Signal Processing and Machine Learning Techniques Based Hybrid Approaches for Decent Fault Classification of Induction Motor

Abstract

Abstract Learning of better feature representation instinctively by Convolutional Neural Networks (CNN) has inspired to address the unsolved issues in the stator current based multi-class fault diagnosis of induction motor drives. Current envelope of stator current acquired using the Hilbert transform is proven to be the effective pre-processing method to handle the complex data pattern of motor current and reveal the masked defect information. The self-synthesized quality features through deep convolution layers outperforms and reaches an unmatched accuracy level compared to the counterpart-feature engineering scheme. The method of feature engineering is also developed with DHT-DWT based feature extraction process with novel idea of suitable mother wavelet selection scheme. The most notable achievement of this research work is to address the unique advantages of hybridization of signal processing technique and CNN model where the enrichment in feature quality is acquired due to unveiling the buried fault information close to dominating supply frequency. The proposed method is reliable in analyzing multi-class motor fault detection having a good generalization approach. The compact design of hybrid CNN-envelope approach dealt with very low resolution stator current sampled at 1.28 kHz, has reduced the computation intricacies to a great extent and projected it as the right aspirant for real-time applications.