A comparative case study between shallow and deep neural networks in induction motor's fault diagnosis

Abstract

Here, performance of auto-encoder deep neural networks in detection and isolation of induction motor states (healthy, bearing outer race fault, stator winding short circuit and broken rotor bar) in the presence of unbalanced power supply and electro-pump dry running disturbances is investigated. Easily available three-phase electrical current signals are denoised using independent component analysis, and then the frequency-domain signal is used to train a neural network. A comparison is made between shallow and deep neural networks and also between the conventional structure of deep methods and the encoder–decoder structure in terms of training and test accuracy and robustness. In fact, the depth is increased and the effectiveness is investigated. At the end, it is shown that an encoder–decoder structure leads to the best result in terms of accuracy and robustness. The algorithms are examined experimentally, and the results show that the auto-encoder deep neural network can detect the aforementioned faults with a high reliability in the presence of disturbances.