Adaptive Linear System Identification over Simulated Wireless Environment

Wireless technologies have become one of the basic industrial pillars, whereas system
identification represents an important tool in many practical engineering circumstances
and thus sooner or later both wireless technologies and system identification should be
linked together in sense of having an identifier that is able to reliably identify a system
over wireless links. It is well known that wireless links are considered as unreliable
medium and therefore the loss of the system observations across them is unavoidable.
The system observations represent the main element in the identification process since
the identifier relies only on these observations in order to identify the underlying function
of the system as they are the only information available to tell about the system
dynamics, for this reason vast amount of literature in the context of system identification
is written about the way the excitation signal is chosen to force the system to show its
dynamic and also about the way the sampling process is carried out to obtain informative
observations in order to construct a satisfactory model for the system. This shows that the
random loss of these observations (which are vital and core element of identification
process) might deter the system modeling process. Experience shows that well sampled
observations over regular intervals during observations loss could not guarantee a
satisfactory model for the system. This thesis looks into the concepts of system
identification and the behavior of the identifier components when placing wireless links
between the system and the identifier. The thesis investigates the possibility of
performing system identification over wireless network for both on-line and off-line
system identification approaches. This research work studies the effects of observations
loss on the performance of the learning algorithms and it focuses only on first order
autoregressive with exogenous input (ARX) model structure adopted on linear network.
The work looks thoroughly on three forms of instantaneous learning algorithms which
are: first order algorithms (e.g. least mean square (LMS)), second order algorithms (e.g.
recursive least squares (RLS)) and finally high order or sliding window (SW) algorithms
(e.g. moving average (MA)).