Interpretable CausalFormer Prediction Model for Abnormal Vibration Recognition in Sluice Gates

Abnormal vibration in sluice gates can precede seal wear, guide-slot impact, cavitation-related excitation, and actuator faults, yet field records are difficult to interpret because flow head, opening degree, sediment load, and operation commands change together across a 30-day monitoring period. This study develops an interpretable CausalFormer prediction model that combines regime-aware temporal attention, causal masking, residual-based abnormal recognition, and operator-readable attribution. A 30-day monitoring set containing 1,296,000 vibration samples from four accelerometers was organized with opening, upstream level, downstream level, flow estimate, motor current, and temperature. The proposed model reached a vibration forecasting MAE of 0.031 g and an RMSE of 0.047 g, reducing RMSE by 27.7 percent compared with a standard Transformer and by 35.6 percent compared with LSTM under the same monitoring split. For abnormal recognition, it achieved 96.1 percent F1-score, 95.4 percent precision, and 96.9 percent recall under mixed gate-opening conditions. Causal attention pruning reduced explanation instability by 41.3 percent and kept 88.6 percent of high-risk decisions aligned with hydraulic or electromechanical trigger variables. The results indicate that prediction residuals become more reliable when temporal dependence is separated from operating-condition causality. The model offers a practical route for sluice-gate condition monitoring where operators need both early abnormal-warning accuracy and traceable evidence for maintenance decisions.