TransDACLE: Transformer-based Data Cleansing Framework for Digital Twin-Enabled Wastewater Treatment Plants

The digital transformation of wastewater treatment plants (WWTPs) through digital twin technology promises significant operational improvements, yet faces a fundamental challenge: the pervasive issue of data quality degradation. Common issues such as equipment malfunctions, sensor fouling, blockages and drifts frequently contaminate measurement data, compromising the reliability of digital solutions. Current anomaly detection methods struggle with complex temporal dependencies and strong non-linearities inherent in wastewater treatment data, leading to high false alarm rates and missed detections. To address this challenge, we propose TransDACLE (Transformer-based Data Anomaly Cleansing and Learning Enhancement), which consists of three key innovations: stacked sparse autoencoders (SSAE) for deep feature extraction, a modified Transformer architecture with dual-phase training for enhanced process variation robustness, and extreme value theory (EVT) for dynamic threshold optimization. Testing on the BSM2 benchmark dataset demonstrates TransDACLE's superior performance, achieving 96.0% accuracy with significantly reduced false alarm rates (3.3%) and missed alarm rates (9.4%) compared to eleven state-of-the-art methods. The framework's ability to ensure reliable input data quality provides a foundational stepping stone for the widespread adoption of digital twins in WWTPs, catalyzing the industry's digital transformation and advancing sustainable wastewater management.