Fluorescent fingerprints-based anomaly detection in drinking water quality and identification of contributing features by explainable AI

Abstract Water quality monitoring is essential for ensuring the safety and security of our drinking water supplies. However, conventional methods, which assume the use of pristine water sources, rely on a limited set of regulated indicators and may overlook unregulated contaminants, especially when such clean sources are not available. In this study, we developed a complementary water quality monitoring tool for water sources potentially influenced by anthropogenic activities, using anomaly detection model Deep Support Vector Data Description (Deep SVDD), based on excitation-emission matrix (EEM) data of drinking water samples collected across Japan. The model effectively identified deviations from baseline water quality (i.e., tap water and spring water for potable use) in a variety of non-drinking water samples, including river water, treated wastewater, chemically spiked water, and diluted industrial discharge, demonstrating high sensitivity to low-level contamination. Although a similar model was also developed using high-resolution mass spectrometry (HRMS) data, it showed limited ability to characterize the sample origins. The explainable AI was further applied to the EEM-based model, revealing that fluorescence features associated with the transphilic neutral organic matter contributed significantly to anomalies. These results demonstrate that fluorescent fingerprints-based anomaly detection, enhanced by explainable AI, offers a rapid approach for identifying subtle anthropogenic sign in water sources, with potential applications in early warning in water quality monitoring systems.