Signal Preprocessing and Deep Learning Applied to Inter-Turn Fault Classification in Three-Phase Induction Motors

Abstract

Three-phase induction motors (TIMs) are the primary source of mechanical energy in industrial applications. With advances in power electronics, TIMs can be easily controlled for various applications, including pumps, treadmills, exhausters, compressors, and fans. However, their widespread use requires efficient and rapid maintenance strategies. Performing a timely diagnosis of faults is crucial, as unexpected downtime can cause significant economic losses for industries. One of the most common faults in TIMs is short circuits in the stator winding, which can cause overload, excessive heat generation, increased vibrations, and insulation degradation. Modern signal processing techniques, combined with advances in neural networks, offer new approaches to fault detection and classification with high speed and accuracy. The integration of electro-acoustic (EA) sensors further enhances fault detection capabilities, making them a powerful tool for improving motor maintenance. Therefore, this paper presents a method to identify and classify the severity of short circuit faults in TIMs, applying discrete wavelet transform (DWT) as a preprocessing of signal, and then the extracted features is compacted as a tensor, which is used to train and validate a deep neural network (DNN).