Rockfall Monitoring for Highways Based on Distributed Fiber Optics and Model Merging

Abstract

Rockfall hazards have increasingly drawn global attention due to their devastating impacts on human lives, infrastructure, and the environment. Traditional monitoring approaches for these events typically rely on ground-based radar, video surveillance, and point sensors such as accelerometers and strain gauges. Although these methods provide valuable insights into ground motion and event characteristics, they are often constrained by sparse spatial coverage, limited resolution, susceptibility to environmental interference, and high installation and maintenance expenses. Distributed Acoustic Sensing (DAS) technology mitigates these issues by converting optical fiber cables, extending tens of kilometers, into a continuous array of virtual sensors with meter spatial resolution. Previous studies have demonstrated that DAS can capture high-frequency seismic signals generated by artificially triggered rockfalls, exhibiting strong correlation with conventional seismic records. In this study, we conducted an extensive DAS data collection campaign that captured routine operational signals from various highway segments, diverse vehicle types, and different time periods. The testing section, approximately 1 km long, is situated along an highway in Quzhou, Zhejiang Province, China. Subsequently, we manually simulated rockfall events by allowing wooden stakes of varying weights to fall naturally at different locations, thereby emulating the characteristics of real rockfall incidents. Given the substantial daily sample size and the rigorous requirements for a low false alarm rate, a single model proved insufficient. Therefore, by integrating features from the time-frequency domain and spatial resolution, we fused an audio classification model- PANNs with Self-Attention and Convolution to develop a DAS-based long-range rockfall monitoring and early warning system. This model delivers accurate, real-time monitoring over extensive distances, achieving a recall rate exceeding 92% while maintaining a false alarm rate of only 0.058%. Moreover, the system can provide timely alerts regarding incidents, offer recommendations to road maintenance departments, and significantly mitigate secondary disasters such as traffic accidents and congestion triggered by rockfalls.