Extraction and precise source tracing of mixed wastewater three-dimensional fluorescence spectra based on machine learning techniques

Accurately identifying wastewater components and sources in industrial park underground pipelines is key for wastewater discharge management and risk early warning. Traditional three-dimensional excitation-emission matrix fluorescence spectroscopy (EEMs) can identify these but can't quantify different sources' contributions. To tackle the challenge of quantifying multi-component mixed pollution sources in industrial park wastewater, this study, based on prior EEMs research on industrial wastewater characteristic peaks, chose lubricating oil, Reactive Blue 19, and cefradine as characteristic markers for petroleum (PI), printing and dyeing (PADI), and cephalosporin pharmaceutical (CPI) industry wastewater, respectively. Through simulated experiments with different proportions, it used EEMs-based machine learning algorithms to build Random Forest (RF) classification and regression models for efficient wastewater component identification and quantitative analysis. Results showed the RF classification model had 100% accuracy in identifying mixed samples in the lab, and the RF regression model had an average correlation coefficient of 0.9599 across various industries' wastewater. When combined, the correlation coefficients for PI, PADI, and CPI rose to 0.9378, 0.9663, and 0.9911. The error between parsed and actual wastewater component proportions using EEMs-based machine learning was 7.6%, indicating high feasibility for quantifying industrial mixed wastewater sources. Yet, it's still necessary to address interference from the complexity of actual mixed wastewater on parsing results. Graphical abstract