An improved prediction model for COD measurements using UV-Vis spectroscopy.

Abstract

In the 21st century, although water quality has been improved in the last two decades, water pollution by organic contaminants has remained a non-negligible issue in China, so Chemical-Oxygen Demand (abbreviated as COD, unit: mg L-1) is often used as the main index to measure the degree of surface water pollution. UV-Vis spectroscopy, as a sensitive and rapid analytical technique, is a green detection technology suitable for automatic online COD detection equipment. However, due to the complex composition of surface water, the interference degree of the UV-Vis spectrum caused by turbidity is strongly correlated with the size, type and color of particulate matter in the solution, which results in noise sensitivity and poor generalization of the current detection model. Therefore, the main purpose of this research is to improve the traditional detection model performance by using deep learning and a spectrum preprocessing algorithm. Firstly, we used an improved noise filter based on discrete wavelet transforms to solve the noise sensitivity. Secondly, we proposed a novel COD detection network to address poor generalization. Thirdly, we collected a total of 2259 water samples' UV-Vis absorption spectra and corresponding COD as a dataset. Then, we pipelined the improved noise removal algorithm and proposed COD detection network, as a complete COD prediction model. Finally, the experiment on the dataset shows that the COD prediction model has a good performance in terms of both noise tolerance and accuracy.