RPCNNet: A Deep Learning Approach to Sense Minor Stator Winding Interturn Fault Severity in Induction Motor Under Variable Load Condition

Abstract

In this article, a deep learning-based automated framework has been proposed to sense minor stator winding interturn (MSIT) fault in an induction motor (IM) without having load level information. For this purpose, an experimental setup has been configured in laboratory through which three phase current signal of a three-phase squirrel cage IM has been recorded at different MSIT fault, different loading level and external voltage unbalance condition. In this proposed framework, three phase stator current profiles firstly are converted to 1-D Park’s vector modulus (PVM) signal using Park’s transformation and extended Park’s vector approach (EPVA). Thereafter, PVM of three phase stator currents have been encoded into RGB image using recurrence plot (RP) technique and have been fed to a customized convolutional neural network module (RPCNNet) for detection of MSIT fault severity. The recognition performance of the RPCNNet is also compared with four benchmark convolutional neural network (CNN) models namely “VGG16,” “AlexNet,” “ResNet50,” and “DenseNet201,” respectively. From the result, it is observed that RPCNNet has shown comparable performance and offers significantly reduced computational cost compared to the benchmark CNN models. Proposed framework has also outperformed the existing approaches by a significant margin and shown promising result under both noiseless and noisy environment. The proposed framework is also capable of differentiating MSIT fault condition from unbalanced supply voltage condition. Hence, proposed framework is very much suitable for practical application.