Predictive maintenance for induction motors: a novel synergy of deep learning and machine learning techniques

Abstract

Condition monitoring of induction motors is vital for preventing unexpected downtimes and minimizing the maintenance costs in industrial settings. A predictive maintenance model for early detection of faults is proposed. The motor current, flux, vibration, thermal and acoustic emission signatures are commonly used for fault detection as these signals reveal fault specific frequency components. Signal processing techniques like wavelet transform and Hilbert transform are used to identify faults at incipient stages. Machine learning and deep learning models are used nowadays to extract features and classify the faults accurately. The current and flux signals from healthy and faulty motors for inter turn faults were analyzed in this work and features were extracted through various signal processing methods. These features were then used to train models, including deep learning architectures, to classify motor faults. While the classical machine learning models provided a reasonably accurate fault classification, the convolutional neural network provided a very good classification accuracy. The findings show that deep learning models excel in detecting faults, especially under noisy and varying operational conditions, outperforming the traditional methods. These models offer a scalable, real-time solution for improving the reliability and efficiency of induction motor and facilitate more reliable assessments and contribute towards energy efficiency and extended lifetime of equipment.