摘要
A new model for analyzing multi-component overlapping fluorescence spectra. • Simultaneously achieving qualitative and quantitative analysis of each component. • Proposed multitask HRNet-MobileNetV4 for analysis EEM spectra. • Accurate single-component characterization and concentration prediction in mixtures. • Achieve component resolution capabilities comparable to PARAFAC. In water environmental monitoring, excitation–emission matrix (EEM) fluorescence spectroscopy is a powerful tool, yet its application is often hindered by the coexistence of multicomponent analytes with highly overlapping spectral signals. This study presents a novel, end-to-end deep learning approach for simultaneous qualitative and quantitative analyses. We constructed a diversity-enhanced dataset by combining experimental measurements, public data, and simulated spectra to include a wider variety of substance types and concentration ranges, thereby improving the model's generalization capabilities. Subsequently, a customized multitask HRNet–MobileNetV4 network was developed to characterize spectral features and predict concentrations simultaneously. Validation results demonstrate high precision: the mean relative error for concentration prediction ranges from 1.18% to 4.28%, and the cosine similarity between reconstructed and actual spectra exceeds 0.99. Compared to PARAFAC, the proposed method automatically determines the number of components, significantly accelerating the analysis process while maintaining comparable prediction accuracy. Furthermore, once trained, the model enables rapid, independent single-sample analysis. This approach offers a robust and efficient strategy for real-time, in-situ monitoring of complex water pollutants, significantly advancing smart environmental perception technologies. Download: Download high-res image (188KB) Download: Download full-size image
危险品化学品经营许可证(带存储)