IMTC 2006 - Instrumentation and Measurement
Technology Conference
Sorrento, Italy 24-27 April 2006
Incipient Fault Diagnosis in Electrical Drives by Tuned Neural Networks
A. Azzini
1
, L. Cristaldi
2
, M. Lazzaroni
1
, A. Monti
3
, F. Ponci
3
, A.G.B. Tettamanzi
1
1
Dipartimento di Tecnologie dell’Informazione,
via Bramante, 65 – 26013 Crema, Italy
Phone: +39 02 50330058, Fax: +39 02 50330010, E–mail: azzini,lazzaroni,tettamanzi@dti.unimi.it.
2
Dipartimento di Elettrotecnica,
Politecnico di Milano,
Piazza Leonardo Da Vinci, 32 – 20133 Milano, Italy
Phone: +39 02 23993702, Fax: +39 02 23993703, E–mail: loredana.cristaldi@polimi.it.
3
Department of Electrical Engineering, University of South Carolina,
Swearingen Center – Columbia, SC 292082-0133 – USA
Phone: +1 8037772722, Fax: +1 8037778045, E–mail: monti,ponci@engr.sc.edu.
Abstract – In order to identify any decrease in efficiency and any loss
in industrial application a suitable monitoring system for processes
is often required. With the proposed approach useful diagnostic in-
dications can be obtained by a low-cost extension of the monitoring
activity. In this way, the reliability of the obtained indications can be
significantly increased considering the combination of advanced time-
frequency transform, or time - scale, such as wavelets, and a new evo-
lutionary optimisation approach based on Artificial Neural Networks
(ANNs). This paper describes an approach to the joint optimization of
neural network structure and weights which can take advantage of the
backpropagation algorithm as a specialized decoder. The presented
approach has been successfully applied to a real-world machine fault
diagnosis problem.
Keywords – Evolutionary Algorithms, Pattern Recognition, Diagnos-
tic, Testing, Neural Networks.
I. INTRODUCTION
Industrial applications normally require suitable monitoring
systems of the main relevant processes in order to identify any
decrement in the efficiency involving in economical losses. A
monitoring system consists of a set of devices, procedures and
diagnostic tools that follows every single step of a process.
Early detection of operating conditions of the apparatus that
deviate from the optimal ones may avoid subsequent failures,
or even faults.
A diagnostic tool for an AC drive is here considered. In
particular, it is reasonable to assume that the only accessible
points of the system are the AC input terminals having, for ex-
ample, in mind the recent trend to more and more integrated
systems where the drive can be considered as a black-box or in
case of non accessibility of the inner part of the electrical sys-
tem. In particual, this second opportunity can be considered
very useful when the electrical system is located in restricted
or non accessible area. The opportunity of combining diag-
nostic and monitoring operations on an AC motor drive with-
out using dedicated sensors cannot achieve a diagnosis as reli-
able as that provided by totally customized systems as shown
in literature. However, in many applications the more impor-
tant diagnostic result is the possibility to recognize fault and
non fault case without any further information. Moreover, the
possibility of recognizing incipient faults in the beginning of
the fault situation is the more importan task in many appli-
cations. Starting from this point of view, useful diagnostic
indications are obtained by a low-cost extension of the mon-
itoring activity, and the reliability of the obtained indications
are significantly increased considering the combination of ad-
vanced time-frequency transform, or better time - scale, such as
wavelets, and a new evolutionary optimisation approach based
on Artificial Neural Networks (ANNs). The output of the ANN
is, in the present work, an Index, called ANNI that can be con-
sidered a risk coefficient giving the user the likelihood of be-
ing in fault conditions. This paper describes an approach to
the joint optimization of neural network structure and weights
which can take advantage of the backpropagation algorithm as
a specialized decoder. The approach is applied to the presented
real world machine fault diagnosis problem. However, the pro-
posed methodology here used for a particular application can
be used when classification tasks is required. The results are
successfully compared with those obtained with a traditional
neural network implementation [1], [2], [3], [4], [5].
II. EVOLVING ARTIFICIAL NEURAL NETWORKS
The evolution of an Artificial Neural Networks using ge-
netic algorithms has shown great promise in complex re-
inforcement learning tasks. Neuro-Genetic systems search
through the space of behaviors for a network that performs
well at a given task. These particular systems are biologically-
inspired computational models that use evolutionary algo-
Universita degli Studi di Milano, `
1284 0-7803-9360-0/06/$20.00 ©2006 IEEE