DEEP SVDD BASED MODELING OF FLUORESCENCE FEATURE OF DRINKING WATERS FOR EVALUATION OF RECLAIMED WATER QUALITY

Abstract

 Various known and unknown chemicals potentially present in treated wastewaters have raised un urgent needs for a more comprehensive water quality assessment and monitoring tool to ensure the safety and reliability of water reuse systems. In this study, we developed a model of the fluorescence feature of drinking waters using machine learning based anomaly detection, and the similarity between the water quality or the model and those of effluents from advance wastewater treatment processes was evaluated quantitatively. Total 387 samples of actual drinking waters (i.e., tap water, spring water, bottled water) were collected from all over Japan, and their three dimensional fluorescence spectra data were learned by Deep SVDD, one of the machine learning based anomaly detection method. The abnormality (deviation from the model water quality) calculated for drinking water and reclaimed water samples demonstrated that actual drinking water quality itself was not unform, and varied by sampling dates or location. Although the abnormality of the RO permeate slightly exceed the range of variation observed for the actual drinking waters, the deviation was much smaller than the diluted MBR effluents. The inline LC-EEM analysis further revealed that the molecular weight distribution of the fluorescent compounds in the reclaimed waters were smaller than the actual drinking waters. This study first presented the effectiveness of machine learning based anomaly detection method for extracting useful information from the non-target analysis data.