Rolling bearings are critical in rotating machinery. Their health directly affects operational safety and efficiency. Accurately predicting their remaining useful life is essential for predictive maintenance and reducing unplanned downtime. Similarity-based methods are attractive due to their intuitive nature and avoidance of complex physical models. However, the scarcity of full-lifecycle data limits reference trajectory completeness, and global metrics struggle to capture non-stationary local fluctuations during degradation. To address these challenges, this article proposes a dynamic similarity metric method integrating probabilistic data augmentation and local fluctuation perception. Root mean square features are first extracted from vibration signals, and a two-stage degradation model constructs health indicators. Then, key degradation parameters are then modeled using Gaussian, Gamma, and Weibull mixture distributions to generate an enhanced trajectory library that preserves real statistical characteristics. An adaptive similarity metric, based on sliding window local fluctuation analysis, dynamically quantifies fluctuation intensity and incorporates time decay weights to align with current degradation dynamics. Extensive experiments on the XJTU-SY and self-collected datasets show that the proposed method significantly improves prediction accuracy under small-sample conditions, outperforming several existing approaches, with ablation studies validating the contribution of each module.