model, which is built up here by the application of autoregres-
sive integrated (ARI-) processes. The time-characteristic of
each orientation parameter will be modelled by a specific
representation of an ARI-process.
In chapter 2 the theoretical background of the time-series
model will be outlined. Chapter 3 will present some applica-
tions of the theory : 1. modelling the position parameters of
an aircraft flight measured by the Global Positioning System
GPS, and 2. modelling the attitude parameters of an aircraft
and of a space shuttle measured by an Inertial Navigation
System INS.
2. Theoretical background
2.1 Autoregressive integrated processes
Autoregressive integrated (ARI-) processes are a widely applied
class of stochastic processes for various kinds of time-series
(Haykin 1979, Kay and Marple 1981).
An autoregressive (AR-) process of order p describes a statio-
nary time-series x: by
P
Xt = — I aı-Xt-ı + et V(et) = ce? (1)
1=1
The time-series considered in our applications are generally
not stationary. These time-series have to be transformed into
stationary time-series by the elimination of trends. A usual
way is to take derivatives of the time-series. The autoregres-
sive integrated process of order (p,d) is achieved if the d-th
derivative of the original time-series can be described by a
stationary AR(p) process. Formally, the derivation can be ex-
pressed by an additional number d of process parameters ai in
eq.(1). Then the ARI(p,d) process fully describes the dynamic
behaviour of any orientation parameter by a number of (p+d)
process parameters ai and by the variance ce 2 of the prediction
errors.
2.2 The filtering algorithm
Two kinds of equations build up a Gauss-Markov model to esti-
mate the filtered time-series x: from the observed series y:
E(Ys) ext V(yt) = On 2 (2)
ai *Xt-1 Viet) Oe 2 (3)
T
o
1
"
M
E(et )
Eq.(2) expresses the observation process with on 2 being the
variance of the observation noise. Eq.(3) presents the ARI
model to which the unknown time-series xt has to correspond.
The estimation of the filtered time-series Xt presumes the
knowledge of the stochastic part of the Gauss-Markov model,
i.e. the variances on 2 and oe 2. As the variances in general are
a priori unknown, they must be estimated by applying the
variance component estimation (VCE) technique. The formulation
in the frequency domain published by Fórstner (1984) is recom-
mended for our application.
a3