A study on prototypical networks for DAS acoustic event classification in railway trackside safety monitoring

Abstract

To achieve intelligent perception of environmental risks in railway transportation safety, we investigate acoustic event classification models using a real-world distributed acoustic sensing (DAS) sound dataset. This dataset contains five representative vibration categories: background noise, hammering, rockfall, prying, and train passage. In this article, we design a prototypical network-based model named Proto-LightCNN and employ K-shot experiments to identify its optimal configuration, demonstrating the distinct advantages of meta-learning in few-shot scenarios. To validate its effectiveness on scarce datasets, we conduct a comprehensive comparative study using real-world railway collected data, benching our proposed Proto-LightCNN against traditional machine learning and conventional deep learning classification methods. Results indicate that our framework achieves an accuracy of 94.74%, confirming its strong capability in feature optimization. Furthermore, the recognition robustness against similar disturbances is significantly enhanced across a 20 km operational railway line. This study establishes a robust and reliable technical framework for intelligent intrusion detection, offering significant practical value for DAS-based railway safety monitoring.