Prediction of open-pit slope displacement and time-to-failure based on hybrid deep learning

Slope failure represents a major safety hazard in open-pit mining operations, where accurate prediction of slope displacement and remaining time-to-failure (TTF) is critical for effective early warning. This study proposes a Hybrid Deep Learning and Bayesian regression framework (HDL-BR) for continuous monitoring of open-pit mine rock slopes. HDL-BR comprises two independent modules. The first module employs a hybrid TCN–BiLSTM–Attention (TLA-Bi) model for displacement prediction. TCN extracts multi-scale deformation features, BiLSTM captures long-range temporal dependencies, and the attention mechanism weights critical time steps. The second module, Probabilistic Time-To-Failure (PTTF), combines a four-criterion onset-of-acceleration (OOA) detection algorithm with Bayesian Regression (BR) applied to the inverse velocity (INV) method, providing probabilistic TTF estimation with 95% CI. HDL-BR is validated using two case studies from the Changshanhao open-pit gold mine in Inner Mongolia. The S1 area represents a failure case with three-stage creep behavior. The S3 area represents a non-failure case with transient acceleration followed by re-stabilization. For displacement prediction, TLA-Bi outperformed six baseline models. For S1, HDL-BR successfully predicted failure time with quantified uncertainty intervals. For S3, no valid OOA was detected, and PTTF was not activated. The HDL-BR framework provides integrated slope monitoring with accurate displacement prediction and probabilistic failure time estimation, offering quantified uncertainty for risk-informed early warning decision-making in open-pit mines.