Rapid Prediction of Wastewater Index Using CNN Architecture and PLS Series Statistical Methods

Abstract

Chemical oxygen demand (COD) is an important index to measure the degree of water pollution. In this paper, near-infrared technology is used to obtain 148 wastewater spectra to predict the COD value in wastewater. First, the partial least squares regression (PLS) model was used as the basic model. Monte Carlo cross-validation (MCCV) was used to select 25 samples out of 148 samples that did not conform to conventional statistics. Then, the interval partial least squares (iPLS) regression modeling was carried out on 123 samples, and the spectral bands were divided into 40 subintervals. The optimal subintervals are 20 and 26, and the optimal correlation coefficient of the test set (R T ) is 0.58. Further, the waveband is divided into five intervals: 17, 19, 20, 22 and 26. When the number of joint intervals under each interval is three, the optimal R T is 0.71. When the number of joint subintervals is four, the optimal R T is 0.79. Finally, convolutional neural network (CNN) was used for quantitative prediction, and R T was 0.9. The results show that CNN can automatically screen the features inside the data, and the quantitative prediction effect is better than that of iPLS and synergy interval partial least squares model (SiPLS) with joint subinterval three and four, indicating that CNN can be used for quantitative analysis of water pollution degree.