The data used in this study.

Slope reliability analysis often assumes isotropic or anisotropic random fields with horizontal orientation to characterize the spatial variability of soil parameters. However, this neglects the influence of rotated anisotropic spatial variability, leading to conservative and unrealistic failure probability estimates. To overcome this limitation, we propose a novel method based on the Hierarchical Recurrent Highway Network (HRHN) with attention mechanisms. This method is applied to a time-dependent reliability assessment of the Baishuihe landslide. By incorporating the spatial variability of geotechnical properties—especially the direction of maximum fluctuation—the study constructs both the most adverse and favorable extreme scenarios, enabling the exploration of failure probability evolution under reservoir drawdown and rainfall infiltration. Compared with traditional horizontally anisotropic random fields, the proposed model produces a range of failure probabilities rather than a single curve. This interval range—formed by maximum and minimum failure probabilities—better captures the uncertainty of the model and accounts for external factors such as rainfall and geological changes. Our approach offers a more comprehensive and realistic perspective for geotechnical risk assessment.边坡可靠性分析通常借助水平取向的各向同性或各向异性随机场，表征土性参数的空间变异性。然而此类方法忽略了旋转各向异性空间变异性的影响，会得到偏保守且脱离实际的破坏概率预估结果。为突破这一局限，本文提出一种基于带注意力机制的分层循环高速公路网络（Hierarchical Recurrent Highway Network, HRHN）的新型方法，并将其应用于白水河滑坡的时变可靠性评估。该研究纳入岩土体特性的空间变异性——尤其是最大波动方向，构建了最不利与最有利两类极端工况，以此探究水库水位消落与降雨入渗条件下破坏概率的演化规律。相较于传统的水平各向异性随机场方法，所提模型可生成一系列破坏概率而非单一曲线；该由最大与最小破坏概率构成的区间范围，能够更精准地捕捉模型不确定性，并兼顾降雨、地质变化等外部影响因素。本研究为岩土工程风险评估提供了更为全面且贴合实际的研究视角。