Multisensing for fault identification in electrical machines using quadratic distributions

Abstract

Monitoring the performance of Induction Motors (IM) is crucial to prevent unexpected downtime in the industry. Since manufacturing processes are typically dynamic, the diagnostic strategy should consider non-stationary situations, which require techniques that provide time-frequency analyses. However, the most well-known tools for this purpose are the Short-Time Fourier Transform and the Wavelet Transform, which are not sufficient to address many diagnostic problems. In contrast, the Wigner-Ville Distribution (WVD) stands out for its resolution in both time and frequency. However, interferences known as cross-terms are intrinsically generated during calculation and can hinder its application. For this reason, using the WVD is often avoided in studies despite its characteristics for handling transients. Modified versions of the WVD, also known as quadratic distributions or Cohen class Bilinear Distributions (CCBD), were created to suppress cross-terms, but they are little explored within the context of electric machine diagnostic research. To address this problem, the present work investigates and compares various CCBDs applied to processing vibration, audio, and current signals for bearing, stator and rotor fault identification in IM. To ensure the robustness of the methodology, the collected data set comprises circumstances that emulate some machine operating conditions in the industrial environment, including 1) line and inverter power supply, 2) voltage unbalances in the grid, and 3) load oscillations. The application of machine learning techniques demonstrates the effectiveness of the proposed approach with performance equivalent to or even superior to those found in the literature, even under adverse motor operating conditions. For stator fault, the study found that the applied methodology is effective under ideal conditions using the fifth harmonic of the line current. Broken rotor bar detection proved robust across various conditions, with Convolutional Neural Networks and Zhao-Atlas-Marks achieving over 99% accuracy. Several machine learning methods often exceeded 99% for bearing failures, with vibration performing best, while current and audio showed satisfactory results. Thus, the study presents an alternative methodology using CCBDs for monitoring and diagnosing faults in three-phase induction motors.