Unsupervised learning approach for pH anomaly detection in wastewater treatment plants

Abstract

Sustainability has been a concern for society over the past few decades, preserving natural resources being one of the main themes. Among the various natural resources, water was one of them. The treatment of residual waters for future reuse and release to the environment is a fundamental task performed by Wastewater Treatment Plants (WWTP). Hence, to guarantee the quality of the treated effluent in a WWTP, continuous control and monitoring of abnormal events in the substances present in this water resource are necessary. One of the most critical substances is the pH that represents the measurement of the hydrogen ion activity. Therefore, this work presents an approach with a conception, tune and evaluation of several candidate models, based on two Machine Learning algorithms, namely Isolation Forests (iF) and One-Class Support Vector Machines (OCSVM), to detect anomalies in the pH on the effluent of a multi-municipal WWTP. The OCSVM-based model presents better performance than iF-based with an approximate 0.884 of Area Under The Curve - Receiver Operating Characteristics (AUC-ROC).