Machine learning-enabled fluorescence spectrum source tracing of typical pollutants in industrial wastewater.

Abstract

Complex and difficult-to-trace pollutants in industrial wastewater are posing major challenges for regulation and treatment. The known source identification methods rely on rough water quality data and models, which limits the accuracy of source tracing. In this study, a convolutional neural network-principal component analysis-random forest (CNN-PCA-RF) fusion framework that directly utilizes the three-dimensional fluorescence excitation-emission matrix (EEM) data was established. An EEM database with five sectors (chemical, smelting, fine chemical, pesticide manufacturing, biotechnology) and 207 datasets was constructed. The 2D-CNN extracts spatial-spectral EEM features, PCA performs dimensionality reduction, and RF provides stable classification. Under a multi-concentration data training setting, identification accuracies reached 95.99% at the outlet level and 96.85% at the sector level. Under validation with previously unseen random dilution factors, accuracies were 90.35% and 97.37%, respectively.