System Design for the Detection of Induction Motor Faults using Multi-Domain Feature Fusion

Abstract

Accurate and early detection of the bearing faults would make it possible to solve the problem at a lower cost by replacing relevant parts. In this article, a machine learning technique has been proposed for the detection of inner-race and outer-race bearing faults of the motor. A test bench is developed for experimentation in which easy insertion and removal of the faulty and healthy bearing of an induction motor is possible. A vibration sensor was employed to diagnose the motor faults. A dataset of 3-dimensional vibration signals is acquired from the motor representing inner-race and outer-race bearing faults. Vibrational signal data is pre-processed using empirical mode decomposition (EMD) to decompose the raw signals into its integral components called Intrinsic Mode Functions (IMFs) and to find the region of interest. A new scheme for the selection of discriminated features in bearing fault detection application is presented here in which firstly, Multi-domain features were extracted from pre-processed, and were serially fused. Then, Principle Component Analysis (PCA) technique was employed for the reduction of feature vector dimension which resulted a set of only ten features. The proposed features have positive repercussions on the performance of various existing classifiers used in motor bearing fault detection applications. However, Support Vector Machine with a linear kernel (SVM-L) is proposed to recognize the healthy, inner-race fault and outer-race fault states of the motor bearing due to its linear basis. The results of extensive experimentation appreciate the performance of the proposed technique.