IRCNN+: An enhanced Iterative Residual Convolutional Neural Network for non-stationary signal decomposition

Abstract

Time–frequency analysis is a crucial yet challenging task in numerous applications. The mainstream approach involves first decomposing non-stationary signals into quasi-stationary components to enhance the time–frequency feature clarity during analysis. Inspired by deep learning, we proposed the Iterative Residual Convolutional Neural Network (IRCNN) to address non-stationary signal decomposition. Deep learning enables IRCNN not only to achieve more stable decomposition results than existing methods but also to handle batch processing of large-scale signals with low computational costs. However, certain structural components of IRCNN remain relatively rudimentary, leading to limitations in local feature characterization, smoothness constraint, and adaptability requirement for non-stationary signals. This study aims to further refine IRCNN by integrating flexible techniques from deep learning and optimization—such as multi-scale convolutional layers, attention mechanism, and total variation denoising method—to enhance the method and overcome its inherent limitations.