A bearing remaining useful life prediction method integrating cosine similarity change point detection

Abstract Rolling bearings are key wear-prone components whose degradation exhibits strong nonlinearity, phased evolution, and heavy noise. These characteristics increase the difficulty of degradation modeling and remaining useful life (RUL) prediction. To address inaccurate change point detection and experience-dependent model selection in existing studies, this paper proposes an adaptive RUL prediction framework based on data-driven phase identification and segmented modeling. An adaptive particle filter–sliding window cosine similarity (PF–SCS) method is developed to detect degradation trend change points. It couples state recursion with shape similarity to accurately locate the accelerated deterioration point and avoids subjective threshold settings. After change point identification, a fusion fitting criterion (FFC) is constructed by standardizing and combining residual sum of squares and Akaike information criterion, enabling objective model selection for the post-change-point degradation stage. A Wiener-process-based state-space model is further established, and maximum likelihood estimation is used for online parameter updating to obtain a time-varying RUL posterior distribution. Experiments on the XJTU-SY bearing dataset and self-built accelerated life tests verify the proposed framework. The PF–SCS method precisely detects degradation mutations, and the Wiener-based model achieves accurate and reliable RUL prediction.