複雑な産業排水中のフェノールをロバストにモニタリングするための多次元差分分光法と深層学習の相乗的統合

Phenol is a key chemical intermediate and a common, toxic pollutant found in industrial wastewater from processes like coking and oil refining. Its leakage and discharge lead to significant contamination, posing a dual threat to ecological systems and human health, making its accurate, real-time monitoring paramount. Current detection methods have significant drawbacks. Chemical sensors can have long response times or insufficient accuracy. Standard optical methods like chromatography are cumbersome and ill-suited for rapid, on-site use. Simple UV-Vis spectroscopy is severely compromised by overlapping spectral interference and environmental instability in complex water samples.

The strategy first leverages phenol's specific bromination to selectively eliminate its spectral signature. A custom dual-channel, variable-path (1-10 cm) system was built to capture this transformation, generating high-dimensional (3D) differential spectral matrices. This 3D data approach reduced the Mean Absolute Error (MAE) by 45.4 % compared to conventional 1D spectra (when using the same CNN model). For analysis, a novel Parallel Associative Neural Network (PSNN) was developed, which substantially outperformed a standard CNN by achieving a further 66.7 % reduction in MAE. This integrated strategy (3D + PSNN) yielded a final prediction error of 72.09 μg/L under laboratory conditions and an MAE of 172.9 μg/L in field tests against HPLC benchmarks.

This study presents a novel strategy integrating chemical selectivity, multidimensional optical sensing, and a custom Parallel Associative Neural Network (PSNN). The work provides a robust framework for developing highly selective and sensitive intelligent systems for on-line water quality monitoring. The practicality and stability of this integrated approach in complex environments were validated through temperature control experiments and field tests on real industrial wastewater.