Tracking Airto-Air Missile Using Proportional Navigation Model with Genetic Algorithm Particle Filter

The purpose of this paper is to track the air-to-air missile. Here we put forward the PN-GAPF (Proportional Navigation motion model and Genetic Algorithm Particle Filter) method to solve the problem. The main jobs we have done can be listed as follows: firstly, we establish the missile state space model named as the Proportional Navigation (PN) motion model to simulate the real motion of the air-to-air missile; secondly, the PN-EKF and PN-PF methods are proposed to track the missile, through combining PNmotion model with EKF and PF; thirdly, in order to solve the particle degeneracy and diversity loss, we introduce the intercross and variation in GA to the particles resampling step and then the PN-GAPF method is put forward. The simulation results show that the PN motion model is better than the CV and CA motion models for tracking the air-to-air missile and that the PN-GAPF method is more efficient than the PN-EKF and PN-PF.


Introduction
The air-to-air missile is the main weapon in the air combat.The aircraft been chased should maneuver to avoid the attacking missile, after the air-to-air missile was fired by an opposition fighter.The aircraft guidance method for maneuver evasion is based on knowing the missile location and the real-time track [1].However, we cannot acquire the exact location information of the missile because the measurement for the missile location has great error.We need an on-line filter method to eliminate the error and track the air-to-air missile.How to effectively track the missile is the research content in this paper.
This problem is the domain of the single target tracking.The common target motion models such as the CV, CA, and Current Statistical (CS) models cannot well and truly describe the air-to-air missile maneuver because the missile has a good maneuverability and a supersonic speed [2].However, we can research from the navigation law point to establish a smarter motion model because the missile maneuver obeys some navigation law [3].The Proportional Navigation (PN) law is the most common in the air-to-air missiles [4].Therefore we establish a new PN motion model in 3d Cartesian coordinate system by analysis of the PN mechanism.
The state space model for tracking the air-to-air missile is obtained further, through combining with the nonlinear measurement equation in the radar spherical coordinate system.The standard Kalman Filter (KF) cannot be used to track the missile because of the nonlinear measurement equation.The Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) are used to solve the nonlinear problem, but the methods are not efficient enough in practice [5].Recently, the Particle Filter (PF) or Monte Carlo (MC) method is widely applied in the nonlinear and non-Gaussian filter problem because PF is to be able to represent the required unknown probability density function by a scatter of particles sampled from a known probability density function [6].However, the resampling of standard PF has a disadvantage of the loss of particle diversity [7].We are inspired by the evolution idea of Genetic Algorithm (GA) to improve the performance of the PF at resampling step [8].We take the particles and weights in the PF as the chromosomes and adaptability in the GA, respectively.Then the intercross and variation steps in the GA can be adopted to deal with the particles in resampling step for avoiding the particle degeneracy and loss of diversity.In this paper, Genetic Algorithm Particle Filter (GAPF) combined with the PN motion model (PN-GAPF) is used to track the air-to-air missile.PN-GAPF has better the estimate accuracy and the tracking stability, in comparison with other algorithms and models in computer simulation.
This paper is divided into five sections.Section 1 is introduction.Section 2 presents the state space model for tracking the air-to-air missile and the method of establishing the PN motion model.Section 3 describes the nonlinear filter algorithms such as the EKF, PF, and GAPF proposed.Section 4 discusses the performance of the PN motion model and GAPF algorithm on the base of the simulation results.Section 5 summarizes the main research content.

Problem Formulation
2.1.The State Space Model.The state space model for tracking the air-to-air missile consists of the missile state equation and the measurement equation.In order to simplify the problem, the missile state equation is usually modeled linearly in Cartesian coordinate system and the measurement equation is expressed nonlinearly in spherical coordinate system because of the air-to-air missile measurements given in the radar spherical coordinate system [9]..In general, we estimate the state of the moving air-to-air missile with the discrete-time linear state space in 3d Cartesian coordinate system [10]:

The Linear Missile State Equation
where X( + 1) and X() are the state vector at time  = ( + 1) and  =  respectively and  is a sampling period.
A is the state transition matrix.k() is the process noise which is modeled as a zero-mean white Gaussian process with covariance matrix Q().
The CV model and the CA model are the most common motion models for tracking air-to-air missile [11].The state vector X CV () and the state transition matrix A CV of CV model in 3d Cartesian coordinate system are And the state vector X CA () and the state transition matrix A CA of CA model are where (), (), and () are the missile positions along each axis of coordinates.Correspondingly, V  (), V  (), and V  () are the velocities.  (),   (), and   () are the accelerators.
If sampling the measurement in a very short time, we can consider the missiles maneuvering as an approximate CV model or CA model [11].

The Nonlinear Measurement Equation.
In the real air combat, the measurement of air-to-air missile is as follows in the radar spherical coordinate system [12]: where Z() is the measurement vector.() is the distance between the aircraft attacked and the air-to-air missile which can be detected by the airborne laser range finder.() and () are the pitching and azimuth angles, respectively, which can be detected by the aircraft radar warning device [13].
The nonlinear measurement equation is where h(⋅) is the nonlinear measurement function which will be given at the next section.w() is the measurement error which is zero-mean white Gaussian noise with covariance matrix R() [14].

The Proportional Navigation Motion Model.
If the air-toair missile movement is modeled just simply as the CV or CA motion model, it will produce great errors of the missile state estimation [4].It is known that the air-to-air missile is designed to maneuver by the Proportional Navigation law.Therefore, the motion model based on the Proportional Navigation law is a smarter model to reflect the states in the attack process.We describe the relative movement between the aircraft attacked and the air-to-air missile in Figure 1.

The Geographic Coordinate
System.We take "northsky-west" geographic coordinate system [15] as the 3d Cartesian coordinate system to describe the relative movement in Figure 1.The origin of coordinates "" is set at the firing point of the air-to-air missile.The "" axis is tangent with longitude and points to the north.The "" axis points to the sky and the "" axis points to the west.
The relative movement and variables.

The Relative Movement.
We assume that the attacked aircraft position and the missile position are (  ,   ,   ) and (  ,   ,   ) in "" coordinate system, respectively.
(,   ,   ) are the relative distance and the relative pitching and relative azimuth angles between them in the spherical coordinate system.The relationship of those variables is [16] The velocity vectors of the aircraft and the missile are (V  , V  , V  ) and (V  , V  , V  ) in the "" coordinate system; then (V  ,   ,   ) and (V  ,   ,   ) are in the spherical coordinate system.The relationships of those variables are Popularly, the Proportional Navigation law can be decoupled in the "" plane and "" plane [17]: where   and   are the Proportional Navigation coefficients, 2 ≤  ≤ 6. θ  φ  θ  and φ  are the derivations of       and   , respectively.

Modeling the Proportional Navigation State Space Model.
We assume that the velocity V  is a fixed value [18] because the roll booster of air-to-air missile works in a short time.Then we take derivative of ( 7): Combining ( 7) with ( 8), (10) will be changed to Now we take the missile state vector as (11), the continuous-time linear state equation [19] can be established as Then we convert (12) to the discrete linear state equation as The precise analytic expression of A PN cannot be acquired, but we use (15) to get an approximation by taking  = 3 [20]: The measurement equation can directly be modeled in discrete form like (5) because the measurements of the missile are detected discretely.According to (6), h(X()) is?

Nonlinear Filtering Algorithm
The Kalman filter assumptions do not hold because the measurement equation is a nonlinear function in the measurement equation [21].Therefore, for the tracking missile, there are three approximate nonlinear filters to solve the nonlinear problem in the paper.

The Extended Kalman Filter with the PN Motion Model.
The EKF approximates the nonlinear function through utilizing the first term in a Taylor expansion of the nonlinear function [22]: , , φ  and θ  should be estimated on the basis of the prior estimated information because they are unknown variables in state transport matrix: ( The EKF with the PN motion model to track air-to-air missile is shown in Algorithm 1.

The Sequential Importance Sampling (SIS) Algorithm.
The Particle Filter algorithm to deal with the nonlinear tracking problem has been used on a wide consensus [6].Let Z 1: = {Z(1), Z(2), . . ., Z()} denote a sequence of observations and (X() | Z 1: ) denote the posterior density of X().We can acquire Bayesian estimation X() and P() by The calculation of the posterior density requires integrals, but the normalizing integral cannot be evaluated analytically and numerical integration over possibly high-dimensional spaces is infeasible.The ideal of SIS algorithm is that the required posterior density function is represented by a set of random samples with associated weight {x () (),  () ()}

The Resampling
Step.After a few iterations with SIS, the variance of the importance weights increases because most particles will have negligible weights.This is called particle degeneracy.Fixing on an importance sampling proposal distribution, one way to mitigate the particle degeneracy is to adopt the resampling step.The effective sample size  eff denotes an index variable of degeneracy introduced in [6]. eff can be estimated approximately by If  eff falls below a threshold   , the resampling algorithm will sample from (24) to generate a new set {x () * ()}
The standard PF with the PN motion model to track the air-to-air missile is shown in Algorithm 2.  (22) to update A PN Calculate  eff using (29) IF  eff <   Take resampling algorithm End FOR Algorithm 2

The Genetic Algorithm Particle
Filter.The standard Particle Filter can solve the particle degeneracy, but the resampling algorithm relying on the high weight particles leads to the loss diversity [7].The evolution idea of Genetic Algorithm (GA) [8] can be adopted in the resampling step to solve the problem.The chromosomes and adaptability in the GA are corresponding to the particle and weight in the PF, respectively.So we can use the selection, intercross, and variation steps to deal with the particle in resampling step to increase the diversity.
GA for the resampling step is shown in Algorithm 3.
The GAPF uses the GA to replace the standard resampling step in the PN and increases the diversity, which is validated by simulations.

Hardware Condition
An Intel5Core6i3-2100@3.10 GHz CPU is used with 2.00 GB RAM and MATLAB 2010b.
Else leap over the resampling step END IF Algorithm 3

The True Ballistic Trajectory of the Air-to-Air Missile.
We assume that the aircraft attacked flies along a straight line at a constant velocity |V  | = 200 m/s with the pitching angle   = /120 and the azimuth angle   = /6 at the original point (10000, 500, 10000) in north-sky-west geographic coordinate system.We take the firing point of the air-to-air missile as the coordinate origin (0, 0, 0) and assume that the missile moves following the Proportional Navigation law, whose proportional parameter   =   = 4, at a constant speed V  = 500 m/s with the pitching angle   = 0 and the azimuth angle   = 0. Through the missile dynamics equation set (refer to [3]), we can acquire the ballistic trajectory of the missile and the measurement set Z with a second sampling interval in Figure 2.
In Figure 6, we adopt three filter algorithms, such as EKF, PF, and GAPF, to track the missile with the constant PN model.The PN-EKF algorithm is a comparison which is the same in Figures 3 and 4 PN-PF and PN-GAPF.The number of intercross particles is  1 = 0.3  and the number of variation particle is  2 = 0.005  in the PN-GAPF.Figure 6 shows that the GAPF performance is better than EKF and PF.The RSME curves in Figure 7 directly prove it.In Figure 8, we show that the average computation time of per step in the EKF is shortest among them, but GAPF needs more time in intercross and variation steps than the standard resampling algorithm in the PF.Further improvements in performance can be acquired by increasing the number of particles, but the computation time also becomes longer, which heavily leads to not meeting realtime tracking requirement.

Conclusion
This paper aims to solve the problem of tracking the airto-air missile.Firstly we establish the PN motion model based on the PN law to preferably track the air-to-air missile in 3d Cartesian coordinate system.Secondly, the nonlinear filter problem should be solved because of the nonlinear measurement equation in the state space model for tracking the missile.In the nonlinear filter algorithms, we take the standard EKF and PF to track the missile with the PN motion model.However, both filters' results are not good because the linear approximate degree is inadequate for EKF and the diversity of particles is deficient for the standard resampling step.We introduce the intercross and variation steps in the GA to the resampling step in the standard PF for increasing the diversity.The simulation results show that the GAPN, which replaces the standard resampling step with the GA, can obtain a better performance for tracking the missile with PN.Meanwhile, the PN model proposed in this paper is a more suitable model for the air-to-air missile than CV and CA models.

A:
The state transition matrix   (): The accelerator on -axis   (): The accelerator on -axis   (): The accelerator on -axis h(⋅): The nonlinear measurement function H  : The first-order Taylor expansion of the nonlinear function h(⋅)   : The Proportional Navigation coefficient in the "" plane   : The Proportional Navigation coefficient in the "" plane  eff : The effective sample size   : The true sample size : The sampling period P( |  − 1): The prediction state covariance matrix P( − 1): The estimation state covariance matrix (⋅ | ⋅): The conditional probability (⋅): Th e i m p o r t a n c e s a m p l i n g p r o p o s a l distribution Q(): The process noise covariance matrix R(): The measurement noise covariance matrix (): The distance between aircraft attacked and air-to-air missile V  (): Th e p i t c h i n g a n g l e (): Th e a z i m u t h a n g l e w(): The measurement noise k(): Th e p r o c e s s n o i s e .

Figure 2 :
Figure 2: The true ballistic trajectory and the measurements.

Figure 3 :Figure 4 :
Figure 3: The EKF track with the CV CA and PN.

Figure 8 :
Figure 8: Average computation time of per step.
Th e v e l o c i t y o n -axis V  (): Th e v e l o c i t y o n -axis V  (): Th e v e l o c i t y o n -axis X(): Th e s t a t e v e c t o r (): Th e p o s i t i o n o n -axis (): Th e p o s i t i o n o n -axis (): Th e p o s i t i o n o n -axis Z(): Th e m e a s u r e m e n t v e c t o r {x () (),  () ()}   =1 : The set of random samples x () () with associated weight  () () ():