Sei sulla pagina 1di 83

Introduction to Interactive Multiple Model

Dr. A K Sarkar
Scientist
Directorate of Systems
DRDL, Hyderabad-58

Ananthasayanam M R and Sarkar A K: Interacting Multiple


Model
Based State Estimator For Radar and Seeker Application,
Paper No. AIAA-2006-6242-CP (2006).
Workshop on
KALMAN FILTERING AND ITS APPLICATIONS,
conducted by IISc, Bengaluru, at LRDE
01-03-2011
•Introduction

•IMM Algorithm (Physical Concept)

•Different EKF Models

•Simulation Results

•Conclusions and Further Research


Problem Definition

To design a target tracker from onboard seeker


with following characteristics

•Should track a randomly maneuvering target

•Estimator should have minimum tracking error and lag

•Should be robust enough for tracking target with different


levels of maneuver
IMM Algorithm
(An Overview and Historical Perspectives)

•Original Contribution by Y Bar-Shalom (1974) In Context


of Joint Probabilistic Data Association (Sonar Application)

•Probably First Published work on IMM Application for


Radar Tracking Problem By Barshalom and Lee in 1992
(IEEE in AES)

•Bar-Shalom et al. applied IMM algorithm in context of Multi


Target Tracking and Track Data Association

•IMM Successfully Applied in context of Tracking in


Presence of Glint Noise (Wu)
What IMM Is?

•It is an Adaptive Estimator

•Finite number of state models to mimic different target


Dynamical situations

•Different levels of process noise in each model

•All state models running in parallel

•At any instant of time IMM output is weighted average of


All model outputs

•Weight calculation is probabilistic ( innovation based )


(Transition probability)

Physically mode probability means


probability of occurrence of a given model

(posterior mode probability)


Different Steps in IMM
Initialize IMM with Markov Chain Transition Probability Matrix
and Posterior Mode Probability of Different State Models

Initialize Each EKF Model

Start Recursion

•Mix States and Covariances of Different Models Using Above


Mentioned Transition Probability Matrix

•Propagation and Update of Mixed States and Covariances of


Each EKF Model based on Measurements

•Calculation of Maximum Likelihood (ML) Function of Innovation of


Each Model

•Update Mode Probability of Each Model Using ML Function

•Combine Estimate of State and Covariance Based on Estimated


States and Covariances of Each Model and Posterior Mode
Probability Update

Go To Beginning of Recursion


Use both transition
and mode probability

Update mode probability


based on likelihood function
Seeker Measurements and Seeker Filter Design
(Recapitulation)
Schematic Diagram of Pursuer
and Evader Engagement
•Pursuer Body Frame
•Pursuer Fin Frame
•Pursuer Seeker Gimbal
Frame
•Local Vertical Frame
•Inertial Frame

Different Axes System For Seeker


PROBLEM DEFINITION
• Estimation of Relative Position,
Relative velocity and Target acceleration
from noisy (Range, Range rate, Gimbal
angles and Gimbal angle rates )
of Seeker Measurements

CONSTRAINTS
Measurement Data at 10 milli seconds

interval
• Measurement noise is non gaussian due
to eclipsing, glint and thermal noise
•There is an aperiodic data loss in LOS
rates due to eclipsing
Schematic Diagram of Seeker Model (pitch plane)
Different Seeker Measurements
Measurements Eclipsing Zone Non Eclipsing Zone

Range A A

Range Rate A A

Gimbal Angle (yaw) A A

Gimbal Angle (pitch) A A

LOS rate (yaw) NA A

LOS rate (pitch) NA A

NA = Not Available
A = Available
Total 6 and 4 measurements during non eclipsing and
eclipsing time zones respectively
A Bank of Kalman Filters Kept in Parallel

State Equations:
•Constant Velocity (CV) (position, velocity)

•Constant Acceleration (CA)


(position, velocity, acceleration)

•Constant Jerk (CJ)


(position, velocity, acceleration, jerk)

Measurement Equations:

Range, Range Rate, Gimbal angle (yaw),


Gimbal angle (pitch), LOS rate (yaw),
LOS Rate (pitch)
KALMAN FILTER FORMULATION (CP Frame)

•State Equations in Cartesian Frame (Linear)

•Measurement Equations in Polar Seeker


Gimbal Frame (Nonlinear)
State to measurements transformations
State Equations
Xˆ1 (k -1 k -1), P1 (k -1 k -1) Xˆ2 (k -1 k -1), P2 (k -1 k -1)

Interaction / Mixing m (k -1 k -1)

Xˆ01 (k -1 k -1), P01 (k -1 k -1) Xˆ02 (k -1 k -1), P02 (k - 1 k -1)

L1 (k ) L 2(k )
Filter Filter
Z(k)
M1 M2

Xˆ1 (k k ), P1 (k k ) Xˆ2 (k k ), P2 (k k )

Xˆ1 (k k ), P1 (k k )
X̂1 (k k )
Mode
L1 (k )
State estimate
probability m(k k ) and
Xˆ2 (k k ), P2 (k k )
update and

L2 (k ) P1 (k k )
mixing covariance
probability m(k )
calculation m(k ) combination

Fig- 1 : A Schematic Showing One Cycle of 2-Model Estimation Algorithm


Target Random
Random IMM Estimator Estimated States
Kinematics Acceleration
Acceleration
x̂
-
Missile + Predicted State Predicted
Kinematics & Update Measurements

Seeker
Noise + +
+ Seeker
Measurements

Achieved
Acceleration Autopilot & Commanded Acceleration Guidance
Actuator Law

PURSUER EVADER ENGAGEMENT SIMULATION IN CLOSE LOOP


SIMULATION RESULTS
(IMM Consists of Two EKF Models
CV, CA)
Targat latax history along yaw and pitch plane
SYSTEM OUTPUT RESPONSE
( IMM WITH CV and CA EKF
IN CP FRAME)
True, Measured and Estimated Range
with Estimation Error
True, Measured and Estimated
Range Rate and Estimation Error
True, Measured and Estimated
Gimbal Yaw Angle and Estimation Error
True, Measured and Estimated
Gimbal Pitch Angle and Estimation Error
Attenuation From 7-14 sec (20 deg/s to 0.3 deg /s)
During end game 0.2 sec lag

Measured Typical Yaw LOS rates


Time History in Gimbal Frame (Q=0)
Measured Typical Pitch LOS rates Time
History in Gimbal Frame (Q=0)
Measured Typical Pitch LOS rates Time History in Gimbal
Frame (Q-CV = 1, Q-CA=10)
OBSERVER STATE HISTORY
Comparison of IMM Estimated delta-x (m) with true values
And estimation error
Comparison of IMM Estimated delta-y (m) with true values
And estimation error
Comparison of IMM Estimated delta-z (m) with true values
And estimation error
Comparison of IMM Estimated delta-V_x (m/s) with true values
And estimation error
Comparison of IMM Estimated delta-V-y (m/s) with true values
And estimation error
Comparison of IMM Estimated delta-V_z (m/s) with true values
And estimation error
Comparison of IMM Estimated and True
Acceleration Time History
Comparison of Estimated States
from IMM, CV EKF (Model # 1),
CA EKF (Model #2)
Comparison of Estimated delta-x (m) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Comparison of Estimated delta-y (m) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Comparison of Estimated delta-z (m) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Comparison of Estimated delta-Vx (m/s) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Comparison of Estimated delta-Vy (m/s) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Comparison of Estimated delta-Vz (m/s) with true values
And estimation error (IMM, CV-EKF, CA-EKF)
Time History of Different Mode Probabilities
(CV, CA model)

Target Maneuver History


(CA, CJ model)

Target Maneuver History


6 DOF Simulation Results With
IMM (CV + CA) Based Seeker Filter
in Close Loop
Demanded and Feedback Control Surface Deflections ( EKF)
Pitch and Yaw Latax Autopilot Performance
Missile Body Rates, Velocity and Mach No Time History
Missile Position, Angle of Attack and Side Slip Angle History
Missile Target Interception Trajectory in Pitch and Yaw Plane
•IMM Estimator with each consisting of
two state models used

•IMM1 consists of CV EKF and CA EKF


in parallel

•IMM2 consists of CA EKF and CJ EKF


in parallel
INTROSPECTIONS AND CONCLUSION
• Though Constant velocity, Constant acceleration and
Constant jerk models have been tried, IMM with CV + CA
or IMM with CA + CJ models perform close to CJ EKF in
stand alone mode

• Through simulation experience it is felt that only 2


models at a time is good enough to track a randomly
maneuvering target

• In EKF judicious choice of Q-matrix elements is


mandatory. In IMM choice of Q-matrix is less critical.

• Target data has to be generated using Aircraft 6DOF


model to mimic real world scenario in better way

• Glint noise has to be modeled in estimator for


minimizing oscillations during end game
•IMM Algorithm Explained with physical
concept

•Transition and mode probability explainded

•Different EKF Models (CV, CA, CJ) explained

•Simulation Results with randomly


maneuvering target described

•Conclusions brought out


FUTURE WORK
•Different Types of Maneuvers, Low to High Maneuvers and
Coordinated Turn

•Sensitivity study on variation of mode probability matrix


has to be carried out

•Minimization of Lag and High Noise Attenuation

• Real time implementation of present IMM Algorithm

•Implementation of IMM Estimator for radome slope error


estimation and glint noise removal

•Validation of IMM estimator over launch boundary consisting


of range of (6-90) km and altitude of (0.5-15) km of operation

•IMM has to work for different levels of aircraft type of


target maneuver e.g. Coordinated Turn

•Monte Carlo Simulation within complete kill zone using


IMM estimator in close loop 6DOF model
THANK YOU FOR THE
PATIENT HEARING !

Other Contributors
Dr. M R Ananathasayanam

Potrebbero piacerti anche