CN115480582A - Maneuver prediction method for LSTM-based target, electronic device and storage medium - Google Patents

Maneuver prediction method for LSTM-based target, electronic device and storage medium Download PDF

Info

Publication number
CN115480582A
CN115480582A CN202210986452.6A CN202210986452A CN115480582A CN 115480582 A CN115480582 A CN 115480582A CN 202210986452 A CN202210986452 A CN 202210986452A CN 115480582 A CN115480582 A CN 115480582A
Authority
CN
China
Prior art keywords
target
maneuver
prediction
value
unmanned aerial
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210986452.6A
Other languages
Chinese (zh)
Other versions
CN115480582B (en
Inventor
沈沛意
张亮
吕梦鸽
朱光明
宋娟
冯明涛
李宁
高尔扬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Shipbuilding Corp Comprehensive Technical And Economic Research Institute
Xidian University
Original Assignee
China Shipbuilding Corp Comprehensive Technical And Economic Research Institute
Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Shipbuilding Corp Comprehensive Technical And Economic Research Institute, Xidian University filed Critical China Shipbuilding Corp Comprehensive Technical And Economic Research Institute
Priority to CN202210986452.6A priority Critical patent/CN115480582B/en
Publication of CN115480582A publication Critical patent/CN115480582A/en
Application granted granted Critical
Publication of CN115480582B publication Critical patent/CN115480582B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course, altitude or attitude of land, water, air or space vehicles, e.g. using automatic pilots
    • G05D1/08—Control of attitude, i.e. control of roll, pitch, or yaw
    • G05D1/0808—Control of attitude, i.e. control of roll, pitch, or yaw specially adapted for aircraft
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course, altitude or attitude of land, water, air or space vehicles, e.g. using automatic pilots
    • G05D1/10—Simultaneous control of position or course in three dimensions
    • G05D1/101—Simultaneous control of position or course in three dimensions specially adapted for aircraft
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Control Of Position, Course, Altitude, Or Attitude Of Moving Bodies (AREA)

Abstract

本发明公开了一种基于LSTM的目标的机动预测方法、电子设备和存储介质，包括以下步骤；步骤1：构建无人机对战过程目标机机动选择数据集；步骤2：提取目标机动对抗过程空间态势特征；步骤3：对数据集进行分割和预处理；步骤4：建立基于LSTM的目标的机动预测网络模型；步骤5：利用训练集

Figure DDA0003802111100000011
对基于LSTM的目标的机动预测网络模型进行训练，并利用测试集
Figure DDA0003802111100000012
进行准确率检测。本发明不论是对单个机动控制参数的预测，还是对所有机动控制参数同时预测正确问题都能得了良好的效果。

Figure 202210986452

The invention discloses an LSTM-based target maneuver prediction method, electronic equipment and a storage medium, comprising the following steps: Step 1: constructing a target machine maneuver selection data set during the UAV battle process; Step 2: extracting the target maneuver confrontation process space Situation features; Step 3: Segment and preprocess the data set; Step 4: Establish a LSTM-based target maneuver prediction network model; Step 5: Use the training set

Figure DDA0003802111100000011
Train the LSTM-based target maneuver prediction network model and use the test set
Figure DDA0003802111100000012
Perform an accuracy test. The present invention can achieve good results no matter the prediction of a single maneuvering control parameter or the simultaneous prediction of all maneuvering control parameters.

Figure 202210986452

Description

基于LSTM的目标的机动预测方法、电子设备和存储介质LSTM-based target maneuver prediction method, electronic equipment and storage medium

技术领域technical field

本发明涉及飞行器飞行控制和人工智能技术领域，具体涉及一种基于LSTM的目标的机动预测方法、电子设备和存储介质。The invention relates to the technical fields of aircraft flight control and artificial intelligence, in particular to an LSTM-based target maneuver prediction method, electronic equipment and a storage medium.

背景技术Background technique

近年来，随着无人机相关技术的快速发展，无人机在军事领域的应用越来越广泛，对目标无人机下一阶段的机动动作进行预测问题逐渐引起人们的关注。在空战对抗过程中，双方无人机作为智能体进行策略选择下一阶段的机动动作，从而让己方无人机在空间占位上处于优势状态，而如何提前预知目标无人机下一阶段将会采取的机动动作，是己方机动动作选择策略的关键问题。In recent years, with the rapid development of UAV-related technologies, UAVs have become more and more widely used in the military field, and the problem of predicting the maneuvering actions of target UAVs in the next stage has gradually attracted people's attention. In the process of air combat confrontation, the UAVs of both sides act as intelligent agents to choose the maneuvering action of the next stage, so that the UAVs of one’s own side are in an advantageous state in terms of space occupation, and how to predict in advance the target UAV’s next stage will be The maneuvering action that will be taken is the key issue in the selection strategy of one's own maneuvering action.

目前，对目标无人机的机动动作进行预测的方法将目标机的机动动作选择看作固定的几种机动动作的加权和，各个机动动作的权重由专家知识确定。然而无人机进行机动选择是一个连续动态地对空间态势占位寻优的过程，固定的机动动作加权和的预测方法没有考虑目标机动选择的空间态势特征，忽视了目标机动动作的动态性与时序性特点，而且没有考虑目标机对我方无人机机动的感知情况，所以很难准确描述目标机动动作选择的规律，导致对目标机未来的机动动作预测不准确。At present, the method of predicting the maneuvering action of the target UAV considers the maneuvering action selection of the target aircraft as the weighted sum of several fixed maneuvering actions, and the weight of each maneuvering action is determined by expert knowledge. However, the UAV’s maneuver selection is a continuous and dynamic process of optimizing the space situation. The fixed maneuver weighted sum prediction method does not consider the space situation characteristics of the target maneuver selection, and ignores the dynamics of the target maneuver. It is difficult to accurately describe the law of target maneuver selection, which leads to inaccurate prediction of the target aircraft's future maneuvers.

发明内容Contents of the invention

为了克服以上技术的缺点，本发明的目的在于提供一种基于LSTM的目标的机动预测方法、电子设备和存储介质，不论是对单个机动控制参数的预测，还是对所有机动控制参数同时预测正确问题都能得良好的效果。In order to overcome the shortcomings of the above technologies, the object of the present invention is to provide a maneuver prediction method, electronic equipment and storage medium based on LSTM targets, whether it is the prediction of a single maneuver control parameter, or the simultaneous prediction of all maneuver control parameters. Can get good results.

为了实现上述目的，本发明采用的技术方案是：In order to achieve the above object, the technical scheme adopted in the present invention is:

一种基于LSTM的目标的机动预测方法，包括以下步骤；A maneuver prediction method based on an LSTM target, comprising the following steps;

步骤1：构建无人机对战过程的目标机动选择数据集S，所述目标指与我方无人机对战的目标无人机；Step 1: Construct the target maneuver selection data set S of the UAV battle process, and the target refers to the target UAV that is fighting with our UAV;

步骤2：提取目标机动对抗过程空间态势特征；Step 2: Extract the space situational characteristics of the target maneuvering confrontation process;

步骤3：对步骤1的目标机动选择数据集S进行分割和预处理；Step 3: Segment and preprocess the target maneuver selection data set S in step 1;

步骤4：建立基于LSTM的目标的机动预测网络模型；Step 4: Establish a maneuver prediction network model based on the LSTM target;

步骤5：利用训练集

Figure BDA0003802111080000021
对基于LSTM的目标的机动预测网络模型进行训练，并利用测试集
Figure BDA0003802111080000022
进行准确率检测。Step 5: Utilize the training set
Figure BDA0003802111080000021
Train the LSTM-based target maneuver prediction network model and use the test set
Figure BDA0003802111080000022
Perform an accuracy test.

所述步骤1利用某一对一空战仿真系统，对战双方设置不同初始状态进行多次对战实验，每局对战每隔20ms记录一次对战双方无人机的状态信息st和机动动作控制信息ct，形成空战对抗数据，并且每局对战过程以某一方的胜利或者对战时长超过2分钟时终止。The step 1 utilizes a certain one-to-one air combat simulation system, and the two parties set up different initial states to conduct multiple battle experiments, and each game records the status information st and maneuver control information c t of the drones of the two parties every 20 ms , to form air combat confrontation data, and the process of each battle ends when one side wins or the battle lasts longer than 2 minutes.

选取280组空战对抗数据作为目标机动预测的数据集S，在任意t时刻，对战双方无人机的状态信息st和机动控制信息ct均包含以下多个维度的信息，如式所示：Select 280 sets of air combat confrontation data as the data set S for target maneuver prediction. At any time t, the status information s t and maneuver control information c t of the UAVs of both sides in the battle contain information of the following multiple dimensions, as shown in the formula:

st＝[Xt,Yt,Zt,EXt,EYt,EZt,AXt,AYt,AZt,Vt]s t ＝[X t , Y t , Z t , EX t , EY t , EZ t , AX t , AY t , AZ t , V t ]

ct＝[at,pt,yt]c t =[a t ,p t ,y t ]

式中，Xt,Yt,Zt分别表示t时刻无人机的三维空间位置坐标，EXt,EYt,EZt分别表示无人机t时刻的姿态欧拉角，(AXt,AYt,AZt)组成了无人机t时刻的机头朝向向量，Vt表示t时刻无人机的速度大小信息，st包含10个维度的状态信息，at,pt,yt分别表示控制无人机速度和方向的加速度信息，俯仰角信息和偏航角信息，ct包含3个维度的机动控制信息，任意t时刻我方无人机的状态信息用

Figure BDA0003802111080000031
表示，机动控制信息用
Figure BDA0003802111080000032
表示；目标无人机的状态信息用
Figure BDA0003802111080000033
表示，机动控制信息用
Figure BDA0003802111080000034
表示；数据集S＝[su,cu,se,ce]。In the formula, X t , Y t , Z t respectively represent the three-dimensional space position coordinates of the UAV at time t, EX t , EY t , EZ t respectively represent the attitude Euler angles of the UAV at time t, (AX t , AY t , AZ t ) constitute the nose orientation vector of the UAV at time t, V t represents the speed information of UAV at time t, s t contains state information of 10 dimensions, a t , p t , y t are respectively Represents the acceleration information, pitch angle information and yaw angle information for controlling the speed and direction of the UAV, c t contains maneuver control information in three dimensions, and the state information of our UAV at any time t is used
Figure BDA0003802111080000031
Indicates that maneuver control information is used
Figure BDA0003802111080000032
Indicates; the status information of the target UAV is represented by
Figure BDA0003802111080000033
Indicates that maneuver control information is used
Figure BDA0003802111080000034
Represents; data set S=[s u , c u , s e , c e ].

所述步骤2在近距空战对抗过程中，无人机获胜的关键是当目标机在己方无人机攻击范围内时，己方无人机角度态势处于优势状态，此时目标机的方位角p与目标机的进入角q取值趋近于0，对目标机的机动进行预测时需要考虑目标机与我机之间的距离d和目标机的角度态势Ta情况，从步骤1得到的目标机动预测数据集S中提取目标机动的距离特征d和角度态势特征Ta：In step 2, in the process of short-distance air combat confrontation, the key to the victory of the UAV is that when the target aircraft is within the attack range of one's own UAV, the angular situation of one's own UAV is in a dominant state. At this time, the azimuth p of the target aircraft is The value of the entry angle q with the target aircraft is close to 0. When predicting the maneuvering of the target aircraft, the distance d between the target aircraft and our aircraft and the angle situation T a of the target aircraft need to be considered. The target aircraft obtained from step 1 Extract the distance feature d and angle situation feature T a of the target maneuver from the maneuver prediction data set S:

Figure BDA0003802111080000035
Figure BDA0003802111080000035

Figure BDA0003802111080000036
Figure BDA0003802111080000036

式中，Xu,Yu,Zu分别表示我方无人机的三维空间位置坐标，Xe,Ye,Ze分别表示目标无人机的三维空间位置坐标。In the formula, X u , Y u , Zu represent the three-dimensional space position coordinates of our UAV respectively, X e , Y e , Z e represent the three-dimensional space position coordinates of the target UAV respectively.

将提取的距离特征和角度态势特征并入目标机动预测数据集S：Merge the extracted distance features and angle situation features into the target maneuver prediction dataset S:

S＝[su,cu,se,ce,d,Ta]S＝[s u , c u , s e , c e , d, T a ]

式中，su,cu分别表示目标机动预测数据集S中我方无人机的状态信息和机动控制信息，se,ce分别表示目标机动预测数据集S中目标无人机的状态信息和机动控制信息。In the formula, s u , c u represent the state information and maneuver control information of our UAV in the target maneuver prediction data set S respectively, s e , c e represent the state of the target UAV in the target maneuver prediction data set S information and maneuver control information.

所述步骤3包括：Said step 3 includes:

3a，将步骤1中目标机动预测数据集S按照训练集和测试集之比为7:3的比例划分为训练集Strain和测试集Stest；3a, the target maneuver prediction data set S in step 1 is divided into a training set S train and a test set S test according to the ratio of the training set and the test set being 7:3;

3b，分别提取训练集Strain和测试集Stest机动控制信息的各个维度数据，并进行标签化预处理，得到标签化的训练集控制信息数据

Figure BDA0003802111080000041
3b. Extract the data of each dimension of the maneuver control information of the training set S train and the test set S test respectively, and perform labeling preprocessing to obtain the labeled training set control information data
Figure BDA0003802111080000041

和测试集状态信息数据

Figure BDA0003802111080000042
无人机通过加速度、俯仰角速度和偏航角速度控制自身的速度和方向，加速度的大小会影响无人机速度的变化，俯仰角速度和偏航角速度的存在控制着无人机的方向，为了模拟空战中无人机的不同机动动作，仿真系统设置加速度大小一共有4种取值情况{-30,-10,40,60}(单位为：米/秒平方)，每一取值表示无人机进行不同的加速或减速机动，俯仰角速度有9种取值情况{-30,-21,-12,-5,0,5,12,21,30}(单位为：度/秒)，每一取值表示无人机以某一角速度进行不同的俯仰机动，偏航角速度有9种取值情况{-60,-34,-18,-8,0,8,18,34,60}(单位为：度/秒)，每一取值表示无人机以某一角速度进行不同的偏离原来航线的机动，所以无人机机动控制参数共有4*9*9＝324种取值组合，每种组合代表无人机的一种机动动作，对机动动作的预测即是对机动控制参数组合取值的预测，将加速度、俯仰角速度和偏航角速度的取值按照映射成类别标签的形式；and test set state information data
Figure BDA0003802111080000042
The UAV controls its own speed and direction through acceleration, pitch rate and yaw rate. The magnitude of acceleration will affect the change of the speed of the drone. The existence of pitch rate and yaw rate controls the direction of the drone. In order to simulate air combat For the different maneuvering actions of the UAV, the simulation system sets the acceleration to a total of 4 value situations {-30,-10,40,60} (unit: m/s square), and each value represents the UAV For different acceleration or deceleration maneuvers, there are 9 values of the pitch angular velocity {-30,-21,-12,-5,0,5,12,21,30} (unit: degree/second), each The value indicates that the drone performs different pitch maneuvers at a certain angular velocity, and there are 9 value situations for the yaw angular velocity {-60,-34,-18,-8,0,8,18,34,60} (unit is: degree/second), each value indicates that the UAV performs different maneuvers deviating from the original route at a certain angular velocity, so the maneuver control parameters of the UAV have 4*9*9=324 value combinations, each The combination represents a kind of maneuvering action of the UAV. The prediction of the maneuvering action is the prediction of the value of the combination of maneuvering control parameters, and the values of acceleration, pitch rate and yaw rate are mapped into the form of category labels;

3c，分别对训练集Strain和测试集Stest各个维度的数据进行归一化预处理，得到归一化后的训练集数据

Figure BDA0003802111080000043
和测试集数据
Figure BDA0003802111080000044
3c, perform normalized preprocessing on the data of each dimension of the training set S train and the test set S test respectively, and obtain the normalized training set data
Figure BDA0003802111080000043
and test set data
Figure BDA0003802111080000044

Figure BDA0003802111080000045
Figure BDA0003802111080000045

式中，xi表示训练集或测试集中第i维的数据，

Figure BDA0003802111080000046
表示第i维数据的最小值，
Figure BDA0003802111080000047
表示第i维数据的最大值，
Figure BDA0003802111080000048
是第i维数据归一化后的结果；In the formula, x i represents the i-th dimension data in the training set or test set,
Figure BDA0003802111080000046
Indicates the minimum value of the i-th dimension data,
Figure BDA0003802111080000047
Indicates the maximum value of the i-th dimension data,
Figure BDA0003802111080000048
is the normalized result of the i-th dimension data;

归一化后的训练集数据

Figure BDA0003802111080000051
和测试集数据
Figure BDA0003802111080000052
均按照(batch,seq,input)的维度格式存储，其中batch表示网络模型批处理的维度，seq表示用来预测目标下一时刻机动动作的时间序列长度，input表示归一化后的数据集中每条数据的维度。Normalized training set data
Figure BDA0003802111080000051
and test set data
Figure BDA0003802111080000052
They are all stored in the dimensional format of (batch, seq, input), where batch represents the dimension of batch processing of the network model, seq represents the length of the time series used to predict the maneuvering action of the target at the next moment, and input represents each normalized data set Dimensions of the data.

所述步骤4具体为：The step 4 is specifically:

基于LSTM的目标的机动预测网络模型包括1个具有2层隐藏层的LSTM网络、3个全连接层和3个交叉熵损失函数，LSTM网络模型的输入为连续seq个时刻归一化后的训练集数据序列，输入格式为(batch,seq,input)；输出为封装的近seq时间段内目标机机动的隐藏状态

Figure BDA0003802111080000053
输出格式为(batch,seq,hidden)；The LSTM-based target maneuver prediction network model includes an LSTM network with 2 hidden layers, 3 fully connected layers, and 3 cross-entropy loss functions. The input of the LSTM network model is the normalized training of consecutive seq moments Set the data sequence, the input format is (batch, seq, input); the output is the hidden state of the target machine maneuvering in the encapsulated time period near seq
Figure BDA0003802111080000053
The output format is (batch, seq, hidden);

将近seq时间段内目标机机动的隐藏状态

Figure BDA0003802111080000054
分别输入3个全连接层，结合3个交叉熵损失函数La，Lp，Ly，分别获得加速度、俯仰角速度和偏航角速度在seq+1时刻取每类值的概率情况：The hidden state of the target machine's maneuver in the near seq time period
Figure BDA0003802111080000054
Input 3 fully-connected layers respectively, combined with 3 cross-entropy loss functions L a , L p , L y , respectively obtain the probability of each type of value of acceleration, pitch rate and yaw rate at seq+1:

Figure BDA0003802111080000055
Figure BDA0003802111080000055

Figure BDA0003802111080000056
Figure BDA0003802111080000056

式中，L表示交叉熵函数计算公式，N表示样本数量，M表示类别数量，yic为符号函数，若样本i的真实类别为c，yic取1，否则yic取0；La，Lp，Ly分别表示对加速度、俯仰角速度和偏航角速度使用交叉熵损失函数L计算得到的加速度交叉熵损失，俯仰角交叉熵损失和偏航角交叉熵损失；Pa表示目标加速度在seq+1时刻取每类值的概率情况，

Figure BDA0003802111080000057
表示目标加速度在seq+1时刻取第m类值的概率情况，Pp表示目标俯仰角速度在seq+1时刻取每类值的概率情况，
Figure BDA0003802111080000058
表示目标俯仰角速度在seq+1时刻取第n类值的概率情况，Py表示目标偏航角速度在seq+1时刻取每类值的概率情况，
Figure BDA0003802111080000061
表示目标偏航角速度在seq+1时刻取第n类值的概率情况。In the formula, L represents the calculation formula of the cross-entropy function, N represents the number of samples, M represents the number of categories, and y ic is a sign function. If the real category of sample i is c, y ic takes 1, otherwise y ic takes 0; L a , L p , L y respectively represent the acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss calculated by using the cross entropy loss function L for acceleration, pitch rate and yaw rate; P a represents the target acceleration in seq The probability of taking each type of value at +1 time,
Figure BDA0003802111080000057
Indicates the probability that the target acceleration takes the mth class value at the time seq+1, and P p represents the probability that the target pitch angular velocity takes each class value at the time seq+1,
Figure BDA0003802111080000058
Indicates the probability of the target pitch rate taking the nth type of value at the time seq+1, P y represents the probability of the target yaw rate taking each type of value at the time seq+1,
Figure BDA0003802111080000061
Indicates the probability that the target yaw angular velocity takes the nth type value at the time seq+1.

分别将目标加速度a、俯仰角速度p和偏航角速度y取值概率最大的类别值作为目标机seq+1时刻加速度a、俯仰角速度p和偏航角速度y的预测类别值，Pa pre，Pp pre，Py pre：The category value with the highest probability of taking target acceleration a, pitch angular velocity p and yaw angular velocity y is used as the predicted category value of acceleration a, pitch angular velocity p and yaw angular velocity y of the target machine seq+1 respectively, P a pre , P p pre , P y pre :

Figure BDA0003802111080000062
Figure BDA0003802111080000062

加速度a、俯仰角速度p和偏航角速度y的预测类别值组合在一起即得到机动动作的预测类别标签值，至此，无人机的机动预测问题转变为目标加速度a、俯仰角速度p和偏航角速度y的分类预测问题。The predicted category values of acceleration a, pitch angular velocity p, and yaw angular velocity y are combined to obtain the predicted category label value of the maneuver. At this point, the maneuver prediction problem of the UAV is transformed into target acceleration a, pitch angular velocity p, and yaw angular velocity The classification prediction problem of y.

所述步骤5在模型训练的时候，对加速度交叉熵损失La，俯仰角交叉熵损失Lp和偏航角交叉熵损失Ly进行加权融合，将融合后的损失函数值作为整个网络模型的训练损失Ltrain：In the step 5, when the model is trained, the acceleration cross-entropy loss L a , the pitch angle cross-entropy loss L p and the yaw angle cross-entropy loss L y are weighted and fused, and the fused loss function value is used as the value of the entire network model Training loss L train :

Ltrain＝waLa+wpLp+wyLy L train ＝w a L a +w p L p +w y L y

式中，wa，wp，wy分别表示加速度交叉熵损失La，俯仰角交叉熵损失Lp和偏航角交叉熵损失Ly的权值。In the formula, w a , w p , w y represent the weights of acceleration cross-entropy loss L a , pitch angle cross-entropy loss L p and yaw angle cross-entropy loss L y respectively.

一种电子设备，包括处理器、存储器和通信总线，其中，处理器、存储器通过通信总线完成相互间的通信；An electronic device, including a processor, a memory and a communication bus, wherein the processor and the memory complete mutual communication through the communication bus;

存储器，用于存放计算机程序；memory for storing computer programs;

处理器，用于执行存储器上所存放的程序时，实现上述所述的一种基于LSTM的目标的机动预测方法。When the processor is used to execute the program stored in the memory, it realizes the above-mentioned LSTM-based target maneuver prediction method.

一种计算机可读存储介质，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述所述的一种基于LSTM的目标的机动预测方法。A computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the aforementioned LSTM-based target maneuver prediction method is implemented.

本发明的有益效果：Beneficial effects of the present invention:

本发明在预测目标机机动时，提取目标机的空间态势占位特征，并与我方无人机的状态和机动信息、目标机的状态和机动信息共同作为目标机机动预测的依据，能够充分挖掘空战对抗数据中目标机进行机动选择的规律，使得对目标机机动的预测更准确The present invention extracts the spatial situation occupancy characteristics of the target aircraft when predicting the maneuvering of the target aircraft, and uses the state and maneuver information of our UAV and the state and maneuver information of the target aircraft together as the basis for the maneuver prediction of the target aircraft, which can fully Mining the law of target aircraft maneuver selection in air combat confrontation data makes the prediction of target aircraft maneuver more accurate

本发明分析了无人机的机动动作与无人机的加速度、俯仰角速度和偏转角速度的关系，将加速度、俯仰角速度和偏转角速度的取值标签化，把目标机机动动作选择的预测问题转换为对加速度、俯仰角速度和偏转角速度三个机动控制变量的分类预测问题，能够准确预测目标机未来的机动动作，为目标机机动预测方法的研究提供了新思路The present invention analyzes the relationship between the maneuvering action of the UAV and the acceleration, pitch angular velocity and yaw angular velocity of the UAV, labels the values of acceleration, pitch angular velocity and yaw angular velocity, and converts the prediction problem of target aircraft maneuver selection into For the classification and prediction of the three maneuver control variables of acceleration, pitch angular velocity and yaw angular velocity, it can accurately predict the future maneuvering actions of the target aircraft, and provides a new idea for the research of target aircraft maneuver prediction methods

本发明从目标机机动具有时序性的特点出发，建立基于LSTM的目标机机动预测网络模型，利用LSTM对时序数据强大的预测能力，根据一段时间内我方无人机的状态信息和机动信息、目标机的状态信息和机动信息以及目标机在该段时间内的空间态势占位特征，预测目标机下一时刻的机动动作，相比其它网络模型取得了更好的预测效果，实现了对目标机机动动作的准确预测。The present invention starts from the timing characteristics of target aircraft maneuvers, establishes a target aircraft maneuver prediction network model based on LSTM, utilizes the powerful prediction ability of LSTM for time series data, according to the state information and maneuver information of our UAV within a period of time, The status information and maneuvering information of the target machine and the space situation occupancy characteristics of the target machine during this period of time can predict the maneuvering action of the target machine at the next moment. Accurate prediction of maneuvering maneuvers.

附图说明：Description of drawings:

图1是空战对抗过程中我方无人机与目标机方位角示意图。Figure 1 is a schematic diagram of the azimuth angle between our UAV and the target aircraft in the process of air combat confrontation.

图2是无人机机动控制参数标签化处理示意图。Figure 2 is a schematic diagram of labeling processing of UAV maneuver control parameters.

图3是基于LSTM的目标的机动预测方法示意图。Figure 3 is a schematic diagram of the LSTM-based target maneuver prediction method.

具体实施方式detailed description

下面结合实施例对本发明作进一步详细说明。The present invention is described in further detail below in conjunction with embodiment.

基于LSTM的目标的机动预测方法，包括以下步骤：The LSTM-based target maneuver prediction method includes the following steps:

1，构建无人机对战过程目标机机动选择数据集；1. Construct a target machine maneuver selection data set during the UAV battle process;

利用某一对一空战仿真系统，对战双方设置不同初始状态进行多次对战实验，每局对战每隔20ms记录一次对战双方无人机的状态信息st和机动动作控制信息ct，形成空战对抗数据，并且每局对战过程以某一方的胜利或者对战时长超过2分钟时终止。本实例中选取280组对战实验共包括14397712条对战机动数据作为目标机动预测的数据集S。在任意t时刻，对战双方无人机的状态信息st和机动控制信息ct均包含以下多个维度的信息，如式所示：Using a one-to-one air combat simulation system, the two sides set different initial states to conduct multiple battle experiments, and record the state information s t and maneuver control information c t of the drones of the two sides every 20ms in each game to form an air combat confrontation Data, and the process of each game ends when one side wins or the game lasts longer than 2 minutes. In this example, 280 sets of battle experiments are selected, including 14,397,712 pieces of battle maneuver data as the data set S for target maneuver prediction. At any time t, the state information s t and the maneuver control information c t of the UAVs of the opposing sides both contain the following multi-dimensional information, as shown in the formula:

st＝[Xt,Yt,Zt,EXt,EYt,EZt,AXt,AYt,AZt,Vt]s t ＝[X t , Y t , Z t , EX t , EY t , EZ t , AX t , AY t , AZ t , V t ]

ct＝[at,pt,yt]c t =[a t ,p t ,y t ]

式中，Xt,Yt,Zt分别表示t时刻无人机的三维空间位置坐标，EXt,EYt,EZt分别表示无人机t时刻的姿态欧拉角，(AXt,AYt,AZt)组成了无人机t时刻的机头朝向向量，Vt表示t时刻无人机的速度大小信息，st包含10个维度的状态信息。at,pt,yt分别表示控制无人机速度和方向的加速度信息，俯仰角信息和偏航角信息，ct包含3个维度的机动控制信息。本实例中任意t时刻我方无人机的状态信息用

Figure BDA0003802111080000081
表示，机动控制信息用
Figure BDA0003802111080000082
表示；目标无人机的状态信息用
Figure BDA0003802111080000083
表示，机动控制信息用
Figure BDA0003802111080000084
表示；数据集S＝[su,cu,se,ce]。In the formula, X t , Y t , Z t respectively represent the three-dimensional space position coordinates of the UAV at time t, EX t , EY t , EZ t respectively represent the attitude Euler angles of the UAV at time t, (AX t , AY t , AZ t ) constitute the nose orientation vector of the UAV at time t, V t represents the speed information of the UAV at time t, and st t includes state information in 10 dimensions. a t , p t , and y t represent the acceleration information, pitch angle information, and yaw angle information for controlling the speed and direction of the UAV, respectively, and c t includes maneuver control information in three dimensions. In this example, the status information of our UAV at any time t is used
Figure BDA0003802111080000081
Indicates that maneuver control information is used
Figure BDA0003802111080000082
Indicates; the status information of the target UAV is represented by
Figure BDA0003802111080000083
Indicates that maneuver control information is used
Figure BDA0003802111080000084
Represents; data set S=[s u , c u , s e , c e ].

2，提取目标机动对抗过程空间态势特征；2. Extract the space situational characteristics of the target maneuver confrontation process;

在空战对抗过程中，无人机进行连续机动的目的是与目标无人机进行博弈对抗，从而让自己的空间占位处于优势状态，达到攻击目标机的目的。所以，空间态势信息是无人机进行机动选择的关键因素。在近距空战对抗过程中，无人机获胜的关键是当目标机在己方无人机攻击范围内时，己方无人机角度态势处于优势状态，此时目标机的方位角p与目标机的进入角q取值趋近于0，如图1所示。所以对目标机的机动进行预测时需要考虑目标机与我机之间的距离d和目标机的角度态势Ta情况，从目标机动预测数据集S中提取目标机动的距离特征d和角度态势特征Ta：In the process of air combat confrontation, the purpose of the drone's continuous maneuver is to compete with the target drone, so that its own space occupation is in an advantageous state, and the purpose of attacking the target aircraft is achieved. Therefore, space situation information is a key factor for UAVs to choose maneuvers. In the process of close-range air combat confrontation, the key to UAV's victory is that when the target aircraft is within the attack range of its own UAV, the angle situation of its own UAV is in an advantageous state. The value of the entry angle q is close to 0, as shown in Figure 1. Therefore, when predicting the maneuver of the target aircraft, it is necessary to consider the distance d between the target aircraft and our aircraft and the angle situation T a of the target aircraft, and extract the distance feature d and angle situation feature of the target maneuver from the target maneuver prediction data set S T a :

Figure BDA0003802111080000091
Figure BDA0003802111080000091

Figure BDA0003802111080000092
Figure BDA0003802111080000092

式中，Xu,Yu,Zu分别表示我方无人机的三维空间位置坐标，Xe,Ye,Ze分别表示目标无人机的三维空间位置坐标。In the formula, X u , Y u , Zu represent the three-dimensional space position coordinates of our UAV respectively, X e , Y e , Z e represent the three-dimensional space position coordinates of the target UAV respectively.

将提取的距离特征和角度态势特征并入目标机动预测数据集S：Merge the extracted distance features and angle situation features into the target maneuver prediction dataset S:

S＝[su,cu,se,ce,d,Ta]S＝[s u , c u , s e , c e , d, T a ]

式中，su,cu分别表示目标机动预测数据集S中我方无人机的状态信息和机动控制信息，se,ce分别表示目标机动预测数据集S中目标无人机的状态信息和机动控制信息。In the formula, s u , c u represent the state information and maneuver control information of our UAV in the target maneuver prediction data set S respectively, s e , c e represent the state of the target UAV in the target maneuver prediction data set S information and maneuver control information.

3，对数据集进行分割和预处理。3. Segment and preprocess the dataset.

3a，将目标机动预测数据集S按照训练集和测试集之比为7:3的比例划分为训练集Strain和测试集Stest；3a, the target maneuver prediction data set S is divided into a training set S train and a test set S test according to a ratio of 7:3 between the training set and the test set;

3b，分别提取训练集Strain和测试集Stest机动控制信息的各个维度数据，并进行标签化预处理，得到标签化的训练集控制信息数据

Figure BDA0003802111080000093
3b. Extract the data of each dimension of the maneuver control information of the training set S train and the test set S test respectively, and perform labeling preprocessing to obtain the labeled training set control information data
Figure BDA0003802111080000093

和测试集状态信息数据

Figure BDA0003802111080000094
无人机通过加速度、俯仰角速度和偏航角速度控制自身的速度和方向，加速度的大小会影响无人机速度的变化，俯仰角速度和偏航角速度的存在控制着无人机的方向。为了模拟空战中无人机的不同机动动作，仿真系统设置加速度大小一共有4种取值情况{-30,-10,40,60}(单位为：米/秒平方)，每一取值表示无人机进行不同的加速或减速机动，俯仰角速度有9种取值情况{-30,-21,-12,-5,0,5,12,21,30}(单位为：度/秒)，每一取值表示无人机以某一角速度进行不同的俯仰机动，偏航角速度有9种取值情况{-60,-34,-18,-8,0,8,18,34,60}(单位为：度/秒)，每一取值表示无人机以某一角速度进行不同的偏离原来航线的机动，所以无人机机动控制参数共有4*9*9＝324种取值组合，每种组合代表无人机的一种机动动作，对机动动作的预测即是对机动控制参数组合取值的预测。将加速度、俯仰角速度和偏航角速度的取值按照映射成类别标签的形式，如图2所示；and test set state information data
Figure BDA0003802111080000094
The UAV controls its own speed and direction through acceleration, pitch rate and yaw rate. The magnitude of acceleration will affect the change of the speed of the drone. The existence of pitch rate and yaw rate controls the direction of the drone. In order to simulate different maneuvers of UAVs in air combat, the simulation system sets the acceleration to a total of 4 value situations {-30,-10,40,60} (unit: m/s square), each value represents The UAV performs different acceleration or deceleration maneuvers, and the pitch angular velocity has 9 values {-30,-21,-12,-5,0,5,12,21,30} (unit: degree/second) , each value indicates that the UAV performs different pitch maneuvers at a certain angular velocity, and there are 9 value situations for the yaw angular velocity {-60,-34,-18,-8,0,8,18,34,60 } (unit: degree/second), each value indicates that the UAV performs different maneuvers deviating from the original route at a certain angular velocity, so there are 4*9*9=324 value combinations for the maneuver control parameters of the UAV , each combination represents a maneuver of the UAV, and the prediction of the maneuver is the prediction of the value of the maneuver control parameter combination. The values of acceleration, pitch rate and yaw rate are mapped into the form of category labels, as shown in Figure 2;

3c，分别对训练集Strain和测试集Stest各个维度的数据进行归一化预处理，得到归一化后的训练集数据

Figure BDA0003802111080000101
3c, perform normalized preprocessing on the data of each dimension of the training set S train and the test set S test respectively, and obtain the normalized training set data
Figure BDA0003802111080000101

和测试集数据

Figure BDA0003802111080000102
and test set data
Figure BDA0003802111080000102

Figure BDA0003802111080000103
Figure BDA0003802111080000103

式中，xi表示训练集或测试集中第i维的数据，

Figure BDA0003802111080000104
表示第i维数据的最小值，
Figure BDA0003802111080000105
表示第i维数据的最大值，
Figure BDA0003802111080000106
是第i维数据归一化后的结果；In the formula, x i represents the i-th dimension data in the training set or test set,
Figure BDA0003802111080000104
Indicates the minimum value of the i-th dimension data,
Figure BDA0003802111080000105
Indicates the maximum value of the i-th dimension data,
Figure BDA0003802111080000106
is the normalized result of the i-th dimension data;

归一化后的训练集数据

Figure BDA0003802111080000107
和测试集数据
Figure BDA0003802111080000108
均按照(batch,seq,input)的维度格式存储，其中batch表示网络模型批处理的维度，seq表示用来预测目标下一时刻机动动作的时间序列长度，input表示归一化后的数据集中每条数据的维度。本实例中，batch设置为256，seq设置为14，input为28。Normalized training set data
Figure BDA0003802111080000107
and test set data
Figure BDA0003802111080000108
They are all stored in the dimensional format of (batch, seq, input), where batch represents the dimension of batch processing of the network model, seq represents the length of the time series used to predict the maneuvering action of the target at the next moment, and input represents each normalized data set Dimensions of the data. In this example, batch is set to 256, seq is set to 14, and input is set to 28.

4，建立基于LSTM的目标的机动预测网络模型；4. Establish a maneuver prediction network model based on LSTM targets;

如图3所示，基于LSTM的目标的机动预测网络模型包括1个具有2层隐藏层的LSTM网络、3个全连接层和3个交叉熵损失函数。LSTM网络模型的输入为连续seq个时刻归一化后的训练集数据序列，输入格式为(batch,seq,input)；输出为封装的近seq时间段内目标机机动的隐藏状态

Figure BDA0003802111080000111
输出格式为(batch,seq,hidden)，本实例中hidden设置为128；As shown in Figure 3, the LSTM-based target maneuver prediction network model includes 1 LSTM network with 2 hidden layers, 3 fully connected layers and 3 cross-entropy loss functions. The input of the LSTM network model is the normalized training set data sequence of consecutive seq moments, and the input format is (batch, seq, input); the output is the hidden state of the target machine maneuver in the encapsulated near-seq time period
Figure BDA0003802111080000111
The output format is (batch, seq, hidden), and hidden is set to 128 in this example;

将近seq时间段内目标机机动的隐藏状态

Figure BDA0003802111080000112
分别输入3个全连接层，结合3个交叉熵损失函数La，Lp，Ly，分别获得加速度、俯仰角速度和偏航角速度在seq+1时刻取每类值的概率情况：The hidden state of the target machine's maneuver in the near seq time period
Figure BDA0003802111080000112
Input 3 fully-connected layers respectively, combined with 3 cross-entropy loss functions L a , L p , L y , respectively obtain the probability of each type of value of acceleration, pitch rate and yaw rate at seq+1:

Figure BDA0003802111080000113
Figure BDA0003802111080000113

Figure BDA0003802111080000114
Figure BDA0003802111080000114

式中，L表示交叉熵函数计算公式，N表示样本数量，M表示类别数量，yic为符号函数，若样本i的真实类别为c，yic取1，否则yic取0；La，Lp，Ly分别表示对加速度、俯仰角速度和偏航角速度使用交叉熵损失函数L计算得到的加速度交叉熵损失，俯仰角交叉熵损失和偏航角交叉熵损失；Pa表示目标加速度在seq+1时刻取每类值的概率情况，

Figure BDA0003802111080000115
表示目标加速度在seq+1时刻取第m类值的概率情况，Pp表示目标俯仰角速度在seq+1时刻取每类值的概率情况，
Figure BDA0003802111080000116
表示目标俯仰角速度在seq+1时刻取第n类值的概率情况，Py表示目标偏航角速度在seq+1时刻取每类值的概率情况，
Figure BDA0003802111080000117
表示目标偏航角速度在seq+1时刻取第n类值的概率情况。In the formula, L represents the calculation formula of the cross-entropy function, N represents the number of samples, M represents the number of categories, and y ic is a sign function. If the real category of sample i is c, y ic takes 1, otherwise y ic takes 0; L a , L p , L y respectively represent the acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss calculated by using the cross entropy loss function L for acceleration, pitch rate and yaw rate; P a represents the target acceleration in seq The probability of taking each type of value at +1 time,
Figure BDA0003802111080000115
Indicates the probability that the target acceleration takes the mth class value at the time seq+1, and P p represents the probability that the target pitch angular velocity takes each class value at the time seq+1,
Figure BDA0003802111080000116
Indicates the probability of the target pitch rate taking the nth type of value at the time seq+1, P y represents the probability of the target yaw rate taking each type of value at the time seq+1,
Figure BDA0003802111080000117
Indicates the probability that the target yaw angular velocity takes the nth type value at the time seq+1.

分别将目标加速度a、俯仰角速度p和偏航角速度y取值概率最大的类别值作为目标机seq+1时刻加速度a、俯仰角速度p和偏航角速度y的预测类别值，Pa pre，Pp pre，Py pre：The category value with the highest probability of taking target acceleration a, pitch angular velocity p and yaw angular velocity y is used as the predicted category value of acceleration a, pitch angular velocity p and yaw angular velocity y of the target machine seq+1 respectively, P a pre , P p pre , P y pre :

Figure BDA0003802111080000118
Figure BDA0003802111080000118

加速度a、俯仰角速度p和偏航角速度y的预测类别值组合在一起即得到机动动作的预测类别标签值，根据图2所示类别标签与机动控制参数值的映射关系可得相应的机动控制变量取值。至此，无人机的机动预测问题转变为目标加速度a、俯仰角速度p和偏航角速度y的分类预测问题。The predicted category values of acceleration a, pitch rate p, and yaw rate y are combined to obtain the predicted category label value of the maneuver. According to the mapping relationship between the category label and the maneuver control parameter value shown in Figure 2, the corresponding maneuver control variables can be obtained value. So far, the maneuver prediction problem of UAV has been transformed into the classification prediction problem of target acceleration a, pitch angular velocity p and yaw angular velocity y.

5，利用训练集

Figure BDA0003802111080000121
对基于LSTM的目标的机动预测网络模型进行训练，并利用测试集
Figure BDA0003802111080000122
进行准确率检测；5. Using the training set
Figure BDA0003802111080000121
Train the LSTM-based target maneuver prediction network model and use the test set
Figure BDA0003802111080000122
Perform accuracy testing;

在模型训练的时候，对加速度交叉熵损失La，俯仰角交叉熵损失Lp和偏航角交叉熵损失Ly进行加权融合，将融合后的损失函数值作为整个网络模型的训练损失Ltrain：During model training, the acceleration cross-entropy loss L a , the pitch angle cross-entropy loss L p and the yaw angle cross-entropy loss L y are weighted and fused, and the fused loss function value is used as the training loss L train of the entire network model :

Ltrain＝waLa+wpLp+wyLy L train ＝w a L a +w p L p +w y L y

式中，wa，wp，wy分别表示加速度交叉熵损失La，俯仰角交叉熵损失Lp和偏航角交叉熵损失Ly的权值，本实例中，通过实验确定设置wa，wp，wy的值分别为0.2，0.4，0.4。In the formula, w a , w p , w y represent the weights of acceleration cross-entropy loss L a , pitch angle cross-entropy loss L p and yaw angle cross-entropy loss L y respectively. In this example, the setting w a is determined through experiments , w p , w y are 0.2, 0.4, 0.4 respectively.

为了防止模型训练过程中出现过拟合现象，同时使网络模型更加健壮，在LSTM网络层加入dropout机制，并且本实例中把dropout的值设置为0.4；本实例中网络模型学习率超参数learing rate设置为0.0001，训练轮数epochs设置为500。对于每轮训练保存模型的网络参数。In order to prevent overfitting during model training and make the network model more robust, a dropout mechanism is added to the LSTM network layer, and the value of dropout is set to 0.4 in this example; the network model learning rate hyperparameter learning rate in this example Set to 0.0001 and the number of training epochs to 500. Save the network parameters of the model for each round of training.

利用保存的网络模型参数对测试集数据进行测试验证，本模型采用的评价指标是准确率Accuracy，准确率Accuracy越接近于1，模型的分类预测越精准。本发明分别统计了测试集中加速度、俯仰角速度和偏航角速度的预测准确率Acc(a)，Acc(p)，Acc(y)以及在同一时刻加速度、俯仰角速度和偏航角速度同时预测正确的准确率Acc(a,p,y)，同时将本模型与BP神经网络，RNN神经网络进行了对比，并对每个比较的网络模型都做了多次独立实验，实验结果如表1所示：Use the saved network model parameters to test and verify the test set data. The evaluation index used in this model is the accuracy rate. The closer the accuracy rate is to 1, the more accurate the classification prediction of the model is. The present invention respectively counts the prediction accuracy rates Acc(a), Acc(p) and Acc(y) of acceleration, pitch angular velocity and yaw angular velocity in the test set and the accuracy of simultaneous prediction of acceleration, pitch angular velocity and yaw angular velocity at the same time. rate Acc(a,p,y), and compared this model with BP neural network and RNN neural network, and conducted several independent experiments on each compared network model. The experimental results are shown in Table 1:

表1不同网络模型对目标机机动预测准确率比较Table 1 Comparison of accuracy rate of target aircraft maneuver prediction by different network models

Figure BDA0003802111080000131
Figure BDA0003802111080000131

从表1的结果可以看出，在本实例中，与BP神经网络和RNN相比，本发明的网络模型不论是对单个机动控制参数的预测，还是对所有机动控制参数同时预测正确问题都取得了良好的效果，其中所有机动控制参数同时预测正确表示机动控制参数组合值预测正确，即某一时刻的目标机动预测正确。As can be seen from the results in Table 1, in this example, compared with the BP neural network and RNN, the network model of the present invention achieves the correct problem no matter it is the prediction of a single maneuvering control parameter or the simultaneous prediction of all maneuvering control parameters. The prediction of all maneuver control parameters at the same time is correct, which means that the prediction of the combined value of the maneuver control parameters is correct, that is, the target maneuver prediction at a certain moment is correct.

本发明分析了现有空战对抗过程目标机动动作预测方法，发现现有目标机动预测方法仅从目标自身出发，将目标机的机动动作分为固定的几种动作的加权组合，每种动作的权值表示目标机采取此种机动动作的概率。没有考虑其机动动作与其所占空间态势的关系；以及没有考虑我方无人机状态对目标机进行机动的影响情况；所以本发明在预测目标机机动时，提取目标机的空间态势占位特征，并与我方无人机的状态和机动信息、目标机的状态和机动信息共同作为目标机机动预测的依据，这能够充分挖掘空战对抗数据中目标机进行机动选择的规律，使得对目标机机动的预测更准确The present invention analyzes the existing target maneuver prediction method in the air combat confrontation process, and finds that the existing target maneuver prediction method only starts from the target itself, divides the target aircraft's maneuvers into fixed weighted combinations of several actions, and the weight of each action The value represents the probability that the target aircraft takes such a maneuver. The relationship between its maneuvering action and the space situation it occupies is not considered; and the influence of the state of our UAV on the maneuvering of the target machine is not considered; so the present invention extracts the space situation occupancy characteristics of the target machine when predicting the maneuvering of the target machine , and together with the state and maneuver information of our UAV, the state and maneuver information of the target aircraft as the basis for target aircraft maneuver prediction, this can fully tap the law of target aircraft maneuver selection in air combat confrontation data, making the target aircraft Maneuvering predictions are more accurate

本发明分析了无人机的机动动作与无人机的加速度、俯仰角速度和偏转角速度的关系，将加速度、俯仰角速度和偏转角速度的取值标签化，把目标机机动预测问题转换为对加速度、俯仰角速度和偏转角速度三个机动控制变量的分类预测问题，能够准确预测目标机未来的机动动作，为目标机机动预测方法的研究提供了新思路The present invention analyzes the relationship between the maneuvering action of the UAV and the acceleration, pitch angular velocity and yaw angular velocity of the UAV, labels the values of the acceleration, pitch angular velocity and yaw angular velocity, and converts the maneuver prediction problem of the target aircraft into the acceleration, pitch angular velocity and yaw angular velocity The classification prediction problem of the three maneuver control variables of pitch rate and yaw rate can accurately predict the future maneuvering actions of the target aircraft, and provides a new idea for the research of target aircraft maneuver prediction methods

本发明从目标机机动具有时序性的特点出发，建立基于LSTM的目标机机动预测网络模型，利用LSTM对时序数据强大的预测能力，根据一段时间内我方无人机的状态信息和机动信息、目标机的状态信息和机动信息以及目标机在该段时间内的空间态势占位特征，预测目标机下一时刻的机动动作。由于篇幅限制，本发明展示了某次空战对抗过程中，某段时间目标机实际进行一个batch次共256次机动与本模型相应进行一个batch次共256次预测机动的预测实例，预测效果如表2表示，实验结果进一步验证了本发明的方法对目标机机动预测的有效性。The present invention starts from the timing characteristics of target aircraft maneuvers, establishes a target aircraft maneuver prediction network model based on LSTM, utilizes the powerful prediction ability of LSTM for time series data, according to the state information and maneuver information of our UAV within a period of time, The status information and maneuvering information of the target aircraft and the occupancy characteristics of the space situation of the target aircraft during this period can predict the maneuvering action of the target aircraft at the next moment. Due to space limitations, the present invention shows a prediction example in which the target aircraft actually performs a batch of 256 maneuvers and this model performs a batch of 256 predicted maneuvers in a certain period of time during an air combat confrontation process. The prediction results are shown in the table 2, the experimental results have further verified the effectiveness of the method of the present invention for target aircraft maneuver prediction.

Figure BDA0003802111080000141
Figure BDA0003802111080000141

Figure BDA0003802111080000151
Figure BDA0003802111080000151

Figure BDA0003802111080000161
Figure BDA0003802111080000161

Figure BDA0003802111080000171
Figure BDA0003802111080000171

Figure BDA0003802111080000181
Figure BDA0003802111080000181

Figure BDA0003802111080000191
Figure BDA0003802111080000191

Figure BDA0003802111080000201
Figure BDA0003802111080000201

本说明书中的各个实施例均采用相关的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于系统实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。Each embodiment in this specification is described in a related manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for relevant parts, refer to part of the description of the method embodiment.

以上所述仅为本发明的较佳实施例而已，并非用于限定本发明的保护范围。凡在本发明的精神和原则之内所作的任何修改、等同替换、改进等，均包含在本发明的保护范围内。The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention are included in the protection scope of the present invention.

Claims (10)

1. A maneuver prediction method for an LSTM-based target, comprising the steps of;
step 1: constructing a target maneuvering selection data set S in the unmanned aerial vehicle fighting process, wherein the target is a target unmanned aerial vehicle in fighting with the unmanned aerial vehicle of one party;
step 2: extracting space situation characteristics of a target maneuver countermeasure process;
and step 3: segmenting and preprocessing the target mobile selection data set S in the step 1;
and 4, step 4: establishing a maneuvering prediction network model of the target based on the LSTM;
and 5: using training sets
Figure FDA0003802111070000011
Training a maneuver prediction network model for an LSTM-based target and utilizing a test set
Figure FDA0003802111070000012
And detecting the prediction accuracy.
2. The LSTM-based target maneuver prediction method of claim 1, wherein step 1 utilizes a one-to-one air combat simulation system, wherein the two parties set different initial states for carrying out a plurality of combat experiments, and the state information s of the unmanned aerial vehicles of the two parties is recorded every 20ms during each combat t And maneuver control information c t Forming air combat confrontation data, and stopping when the winning or the fighting duration of a certain party exceeds 2 minutes in each combat process;
selecting 280 groups of air combat countermeasure data as a data set S of target maneuver prediction, and at any time t, obtaining state information S of unmanned aerial vehicles of both parties of the battle t And maneuver control information c t Each contains the following information in multiple dimensions, as shown in the formula:
s t ＝[X t ,Y t ,Z t ,EX t ,EY t ,EZ t ,AX t ,AY t ,AZ t ,V t ]
c t ＝[a t ,p t ,y t ]
in the formula, X t ,Y t ,Z t Respectively representing the three-dimensional spatial position coordinates, EX, of the drone at time t t ,EY t ,EZ t Respectively representing the Euler angles of the postures of the unmanned aerial vehicle at the t moment (AX) t ,AY t ,AZ t ) Form the aircraft nose orientation vector at moment t of unmanned aerial vehicle, V t Speed information, s, representing the unmanned plane at time t t Containing status information of 10 dimensions, a t ,p t ,y t Acceleration information, pitch angle information and yaw angle information, c, representing the speed and direction of the controlling drone, respectively t Contains maneuvering control information of 3 dimensions, and is used for state information of unmanned aerial vehicle at any time t
Figure FDA0003802111070000021
For indicating, maneuvering control information
Figure FDA0003802111070000022
Represents; state information of target unmanned aerial vehicle
Figure FDA0003802111070000023
For indicating, maneuvering control information
Figure FDA0003802111070000024
Represents; data set
Figure FDA0003802111070000025
3. The method of claim 1 for maneuver prediction of LSTM-based targets, wherein the method further comprisesIn the step 2, in the process of the short-distance air combat countermeasure, the key of the unmanned aerial vehicle winning is that when the target machine is in the own unmanned aerial vehicle attack range, the own unmanned aerial vehicle angle situation is in the dominant state, the value of the azimuth angle p of the target machine and the value of the entrance angle q of the target machine at the moment are close to 0, and the distance d between the target machine and the own machine and the angle situation T of the target machine need to be considered when the maneuver of the target machine is predicted a Extracting the distance feature d and the angle situation feature T of the target maneuver from the target maneuver prediction data set S obtained in the step 1 a ：
Figure FDA0003802111070000026
Figure FDA0003802111070000027
In the formula, X u ,Y u ,Z u Respectively representing the three-dimensional spatial position coordinates, X, of my unmanned aerial vehicle e ,Y e ,Z e Respectively representing the three-dimensional spatial position coordinates of the target drone.
4. The method of claim 3, wherein the extracted distance features and angular situation features are incorporated into the target maneuver prediction dataset S:
S＝[s u ,c u ,s e ,c e ,d,T a ]
in the formula, s u ,c u Respectively representing the state information and maneuver control information, S, of the unmanned aerial vehicle of the same party in the target maneuver prediction data set S e ,c e And respectively representing the state information and the maneuver control information of the target unmanned aerial vehicle in the target maneuver prediction data set S.
5. The method of claim 1, wherein the step 3 comprises:
3a, dividing the target maneuver prediction data set S in the step 1 into a training set S according to the proportion that the ratio of the training set to the test set is 7:3 train And test set S test ；
3b, respectively extracting a training set S train And test set S test All dimension data of the maneuvering control information are subjected to labeling pretreatment to obtain labeled training set control information data
Figure FDA0003802111070000031
And test set status information data
Figure FDA0003802111070000032
The unmanned aerial vehicle controls the speed and the direction of the unmanned aerial vehicle through acceleration, pitch angle speed and yaw angle speed, the change of the speed of the unmanned aerial vehicle can be influenced by the magnitude of the acceleration, the direction of the unmanned aerial vehicle is controlled by the pitch angle speed and the yaw angle speed, and in order to simulate different maneuvers of the unmanned aerial vehicle in an air battle, the simulation system sets 4 value conditions { -30, -10,40,60} in the magnitude of the acceleration (the unit is: meter/second square), each value-taking value indicates that the unmanned aerial vehicle performs different acceleration or deceleration maneuvers, the pitch angle speed has 9 value-taking conditions { -30, -21, -12, -5,0,5,12,21,30} (the unit is: degree/second), each value-taking value indicates that the unmanned aerial vehicle performs different pitching maneuvers at a certain angular speed, the yaw angular speed has 9 value-taking conditions { -60, -34, -18, -8,0,8,18,34,60} (the unit is: degree/second), each value-taking value indicates that the unmanned aerial vehicle performs different maneuvers deviating from an original flight path at a certain angular speed, so that the maneuvering control parameters of the unmanned aerial vehicle have 4 value-taking combinations of 9 × 9=324, each combination represents one maneuvering action of the unmanned aerial vehicle, the maneuver action prediction of the maneuvering action is the prediction of the maneuvering control parameter combination value, and the values of the acceleration, the pitch angle speed and the yaw angular speed are mapped into a type label form;
3c, respectively aligning the training sets S train And test set S test Carrying out normalization pretreatment on the data of each dimension to obtain normalized training set data
Figure FDA0003802111070000033
And test set data
Figure FDA0003802111070000034
Figure FDA0003802111070000041
In the formula, x i Data representing the ith dimension in a training set or a test set,
Figure FDA0003802111070000042
represents the minimum value of the ith-dimension data,
Figure FDA0003802111070000043
represents the maximum value of the ith-dimension data,
Figure FDA0003802111070000044
is the result of the normalization of the ith dimension data;
normalized training set data
Figure FDA0003802111070000045
And test set data
Figure FDA0003802111070000046
Are stored in a dimensional format of (batch, seq, input), wherein batch represents the dimension of the network model batch processing, seq represents the length of the time series used for predicting the maneuver of the target at the next moment, and input represents the dimension of each piece of data in the normalized data set.
6. The method for maneuver prediction of LSTM-based targets according to claim 1, wherein said step 4 is specifically:
the model of the mobile prediction network of the LSTM-based target comprises 1 LSTM network with 2 hidden layers, 3 fully connected layers and3 cross entropy loss functions, wherein the input of the LSTM network model is a training set data sequence normalized at continuous seq time, and the input format is (batch, seq, input); the output is the hidden state of the target machine maneuver in the packaged near seq time period
Figure FDA0003802111070000047
The output format is (batch, seq, hidden);
hidden state of target machine maneuver within near seq time period
Figure FDA0003802111070000048
Respectively inputting 3 full-connection layers, combining 3 cross entropy loss functions L a ，L p ，L y Respectively obtaining the probability condition of each type of value of the acceleration, the pitch angular velocity and the yaw angular velocity at seq + 1:
Figure FDA0003802111070000049
Figure FDA00038021110700000410
in the formula, L represents a cross entropy function calculation formula, N represents the number of samples, M represents the number of categories, y ic Is a symbolic function, if the true class of the sample i is c, y ic Get 1, otherwise y ic Taking 0; l is a ，L p ，L y Respectively representing acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss which are obtained by calculating acceleration, pitch angle speed and yaw angle speed by using a cross entropy loss function L; p a Representing the probability that the target acceleration takes each class of value at the time seq +1,
Figure FDA0003802111070000051
representing the probability that the target acceleration takes the mth class value at seq +1, P p Representing target pitch angle velocity at seq +1 time takes the probability case of each class of value,
Figure FDA0003802111070000052
represents the probability that the target pitch angular velocity takes the nth value at seq +1 y Representing the probability case that the target yaw rate takes each type of value at time seq +1,
Figure FDA0003802111070000053
the probability that the target yaw rate takes the nth-class value at seq +1 is shown.
7. The LSTM-based target maneuvering prediction method as recited in claim 6, characterized in that the category values with the maximum value probability of the target acceleration a, the pitch angle velocity P and the yaw angle velocity y are respectively used as the predicted category values of the target machine seq +1 moment acceleration a, the pitch angle velocity P and the yaw angle velocity y, P a pre ，P p pre ，P y pre ：
Figure FDA0003802111070000054
And combining the predicted category values of the acceleration a, the pitch angle speed p and the yaw angle speed y to obtain a predicted category label value of the maneuvering action, and converting the maneuvering prediction problem of the unmanned aerial vehicle into a classification prediction problem of the target acceleration a, the pitch angle speed p and the yaw angle speed y.
8. The method of claim 1, wherein step 5 is implemented for cross-entropy loss L of acceleration during model training a Pitch angle cross entropy loss L p Sum yaw angle cross entropy loss L y Performing weighted fusion, and taking the fused loss function value as the training loss L of the whole network model train ：
L train ＝w a L a +w p L p +w y L y
In the formula, w a ，w p ，w y Respectively representing the acceleration cross entropy loss L a Pitch angle cross entropy loss L p Sum yaw angle cross entropy loss L y The weight of (2).
9. An electronic device is characterized by comprising a processor, a memory and a communication bus, wherein the processor and the memory are communicated with each other through the communication bus;
a memory for storing a computer program;
a processor, configured to execute a program stored in a memory, to implement a method of maneuver prediction based on the LSTM objective of any of claims 1 to 8.
10. A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor, implements a method of maneuver prediction for LSTM-based objectives according to any of claims 1 to 8.
CN202210986452.6A 2022-08-17 2022-08-17 LSTM-based target maneuver prediction method, electronic device and storage medium Active CN115480582B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210986452.6A CN115480582B (en) 2022-08-17 2022-08-17 LSTM-based target maneuver prediction method, electronic device and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210986452.6A CN115480582B (en) 2022-08-17 2022-08-17 LSTM-based target maneuver prediction method, electronic device and storage medium

Publications (2)

Publication Number Publication Date
CN115480582A true CN115480582A (en) 2022-12-16
CN115480582B CN115480582B (en) 2024-12-03

Family

ID=84423028

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210986452.6A Active CN115480582B (en) 2022-08-17 2022-08-17 LSTM-based target maneuver prediction method, electronic device and storage medium

Country Status (1)

Country Link
CN (1) CN115480582B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116227952A (en) * 2023-05-09 2023-06-06 中国人民解放军海军潜艇学院 Method and device for selecting attack target defense strategy under key information deficiency
CN116383731A (en) * 2023-03-06 2023-07-04 南京航空航天大学 Tactical maneuver identification method, tactical maneuver identification system, electronic equipment and storage medium
CN117150738A (en) * 2023-08-10 2023-12-01 中国船舶集团有限公司第七〇九研究所 Action direction pre-judging method under complex scene

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112001433A (en) * 2020-08-12 2020-11-27 西安交通大学 A track association method, system, device and readable storage medium

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112001433A (en) * 2020-08-12 2020-11-27 西安交通大学 A track association method, system, device and readable storage medium

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
CARINA VOGL 等: "Frenet Coordinate Based Driving Maneuver Prediction at Roundabouts Using LSTM Networks", CSCS \'20: COMPUTER SCIENCE IN CARS SYMPOSIUM, 4 December 2020 (2020-12-04) *
MUHAMMAD WASEEM ASHRAF 等: "Dogfight: Detecting Drones from Drones Videos", 2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, 2 November 2021 (2021-11-02) *
杨星鑫 等: "基于LSTM的无人机轨迹识别技术研究", 现代计算机, 15 February 2020 (2020-02-15) *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116383731A (en) * 2023-03-06 2023-07-04 南京航空航天大学 Tactical maneuver identification method, tactical maneuver identification system, electronic equipment and storage medium
CN116383731B (en) * 2023-03-06 2023-11-14 南京航空航天大学 Tactical maneuver identification method, tactical maneuver identification system, electronic equipment and storage medium
CN116227952A (en) * 2023-05-09 2023-06-06 中国人民解放军海军潜艇学院 Method and device for selecting attack target defense strategy under key information deficiency
CN116227952B (en) * 2023-05-09 2023-07-25 中国人民解放军海军潜艇学院 Method and device for selecting attack target defense strategy under key information deficiency
CN117150738A (en) * 2023-08-10 2023-12-01 中国船舶集团有限公司第七〇九研究所 Action direction pre-judging method under complex scene
CN117150738B (en) * 2023-08-10 2024-05-10 中国船舶集团有限公司第七〇九研究所 Action direction pre-judging method under complex scene

Also Published As

Publication number Publication date
CN115480582B (en) 2024-12-03

Similar Documents

Publication Publication Date Title
CN115480582A (en) Maneuver prediction method for LSTM-based target, electronic device and storage medium
US11669110B2 (en) Control system based on multi-unmanned aerial vehicle cooperative strategic confrontation
EP2137481B1 (en) A method and a system for estimating the impact area of a military load launched from an aircraft
CN113190037A (en) Unmanned aerial vehicle optimal path searching method based on improved fluid disturbance and sparrow algorithm
TWI794516B (en) Training and/or assistance platform for air management via air traffic management electronic system, associated method
Zhang et al. Maneuver decision-making of deep learning for UCAV thorough azimuth angles
CN114676743B (en) Low-speed small target track threat identification method based on hidden Markov model
CN112749761A (en) Enemy combat intention identification method and system based on attention mechanism and recurrent neural network
CN114690790B (en) GRU-based rapid identification method for guidance laws of intercepted aircraft
CN117520950B (en) Multi-objective UAV fault diagnosis method based on attention knowledge sharing network
CN113625569B (en) Small unmanned aerial vehicle prevention and control decision method and system based on hybrid decision model
CN114266355A (en) Tactical intention identification method based on BilSTM-Attention
CN116796284A (en) Unmanned aerial vehicle cluster-oriented intention recognition method applied to countermeasure environment
CN115661576A (en) A method for aircraft group intent recognition under sample imbalance
Meftah et al. Improving autonomous vehicles maneuverability and collision avoidance in adverse weather conditions using generative adversarial networks
CN107832939A (en) Unmanned platform aerial opposition deduction method and device
CN118277800B (en) Non-cooperative unmanned aerial vehicle flight intention recognition method based on similarity measurement
CN118171742B (en) Knowledge-data driven air combat target intention reasoning method and system based on residual estimation
CN115035389A (en) Fine-grained Image Recognition Method and Device Based on Reliability Evaluation and Iterative Learning
CN119475984A (en) Method, device and storage medium for generating multi-vehicle interactive test scenarios for autonomous driving
CN118070967A (en) Ship track prediction method and system for self-adaptive multi-mode data
CN117572355A (en) Intelligent deception method and device for target recognition network model
CN117930860A (en) Action control agent training method based on reinforcement learning
CN117493986A (en) Unmanned plane bee colony interception decision-making method based on support vector machine
Han et al. Dp-gan: A novel generative adversarial network-based drone pilot identification scheme

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant