Supervised machine learning based system for automatic fault-detection in water-quality sensors

Online water-quality sensors installed in Wastewater Treatment Plants (WWTPs) are prone to process disturbances that generate erroneous data. Faulty sensor data can disrupt automation systems and result in sub-optimal performance of WWTPs. This paper presents a machine-learning-based system for real-time detection and the subsequent correction of faulty sensor data installed in a full-scale municipal WWTP. The fault detection system is developed by training a k-nearest neighbour (kNN) classifier with labelled historical data. The trained kNN classifier is then deployed in the WWTP's web-based Supervisory Control And Data Acquisition (SCADA) system to assess the performance in real-time. A qualitative comparison between raw and corrected sensor data demonstrates the system's potential to detect sensor faults and provide stable and reliable surrogate values.