Optimized Time–Frequency Analysis for Induction Motor Fault Detection Using Hybrid Differential Evolution and Deep Learning Techniques

The dynamic operation of induction motors in industrial environments requires techniques for analyzing data in the time–frequency domain to monitor machine conditions. The Wigner–Ville distribution is widely recognized for its high time–frequency resolution but suffers from the problem of cross‐terms. The smooth pseudo Wigner–Ville (SPWVD) applies two‐dimensional filters to suppress these artifacts. However, this process may compromise the readability of legitimate components and usually involves using empirical methods to determine acceptable values for the filter parameters. Then, determining the combination of parameters that produce the most readable representations is time‐consuming and relies on a subjective qualitative assessment. To address this issue, this work proposes a novel approach to find filter parameter values that optimize the readability of the representation by applying differential evolution (DE) joint with convolutional neural networks (CNN) to create a workflow for detecting inter‐turn short circuits in induction motors. The experimental results demonstrate that the proposed approach achieves robust and accurate classification and localization of multiple fault severities, reaching an accuracy of more than 96%. The method was evaluated under diverse and challenging operating conditions, including a wide range of load torques and voltage unbalance levels. Furthermore, comparisons with alternative approaches using particle swarm optimization (PSO) and gated recurrent units (GRU)‐based models highlight the superior performance and reliability of the proposed DE + CNN‐based solution.