Open-Set Recognition for Unknown Organic Pollutants in Drinking Water With 3-D Fluorescence Spectroscopy

Abstract

Most organic pollutants analysis and detection in water using three-dimensional fluorescence spectroscopy usually rely on the comparison of new samples with a given set of benchmark pollutants. In our real world, however, tested samples collected from in-situ monitoring system may contain unknown organic pollutants. The conventional classification approaches were forced to choose from one of the benchmarks, which may lead to poor detection performance. In this paper, an open-set recognition approach was proposed using the wavelength-coding feature extraction network incorporated into extreme value machine (EVM) to overcome the shortages. Convolutional neural network (CNN) coupled with wavelength-coding module of spectra was designed to obtain the feature vectors from excitation-emission matrices (EEMs) data. The probability distribution for each known class can be achieved by training the EVM with the extracted feature vectors. New samples then can be identified as known or unknown according to the threshold. An online simulation pilot device was set up to simulate the in-situ monitoring conditions and verify the effectiveness of the proposed method. Twelve kinds of organic pollutants were collected and tested. Compared with conventional methods and other open-set methods, the experimental results showed that the proposed method achieved more precision regardless of whether the unknown pollutants appeared or not. The concept of open-set recognition in water quality detection and the proposed method described in this work can be applied to the online identification system with the increasing categories of pollutants, which has great potential to ensure the safety of drinking water.