Bearing fault diagnosis for electric machines using motor current signals and state classification

Abstract

Procedures for diagnosing bearing failures of induction motors based on motor current signals in published methods commonly denoise signals acquired from current sensors, then extract the typical characteristics from the denoised signals, and use a classifier to discriminate the state of bearing. However, the current signals in practice could be affected by surrounding noises, creating extraordinary peaks in the signals which may lead to an inaccurate diagnostic result. Thus, the traditional methods may not be very effective in diagnosing failures for induction motors using motor current signals in real-time. To mitigate these issues, this work introduces a new technique, which consists of characterizing the bearing health as a state vector composed of signal features, evaluating the real bearing status from the characteristic-space using a Kalman filter and a k-NN classifier. This technique still achieves quite high precision even in noisy condition. Experimental results with noise-adding signals demonstrate that the proposed technique has a mean proper identification proportion of 92.06% and a mean false proportion of 7.94%, while the conventional ones turn out a maximum mean true identification proportion of 53.12 % and a minimum mean mistake proportion of 46.88%.