Dual-path knowledge-guided gating network for 3D fluorescence spectra recognition of fluorescent dissolved organic matter in drinking water

Abstract

Three-dimensional excitation–emission matrix (3D-EEM) fluorescence spectroscopy is widely used to monitor fluorescent dissolved organic matter (FDOM) in drinking water. However, its application in drinking water distribution systems (DWDSs) remains constrained by severe spectral overlap, background FDOM fluorescence fluctuations, and the limited extensibility of FDOM fingerprint libraries. In this study, we constructed a 3D-EEM FDOM dataset using authentic distributed drinking water as the solvent and developed a Dual-path Knowledge-guided Gating Network (DuKG-Net). The network integrates 3D-EEM image features extracted by MobileNetV4-Small with 17 domain-knowledge indicators through a cross-modal bilinear gating mechanism, enabling multi-label recognition of mixed-FDOM samples. On the independent test set, DuKG-Net achieved an overall exact match ratio (EMR) of 0.9786, with EMRs of 1.0000 and 0.9700 for unary and binary samples, respectively, and a macro-F1 score of 0.9923. It also outperformed eight representative deep learning architectures and seven conventional 3D-EEM analysis methods. Across three additional distributed drinking-water backgrounds with distinct background FDOM fluorescence signals, the overall EMR remained at or above 0.9429, indicating cross-background transferability. A centroid-distance-based open-set recognition (OSR) strategy achieved an area under the receiver operating characteristic curve (AUROC) of 0.9869 for four out-of-library (OOL) substances. These results support the use of knowledge-guided 3D-EEM models for rapid FDOM screening and the progressive expansion of FDOM fingerprint libraries.