Feasibility of Time Domain Analysis for Classification of Non-Stationary Signals

Extensive research is carried out in classifying non stationary signals as normal or abnormal. Considerable research is carried out in applying various transforms namely Discrete Wavelet Transform, Discrete framelet Transform, Stockwell transform, Hilbert Transform for converting the signals from time domain to spectral diomain. On the spectral domain features, statistical parameters are obtained and classification is performed based on these parameters by using classifiers. In recent years, deep learning networks are sued for classification. Inherent to its architecture, computational complexity of these networks is more than the other classifiers and hence these techniques cannot be used for online classification. Hence it is necessary to propose a technique that involves less computational complexity thereby making it suitable for real time application. In this research work, feasibility of less complex, time domain based analysis is studied. Two different non-stationary signals namely burst signals and sawtooth signals are considered. From the proposed methodology, the maximum sensitivity, specificity and accuracy are 96.12%, 83.56% and 81.12% respectively.