the incidence angle φ of the projectile is greater than the critical ricochet angle.
[0025] Further, the physics-driven unit does not contain any trainable parameters, and is used to receive the same projectile parameters, target parameters and projectile-target encounter conditions as in step (1), and directly calculates the theoretical prediction value of the projectile's post-target movement state according to the projectile penetration resistance model.
[0026] Further, the feature fusion layer introduces a learnable weight gating mechanism, which is a trainable fully connected neural network layer. The weight value is dynamically generated according to the input projectile-target parameters, and is used to represent the credibility ratio of the data-driven prediction result and the physics-driven prediction result under the current input condition; including in the complex broken zone condition, more believing the prediction of the data-driven branch, in the rigid penetration stage, more believing the prediction of the physics-driven branch.
[0027] Further, in step (4), the distance between the tip of the projectile and the target surface of the second layer target when the projectile moves to the distance l between the center of mass and the target is calculated:
[0028] ,
[0029] After judging the distance, the angle of incidence of the projectile penetrating the next layer of target is obtained And the angle of attack .
[0030] Further, the distance judgment method in step (4) is as follows:
[0031] When d ≤d0, the current angle of incidence and angle of attack are directly output; if d >d0, enter the time iteration loop, the iteration step is Δt, the projectile position is updated and d is recalculated in each iteration, if the distance d is still negative and the absolute value is greater than d0, continue to increase the time until d ≤ d0, the current angle of incidence and angle of attack are output; if d changes from negative to positive, that is, the tip of the projectile penetrates the target surface of the second layer target to the right side of the target surface, stop iteration, and output the angle of incidence and angle of attack of the last position of the tip of the projectile on the left side of the target surface in the last step.
[0032] Compared with the prior art, the present application has the following advantages:
[0033] (1) The model has high solving precision, can not only learn the complex nonlinear relationship in the data, but also strictly follow the basic physical law. The key parameters such as the attitude, trajectory and speed attenuation of the projectile can be accurately predicted;
[0034] (2) The model has high calculation efficiency, and the trajectory prediction can be completed in a few minutes;
[0035] (3) The model has good generalization performance and can predict different projectile structures, target structures and projectile-target encounter conditions;
[0036] (4) A reliable prediction model can be established based on existing numerical simulation or historical test data, which effectively reduces the research cost. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 is the process flow chart of the multi-layer concrete penetration trajectory prediction method of the present application fusing physical information CNN.
[0038] Figure 2 is the schematic diagram of the projectile-target flight process of the multi-layer concrete penetration trajectory prediction method of the present application fusing physical information CNN.
[0039] Figure 3 is the trajectory test result diagram of the multi-layer concrete penetration trajectory prediction method of the present application fusing physical information CNN.
[0040] Figure 4 Trajectory calculation result diagram of the multi-layer concrete penetration trajectory prediction method of the fusion physical information CNN of the application. DETAILED DESCRIPTION
[0041] The application will be described in further detail below with reference to the drawings.
[0042] A multi-layer concrete penetration trajectory prediction method of a fusion physical information CNN, which divides the process of a projectile penetrating a multi-layer spaced concrete target into two independent physical stages of single-layer target penetration and inter-target flight, and constructs a systematic prediction framework including a data-driven part, a physics-driven part, a feature fusion layer, and a projectile inter-target flight process calculation part. First, a single-layer concrete target penetration dataset is constructed, and a single-layer target penetration prediction model that fuses physical information is constructed based on a one-dimensional convolutional neural network; at the same time, a physics-driven part is introduced to form a double-branch prediction architecture. Through a designed learnable gating fusion mechanism, the confidence of the results of data-driven and physics-driven is dynamically balanced, thereby constructing a single-layer target penetration prediction model that has both physical consistency and data adaptability. Further combined with a projectile inter-target free flight dynamics model, a multi-layer spaced concrete thin target trajectory prediction model that has both rapidity and high accuracy is established. The radius of the projectile is r, the curvature radius of the projectile head is R, the thickness of the target body is H, the multi-layer targets are arranged in parallel with a spacing of H, the target body inclination angle θ is the angle between each target body and the vertical direction, the projectile attitude angle β is the angle between the projectile axis and the horizontal line, which is positive when the projectile direction is above the horizontal line, the attack angle α is the angle between the projectile axis and the velocity direction, which is positive when the projectile direction is above the velocity direction, and the incidence angle φ is the angle between the projectile axis and the target normal, which is positive when the projectile direction is above the target normal. And it is assumed that: (1) the projectile is not affected by external forces when flying between targets after penetrating the target; (2) the projectile rotates around the center of mass. t
[0043] To eliminate the influence of different variables caused by the difference in dimension and improve the efficiency of the gradient descent algorithm to find the optimal solution, the maximum and minimum value strategy is used to normalize the features and labels of the data set of projectile penetrating single-layer concrete thin target. The obtained data is randomly shuffled, and then divided into training set and test set according to the proportion of 80% and 20%, and the repeatability of training is ensured by setting a fixed random seed. The prediction model of the ballistic characteristics of the projectile penetrating the single-layer concrete thin target is based on the 1D-CNN architecture. The input quantities include the projectile features (5), the target features (2), and the projectile-target encounter features (4), and the output quantities are the projectile motion parameters (5). Two one-dimensional convolution layers are set. The first convolution layer uses 32 convolution kernels with a size of 2, and padding is performed at both ends of the input data to ensure that the information at all positions is treated equally. To improve the model's ability to learn complex coupled relationships, a ReLU (Rectified Linear Unit) activation function is used to introduce non-linear features. After the first convolution, the maximum pooling method is selected, and the feature map is down-sampled through the pooling layer with a size and step of 2. The output of the previous layer is input into the second convolution layer, which uses 64 convolution kernels with a size of 2. Then it is processed by the ReLU activation function and down-sampled through the same maximum pooling layer (kernel size and step are both 2). After flattening the multi-dimensional feature map into a one-dimensional vector of length 64, it is input into the fully connected network, which contains two fully connected layers: the first layer linearly maps the 64-dimensional input vector to 128 hidden units, and then uses the ReLU activation function to enhance the non-linear representation ability; the second layer as the output layer, further linearly maps the 128-dimensional features to 5 output nodes, corresponding to the five prediction targets. In the training stage, the weights and biases of the convolution layers and fully connected layers are first initialized, the model output is obtained through forward calculation, and the loss function used is the combined loss function, whose expression is: . Where, is the mean square error between the predicted value and the true value; is the physical constraint penalty term, which is used to punish the prediction results that violate the physical laws; is the weight coefficient, with a value range of 0.05 to 0.1. In the backpropagation process, the gradient is calculated layer by layer in the opposite order of forward propagation, and the Adam optimizer is used to update the model parameters until 500 training cycles are completed. Thus, the ballistic prediction model of the projectile penetrating the single-layer concrete thin target is established.
[0044] After inputting the initial parameters into the single-layer target ballistic prediction model, the target post projectile motion parameters are output, and if there is a next layer of target, the projectile inter-target flight process is further calculated. When the projectile penetrates through the single-layer concrete target and the tail of the projectile just emerges from the target, the attitude angle is β1, and the angle between the velocity and the horizontal direction is γ:
[0045] (1)
[0046] When the velocity direction is above the horizontal line, γ is positive. The velocity v of the projectile after leaving the target can be decomposed into horizontal velocity v z and vertical velocity v y The center of mass of the projectile flies along the velocity direction after leaving the target, and the distance S1 between the center of mass of the projectile after leaving the target and the back of the previous layer of target is:
[0047] (2)
[0048] The flight distance S of the center of mass of the projectile before the distance l from the target is:
[0049] (3)
[0050] The flight time t of the center of mass of the projectile before the distance l from the target is:
[0051] (4)
[0052] The attitude angle β' of the projectile when moving to the distance l from the target is:
[0053] (5)
[0054] The incidence angle φ' of the projectile at this time is:
[0055] (6)
[0056] The attack angle α' of the projectile is:
[0057] (7)
[0058] A coordinate system is established with the contact point of the first layer of target with the ground as the origin, the horizontal velocity direction of the projectile as the x-axis, and the vertical velocity direction as the y-axis. The vertical distance h between the center of the projectile tail and the bottom of the target when the projectile leaves the target is h, and the coordinates of the center of mass of the projectile when it just leaves the target are:
[0059] (8)
[0060] The coordinates of the center of mass of the projectile when it moves to the distance l from the target are:
[0061] (9)
[0062] The coordinates of the tip of the projectile at this time are:
[0063] (10)
[0064] The equation of the back of the first layer of target is:
[0065] (11)
[0066] The second layer target surface equation is:
[0067] (12)
[0068] The distance d of the bullet tip to the second layer target surface when the bullet moves to the distance l of the mass center to the target body is:
[0069] (13)
[0070] When d ≤ d0, the current incidence angle and attack angle are directly output; if d > d0, a time iteration cycle is entered, the iteration step is Δt, the bullet position is updated and d is recalculated in each iteration, if the distance d is still negative and the absolute value is greater than d0, the time is continuously increased, until d ≤ d0, the current incidence angle and attack angle are output; if d changes from negative to positive (i.e. the bullet tip penetrates the second layer target surface to the right side of the target surface), the iteration is stopped, and the incidence angle and attack angle of the last position (i.e. the bullet tip is located on the left side of the target surface) of the previous step are output. The speed, incidence angle, attack angle and angular velocity of the bullet impacting the next layer target body are finally obtained, and the bullet-target encounter conditions are updated to further calculate the motion posture of the bullet after penetrating the target body.
[0071] Embodiment
[0072] In an embodiment of the present application, a multi-layer concrete penetration trajectory prediction method based on a physical information CNN is provided. In this embodiment, the trajectory characteristics of a sharp-ovate bullet penetrating a two-layer spaced concrete target are predicted. As shown in Figure 1 , the method comprises the following steps.
[0073] Step 1: A single-layer concrete target penetration data set is constructed, containing 127 groups of input features and corresponding output parameters. The input features include bullet parameters, target parameters and bullet-target encounter conditions, and the output parameters are the motion state parameters of the bullet after penetrating the single-layer target; through , the input and output parameters of the data set are normalized and preprocessed, and the parameter values are mapped to the interval [0, 1], and the training set and the test set are divided according to the proportions of 80% and 20%.
[0074] Step 2: Based on the single-layer concrete target penetration data set, a deep learning network single-layer target prediction model integrating physical mechanisms is constructed and trained, a total of 500 training cycles are set, and during the training process , the training average value is always 0, indicating that all the prediction results do not violate the physical laws; , the training average value is stably around 0.0012, and the test average value is stably around 0.0019, indicating that the model has good prediction performance.
[0075] Step 3: Re-instantiate the CNN model with the same structure as in the training phase, load the saved best model parameter file, and ensure that the model structure, parameters, and data preprocessing method are consistent with the training phase. The initial parameters including the projectile diameter d = 30 mm, length L = 180 mm, CRH = 4, mass m = 520 g, projectile tip to mass center distance l = 95.58 mm, target body strength f c = 40 MPa, thickness H t = 100 mm, speed v = 705 m / s, angular velocity ω = 0° / ms, angle of attack α = -0.7°, and incidence angle φ = 15.4° are input into the single-layer target ballistic prediction model to output the projectile motion parameters after penetrating the target, and the horizontal velocity v z = 628 m / s, vertical velocity v y = -7.58 m / s, attitude angle β = -2.8°, target post-angle velocity ω' = -0.77° / ms, and the vertical distance h y = 3.84 mm between the projectile tail after penetrating the single-layer target and the projectile tip before penetrating is calculated.
[0076] Step 4: The target body spacing L = 470 mm, the projectile moves to the critical distance d0 = 5 mm from the second layer target surface before the mass center distance l from the target body, and the iteration step is Δt = 0.01 ms. The projectile-target encounter condition parameters of speed v = 628 m / s, angular velocity ω = -0.58° / ms, angle of attack α = 2.69°, and incidence angle φ = 11.02° are calculated through the inter-target free flight part. The initial parameters are input into the single-layer target ballistic prediction model to calculate the horizontal velocity v z = 556 m / s, vertical velocity v y = -59.6 m / s, attitude angle β = -8.6°, target post-angle velocity ω' = 4.52° / ms, and the vertical distance h y = 3.45 mm between the projectile tail after penetrating the second layer target and the projectile tip before penetrating is calculated.
[0077] Step 5: The speed v = 559 m / s and attitude angle β = -3.31°
[0078] The above-described preferred embodiments of the present application are not intended to limit the present application, and various modifications and changes can be made by those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A multi-layered concrete penetration trajectory prediction method fusing physical information CNN, characterized in that, The method comprises the following steps: Step (1): constructing a single-layer concrete target penetration dataset, including multiple sets of input features and corresponding output parameters; the input features include projectile parameters, target body parameters, and projectile-target encounter conditions, and the output parameters are the motion state parameters of the projectile after penetrating the single-layer target; Step (2): based on the single-layer concrete target penetration dataset, a deep learning network single-layer target prediction model is constructed and trained, which integrates physical mechanisms; Step (3): the trained single-layer target prediction model is used for single-layer target penetration prediction to predict the motion state parameters of the projectile after penetrating the i-th target; a rigid body dynamics-based projectile-target free flight model is established; Step (4): for a given multi-layer spaced concrete target configuration and projectile initial condition, the post-target motion state parameters of the i-th target are input into the projectile-target free flight model to calculate the projectile-target encounter condition parameters before the projectile hits the i+1-th target; Step (5): the calculated projectile-target encounter condition parameters and the projectile and target body parameters are input into the single-layer target prediction model to predict the motion state parameters of the projectile after penetrating the i+1-th target; Step (6): steps (4) and (5) are repeatedly executed until the final projectile motion state after the projectile penetrates all preset target layers is predicted.
2. The prediction method of claim 1, wherein, The parameters of the projectile in step (1) include diameter d, length L, CRH, mass m and the distance between the tip of the projectile and the center of mass l; the parameters of the target include strength f c and thickness H t ; the conditions of the projectile-target encounter include velocity v, angular velocity ω, angle of attack α and angle of incidence φ; the parameters of the projectile after the target include horizontal velocity v z , vertical velocity v y , attitude angle β, angular velocity ω' after the target and the vertical distance h y between the tail of the projectile after penetrating the single-layer target and the tip of the projectile before penetration.
3. The method of claim 1, wherein, The input and output parameters of the data set in step (1) are normalized and preprocessed, and each parameter value is mapped to the interval [0, 1], and the normalization formula is: , wherein x' is the normalized value of the feature and label, x is the original feature and label value, x min and x max are the minimum and maximum values of the input value and the output value; and the training set and the test set are divided according to the proportions of 80% and 20%.
4. The method of claim 1, wherein, The single-layer target prediction model in step (2) comprises a data-driven unit, a physically-driven unit and a feature fusion layer.
5. The method of claim 4, wherein, The data-driven unit is a pure deep learning path, which is a one-dimensional convolutional neural network 1D-CNN model with an attention mechanism, comprising: an input layer for receiving an input tensor with a shape of (N, 11, 1), wherein N is the batch size; a first one-dimensional convolutional layer using 32 convolutional kernels with a size of 2 and a step of 1, and adopting a two-end zero padding strategy, followed by a ReLU activation function and a 2x1 max pooling layer; a second one-dimensional convolutional layer using 64 convolutional kernels with a size of 2 and a step of 1, and adopting a two-end zero padding strategy, followed by a ReLU activation function and a 2x1 max pooling layer; a flattening layer for converting the feature map into a feature vector; a first fully connected layer containing 128 neurons and using a ReLU activation function; an output layer containing 5 neurons and using a linear activation function for outputting the predicted 5 post-target motion state parameters.
6. The method of claim 5, wherein, The loss function used for training the one-dimensional convolutional neural network model of the data-driven unit is a combined loss function, and its expression is: ， wherein, is the mean square error between the predicted value and the true value; is a physical constraint penalty term for penalizing the predicted result that violates the physical law; is a weight coefficient, and the value range is 0.05 to 0.
1. Physical constraint penalty term At least includes the penalty for the following cases: the bullet out of the target level speed v z The absolute value of the bullet out of the target angular velocity ω' is greater than the preset threshold, or the bullet incidence angle φ is greater than the critical skip angle.
7. The method of claim 6, wherein, The physically-driven unit does not contain any trainable parameters, and is used to receive the same projectile parameters, target body parameters and projectile-target encounter conditions as in step (1), and directly calculates the theoretical prediction value of the post-projectile motion state according to the projectile penetration resistance model.
8. The method of claim 7, wherein, The feature fusion layer introduces a learnable weight gating mechanism, which is a trainable fully connected neural network layer, and the weight value is dynamically generated according to the input projectile-target parameters, representing the credibility ratio of the data-driven prediction result and the physically-driven prediction result under the current input condition; including more believing in the prediction of the data-driven branch under complex broken zone conditions, and more believing in the prediction of the physically-driven branch in the rigid penetration stage.
9. The method of claim 8, wherein, In step (4), the distance between the tip of the projectile and the second layer target surface when the projectile moves to the distance l between the center of mass and the target is calculated: ， the incidence angle of the projectile penetrating the next layer of target body after distance judgment and the angle of attack .
10. The method of claim 9, wherein, The distance judgment method in step (4) is as follows: When d ≤ d0, the current incidence angle and attack angle are directly output; if d > d0, the time iteration loop is entered, the iteration step is Δt, the projectile position is updated and d is recalculated in each iteration, if the distance d is still negative and the absolute value is greater than d0, the time is continuously increased, until d ≤ d0, the current incidence angle and attack angle are output; if d changes from negative to positive, that is, the tip of the projectile penetrates the second layer target surface to the right side of the target surface, the iteration is stopped, and the incidence angle and attack angle of the last position of the tip of the projectile on the left side of the target surface in the previous step are output.