Deep residual shrinkage networks with adaptively convex global parametric rectifier linear units for fault diagnosis

Abstract

In response to the challenge posed by traditional deep learning methods, which apply uniform nonlinear transformations to all vibration signals and thus struggle to address fault diagnosis under variable working conditions, a novel activation function called the convex global parametric rectifier linear unit (CGPReLU) is developed based on our prior research. Initially, an analysis of the numerical patterns governing the adaptive derivation process of GPReLU’s two slope parameters revealed the surprising observation that these convex parameter combinations invariably appear in pairs. This discovery serves as the primary motivation for the development of CGPReLU. Leveraging this convex regularity, we subsequently redesigned a lightweight convex sub-network for the adaptive derivation of the CGPReLU’s slope. Simultaneously, a deep residual shrinkage network with CGPReLU is constructed for fault diagnosis. Furthermore, we introduce an innovative evaluation metric designed to measure the collective influence of diagnostic accuracy and computational complexity after the process of model lightweight. Finally, it is experimentally demonstrated that the developed method can maintain a better diagnostic performance while greatly improving the diagnostic efficiency under variable operating conditions compared to our previous work.