Abstract:
The study represents the diagnosis of the low
severity inter turn short circuit (ITSC) fault in the permanent
magnet synchronous motors (PMSMs). A pre-recorded in-house
three-phase current data is used to distinguish between the
healthy, 2% ITSC, and 5% ITSC fault detection in the stator
winding at a fixed speed and load. This analysis is targeted at
condition monitoring and predictive maintenance of the system.
The supervised machine learning techniques take recorded time
series of three-phase current data as a key ingredient for
classification. Specifically, four supervised machine learning
(ML) algorithms decision tree (DT), random forest (RF),
extreme gradient boosting (XGB), and support vector machine
(SVM) are used to evaluate the performance of the classification
for different fault severity level with the healthy condition. For
this classification, statistical features like crest, impulse,
kurtosis, etc. factors with a total of 9 other statistical features
are determined to serve the purpose. Wrapper-based feature
selection (WFS) technique is employed for finding the optimal
subset of features to get the maximum accuracy while reducing
computational complexity using the mentioned supervised
machine learning algorithms. It is vital to note that the
combination of root mean square (RMS) factor, standard
deviation factor, and crest factor is the optimal performing
feature for both classifications between healthy to 2% ITSC and
healthy to 5% ITSC fault with the accuracy percentage of
around 85% and 98.33% in level 3 of SVM algorithm.
