Lightweight Thermal Fault Diagnosis in Induction Motors: A Comparative Study of CNNs and Depthwise-Separable CNNs

This paper investigates lightweight thermal fault detection and diagnosis (FDD) for induction motors, utilising Infrared Thermography (IRT) and compact deep neural networks. We use consistent pipeline to compare a standard convolutional neural network (CNN) with a depthwise-separable convolutional neural network (DS-CNN) across depths d ∈{2,3,4}, employing purely geometric augmentation to maintain radiometric integrity, and evaluate using two protocols: a fixed 50 / 50 train-test split that simulates limited training data and K-fold cross-validation. Experiments are performed on a publicly available motor thermography dataset featuring eleven different operating and fault conditions. At a consistent depth of d=4, the CNN reaches an accuracy of 96%, while the DS-CNN attains 95%, using approximately 25% of the parameters - similar results are shown in the cross-evaluation. Analysis of the confusion matrix reveals clear class separation overall, with some residual confusion mainly in low-severity inter-turn short-circuit fault (ITSC) cases. The results show that using DS-CNN preserves essential thermography cues, while significantly reducing the size of the model.