Anomaly Detection in Wastewater Treatment Plants Using Unsupervised Machine Learning Algorithms

The timely detection of unusual wastewater influent contaminants or treatment defaults can significantly mitigate the release of sub-quality treated sewage effluent. Hence, protecting the marine life and the ecosystem from adverse impacts resulting from water pollution. This paper studies and recommends the best anomaly detection algorithm in the wastewater treatment industry. The study was conducted using several unsupervised Machine Learning (ML) algorithms, precisely the Isolation Forest Algorithm (IFA), Fast Minimum Covariance Determinant algorithm (FMCD), Local Outlier Factor algorithm (LOF), and One-Class Support Vector Machines (OCSVM) algorithm. Along with studying the effect of applying Principle Components Analysis on improving anomaly detection performance. In addition, it compared two strategies for applying anomaly detection. The anomaly detection results showed that the FMCD is the most robust algorithm for identifying anomalies in various dimensions, scoring a recall and F1Score using a univariate anomaly detection strategy exceeding 0.90 in four wastewater parameters. Similarly, it scored using a multivariate anomaly detection strategy recall and an F1Score of 0.76. The study found that multivariate anomaly detection and removal strategy is preferred for modeling tasks. However, a univariate anomaly detection strategy is recommended for plant operations to avoid the high dimensionality curse, which deteriorates the model detection performance. The recommended anomaly detection algorithm and strategy can be deployed in an integrated framework of remote sensing and ML to detect wastewater anomalies at wastewater treatment plants, leading to an effective wastewater treatment control that can contribute to preventing environmental pollution.