A Reduced-Order Model


Integrated GPS/INS
Xiufeng He, Yongqi Chen and H.B. Iz The Hong Kong Polytechnic University


The dominant factor in determining the computation time of the Kalman filter is the dimension n of the model state vector. The number of computations per iteration is on the order of n3. Any reduction in the number of states will benefit directly in terms of increased computation time. In this paper, a high order model in integrated GPS/INS is described first, then a reduced-order model based on the high-order model, is developed. Finally, a faster tracking approach for Kalman filters is discussed. A typical aircraft trajectory is designed for a complex high-dynamic aircraft flight experiment. A Monte Carlo analysis shows that the reduced order model presented in this paper provides satisfactory accuracy for aircraft navigation.

Kalman filters have been widely used in the integration of the GPSDNS system, both the dynamic model and observation model are required to describe the integrated system. The dynamic model is usually described

by a linear differential equation involving the system errors of the INS and GPS systems. The observation models are obtained from a combination of INS and GPS measurements. The computation time of the Kalman filter depends on the dimension n of the integrated system model state vector. The number of computations per recurrence is on the order of n3.Any reduction in the number of states will benefit directly in computation time. Moreover, the reduction of an integrated system model will bring many benefits to engineering realization. Since most currently available control design methods only work on small dimension systems; the complexity of a higher order model often makes it difficult to obtain a stable system. The objective of model order reduction is to find a lower order model which preserves the dynamics of more complex high-order systems. In this paper, a 21-states high-order integrated model is investigated first, based on this, a 15-states reduced-order model is developed by eliminating states which are unobservable or weakly observable. A Monte Carlo analysis shows that the reduced-order model presented in this paper provides satisfactory accuracy for aircraft navigation.
DESCRIPTION OF THE INTEGRATED GPS/INS SYSTEM The Dynamic Error Model In this paper, an INS platform is considered as a local-level, NEU (north, east, and upper), wander-azimuth system. A high-order dynamic error model of Integrated GPS/INS can be written as:

Authors’ Current Address: Department of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Kowloon, Hong Kong. Manuscript received May 1, 1997 0885-8985/98/ $10.00 0 1998 IEEE


IEEE AES Systems Magazine, March 1998

is x-velocity error. is y-velocity error.X ( t ) = A ( t ) X ( t )+ G ( t ) W ( t ) (1) where X(t) is derived of the state vector X(t). where (6x1. is y-gyro constant drift rate. . In the cascaded method. is transformed from a wander-azimuth system. is error in east longitude. it contains eighty percent zero elements and is a sparse matrix. 6zl) is the positioning error of the platform INS. In the observation model of inltegrated GPS/INS. =[b-XJ2 +o/ -&I2 +(z -Z\/)'l' +&/ (3) where (x. Zi) in the ECEF coordinate system. vj is the GPS receiver measurement noise. March I998 41 . is y-accelerometer zero bias. y. yi. is the distance error. Then with is attitude (errorangles. They are sparse matrices. Eq. 6. Assume that the platform INS position ( Xi. mainly caused by clock biases: 6pj = 6 . Ysj.(O = PI. H(t) is the integrated system measurement matrix. + v j where 6iu is a clock bias. is z-acce1e:rometer zero bias. Pij. is error in altitude. it contains eighty-five percent zero elements. V(t) is the GPS receiver measurement noise. is clock bias and clock drift rate. Then (4) 1 XI = x +sx. z) is the true position of the aircraft in the Earth Centered Earth Fixed (ECEF) coordinate system. respectively. its own advantages and disadvantages. Let GPS pseudorange measurements be Pji then the observation equation can be expressed as Z.)* +b. (5) YI = Y + & ZI = z+&. is z-gyro first-order Marov drift. is x-gyro first-order Marov drift. In ' where 62(4) is a correction to the difference between the GPS and INS measurement. the measurement consists of the navigation solution which is three-dimensional position and velocity. (2) where zpJ(t) is the discrepancy between the measurements given by the two systems. (Xsj. this example the tightly (coupledmethod is used. is azimuth error angles.)21i (6) sz(t)= H(t)x(t)+ V ( t ) (7) The Observation Model Two traditional methods of GPS/INS integration are the tightly coupled and the cascaded approach. From the platform INS position. 6y1. is z-velociiy error. ZSJ ) is the jth satellite position in the ECEF coordinate system. The matrix G(t) (21 by 11) is the coefficient matrix. The vector of dynamics noise W(t) is given by I P.)2 The matrix A(t) (21 by 21) is the integrated system dynamic matrix. (2) linearized form is: +(Z/ -z. is x-gyro constant drift rate. and X(t) consists of various errors such that platform INS and GPS. the measurement vectors consist of combinations of the IEEE AES System Magazine. is wander-(pzimuthangle error. is z-gyro constant drift rate. the distance plj between the aircraft and jth satellite can be calculated. is error in north latitude. For a more detailed discussion see [ 5 ] . satellite pseudoranges are used as measurements by the integration filter. is y-gyro first-order Marov drift. 'KX/ --T. is then written as P!. -Y\. In tightly coupled methods. is x-accelerometer zero bias. it contains ninety-six percent zero elements. MODEL ORDER REDUCTION METHOD INVESTIGATION Controllability and Observability The ideas of system controllability and observability play a key role in the presentation and understanding of model reduction since the stability of the Kalman filter is determined by these two basic characteristics.P. Each of the methods ha!.

if A(t) has a 1 of its eigenvalues in the open left half of the plane then P and are the unique symmetric positive semi-definite matrices which satisfy the following Lyapunov equation lis +HTH=O The connection of system controllability and observability to and 6 is said to be [4]: a) f. Reduced Order Model for Integrated GPS/INS Model order reduction methods have been presented in numerous research papers. that is. H) is completely observable.. the observability of system (1) and (7) can be determined by computing the singular values of the matrix M. which are considered white noise. For integrated GPS/INS systems. Here we consider observability. The observability of the system can usually be determined by examining the observability matrix M [l]: of singular values is less than the dimension n. These methods can be divided into two groups. in system (l). LH(n)A(n)A(n-l). If a dynamic system is observable. Modal methods. and only if. they cannot be eliminated and M= . system (1) and (7) are observable. Thus. Otherwise they are unobservable. Since the use of the clock and clock drift bias states are required in the tightly coupled case.The concepts of controllability and observability are embodied in the system controllability and observability grammians. three states of gyros constant drift are removed. we can determine which states are observable and which are unobservable using a Kalman filter. these states may be eliminated while still preserving the basic performance of the dynamics system. VTV = I are the singular values of the matrix M. Since matrices A(t) and H(t) have special properties (sparse). is positive definite if.. those states which are unobservable contribute little to the response of the system. such as Aggregation methods. Contrarily. if the number 42 IEEE AES Systems Magazine. that is. If the number of nonzero singular values is equal to the dimension n. the system is unobservable. Mathematically. and three states of accelerometers zero bias. Consider singular value decomposition (SVD) of the matrix M such that M = UWT = where UTU = I. b) 6 is positive definite if. According to [2].. A more detailed description of this approach is found in [7]. but not suggested in practice because of the complex calculation. For integrated systems (1) and (7). For system (1) and (7). the systems (1) and (7) are observable. are eliminated. the controllability and observability grammians can be defined as [4]: The controllability grammian is m = l e A ' G GT e A T ' d t 0 rcJ The observability grammian is e= [ e A r ' H H T e A' d t 0 (9) For system (1) and (7). to eliminate some states in order to reduce computational difficulties. since the methods of model order reduction are applied in engineering practice. however. G) is completely controllable. we are interested in the first group of methods. (A. Lyapunov function methods and Perturbation methods [6]. it is unobservable.A(O)] If M is a full-rank matrix. Another approach is based on applying an identification procedure to input-output data obtained by driving the original system with a specific input. An approach to eliminating states is based on the link between the physical system and its mathematical models. The first group attempts to retain the dominant modes of the original system. studying the possibility of estimating the state from the output. and only if. (A. an approach called the classical manner of exploiting sparsity can be employed so as to carry out the singular value decomposition of matrix M.. the controllability and observability of the system can be dealt with using the same method [l]. A rough approach is used to examine the observability of integrated systems. Instead of determining the rank of matrix M. accelerometers zero bias and user clock and clock drift bias states cannot be observed by the Kalman filter. Thus. if a state cannot be estimated from the output. all the states of the dynamic equation can be estimated at the output. March 1998 . Gyros constant drift. The above method of examining the observability of integrated systems is workable theoretically. It is not easy to examine observability in systems (1) and (7) because of the difficulty in determining the rank of matrix M.

respectively. this reduces the number of multiplications.4 7Kp7sv. respectively. 43' The matrix A ~ ( t ) ( 1 5 by IS) is the reduced-order integrated system dynamic matrix. HR(t) is the reduced order integrated system measurement matrix. y(k) and H(k) are sparse matrices. March 1998 43 . most of the computation time concentrates on Q. k- +r(k)@)rf (k) (17) q k ) = q k / k-l)HT(k)[l@jl'fk/ k-l)@(k)+hfk)]-' f l k / k) = [I-K(k)H(k)llik/ k -1)[1. A Kalman filtering algorithm with closed loop control is employed.(k)P(k/ k)@T(k) in Equation (17) [3]. only L2 . the conirol aim is to eliminate these errors so a direct control law can be considered as where P(k/k) and P(k/k-1) are update covariance matrices and redicted covariance matrices.are treated as consider states increasing the filter's robustness [5]. It is suggested individual elements be used for the Kalman filter. c (k) is a state transition matrix. which require 2n3 = 2 x 153 = 6750 multiplications.$. 9 C W .~ ( k ) longitude error@) Height i(k/k-l)=O (22) (15) (16) Z(k) = H(k) ' X(k ) + V ( k ) where Q. y (k) is the coefficient matrix.%p.A reduced-order dynamic of the integrated GPSDNS can be written .E. that is. The algorithm is summarised as q k i k-1) =qk)p(k-i. the matrix G ~ ( t ) ( 1 5 by 11) is the coefficient matrix. H(k) lis the measurement matrix.*(k / k) and (k / k -1) are updated state vector and predicated stare vector.x(k)+ r ( k ) . Since state transition @(k) is a sparse matrix. w(k) and v(k) are white noise with associated variance matrices Q(k) and R(k).K Tp SIMULATION (21) To investigate the behaviour of the reduced-order integrated system as compared to the i( k / k ) = K( k)Z( k ) IEEE AES Systems Magazine.4. K(k) is gain matrix. respectively.p. instead of treating matrix @ ( k ) .8 = 2961 multiplications are now required.J2 2 2 = 55 . Thus a large com utational savings can be obtained. A FASTER TRACKING APPROACH FOR KALMAN FILTERJNG Consider the discrete form of the above reduced order integrated GPS/INS system as: x(k+ 1) = q k ) e . The observation equation can be written as: where 6ZR(t) is a correction to the difference between GPS and INS measurements. V(t) is the measurement noise. . . For the above Kalman filter algorithm. where L and J denote the number of non-zero elements and 1 elements in the state transition matrix. Navigation Errors Comparing High-Order Model (21 States) with Reduced-Order Model (15 States) X/&) = k4. respectively. If the multiplication with non-zero elements need be performed.as: Table 1. For @(k) P (k/k)@ (k) calculations. approximately seventy-five percent of the elements of the matrix are zero.@7a&& E. cP(k).qk)H(k)]' (18) Iwww) U(k ) = -2( kI Ik ) State estimate vector can be expressed as: (19) Since the state vectors of integrated systems are error elements.

This is caused by simplifying the gyro and accelerometer models.06 s ..A z i m u t h . A comparison of Figures 1 and 2 show no difference in position error using high-order and reduced-order integrated models.5 I 0 I 2000 4000 Tim e ( s e c I 6000 Fig. pitching. shows aircraft navigation error accuracy using high-order and reduced-order integrated models. Fig.02 s 2000 4000 6000 Tim e ( s e c ) Yes high-order system. and maneuvers for a total time 5300 seconds.3 0.1 E t 0 2000 . but the big difference in azimuth error angle $2 and z-velocity error occur using high-order and reduced-order models. respectively. Table 2 shows computation time comparing the high-order with reduced-order integrated model.- ~. Aircraft Velocity Error Curves (%States) 1) Reduced-order integrated models can provide satisfactory accuracy for aircraft navigation.15 s 6000 T im e ( s e c ) ~~ 15-states 0.Roll . _ Pitch . Based on above simulation results. Computation Time Comparison of 15 States Reduced Model with and without Faster Tracking Kalman Filter h U C 800 600 Faster Tracking Computation Time in Each Step m u) L 0 U 400 200 2 Approach for Kalman Filter No t U a l 0 -200 3 c . CONCLUSION Fig. It is assumed that an aircraft takes off..1 -0.E 800 600 . 2B. flies.3 L P .8 >” 0 -0. and turning.....8Vxp. previous page.- 2 a . $y and velocity error div 8VxP. Table 3 shows computation of reduced-integrated model with and without faster tracking approach. Table 1. There is a small difference in attitude error angles $x. the following conclusions can be drawn: 44 IEEE AES Systems Magazine. a Monte Carlo simulation test is carried out.1 -0. Aircraft Attitude Error Curves (15 States) U) 0. Aircraft Velocity Error Curves (21 States) 0.. rolling.06 s 0. An aircraft trajectory is designed for actual and complex high-dynamic aircraft flight.3 5 0.- k i I : -0 1 -0 3 -0 5 P i 3 4000 6000 Fig. The maneuvers include climbing. 1A. 1B.s U 0 0. Integrated system state Computation time in each step 0 2000 4000 2 1-states 0.. 2A... March 1998 . Figures 1-3 show navigation error curves of high-order and reduced-order integrated GPSDNS. Aircraft Attitude Error Curves (21 States) Table 3.

. E. -- . Jeffrey Eric. 67%.y-psi ~ I 4000 6000 ~- __ . which brings benefit to engineering 161 Zhou Tong.edu Conference WWW address: www.1998 Snowmass at Aspen. Aircraft Position Error Curves (15 States) 2) The computation time of the Kalman filter is proportional to the cube of the number of integrated system state vector. No. 119-121.- O - U) 2 ! 0 0 0 x-posi ~ 4000 - 6000 [4] Mason. Wiley-Interscience Publication.y-posi . Kluwer Academic Publishers. CA 90077-2222.ucla. in fact. 1967. [3] Frank L.. Automatic Control. 1988.org IEEE AES Systems Magazine. Linear System Theory and Design. University of California. it reduced computation time approx. Fig. 1991. USA Telephone: 310-472-8019 E-mail: johnson@ee.aeroconf. Karatsinides. 3B. . Improving Design Assessment and Simulation of Large-scale Dynamic Systems.. March I998 45 . REFERENCES [ l ] Chen. Colorado Sponsor: Aerospace & Electronics Systems Society of the IEEE Contact: Mike Johnson or Beth Leitereg 2225 Roscomare Road Los Angeles.- [2] Davison. Vol. IEEE Transactions on Aerospace and Electronic Systems..20 1 - n -20 I 0 2000 Tim e(sec ) x-psi . A Method for SimplifyingLinear Dynamic System. 3A. IEEE Trans. Enhancing Filter Robustness in Cascaded GPS-INS Integrations.. pp.. Ph. Rinehart and Winston... The time reduction was approximately 60% by the model reduction from 21 Order to 15 Order. Computational Methods for General Sparse Matrices. 3) The constant drift rate of gyros is removed in reduced-order integrated systems.. 30. C. [5] Spiro P.. Fig. Michigan State University. Lewis..T... 1984. The faster tracking approach for Kalman filters presented in this paper is efficient for the integrated system. 1998 IEEE Aerospace Conference March 21-28. AC-12. 4.z-psi realization. Optimal Estimation With an Introduction to Stochastic Control Theory. Ph. Aircraft Position Error Curves (21 States) A E 20 10 e L a C 0 -10 0 .D thesis.J. Berkeley.D thesis.. Holt. 1986. 171 Zahari Zlatev. October 1994.2-posl Identification Using Low Order Models. 1988. the constant drift rate of gyros can be measured and precompensated in order to further improve navigation accuracy..