Abstract
<jats:p>Nonlinear cyclic degradation processes widely occur in monitoring scenarios such as battery health assessment, rotating machinery diagnostics, and industrial equipment reliability management. These processes generate multivariate time-series signals characterized by evolving cyclic structures, phase misalignment, and gradual regime drift, which significantly complicates remaining useful life (RUL) prediction for conventional sequence models that assume stationary or weakly varying temporal patterns. To address these challenges, this paper proposes a cycle-aware representation learning framework that explicitly models cyclic degradation dynamics. The proposed approach first converts variable-length degradation cycles into phase-normalized windows to mitigate temporal misalignment, and then learns cycle-consistent latent representations through phase-aware encoding and robustness-oriented regularization. This design enables the model to capture both intra-cycle morphological patterns and inter-cycle degradation trends while maintaining stability under irregular cycle realizations. Extensive experiments conducted on public cyclic degradation datasets demonstrate that the proposed framework consistently improves prediction accuracy, robustness, and generalization performance compared with several cycle-agnostic baselines. The experimental results confirm that cycle-aware representation learning provides a more reliable and stable solution for RUL prediction under nonlinear cyclic degradation conditions.</jats:p>