Automatic Identification of Diverse Tunnel Threats With Machine Learning–Based Distributed Acoustic Sensing

Abstract

As the backbone of modern urban underground traffic space, tunnels are increasingly threatened by natural disasters and anthropogenic activities. Current tunnel surveillance systems often rely on labor-intensive surveys or techniques that only target specific tunnel events. Here, we present an automated tunnel monitoring system that integrates distributed acoustic sensing (DAS) technology with ensemble learning. We develop a fiber-optic vibroacoustic dataset of tunnel disturbance events and embed vibroscape data into a common feature space capable of describing diverse tunnel threats. On the scale of seconds, our anomaly detection pipeline and data-driven stacking ensemble learning model enable automatically identifying nine types of anomalous events with high accuracy. The efficacy of this intelligent monitoring system is demonstrated through its application in a real-world tunnel, where it successfully detected a low-energy but dangerous water leakage event. The highly generalizable machine learning model, combined with a universal feature set and advanced sensing technology, offers a promising solution for the autonomous monitoring of tunnels and other underground spaces.