Sensor drift detection framework for water systems with probabilistic machine learning

• Main Author: Hoang, Thu Minh
• Other Authors: Law Wing-Keung, Adrian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Environmental engineering::Water treatment
• Online Access: https://hdl.handle.net/10356/167026
• Institution: Nanyang Technological University

Smart sensors and meters have facilitated the control and optimization of water and wastewater treatment processes to achieve an acceptable effluent with cost efficiency. However, sensor drift is a challenge for plant operators as it can induce significant uncertainty in process control and thus reduce productivity. Current mitigation methods are tedious and time-consuming and require external intervention in the treatment process. Anomaly detection via advanced Machine Learning (ML) models available nowadays can be a solution to this sensor drift problem because the models are fast, accurate, and able to give long-term prediction. The focus of this project is to develop a real-time framework for drift detection using probabilistic ML models with anomaly detection. Experiments were conducted to obtain pH and DO data for model development. With the available measurements, four anomaly detection models (i.e., Moving LOF, NGBoost, XGBoost Distribution and Prophet) were examined in this project, and the Prophet model showed the best performance and was not affected by data limitations. Subsequently, the Prophet model was further used to construct the drift determination procedure that could send warnings before sensor drift developed significantly.