Keeping things flowing: machine learning for multi-sensor quality in aeration systems

Wastewater treatment plants (WWTPs) rely on dense sensor networks to monitor aeration processes critical for operational efficiency and stability. This study investigates AI-based sensor quality monitoring using a two-year, high-resolution dataset from a full-scale WWTP. We propose a multi-label framework based on sensor-specific Light Gradient Boosting Machine (LightGBM) models, incorporating temporal feature engineering (lagged observations and rolling statistics), cost-sensitive learning and threshold optimisation. To ensure realistic evaluation under temporal dependencies, a nested 5-fold TimeSeriesSplit is used and results are compared against Random Forest and Logistic Regression baselines. The proposed approach achieves a macro F1-score of 0.664 and a micro F1-score of 0.819 for sensors with sufficient failure support (≥1%), outperforming all baselines. Results show that temporal features significantly improve predictive performance, particularly for physical process variables, while performance degrades for sensors with sparse or temporally clustered failure events. These findings highlight the importance of time-aware validation and temporal feature design in AI-driven predictive maintenance systems for industrial processes.