Method for acquiring killing matrix estimation model based on convolutional neural network

Abstract

The invention relates to a method for acquiring a killing matrix estimation model based on a convolutional neural network, belongs to the technical field of damage probability prediction, and solves the problems of complex calculation and low accuracy of the killing probability of a killing matrix in the prior art. Weapon parameters and the type, the environment and the position of a target are set based on the combat mission; setting a calculation range of weapon parameters in the killing matrix based on the types and performance parameters of the weapons; setting various working condition parameters based on the calculation range, and calculating a joint killing probability matrix based on the killing probability matrix of the warhead to the target corresponding to each working condition parameter and the shock wave damage probability matrix of the type of weapons to the target; performing conversion and recombination processing on the joint killing probability matrix corresponding to each working condition parameter to form a killing matrix training sample data set; and training the two-stage network model based on the killing matrix training sample data set to obtain a killing matrix estimation model. And the prediction efficiency and precision of the damage probability in the killing matrix are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of damage probability prediction, and in particular, to a method for obtaining a kill matrix prediction model based on a convolutional neural network. Background Art

[0002] A kill matrix is a matrix used to describe the damage ability of a weapon system to different targets. It usually includes the damage probability or effect caused by the weapon system to the target under different conditions (such as distance, angle, target type, etc.). When calculating the kill matrix, various factors need to be considered, such as the type, quantity, and position of explosives, and the material, structure, and size of the target. The interaction between these factors is often non-linear and difficult to accurately describe using traditional mathematical models.

[0003] In recent years, neural network models have achieved certain applications in the field of damage assessment. In particular, neural networks have powerful non-linear mapping capabilities and self-learning capabilities, and can extract useful features from a large amount of data and establish complex relationships between inputs and outputs. However, in the kill matrix prediction task, traditional neural networks still have certain limitations, such as poor spatial relationships, loss of information features, and low prediction accuracy.

[0004] Therefore, a method for obtaining a kill matrix prediction model based on a convolutional neural network is needed to improve the prediction efficiency and accuracy of the damage probability in the kill matrix. Summary of the Invention

[0005] In view of the above analysis, embodiments of the present invention aim to provide a method for obtaining a kill matrix prediction model based on a convolutional neural network to solve the problems of complex calculation and low accuracy of the kill probability of the existing kill matrix.

[0006] Embodiments of the present invention provide a method for obtaining a kill matrix prediction model based on a convolutional neural network, including: Setting the type, performance parameters, and warhead type of the weapon, as well as the type, environment, and position of the target based on the combat mission; Setting the calculation range of the weapon parameters in the kill matrix based on the type and performance parameters of the weapon, where the calculation range of the weapon parameters includes: altitude range, speed range, pitch angle range, azimuth angle range, and kill level range; Setting a variety of working condition parameters based on the calculation range, calculating the kill probability matrix of the warhead against the target and the shock wave damage probability matrix of this type of weapon against the target corresponding to each working condition parameter, and calculating the joint kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix; Convert and reorganize the combined kill probability matrix corresponding to each operating condition parameter to form a kill matrix training sample data set; Train a two-stage network model based on the kill matrix training sample data set to obtain a kill matrix prediction model.

[0007] Based on a further improvement of the above method, the types of the weapons include missiles, shells, and explosives; the performance parameters include: range, accuracy, and explosion power; the types of warheads include: high-explosive bombs, armor-piercing shells, and cluster bombs; the types of targets include: buildings, vehicles, and personnel clusters.

[0008] Based on a further improvement of the above method, calculating the kill probability matrix of the warhead against the target and the shock wave damage probability matrix of this type of weapon against the target corresponding to each operating condition parameter includes: calculating the fragment power field data corresponding to each operating condition parameter, where the fragment power field data includes: the mass, velocity, and solid angle of the fragment area in different angular domains; calculating the kill probability of this type of warhead against the target at different positions in the kill matrix based on the fragment power field data, and the kill probabilities at each position form a kill probability matrix; calculating the shock wave damage probability of this type of weapon against the target at the corresponding position in the kill matrix based on the shock wave radius at different positions in the kill matrix, and the shock wave damage probabilities at each position form a shock wave damage probability matrix.

[0009] Based on a further improvement of the above method, calculating the combined kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix includes: calculating the combined kill probability based on the kill probability at each position in the kill probability matrix and the shock wave damage probability at this position in the shock wave damage probability matrix, and the combined kill probabilities at all positions form a combined kill probability matrix.

[0010] Based on a further improvement of the above method, calculating the kill probability of this type of warhead against the target at different positions in the kill matrix based on the fragment power field data includes: A1: Taking the position of the target as the origin, establish a target coordinate system; group the fragments in different angular domains according to mass to form multiple fragment mass groups, and the masses of the fragments in each fragment mass group are the same; A2: Perform the following operations for each position: Place the warhead at the corresponding position in the target coordinate system according to the weapon height defined in the current operating condition parameter, and use the shooting line analysis method for each angular domain to obtain the vulnerable area and exposure probability of the target at this position; Calculate the kill probability of this type of warhead against the target at this position based on the vulnerable area, the exposure probability, and each fragment mass group.

[0011] Based on the further improvement of the above method, calculating the kill probability of this type of warhead at this position based on the vulnerable area, the exposure probability, and each fragment mass group includes: B1: Calculate the kill probability of each group of the fragment mass group at this position: , where i is the number of the fragment mass group, is the target vulnerable area corresponding to the i-th fragment mass group, is the target exposure probability corresponding to the i-th fragment mass group, is the number of fragments in the i-th fragment mass group, is the shortest distance from the warhead to the target to be destroyed, is the solid angle of the fragment area; B2: Calculate the kill probability of this type of warhead at this position based on the kill probabilities of all fragment mass groups at this position: , where, , and N is the number of fragment mass groups.

[0012] Based on the further improvement of the above method, calculating the shock wave damage probability of this type of weapon to the target at the corresponding position in the kill matrix based on the shock wave radii at different positions in the kill matrix includes: calculating the shock wave damage probabilities of this type of weapon at different positions in the kill matrix , , where SR is the shock wave radius, , are constants.

[0013] Based on the further improvement of the above method, calculating the joint kill probability based on the kill probability at each position in the kill probability matrix and the shock wave damage probability at this position in the shock wave damage probability matrix includes: .

[0014] Based on the further improvement of the above method, performing a transformation and recombination process on the joint kill probability matrix corresponding to each working condition parameter to form a kill matrix training sample data set includes: For each working condition parameter, kill matrix training sample data can be constructed, and each kill matrix training sample data is , where M and N are the dimensions of the joint kill probability matrix, is the height of the warhead, is the velocity of the warhead, is the azimuth angle of the warhead, is the pitch angle of the warhead, is the abscissa of this type of warhead in the target coordinate system, is the ordinate of this type of warhead in the target coordinate system, is the kill level of the warhead, is at the said warhead in the kill probability.

[0015] Based on the further improvement of the above method, the two-stage network model includes: A one-stage network model, the input data of the one-stage network model is the working condition parameters and the position information in the kill matrix , and the output data is the predicted kill probability corresponding to each predicted kill probability , , ; A matrix generation model, which combines all corresponding predicted kill probabilities to form a one-stage predicted kill matrix; A two-stage network model, the input data of the two-stage network model is the one-stage predicted kill matrix, and the output data is the two-stage predicted kill matrix, and the two-stage predicted kill matrix is used as the final predicted kill matrix; Among them, the one-stage network model is a one-dimensional convolutional neural network, the two-stage network model is a convolutional network based on the attention mechanism, and the loss function of the two-stage network model is: , Among them, , are constants, is the kill probability of the warhead at the said place in the kill matrix training sample dataset , is the two-stage predicted kill probability at the said place in the two-stage predicted kill matrix of the two-stage network model .

[0016] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects: The present invention provides a method for obtaining a kill matrix prediction model based on a convolutional neural network, studies the damage probability of specific weapons to targets, expands the training sample set by constructing sample data under different working conditions, and can further improve the accuracy of the prediction model; by using the neural network prediction method, it can effectively extract the feature representations under different heights, azimuth angles, elevation angles, and projectile velocities, solve the complex non-linear mapping problem from input to output, obtain the kill matrix of targets within a range of 100 meters, and finally realize the rapid prediction of the kill matrix of specific weapons against specific targets under different parameter conditions, solve the problem that the non-linear relationship between various factors in the kill matrix cannot be accurately expressed in the prior art, and at the same time overcome the disadvantages of large computational amount and low accuracy of the existing theoretical calculation method.

[0017] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the following specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings are only for the purpose of showing specific embodiments and are not considered as a limitation of the present invention. Throughout the drawings, the same reference signs represent the same components; Figure 1 It is an example diagram of a method for obtaining a kill matrix prediction model based on a convolutional neural network in an embodiment of the present invention; Figure 2 It is an example diagram of the elevation angle and azimuth angle in an embodiment of the present invention; Figure 3 It is a side view of the fragment scatter angle domain in an embodiment of the present invention; Figure 4 It is an example diagram of the solid angle distribution of the fragment range in an embodiment of the present invention; Figure 5 It is an example diagram of the shooting line grid of the target model in an embodiment of the present invention; Figure 6 It is an example diagram of the relationship between the shock wave radius and the shock wave damage probability in an embodiment of the present invention; Figure 7 It is an example diagram of the two-stage network model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will specifically describe the preferred embodiments of the present invention in conjunction with the drawings. The drawings constitute a part of this application and are used together with the embodiments of the present invention to explain the principle of the present invention, and are not used to limit the scope of the present invention.

[0020] The present invention provides a method for obtaining a kill matrix model based on a convolutional neural network. First, based on the combat mission, the strike scenario is determined. Based on the parameters of a specific weapon, the range of parameters in the kill matrix is set. Based on the above parameters, a sample set for training the convolutional neural network is established. Then, a two-stage convolutional neural network model is trained based on the sample set, and the trained two-stage convolutional neural network model is used as the final kill matrix prediction model. The technical solution proposed by the present invention overcomes the problem in the prior art that the non-linear relationship of each parameter in the kill matrix cannot be accurately described. Moreover, due to the adoption of the convolutional neural network and training sample data that conform to the real scenario, the kill probability predicted by the finally obtained kill matrix model is closer to the real probability, and the calculation cost is reduced.

[0021] A specific embodiment of the present invention discloses a method for obtaining a kill matrix prediction model based on a convolutional neural network, as Figure 1 shown, including: S1: Based on the combat mission, set the type, performance parameters, and warhead type of the weapon, as well as the type, location environment, and position of the target.

[0022] Among them, the types of weapons include missiles, shells, explosives, etc. After determining the weapon type, the performance parameters of the weapon can be determined, including range, accuracy, explosion power, etc. Then, analyze the warhead type of the weapon, such as high-explosive bombs, armor-piercing bombs, cluster bombs, etc. Finally, determine the action mechanism of the weapon on the target (that is, the damage mechanism research on the target based on the fragmentation kill and shock wave kill to the target's damage mechanism). The target in the present invention refers to the specific target to be struck. The types of targets include: buildings, vehicles, personnel clusters, etc. After determining the target, study the location environment of the target, such as the terrain, climate, and whether there are shelters at the target location, study the position of the target, that is, the specific coordinates of the target location, and the attributes of the target, such as parameters such as target size, structure, and material.

[0023] It can be understood that the technical solution of the present invention can finally obtain a kill matrix prediction model for different targets when applied to different combat missions.

[0024] S2: Based on the type of the weapon and the performance parameters, set the calculation range of the weapon parameters in the kill matrix. The calculation range of the weapon parameters includes: altitude range, speed range, pitch angle range, azimuth angle range, and kill level range.

[0025] Among them, for the height parameter, it refers to the vertical distance between the weapon and the local horizontal ground. For example, the vertical distance between a missile in flight and the local horizontal ground. For the speed parameter, it refers to the flight speed or traveling speed of the weapon. For example, the flight speed of a missile. For the pitch angle parameter, it refers to the angle between the longitudinal axis of the weapon and the horizontal ground. For the azimuth angle parameter, it refers to the angle of the weapon relative to a fixed direction (usually the due north direction). Refer to Figure 2 , which gives the positional relationship between the pitch angle and the azimuth angle in the target coordinate system. For the kill level parameter, it is used to evaluate the strike effect of the weapon system on the target. According to different classification criteria, the kill level can have different divisions.

[0026] In the present invention, the kill levels are mainly divided into the following three categories: light kill, medium kill, and heavy kill. Among them, the degree of light kill is usually relatively low. After a light kill is executed, the target can still continue to perform its basic tasks; for medium kill, after a medium kill is executed, the basic functions of the target are significantly degraded and it cannot perform its normal functions for a long time and needs to be repaired. Medium kill can achieve the purpose of affecting the target's mission execution; for heavy kill, after a heavy kill is executed, the main parts of the target are severely destroyed and it loses its basic functions and cannot continue to perform other tasks.

[0027] The present invention is illustrated by taking a certain fragmentation bomb and a certain light armored vehicle as examples. Refer to Figure 2 for the geometric relationship of weapon-target interaction, which represents a working condition of an airburst weapon. It propagates an explosive shock wave and many fragments from a detonation point. Define the point O directly below the center of mass of the warhead as the point of burst projection, that is, the origin of the ground plane coordinate system. Assume that the velocity vector of the weapon coincides with the weapon axis. Extend the weapon axis to the ground plane to get the point R. The straight line OR defines the x-axis (or longitudinal axis), and the y-axis or transverse coordinate is perpendicular to OR. The target characterized by the vulnerable center is located at (x, y) relative to the point of burst projection of the weapon detonation point in the ground plane coordinates. It can be seen from the figure that the target is at a specific deflection angle of the weapon, while the weapon is at a specific elevation angle of the target.

[0028] After determining the specific weapon for the strike target in step S1, the calculation range of the weapon parameters in the kill matrix can be set based on the type and performance parameters of the weapon. Exemplarily, the speed range of the weapon can be set to , and the kill level range is medium kill and heavy kill. For the parameter ranges of different combat missions, researchers can set them flexibly as long as they can meet the requirements of the combat missions.

[0029] S3: Set multiple working condition parameters based on the calculated range, calculate the kill probability matrix of the warhead against the target corresponding to each working condition parameter and the shock wave damage probability matrix of this type of weapon against the target, and calculate the joint kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix.

[0030] After determining the effective range of each parameter in step S2, a random generation algorithm can be used to generate multiple groups of working condition parameters within the effective range. It can be understood that the more working condition parameters are generated, the more sample data will be ultimately used to train the two-stage network model, thereby further improving the prediction accuracy of the model. In addition, using a random generation algorithm to generate multiple groups of data is also a conventional means in the art, and the present invention does not make any limitations here, as long as a preset number of working condition parameters can be generated.

[0031] The calculation of the kill probability matrix of the warhead against the target and the shock wave damage probability matrix of this type of weapon against the target corresponding to each working condition parameter includes: S31: Calculate the fragment power field data corresponding to each working condition parameter. The fragment power field data includes: the fragment mass, velocity, and the solid angle of the fragment area in different angular domains.

[0032] The fragment power field data refers to various parameters and characteristics of the fragments when the warhead explodes, including the number of fragments, velocity, scatter angle, scatter direction angle, density distribution, and the fragment perforation area, etc. These data are crucial for evaluating the killing power and damage effect of the warhead. The present invention defines the fragment power field data as the characterization data of the fragment power such as the maximum mass, minimum mass, average mass, maximum velocity, minimum velocity, and average velocity of the fragments.

[0033] Exemplarily, the present invention divides the fragment range generated by the warhead into 16 angular domains. This angle is called the scatter angle, and the range is , divided by as the interval. The fragments are scattered from each slice thin layer on the warhead, rather than from the centroid, as shown in Figure 3 . Analyzing in three-dimensional space, the fragments in these conical regions form a distribution as shown in Figure 4 after flying to the ground. The fragment scatter is distributed along a circle, and the range is . The solid angle of the fragment range in Figure 4 is denoted as . According to the range of the angular domain, the maximum solid angle is denoted as , and the minimum solid angle is denoted as . In each scatter angle interval, fragments with different masses and velocities are randomly generated to characterize the fragment power field generated by the weapon explosion.

[0034] The present invention has developed a visualization program for calculating the fragment power field based on the C++ programming language. The input parameters are the parameters of a certain working condition, and the output parameters are the fragment mass, velocity, and solid angle of different angular regions in the fragment power field. The business logic and calculation method of the visualization program can be adjusted in real time according to the combat plan and are implemented using existing technologies, which will not be elaborated here.

[0035] S32: Calculate the kill probability of the warhead of this type on the target at different positions in the kill matrix based on the fragment power field data. The kill probabilities at each position form a kill probability matrix, including: A1: Establish a target coordinate system with the position of the target as the origin; group the fragments in different angular regions according to mass to form multiple fragment mass groups, and the masses of the fragments in each fragment mass group are the same.

[0036] See Figure 2 , and the target is equivalent to a point, denoted as the vulnerable center, and a coordinate system is established with it as the origin.

[0037] A2: Perform the following operations for each position: A21: Place the warhead at the corresponding position in the target coordinate system according to the weapon height defined in the current working condition parameters, and use the shooting line analysis method for each angular region to obtain the vulnerable area and exposure probability of the target at this position.

[0038] For the fragments in each angular region divided in step S31, use the shooting line analysis method to calculate the vulnerable area and exposure probability data of the target.

[0039] The basic composition of the vulnerable area is the shooting line module, and its action mode is: for the fragments with a given flight direction, cover the geometric model of the target with grids of the same size. Each grid contains a shooting line, and the spacing of the grid lines can be determined by oneself, but the smaller the spacing, the longer the required calculation time. Assume that the weapon explodes at a certain distance from the target, and the flight trajectories of individual fragments are parallel. Assume that the shooting line passes through the center of each grid or cell, and convert the target geometric model into a view as shown in Figure 5 , which can determine which components are inclined to which shooting line. In this way, it can be determined which components (critical and non-critical) the fragments flying along the shooting line encounter and the encounter order. Then, perform the cumulative calculation of the kill probability of each shooting line. The target vulnerable area is the sum of the vulnerable areas of all individual cells containing the target part and is only valid for this azimuth angle and elevation angle. The analysis of the shooting line is based on the penetration ability of the fragment on the component. Use the FATEPEN model to calculate the shooting line corresponding to the fragment (i.e., the movement trajectory of the fragment), count the number of cells covered by this shooting line, denoted as , and then the vulnerable area of the cell shooting line can be calculated using the following equation: , In the formula, represents the area of the target within the cell. Here, the area of the cell is used as the value. Then, by summing up the of all p cells containing the target, the vulnerable area of the target can be obtained. .

[0040] In the line-of-fire method, the exposure probability generally refers to the likelihood of a target being hit under specific conditions. The calculation of this probability usually involves multiple factors such as the geometric shape, position, and motion state of the target, as well as the position of the shooter and the performance of the weapon. The general steps for calculating the exposure probability include: (1) Target model establishment: First, a geometric model of the target needs to be established, including information such as the shape, size, and position of the target; (2) Line-of-fire generation: Based on the position of the shooter and the performance of the weapon, possible lines of fire are generated; (3) Hit judgment: Determine whether each line of fire intersects the target, that is, whether it can hit the target; (4) Probability calculation: Calculate the exposure probability based on the ratio of the number of hitting lines of fire to the total number of lines of fire.

[0041] Based on the general process steps of the exposure probability, the present invention has developed a software for calculating the exposure probability based on C++. The calculation formula for the exposure probability is: , In the formula, is the coverage probability of the power generated by the weapon (such as the hitting accuracy, etc.), is the geometric factor of the target exposure (such as area, arrangement method, etc.), is the environmental correction coefficient (such as terrain shielding, motion complexity, etc.).

[0042] A22: Calculate the kill probability of this type of warhead at this position based on the vulnerable area, the exposure probability, and each fragment mass group, including: B1: Calculate the kill probability of each group of the fragment mass group at this position: , where i is the number of the fragment mass group, is the vulnerable area of the target corresponding to the i-th fragment mass group, is the exposure probability of the target corresponding to the i-th fragment mass group, is the number of fragments in the i-th fragment mass group, is the shortest distance from the warhead to the target to be destroyed ( Figure 2 shown in), is the solid angle of the fragment area; B2: Calculate the kill probability of this type of warhead at this position based on the kill probability of all fragment mass groups at this position: , where , and N is the number of fragment mass groups.

[0043] S33: Calculate the shock wave damage probability of this type of weapon to the target at the corresponding position in the kill matrix based on the shock wave radius of this type of weapon at different positions in the kill matrix. The shock wave damage probabilities at each position form a shock wave damage probability matrix.

[0044] Calculate the shock wave damage probability of this type of weapon at different positions in the kill matrix , , where SR is the shock wave radius, , are constants. As Figure 6 shown, for different SR, their shock wave damage probabilities are different. Combine the shock wave damage probabilities corresponding to each position to form a shock wave damage probability matrix.

[0045] The calculation of the combined kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix includes: Calculate the combined kill probability based on the kill probability at each position in the kill probability matrix and the shock wave damage probability at this position in the shock wave damage probability matrix. The combined kill probabilities at all positions form a combined kill probability matrix. Specifically, the combined kill probability at each position is: , Combine the combined kill probabilities at each position to form a combined kill probability matrix.

[0046] S4: Perform transformation and recombination processing on the combined kill probability matrix corresponding to each working condition parameter to form a kill matrix training sample data set, including: For each working condition parameter, kill matrix training sample data can be constructed. Each piece of kill matrix training sample data is , where M and N are the dimensions of the combined kill probability matrix, is the height of the warhead, is the speed of the warhead, is the azimuth angle of the warhead, is the pitch angle of the warhead, is the abscissa of this type of warhead in the target coordinate system, is the ordinate of this type of warhead in the target coordinate system. is the kill level of the warhead, and is the kill probability of the warhead at

[0047] For each set of operating condition parameters, when the weapon is placed at the specified position in the kill matrix, the kill matrix corresponding to the specific weapon - target under this operating condition can be obtained at the corresponding kill level (since the vulnerable area of the target obtained when calculating the kill matrices of different levels is different, the calculated kill probabilities are also different). Therefore, for a target coordinate system and the kill matrix, one set of operating condition parameters corresponds to a joint kill probability, that is, a kill matrix training sample. Exemplarily, the dimension of the original kill matrix data is , and predictions are made within a range of 100 meters before and after the target as the center. The intervals between different kill probabilities in the matrix are 5 meters, and finally a two - dimensional matrix is formed. One matrix can be split into 400 data, that is, corresponding to 400 kill matrix training samples.

[0048] Preferably, each piece of data in the kill matrix training sample data set is pre - processed, including: data cleaning, which is used to handle missing values and outliers. For missing values, common processing methods include filling in missing values and deleting records containing missing values. For outliers, they can be screened by setting thresholds or processed using statistical methods.

[0049] Data integration refers to the process of integrating data from multiple data sources into a data set. During the data integration process, problems such as data matching, duplicate values, and inconsistent data types need to be solved.

[0050] Data transformation refers to transforming data to meet specific analysis requirements. Common transformation methods include feature scaling, discretization, encoding, etc. Data reduction is to reduce the scale of data by reducing its complexity while maintaining the original data relationship. Common reduction methods include feature selection and feature construction.

[0051] Data verification and quality inspection. When pre - processing data, it is necessary to verify the integrity and consistency of the data to ensure that the data conforms to the expected format and range. Deal with duplicate data and redundant data to improve the data quality.

[0052] Data pre - processing is to further improve the data quality. Technicians can pre - process each piece of data using corresponding pre - processing methods as needed. The present invention does not limit the specific implementation steps of data pre - processing here.

[0053] Preferably, when the kill level is used as an input parameter of the neural network, One-Hot encoding is performed to eliminate the sequential relationship.

[0054] Exemplarily, the present invention has constructed a csv file of 13,016 pieces of kill matrix data, where the names of the csv files respectively represent the id serial number (meaningless in reality), kill level, height, azimuth angle, pitch angle, and speed of the projectile. Among them, each csv file includes a kill matrix, and each value in the matrix is the kill probability.

[0055] S5: Training a two-stage network model based on the kill matrix training sample data set to obtain a kill matrix prediction model.

[0056] The kill matrix training sample data set is divided into: a training set for the preliminary learning of the parameters of the convolutional neural network model; a test set for the further optimization and adjustment of the parameters; and a validation set for evaluating the performance and generalization ability of the model. Exemplarily, the ratio can be 7:2:1. The training process of the two-stage network model is a conventional means in the art and will not be elaborated here.

[0057] The two-stage network model includes: A first-stage network model, the input data of the first-stage network model is the working condition parameters and the position information in the kill matrix , and the output data is the predicted kill probability corresponding to each ; , , ; A matrix generation model, combining all corresponding predicted kill probabilities to form a first-stage predicted kill matrix; A second-stage network model, the input data of the second-stage network model is the first-stage predicted kill matrix, and the output data is the second-stage predicted kill matrix, taking the second-stage predicted kill matrix as the final predicted kill matrix; Among them, the first-stage network model is a one-dimensional convolutional neural network, and the loss function of the first stage is the mean square error MSE between the predicted probability at a certain position and the kill probability of the true kill matrix at the same position (the same x, y position). The second-stage network model is a convolutional network based on the attention mechanism, and the loss function of the second-stage network model is: , where , are constants, is the warhead in the kill matrix training sample data set at The kill probability at is the two-stage predicted kill probability in the two-stage predicted kill matrix of the two-stage network model at .

[0058] Figure 7 The example diagram of the two-stage network model is given. The one-stage network model is a convolutional neural network model, including: an input layer, a convolutional layer, a normalization layer, a pooling layer, and a fully connected layer. The Relu activation function is used, and data samples can be input in batches. For example, 128 data samples are input at a time. The two-stage network model is a convolutional network based on the attention mechanism. The present invention does not make specific limitations on the two-stage network model, as long as the designed network can complete the kill matrix prediction. Refer to Figure 7 , for the two-stage network model, the feature matrix and the true kill matrix are subjected to MSE constraint to obtain ; after the one-dimensional vector obtained by one-dimensionalizing the feature matrix passes through two fully connected layers, a predicted effectiveness factor is obtained (the effectiveness factor is an effectiveness index used to measure the damage effectiveness of a specific weapon on a specific target). The predicted effectiveness factor and the actual effectiveness factor are constrained by the cross entropy loss function to obtain . Finally, after weighted averaging, the total loss function of the two-stage model training is obtained.

[0059] By adopting the technical solution proposed by the present invention, when the model converges, the accuracy rate on the training set reaches about 0.98, and the loss drops to about 0.0032. After one round of training, it is tested on the test set, and the accuracy rate converges to about 0.94, and the loss drops to about 0.0035. The time for one round of testing is about 2s, and the speed is increased by about 90 times compared with the original method. Therefore, compared with the prior art, a method for obtaining a kill matrix prediction model based on a convolutional neural network provided in this embodiment studies the damage probability of a specific weapon to a target, expands the training sample set by constructing sample data under different working conditions, and can further improve the accuracy of the prediction model; by adopting the neural network prediction method, it can effectively extract the feature representations at different heights, azimuth angles, elevation angles, and projectile speeds, solve the complex non-linear mapping problem from input to output, obtain the kill matrix of the target within 100 meters, and finally realize the rapid prediction of the kill matrix of a specific weapon for a specific target under different parameter conditions, solve the problem that the non-linear relationship between various factors in the kill matrix cannot be accurately expressed in the prior art, and at the same time overcome the disadvantages of large calculation amount and low accuracy of the existing theoretical calculation method.

[0060] Those skilled in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory, a random access memory, etc.

[0061] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for obtaining a kill matrix prediction model based on a convolutional neural network, characterized in that Including: Setting the type, performance parameters, and warhead type of the weapon, as well as the type, environment, and location of the target based on the combat mission; Setting the calculation range of the weapon parameters in the kill matrix based on the type of the weapon and the performance parameters, where the calculation range of the weapon parameters includes: altitude range, speed range, pitch angle range, azimuth angle range, and kill level range; Setting a variety of operating condition parameters based on the calculation range, calculating the kill probability matrix of the warhead against the target and the shock wave damage probability matrix of this type of weapon against the target corresponding to each operating condition parameter, and calculating the joint kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix; Performing conversion and recombination processing on the joint kill probability matrix corresponding to each operating condition parameter to form a kill matrix training sample data set; Training a two-stage network model based on the kill matrix training sample data set to obtain a kill matrix prediction model.

2. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 1, characterized in that, Including: The types of the weapons include missiles, shells, and explosives; The performance parameters include: range, accuracy, and explosion power; the types of the warheads include: high-explosive bombs, armor-piercing shells, and cluster bombs; the types of the targets include: buildings, vehicles, and personnel clusters.

3. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 2, where calculating the kill probability matrix of the warhead against the target and the shock wave damage probability matrix of this type of weapon against the target corresponding to each operating condition parameter includes: Calculating the fragment power field data corresponding to each operating condition parameter, where the fragment power field data includes: fragment mass, speed, and fragment area solid angle in different angular domains; Calculating the kill probability of this type of warhead against the target at different positions in the kill matrix based on the fragment power field data, and the kill probabilities at each position form a kill probability matrix; Calculating the shock wave damage probability of this type of weapon against the target at the corresponding position in the kill matrix based on the shock wave radius at different positions in the kill matrix, and the shock wave damage probabilities at each position form a shock wave damage probability matrix.

4. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 3, wherein, The calculating the joint kill probability matrix based on the kill probability matrix and the shock wave damage probability matrix includes: Calculating the joint kill probability based on the kill probability at each position in the kill probability matrix and the shock wave damage probability at this position in the shock wave damage probability matrix, and the joint kill probabilities at all positions form a joint kill probability matrix.

5. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 4, characterized in that The calculating the kill probability of this type of warhead against the target at different positions in the kill matrix based on the fragment power field data includes: A1: Taking the position of the target as the origin, establishing a target coordinate system; grouping the fragments in different angular domains according to mass to form multiple fragment mass groups, and the masses of the fragments in each fragment mass group are the same; A2: Performing the following operations for each position: Placing the warhead at the corresponding position in the target coordinate system according to the weapon altitude defined in the current operating condition parameter, and using the shooting line analysis method for each angular domain to obtain the vulnerable area and exposure probability of the target at this position; Calculate the kill probability of this type of warhead at this position based on the vulnerable area, the exposure probability, and each fragment mass group.

6. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 5, wherein The calculating the kill probability of this type of warhead at this position based on the vulnerable area, the exposure probability, and each fragment mass group includes: B1: Calculate the kill probability of each group of the fragment mass group at this position: ， where \(i\) is the number of the fragment mass group, is the target vulnerable area corresponding to the \(i\)-th fragment mass group, is the target exposure probability corresponding to the \(i\)-th fragment mass group, is the number of fragments in the \(i\)-th fragment mass group, is the shortest distance from the warhead to the target to be destroyed, is the solid angle of the fragment area; B2: Calculate the kill probability of this type of warhead at this position based on the kill probabilities of all fragment mass groups at this position: ， Among them, , N is the number of fragment mass groups.

7. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 6, characterized in that Calculating the shock wave damage probability of the weapon type at the corresponding position in the kill matrix to the target based on the shock wave radius of the weapon type at different positions in the kill matrix, including: calculating the shock wave damage probability of the weapon type at different positions in the kill matrix , ， where SR is the shock wave radius, and are constants.

8. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 7, wherein The calculating the combined kill probability based on the kill probability at each position in the kill probability matrix and the shock wave damage probability at this position in the shock wave damage probability matrix includes: 。 9. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 8, wherein, The converting and reorganizing the combined kill probability matrix corresponding to each working condition parameter to form a kill matrix training sample data set includes: For each working condition parameter, kill matrix training sample data can be constructed. Each piece of kill matrix training sample data is , where M and N are the dimensions of the joint kill probability matrix, is the height of the warhead, is the speed of the warhead, is the azimuth angle of the warhead, is the elevation angle of the warhead, is the abscissa of this type of warhead in the target coordinate system, is the ordinate of this type of warhead in the target coordinate system, is the kill level of the warhead, is the kill probability of the warhead at .​​​​​​​​​​​​​​​​​​​​​​ 10. The method for obtaining a kill matrix prediction model based on a convolutional neural network according to claim 9, wherein The two-stage network model includes: One-stage network model, the input data of the one-stage network model being the operating condition parameters and the position information in the kill matrix , the output data being the corresponding predicted kill probability , , ; Matrix generation model, all corresponding predicted kill probabilities are combined to form a first-stage predicted kill matrix; A two-stage network model, the input data of the two-stage network model is the one-stage predicted kill matrix, the output data is the two-stage predicted kill matrix, and the two-stage predicted kill matrix is used as the final predicted kill matrix; Among them, the one-stage network model is a one-dimensional convolutional neural network, the two-stage network model is a convolutional network based on the attention mechanism, and the loss function of the two-stage network model is: ， Among them, , are constants, is the kill probability of the warhead in the kill matrix training sample dataset at , is the two-stage predicted kill probability in the two-stage predicted kill matrix of the two-stage network model at .

Images