Deep Learning-Based Intrusion Detection and Impulsive Event Classification for Distributed Acoustic Sensing Across Telecom Networks

Our study introduces two pioneering applications leveraging Distributed Fiber Optic Sensing (DFOS) and Machine Learning (ML) technologies. These innovations offer substantial benefits for fortifying telecom infrastructures and public safety. By harnessing existing telecom cables, our solutions excel in perimeter intrusion detection via buried cables and impulsive event classification through aerial cables. To achieve comprehensive intrusion detection, we introduce a label encoding strategy for multitask learning and rigorously evaluate the generalization performance of the proposed approach across various domain shifts. For accurate recognition of impulsive acoustic events, we compare several standard choices of representations for raw waveform data and neural network architectures, including convolutional neural networks (ConvNets) and vision transformers (ViT). We also study the effectiveness of the build-in inductive biases under both high- and low-fidelity sensing conditions and varying amounts of labeled training data. All computations are executed locally through edge computing, ensuring real-time detection capabilities. Moreover, our proposed system can be seamlessly integrated with cameras for video analytics, significantly enhancing overall situation awareness of the surrounding environment.