A novel method for detecting bearing faults in induction motors using acoustic sensors and feature engineering

The study and development of fault diagnosis techniques for induction motors (IMs) have received significant attention, particularly regarding bearing failures. Although acoustic signals have been explored in the recent literature, advances in research about the quantity and model of sensors are still needed, as well as to develop approaches for extracting and selecting information that facilitates fault identification. In this context, this paper presents a new methodology to extract specific acoustic features from IMs and identify early-stage bearing distributed faults, which is unexplored in electric machine faults studies. It involves acoustic signals collected from multiple sensors, which are selected through similarity tests, and identifies the most crucial features to diagnose fault conditions using Recursive Feature Extraction and Random Forest as selectors. These selection methods highlight relevant features for the diagnostic process, improve fault identification, and reduce computational effort, resulting in a final accuracy rate of 98.86%. Identification of distributed and incipient bearing failures is achieved by using the selected features and comparing different classifiers and results in the literature. For this purpose, five sensors were used at a distance of 30 centimeters from the IM housing, and their signals were conditioned using electronic gain with operational amplifiers. The results demonstrate the system's applicability in low-noise environments such as test and inspection laboratories and IM out-of-operation testers. The methodology can be suitable for embedded systems, due to the reduction of attribute sets for fault identification.