Diagnosis of inverter-fed induction motors in short time windows using physics-assisted deep learning framework

Abstract

This article presents a framework for accurate fault diagnostics in inverter-fed induction machinery operating under variable speed and load conditions within very short time windows. Condition indicators based on fault characteristic frequencies observed over the extended Park's vector modulus are fused with deep features extracted using stacked autoencoders to generate a multidimensional feature space for fault classification using support vector machine. The proposed approach is demonstrated in a laboratory setup to detect the most commonly occurring faults, namely, the stator turns fault, broken rotor bars fault and bearing fault with an accuracy > 98% within a short time window of 2–3 seconds.