Efficient & lightweight classification of rotor bar faults in induction motors by convolutional and spiking neural networks

Fault detection and diagnostics for electrical machines have attracted significant research interests over the last two decades. At the same time, recent advances in artificial intelligence and smart technologies for diagnostics are utilised in tandem with machine learning for purposes of condition monitoring. However, most existing methods in this direction rely on resource-intensive large-scale networks and often require additional data preprocessing, such as the extraction of characteristic features in the frequency domain. This work proposes a lightweight framework incorporating an emerging spiking neural network (SNN) and an optimised lightweight convolutional neural network (CNN) for the classification of induction motors suffering from rotor electrical faults, notably broken rotor bar faults, using only raw time-series of sensor signals. The proposed methods are evaluated in several bar fault scenarios and loading levels. The data are acquired from 2D FEA simulation of two induction motors with the same power rating but different numbers of rotor bars, and the results are validated by experimental measurements. The proposed frameworks are used to achieve a reliable diagnostic outcome regardless of the machine load. This includes the no-load case and relies solely on non-intrusive monitoring with minimally processed stator current and stray flux signals. Moreover, this is the first application of a SNN architecture for rotor fault diagnostics, while the proposed lightweight CNN significantly reduces the network scale and resource consumption compared to existing methods.