Uncovering hidden industrial wastewater discharges in urban rivers via deep learning-enhanced fiber-optic distributed acoustic sensing.

Illicit wastewater discharges from concealed outfalls threaten urban river ecosystems, often evading conventional monitoring. This study introduces an intelligent framework that integrates distributed acoustic sensing (DAS) with a Residual Network (ResNet) deep learning model to overcome this challenge. By deploying a fiber-optic sensing cable as a dense acoustic sensor array in a Chinese river, we trained a ResNet model to distinguish discharge signals from complex ambient river noise (including vessel traffic and hydrodynamics), achieving over 92% identification accuracy. This system successfully located a concealed, submerged textile factory outfall with meter-scale precision, demonstrating its capability to intercept high-risk pollutants. Analysis of the outfall's acoustic signature revealed continuous, low-frequency (<10 Hz) discharge patterns, including three brief interruptions attributed to routine equipment inspections. The re-initiation of discharge after these interruptions produced a transient energy peak consistent with jet flow pressure dynamics, showcasing distinct frequency characteristics between transient and steady-state responses. This work demonstrates that fusing DAS with deep learning offers a powerful, data-driven solution for regulators to detect, locate, and characterize hidden pollution sources, ultimately advancing environmental risk control in complex aquatic environments.