Citation: Boushaba, A.; Cauet, S.;
Chamroo, A.; Etien, E.; Rambault, L.
Comparative Study between
Physics-Informed CNN and PCA in
Induction Motor Broken Bars MCSA
Detection. Sensors 2022, 22, 9494.
https://doi.org/10.3390/s22239494
Academic Editors: Kim Phuc Tran,
Athanasios Rakitzis and Khanh T. P.
Nguyen
Received: 21 October 2022
Accepted: 29 November 2022
Published: 5 December 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Comparative Study between Physics-Informed CNN and PCA
in Induction Motor Broken Bars MCSA Detection
Abderrahim Boushaba, Sebastien Cauet * , Afzal Chamroo, Erik Etien and Laurent Rambault
University of Poitiers, ISAE-ENSMA Poitiers, 2 rue Pierre Brousse, TSA41105, CEDEX 9, 86073 Poitiers, France
* Correspondence: sebastien.cauet@univ-poitiers.fr
Abstract:
In this article, two methods for broken bar detection in induction motors are considered
and tested using data collected from the LIAS laboratory at the University of Poitiers. The first
approach is Motor Current Signature Analysis (MCSA) with Convolutional Neural Networks (CNN),
in which measurements have to be processed in the frequency domain before training the CNN to
ensure that the resulting model is physically informed. A double input CNN has been introduced to
perform a 100% detection regardless of the speed and load torque value. A second approach is the
Principal Components Analysis (PCA), in which the processing is undertaken in the time domain.
The PCA is applied on the induction motor currents to eventually calculate the
Q
statistic that serves
as a threshold for detecting anomalies/faults. Even if obtained results show that both approaches
work very well, there are major differences that need to be pointed out, and this is the aim of the
current paper.
Keywords: MCSA; PCA; deep learning; physically informed; PINNS; broken bars; fault detection
1. Introduction
Induction Motors (IM) are commonly used nowadays. Aside from a wide use in
modern industries, the application of IMs can be found in various sectors (military, medical,
nuclear, etc.) [
1
,
2
]. These sectors require IMs to mostly operate in harsh conditions following
rough cycles and long working hours, and these uneasy situations increase the probability
of breakdowns and malfunctions. At the same time, precision and fidelity are essential for
these sectors in order to guarantee optimal performance. As a matter of fact, IMs require
regular monitoring and maintenance [
3
,
4
], a process also known as condition monitoring.
In some cases, intrusive analysis is required. This type of analysis is often a waste of
operating time for industries. In other cases, the diagnosis of faults depends only on the
analysis of some recorded measurements, such as currents and vibrations of the IM [
5
], but
this approach requires a high level of experience, which a machine operator usually does
not have. The faults that occur in IMs are classified into mechanical, rotor based and stator
based. The present paper deals with the rotor faults and, more precisely, the broken rotor
bar (BRB) faults.
Different methods have been developed to detect broken rotor bar (BRB) faults without
intrusion. These methods can be divided into two categories: the first category is based
on statistics, while the second one is based on physical information. Statistical methods
use time domain measures to characterise the healthy state of the motor with a set of
statistical indexers or coefficients. This helps in identifying any abnormality in future
measurements that indicate a fault. Some approaches tend to calculate a diverse set of
coefficients from time domain measurements directly. The authors of [
6
] identify three
statistical coefficients: Measure of Central Tendency (MCT), Measure of Variability (MV),
and Measure of Dispersion (MD). In cases with large data sets, the Principal Components
Analysis (PCA) method is used to reduce the dimensions of the data and focus only on
the important ones by analysing the
Q
statistic (squared prediction error (SPE) in other
Sensors 2022, 22, 9494. https://doi.org/10.3390/s22239494 https://www.mdpi.com/journal/sensors
