Errrors-in-Variables identification
Optimal filtering in presence of observation errors on the process
input
Blind identification and deconvolution of communication
channels
Identification of autoregressive models in presence of additive
noise
Identification of FIR, ARX and ARARX models with additive noise on
input and output observations
Realization and identification of multivariable systems
Fault diagnosis
Development and test of advanced e-learning environments
Errors-in-Variables identification
Most procedures used in the identification of dynamical
processes rely on the assumption that observation errors are
present on the process output but not on the input. These errors,
however, affect often all observations so that previous methods can
prove unreliable in several applications. The identification of
models where errors are assumed as present on both input and output
observations is called “EIV (Errors-in-Variables) identification”
and is remarkably more complex than traditional approaches,
particularly when the variance of the additive observation noises
is not assumed as a priori known. The researches performed in this
area have led to the extension of the Frisch scheme to the dynamic
case and to the development of robust and computationally efficient
procedures
Optimal filtering in presence of observation errors on the
process input
The class of stochastic models usually considered in filtering
applications assume additive noises with known statistics acting on
the state and on the output of the considered process; this is, for
instance, the stochastic context of Kalman filtering. Filtering
consists in computing the minimal variance estimate of the state
and of the output of the system that has generated the
observations. Such a context can be realistic when the process
input has been generated by a controller but is much less realistic
in all other cases. In the context of this research theme the
stochastic environment has been extended to the presence of
observation noise on the process input. After properly redefining
the filtering problem in this extended context, new filtering
methodologies have been introduced and developed in order to
optimize their robustness and computational efficiency. These
methodologies include, as particular cases, both Kalman filtering
and Errors-in-Variables filtering where observation errors on the
input and output but not on the state are considered.
Blind identification and deconvolution of communication
channels
Blind channel identification consists in estimating a channel
model on the only basis of a knowledge of its output affected by
additive noise and without any knowledge on the input; only its
statistics are, in some cases, assumed as known. The obtained model
is then used for reconstructing the unknown input. This problem is
known as blind equalization or deconvolution and plays an important
role not only in telecommunications but also radioastronomy,
seismology, biomedicine etc. In the context of this research area,
new approaches have been developed by exploiting the properties of
errors-in-variables models that require only the knowledge of
second order statistics and do not introduce limitations on the
input. Differently from previous approaches these methodologies can
be applied also to unbalanced channels affected by different
amounts of noise and have shown a remarkable robustness degree also
in conditions of poor signal to noise ratios. The developed
techniques have been applied to the solution of some noise and
reverberation cancellation problems in cooperation with the
Laboratory of Automation and Signal Processing (ENSEIRB) of
Bordeaux University.
Identification of autoregressive models in presence of
additive noise
In several applications like radar, sonar, spectral estimation,
geophysics and speech enhancement, the signals are modeled by means
of autoregressive (AR) processes where also and additive
observation noise is present. In such a context the traditional AR
identification techniques are no longer consistent; new procedures
have been developed by remapping the problem into an EIV
identification one. The obtained procedures allow estimating both
autoregressive coefficients and noise variances. Also new filtering
and interpolation algorithms for AR processes have been developed
and subsequently applied in the reconstruction of speech signals in
cooperation with the Laboratory of Automation and Signal Processing
(ENSEIRB) of Bordeaux University.
Identification of FIR, ARX and ARARX models with additive
noise on input and output observations
This research has considered the extension of classical FIR, ARX
and ARARX equation error models given by the introduction of
additive input and output observation noise. In FIR models the
presence of additive input noise describes errors due to sampling
and quantization. In ARX and ARARX models the considered context is
particularly suitable for diagnosis and fault localization where
the additive noises describe measurement errors (transducers noise)
while the equation error describes process disturbances. These
models find also applications in econometrics.The new
identification procedures that have been proposed are characterized
by an excellent ratio between accuracy and computational
efficiency; moreover they do not require any a priori knowledge of
the ratio between input and output noise variances. These
procedures rely on the properties of the dynamic Frisch scheme and
on those of high order Yule-Walker equations.
Realization and identification of multivariable
systems
The identification of multivariable systems is a complex problem
not widely covered by the existing literature. This derives from
the impossibility of performing simple extensions of SISO
procedures to MIMO systems. Important contributions in this area
concern the introduction of canonical forms for MIMO systems
parameterized by the image of the system model in a set of
independent invariants for the equivalence relation given by a
change of coordinates in the state space. More recently a canonical
representation for a class of nonstationary MIMO models and its
application in realization problems have been introduced. A new
geometric approach to the EIV identification of MIMO processes
based on the association of models to directions in the noise space
has also been developed. This approach allows to overcome the
congruence problems that would otherwise be present in the
estimation of the variances of the additive noises on the different
process inputs and outputs.
Fault diagnosis
Fault diagnosis in complex dynamic processes constitutes a
research area well covered by the literature during the last
decades. A fault can be defined as any abnormal working condition
outside the acceptable range defined for the considered process.
The faults can concern transducers, actuators or other system
components. In this area a new geometric approach based on the
properties of the locus of solutions of the dynamic Frisch scheme
has been introduced. This approach relies on the direction
variations of the vector describing the total noise on the
observations in the noise space. This new methodology has been
applied with success in the diagnosis of the fault of the flaps
actuator in an aircraft where fly-collected data have allowed not
only a fault diagnosis but also to evaluate its incidence.
Development and test of advanced e-learning
environments
Roberto Guidorzi has performed, in 1995, the first structured
experiments in the application of ICT (Information and
Communication Technologies) at Bologna University and,
subsequently, has developed advanced environments based on the
development of platform-independent virtual laboratories used by
the students of several courses at Bologna and in other
universities. In this context he has subsequently coordinated the
project that has led to the development of the e-learning platform
of Bologna University, AlmaChannel. These researches are mainly
based on the constructivism and cooperative-learning paradigmata
and allow students to access flexible learning environments that
can give proper support to students' specific needs. These
researches, which are still going on in the context of the course
Model Identification and Data Analysis, have also allowed to
collect a large amount of feedback from students that have
expressed a wide consense for the available tools and for the used
methodologies.