Anomaly Detection in Fiber-Based Distributed Acoustic Sensing Systems Employing Autoencoders

Abstract

Distributed Acoustic Sensing (DAS) systems have emerged as a high-potential solution for a wide range of applications, including vehicular monitoring. A key challenge in DAS signal processing is the reliable detection of anomalies within high-dimensional data streams, particularly under scarce anomalous training samples. In this work, we propose an autoencoderbased approach for anomaly detection in DAS systems. The method employs a fully connected neural architecture trained on normal traces to model typical system behavior. Reconstruction errors are then used to identify anomalous traces. The approach is validated on an open-access DAS dataset, demonstrating the model's ability to identify anomalous traces.