Advance deep convolution neural network for multiple fault diagnosis of induction motor

Abstract

This work proposes a new hybrid combination of data fusion, time-frequency (TF), and deep convolution neural network techniques in conjunction with Transfer Learning (TL). Deep Learning (DL) models have many layers that require high training and testing period. The work proposes a solution for the simultaneous early detection of multiple faults and reduction in the fault prediction period in deep neural networks. The acquired vibration signal from induction motor under variable loading conditions (no load, half load, full load) and different health conditions such as healthy, bearing inner cage fault, bearing outer cage fault, one broken rotor bar, and three broken rotor bars, is divided using the Average Frequency Band Decomposition (AFBD) approach. The TF approach converts the one-dimensional signal into a two-dimensional data set. Many classification scores were produced for various training/testing sets, to evaluate the proposed fault recognition system. The shallow learning techniques such as Support Vector Machine, Decision Tree, Nearest Neighbor, and Discriminant analysis acquired an accuracy of 54.50%, 74.00%, 74.00%, and 76.50%. In comparison, the DL technique achieved an accuracy of 97.40% because of no manual feature extraction and no loss of important features. The advance deep neural network method outperforms the shallow learning techniques and conventional deep neural networks by achieving an accuracy of 99.50% and a reduced time of approximately 8 minutes since there is no training and testing from scratch using TL This proposed method is compared with similar conventional DL models, which suggests the methodology has tremendous potential to provide a solution for unplanned failures.