A hybrid dual-frequency-informed spider net for RUL prognosis with adaptive IDP detection and outlier correction

Abstract

The present study proposes a novel framework to estimate the Remaining Useful Life (RUL) of bearings operating under variable operating conditions, addressing two critical challenges: early detection of the Initial Degradation Point (IDP) in bearings and correction of outlier values. A unique Spider cell prediction unit with dual-frequency correction is proposed. Firstly, a generalized adaptive method is introduced for early IDP detection, leveraging the slope and intercept, along with coupled t-tests to formulate a "sum of slopes" index for detecting the IDP. Secondly, a degradation feature extraction method is introduced, which utilizes synchronous pseudo speed in combination with sliding window averaging. Outlier correction for degradation feature indicators is achieved using constructed boundary conditions. Thirdly, a variational mode decomposition layer is proposed to decompose the input sample into different mode function components. Finally, a novel RUL prediction correction module, where two types of frequency domain feature extractors with trainable parameters are designed to adjust the prediction results of the Spider net by capturing both global trend changes and local details.