A ground combat victory rate prediction method and system based on deep learning

Abstract

The application discloses a ground combat victory rate prediction method and system based on deep learning, and relates to the technical field of battle prediction. The method comprises the following steps: acquiring time series of battlefield situation data of two parties in a ground combat scene; inputting the time series of the battlefield situation data into a ground combat scene real-time victory rate prediction model to obtain a victory rate prediction value of the two parties; and the ground combat scene real-time victory rate prediction model is obtained by training a double-layer LSTM neural network by using a combat data set. The application improves the efficiency and accuracy of victory rate prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battle prediction, in particular to a ground combat victory rate prediction method and system based on deep learning. BACKGROUND

[0002] In the real environment of ground combat, there are many battle objects, complex coordination relationship, frequent maneuvering, and rapid change of battlefield situation, which brings great challenges to battlefield comprehensive situation assessment. Due to the complexity, partial observability and dynamic real-time change of ground combat, relying on artificial cognition of situation development cannot meet the demand. SUMMARY

[0003] The purpose of the present application is to provide a ground combat victory rate prediction method and system based on deep learning, which improves the efficiency and accuracy of victory rate prediction.

[0004] To achieve the above purpose, the present application provides the following scheme:

[0005] A ground combat victory rate prediction method based on deep learning, comprising:

[0006] Obtaining the battlefield situation data time series of the two parties in combat under the ground combat scene;

[0007] Inputting the battlefield situation data time series into a ground combat scene real-time victory rate prediction model to obtain the victory rate prediction value of the two parties in combat; the ground combat scene real-time victory rate prediction model is obtained by training a double-layer LSTM neural network using a combat data set.

[0008] Optionally, each element in the battlefield situation data time series includes a maneuvering feature; the maneuvering feature includes a basic feature, a battle characteristic feature and an equipment platform feature;

[0009] The basic feature includes the number of equipment, the reconnaissance range and the control range of the two parties in combat; the battle characteristic feature includes the difference between the number of equipment of the two parties in combat and the difference between the number of equipment platforms of the two parties in combat; the equipment platform feature includes the number of tracked armored vehicles, the number of fixed-wing unmanned aerial vehicles and the number of ground unmanned vehicles.

[0010] Optionally, the control range is the union of the striking ranges of tracked armored vehicles, fixed-wing unmanned aerial vehicles and ground unmanned vehicles.

[0011] Optionally, each element in the battlefield situation data time series also includes an equipment statistical feature; the equipment statistical feature includes a cooperation coefficient and an antagonistic score between the equipment of the two parties in combat;

[0012] The cooperation coefficient between the equipment of the two parties in combat is expressed as

[0013] The confrontation score between the equipment of the two opposing sides is represented as

[0014] Wherein, H i represents the weapon equipment set of the red side in the two opposing sides, H j represents the weapon equipment set of the blue side in the two opposing sides, C H represents the cooperation matrix, R H represents the confrontation matrix.

[0015] Optionally, the battlefield situation data time sequence is represented as X = [x t-4 , x t-3 , …, x t ] T , x t-4 represents the feature data at t-4 time, x t-3 represents the feature data at t-3 time, x t represents the feature data at t time.

[0016] The battlefield situation data time sequence X is input into the ground combat scene real-time win rate prediction model, and the obtained is the win rate prediction value at t+1 time.

[0017] Optionally, the training process of the double-layer LSTM neural network comprises:

[0018] The double-layer LSTM neural network is trained by taking the sample data time sequence in the combat data set as input and the win rate prediction value of the two opposing sides as output, and the double-layer LSTM neural network with a prediction error less than a set value is taken as the ground combat scene real-time win rate prediction model.

[0019] Optionally, when the double-layer LSTM neural network is trained, the Adam optimization algorithm is used to adjust the parameters in the double-layer LSTM neural network.

[0020] The application discloses a ground combat win rate prediction system based on deep learning, comprising:

[0021] A battlefield situation data time sequence acquisition module is configured to acquire battlefield situation data time sequences of two opposing sides in a ground combat scene.

[0022] A win rate prediction module is configured to input the battlefield situation data time sequence into a ground combat scene real-time win rate prediction model to obtain a win rate prediction value of the two opposing sides.

[0023] According to the embodiments provided by the application, the following technical effects are disclosed:

[0024] The battlefield situation data time sequence acquired in real time is input into the trained double-layer LSTM neural network, and a win rate prediction value of the two parties in combat is obtained, so that the efficiency and accuracy of win rate prediction are improved. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0026] Figure 1 A ground combat win rate prediction method flowchart based on deep learning is provided for the embodiments of the present application.

[0027] Figure 2 A prediction flowchart of the bilinear model and the double-layer LSTM neural network is provided for the embodiments of the present application.

[0028] Figure 3 A time series prediction schematic diagram of the double-layer LSTM neural network is provided for the embodiments of the present application.

[0029] Figure 4 A ground combat win rate prediction system structure schematic diagram based on deep learning is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0031] The purpose of the present application is to provide a ground combat win rate prediction method and system based on deep learning, which improves the efficiency and accuracy of win rate prediction.

[0032] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0033] Embodiment 1

[0034] The present embodiment provides a ground combat win rate prediction method based on deep learning, as shown in the figure, the method comprises: Figure 1

[0035] ​Step 101: Obtain time series of battlefield situation data of both sides in a ground combat scene.

[0036] The application is a real-time victory rate prediction method for ground combat, and the main equipment of ground combat is composed of various armored vehicles, low-altitude armed helicopters, detection devices (radar and sensors) and several types of missiles.

[0037] The real-time combat behavior of the red and blue sides in the ground combat situation is simulated by using battlefield incentive simulation software, and the real-time battlefield situation data of the ground combat at each time during simulation is obtained.

[0038] In order to reduce data redundancy and simplify model input, battlefield situation data is extracted every 5s, 120 data frames are obtained in 10min, and a total of 2880 frames of data are obtained after 4 hours of simulation. In each frame of data, the number of equipment, damage degree, and amount of ammunition carried are mainly included.

[0039] According to the obtained simulation battlefield situation data, the basic features are directly extracted, and the basic features mainly include some obvious information that is helpful for victory rate prediction, such as equipment quantity, reconnaissance range and control range.

[0040] The reconnaissance range is the union of the ranges that can be detected by various detection devices.

[0041] The application also extracts senior information related to ground combat time, describes the performance of the feature at different time nodes, mainly including two categories: one is the effect of battle data at different time points; the other is the effect of weapon equipment at different time nodes. The battle data type can reflect the difference in all resources obtained by the two teams through the battle, mainly including the difference in the number of equipment survival, platform facility survival and occupied landform at different time nodes. For example, the maximum value, minimum value, mean value and difference in quantity between the current time node and the adjacent time node of the weapon equipment, equipment platform facility in the whole combat scene are needed to be counted. The statistical features include battle characteristics (real-time) features and equipment platform (inherent) features.

[0042] In order to establish an accurate and interpretable real-time battlefield victory rate prediction model, both equipment platform (inherent) information and battle characteristics (real-time) information need to be considered. As shown in the formula (1), the equipment model and the real-time model are combined to predict the combat victory rate, wherein the equipment model is mainly captured by a bilinear model, and the real-time victory rate of the combat is predicted by a long short-term memory network (LSTM) model. Figure 2

[0043] ​In the environment of ground combat, the rationality of weapon equipment configuration determines the development trend of combat to some extent, and a reasonable formation is the basis for winning the war, so the equipment configuration characteristics of red and blue sides and the real-time situation data of both sides are taken as the input of the model. The rationality of weapon equipment configuration can be measured by the single-point blasting capability of weapon equipment, the synergy capability between equipment, the suppression and damage capability to enemy equipment, etc. The real-time data of the battle includes the survival number of equipment, the survival number of platform facilities, etc.

[0044] Each element in the battlefield situation data time sequence includes a maneuvering feature; the maneuvering feature includes a basic feature, a battle characteristic feature and an equipment platform feature.

[0045] The basic feature includes the survival equipment number and survival facility number of the two parties in combat, the reconnaissance range, the control range, the weapon equipment and combination, the weapon type and combination; the battle characteristic feature includes the difference value of the equipment number, the difference value of the equipment platform number, the control range difference and the detection range difference of the two parties in combat; the equipment platform feature includes the number of tracked armored vehicles, the number of fixed-wing unmanned aerial vehicles and the number of ground unmanned vehicles.

[0046] The control range is the union of the striking ranges of tracked armored vehicles, fixed-wing unmanned aerial vehicles, ground unmanned vehicles and missiles.

[0047] Each element in the battlefield situation data time sequence also includes an equipment statistical feature; the equipment statistical feature includes the cooperation coefficient and the confrontation score between the equipment of the two parties in combat.

[0048] The number of given weapon equipment is 1-N, and the set of weapon equipment of the two parties in combat is represented as H={H1, H2, …, HN}. N The weapon equipment of the red and blue teams in each red-blue confrontation is represented as T r h ={H i} and T b h ={H j}. For weapon equipment H i , the feature vector is represented as h i ∈R V , and H∈R N×V represents the feature matrix of the equipment.

[0049] The present application proposes a bilinear model to quantify the cooperation and suppression relationship between ground equipment. First, the correlation between different equipment is extracted.

[0050] The synergy scoring function S C H (i, j) is introduced to calculate the cooperation coefficient of equipment i and equipment j, and the cooperation coefficient is represented as

[0051] wherein C H ∈R V×V is the cooperative matrix of H i and H j , H im denotes the mth element in H i , H jn denotes the nth element in H j , C denotes the element in the mth row and the nth column in C H .

[0052] An adversarial score function S R H (i, j) is introduced to calculate the adversarial score of H i and H j , and the adversarial score is expressed as

[0053] wherein R H ∈R V×V is the adversarial matrix of H i and H j .

[0054] wherein H i denotes the weapon equipment set of the red side in the combat, H j denotes the weapon equipment set of the blue side in the combat, C H denotes the cooperative matrix, R H denotes the adversarial matrix, denotes the element in the mth row and the nth column in R H .

[0055] The cooperative matrix and the adversarial matrix are statistically obtained, for example, the armored vehicles have a series of attributes such as range, cross-country ability, penetration ability, detection ability, the aircraft has a series of attributes such as high-altitude attack ability, stealth penetration ability, high-speed raid ability, and the cooperative and restrictive relationships between two attributes are counted to construct the matrix.

[0056] Step 102: inputting the battlefield situation data time series into the ground combat scene real-time win rate prediction model to obtain the win rate prediction value of the combatant sides; the ground combat scene real-time win rate prediction model is obtained by training a double-layer LSTM neural network using a combat data set.

[0057] Five continuous node data are sequentially selected as input, and the battlefield situation data time series at time node t is expressed as X = [x t-4 , x t-3 , …, x t ] T , x t-4 denotes the feature data at t-4 moment, xt-3 x represents the characteristic data at time t-3. t This represents the characteristic data at time t.

[0058] like Figure 3 As shown, the battlefield situation data time series X is input into the ground combat scenario real-time win rate prediction model, and the result is the win rate prediction value at time t+1.

[0059] Based on the predicted win rate at time t+1, decisions are made regarding the next operational step. Specific decisions include: Real-time assessment of the battle situation can be made based on changes in the predicted win rate. When our win rate continues to rise, commanders need to combine other strength comparisons, threat assessments, intent recognition, and other auxiliary technologies to identify the enemy's true operational intentions. When our win rate continues to fall, commanders need to contact friendly forces for reinforcements or change operational behavior to reduce casualties. When our win rate enters a period of fluctuation, we need to prepare for a protracted war or adjust our operational layout to achieve final victory.

[0060] This invention trains a two-layer LSTM neural network using a simulation dataset (combat dataset) to obtain a real-time win rate prediction model for ground combat. The parameters of this model are then fine-tuned using a historical dataset to obtain a refined real-time win rate prediction model. Finally, this refined model is used for real-time win rate prediction.

[0061] The training process of the dual-layer LSTM neural network includes: using the time series of battlefield situation sample data from the combat dataset as input, and the predicted win rates of the two combat parties corresponding to the time series of battlefield situation sample data as output, the dual-layer LSTM neural network is trained; the dual-layer LSTM neural network with a prediction error less than a set value is used as the real-time win rate prediction model for ground combat scenarios; during the training of the dual-layer LSTM neural network, the Adam optimization algorithm is used to adjust the parameters in the dual-layer LSTM neural network. More specifically, the training process of the dual-layer LSTM neural network also includes:

[0062] Normalize the real-time battlefield situation sample data in the simulation dataset (combat dataset).

[0063] The simulation dataset is divided into a simulation training set and a simulation test set using a staggered segmentation method.

[0064] Training a real-time win rate prediction model for ground combat using a simulation dataset includes:

[0065] The battlefield situation data from the simulation training set is input into the real-time win rate prediction model for ground combat to obtain the real-time win rate prediction for the corresponding battlefield situation data.

[0066] For any real-time battlefield situation data in the simulation training set, the variance of the win rate prediction value corresponding to the battlefield situation data and the actual win rate corresponding to the situation data is calculated to obtain the win rate prediction error of the battlefield situation data.

[0067] According to the prediction error of each battlefield situation data in the simulation training set, the parameters in the ground combat win rate prediction model are corrected to obtain the trained ground combat win rate prediction model.

[0068] The accuracy of the trained ground combat win rate prediction model is tested by using the simulation test set.

[0069] The variance of the simulation prediction win rate and the true win rate is calculated according to the following formula:

[0070]

[0071] Where y pre is the win rate prediction value of the model at time i, y i is the actual win rate at time i, and n is the time span.

[0072] The Adam optimization algorithm is used to correct the parameters in the trained ground combat win rate prediction model; the Adam optimization algorithm is shown in the following formula:

[0073]

[0074] Where momentum t is the momentum term, β1 and β2 are hyperparameters, t represents the time, E is the prediction error, w t is the parameter of the trained win rate prediction model at time t, v t is the velocity term at time t, and learning_rate is the learning rate.

[0075] Embodiment 2

[0076] The embodiment provides a ground combat win rate prediction system based on deep learning, as shown in Figure 4 The system comprises:

[0077] A battlefield situation data time sequence acquisition module 201 is configured to acquire battlefield situation data time sequences of two parties in conflict in a ground combat scene.

[0078] A win rate prediction module 202 is configured to input the battlefield situation data time sequences into a ground combat scene real-time win rate prediction model to obtain win rate prediction values of the two parties in conflict; the ground combat scene real-time win rate prediction model is obtained by training a double-layer LSTM neural network using a combat data set.

[0079] The various embodiments described in this specification are presented for the purpose of illustrating the principles of the application and its best mode of operation. Each of the embodiments described in this specification has been provided for the purpose of illustration and is not intended to limit the application. The same or similar reference numerals in different drawings represent the same or similar elements.

[0080] The principles and operation of the present application have been explained so far with the help of specific examples. The examples have been presented for the purpose of illustration and are not intended to limit the application. The application described in this specification can be implemented in hardware and / or software that is modified to operate in accordance with the principles set forth in this specification. Additionally, the description and drawings are to be regarded as illustrative in nature and their objects are to be understood not only to be the solution suggested in the examples, but to include any and all implementations within the scope of the application.

Claims

1. A deep learning-based method for predicting the win rate in ground combat, characterized in that, include: Acquire time series of battlefield situation data for both sides in ground combat scenarios; The battlefield situation data time series is input into the real-time win rate prediction model for ground combat scenarios to obtain the win rate prediction values ​​of both sides; the real-time win rate prediction model for ground combat scenarios is obtained by training a two-layer LSTM neural network using combat datasets; Each element in the battlefield situation data time series includes maneuver characteristics; the maneuver characteristics include basic characteristics, combat characteristic characteristics, and equipment platform characteristics. The basic characteristics include the number of equipment, reconnaissance range, and control range of both sides; the combat characteristics include the difference in the number of equipment and the difference in the number of equipment platforms of both sides; the equipment platform characteristics include the number of tracked armored vehicles, the number of fixed-wing UAVs, and the number of ground unmanned vehicles. Each element in the battlefield situation data time series also includes equipment statistical features; the equipment statistical features include the cooperation coefficient and confrontation score between the equipment of the two sides in the battle. The cooperation coefficient between the equipment of the two sides in the conflict is expressed as: The score for the equipment of the two sides in combat is represented as follows: Among them, H i H represents the collection of weapons and equipment of the Red side in the war. j C represents the collection of weapons and equipment of the Blue side in the war. H Let R represent the cooperative matrix. H Represents the adversarial matrix; The battlefield situation data time series is represented as X = [x t-4 x t-3 ,…,x t ] T x t-4 x represents the characteristic data at time t-4. t-3 x represents the characteristic data at time t-3. t This represents the characteristic data at time t; The battlefield situation data time series X is input into the ground combat scenario real-time win rate prediction model, and the win rate prediction value at time t+1 is obtained. The training process of the two-layer LSTM neural network includes: Using the time series of sample data from the combat dataset as input and the predicted win rates of both sides as output, a two-layer LSTM neural network is trained. The two-layer LSTM neural network with a prediction error less than a set value is used as a real-time win rate prediction model for ground combat scenarios. When training a two-layer LSTM neural network, the Adam optimization algorithm is used to adjust the parameters of the two-layer LSTM neural network.

2. The deep learning-based ground combat win probability prediction method according to claim 1, characterized in that, The control range is the union of the strike ranges of tracked armored vehicles, fixed-wing drones, and ground unmanned vehicles.

3. A ground combat win probability prediction system based on deep learning, characterized in that, The deep learning-based ground combat victory probability prediction system applies the deep learning-based ground combat victory probability prediction method of claim 1, and the deep learning-based ground combat victory probability prediction system includes: The battlefield situation data time series acquisition module is used to acquire the battlefield situation data time series of both sides in ground combat scenarios. The win rate prediction module is used to input the time series of battlefield situation data into the real-time win rate prediction model of ground combat scenarios to obtain the win rate prediction values ​​of both sides in the battle; the real-time win rate prediction model of ground combat scenarios is obtained by training a two-layer LSTM neural network using combat datasets.

Images