Self-Organizing Map and feature selection for of IM broken rotor bars faults detection and diagnosis

This paper presents a new robust and high performances fault diagnosis scheme for broken bar fault detection and severity evaluation. The aim is to ensure an accurate condition monitoring and reduced false or missed alarms rate for induction motor operating in critical applications. It investigates the combination of features selection methods with the Self-Organizing Maps (SOM) neural network in a fault detection and severity evaluation system. This approach, based on the current analysis, uses multiple features extraction techniques, where the zero crossing times (ZCT) signal and the envelope are extracted from the three-phase stator currents. Then, statistical and frequency domains features are calculated from these extracted signals. The ReliefF feature selection technique is used to select from the extracted features the most sensitive and relevant ones. Next, the SOM neural network is used as a decision-making system. The experimental investigations, conducted using a healthy machine and a machine with broken bars, show the effectiveness of the proposed fault detection technique in terms of the classification accuracy.

I. Introduction

Induction motors are widely used in modern industry and commonly in diverse critical applications such as nuclear reactor coolant pumps, petrochemical turbines, and military applications, where high reliability and efficiency are required. They present 80% of the motors in use [1]. In spite of their robustness and high efficiency, they can be the seat of a wide-ranging of failures that can lead to total motor failure, and will directly influence the safe and reliable operation of the whole plant. In this context, condition monitoring and fault diagnosis systems are developed to guarantee a continuous safe operating, and early fault detection and severity evaluation. This system should be reliable and accurate with no false or missed alarms [2].