Detection of broken rotor bar fault in an induction motor using convolution neural network

Abstract

Induction motors are prime component in the industries. Hence, condition monitoring and fault diagnosis of induction motor are important to avoid shutdowns and unplanned maintenance. A technique based on time-domain grayscale current signal imaging (TDGCI) and convolutional neural network (CNN) is proposed for intelligent fault detection of broken rotor bar in an induction motor. The standard current signal dataset made available by the Aline Elly Treml Western Parana State University is used for analysis. This dataset is acquired by simulating the healthy and broken rotor bar (BRB) fault conditions with the four increasing severity levels (1BRB, 2BRB, 3BRB, and 4BRB) at eight loading conditions varying from no load to full load. Conventional machine learning techniques have the limitations of feature selection, while the proposed technique can automatically extract the features from the given input image. The TDGCIs obtained from the time-domain current signal is used as input to exploit the enormous capability of CNN to carry out the image classification, thereby classifying faults features embedded in the images. The efforts are presented to design CNN parameters to achieve the fault classification accuracy of 99.58% for all cases with optimized computational time. The significant reduction in the computational time for fault classification compared to the peer published work is an important contribution.

Machine learning for fault analysis in rotating machinery: A comprehensive review

“…The acquired data sources are subjected to manual or automatic feature extraction procedures for IFDP procedures. The most feature extraction techniques regarding FD, FDE, FI, or FP procedures of rotating machine components are statistical feature extraction (STFE) [ 10 , 27 , 29 , 101 ], Fast Fourier Transform (FFT) [ 8 , 33 , 34 , 109 ], Wavelet Transform (WWT) [ 20 , 73 , 78 , 134 ], and EMD [ 21 , 41 , 52 ]. There are also some methodologies derived from those techniques including Discrete Fourier Transform (DFT) [ 95 ], Short Time Fourier Transform (STFT) [ 36 ], Wavelet Packet Transform (WPT) [ 37 , 83 ], Continuous Wavelet Transform (CWT) [ 36 , 38 ], Discrete Wavelet Transform (DWT) [ 76 , 122 ], and Ensemble Empirical Mode Decomposition (EEMD) [ 112 , 121 ].…”

“…For this purpose, they deemed vibration [ 11 , 14 , 15 , 21 , 22 ], acoustic [ 11 , 23 , 24 ], thermal [ 13 ], current [ 6 , 7 , 9 , 25 , 26 ], pressure [ 27 ], and other characteristic data [ [27] , [28] , [29] ] as the main source for IFDP of rotating machines. Afterwards, the distinctive features are extracted by employing feature extraction methods such as statistical feature extraction [ [30] , [31] , [32] ], Fourier Transform [ [33] , [34] , [35] ], Wavelet Transform [ [36] , [37] , [38] ], Empirical Mode Decomposition [ 28 , 39 , 40 ] or other techniques [ 6 , 7 , 41 , 42 ]. The features may also be extracted automatically by employing deep learning approaches including convolutional neural networks, autoencoders, long-short term machines, etc.…”