Robust Deep Residual Shrinkage Networks for Online Fault Classification

In this paper, a novel approach to improve signal classification in the presence of noise is presented. Using Stock-well transforms for feature extraction on time-series electromagnetic interference data and deep residual neural networks, containing thresholding functions (shrinkage functions) as non-linear transformation layers for classification. Thresholding functions are commonly used for signal de-noising. Setting thresholds for optimal functionality is often complex and requires expertise, this paper will investigate learned methods of threshold selection along with alternate thresholding functions. Using deep learning methods to select thresholds reduces the dependency on experts for the use of thresholding functions for de-noising and allows for adaptation to alternate noise environments. This paper proposed the novel application of two different threshold functions and introduces an architecture update for learning the threshold parameters for classification in the presence of noise. Several experiments are carried out to compare the performance of the systems with varying signal-to-noise ratio data sets taken from real-world operational high-voltage assets. Experimental results show that the proposed approaches using both Garrote and Firm thresholding achieved improved performance increases over utilizing soft thresholding within deep shrinkage networks in low signal-to-noise ratios.