A Hierarchical GRU-Based Predictive Maintenance Framework for SCADA-Monitored Water Pump Stations

This study investigates predictive maintenance for SCADA-controlled pump stations using multi-sensor data processing and hybrid machine-learning models. The conventional maintenance approaches adopted in practice remain reactive, with little provision for actual early warnings under real-world conditions such as noisy data, class imbalance, or varying sensor dynamics. A data-driven solution is proposed to predict pump tripping events using operational SCADA system data for early warning with useful lead times. The dataset, obtained from a water-utility SCADA system, contained missing values, heavy-tailed sensor distributions, and substantial class imbalance. The preprocessing strategy used time-aware imputation, winsorisation, and a sliding-window configuration informed by the characteristics of the SCADA data. Benchmark machine-learning models achieved PR-AUC values of approximately 0.55 or lower for trip-escalation prediction, highlighting the difficulty of predicting rare trip events directly from SCADA data. The proposed hierarchical GRU-based framework achieved PR-AUC values exceeding 0.80, demonstrating a substantial improvement in predictive performance while maintaining high precision and low false-alarm rates. In addition, a Remaining Useful Life (RUL) component was included to extend the system to support near-term risk forecasting. Even so, long-term forecasts remained uncertain, indicating that further model development is required.