Intrusion Pattern Recognition in DAS System Using Multi-Domain Features and Transformer Network

Distributed Acoustic Sensing (DAS) offers a scalable and resilient solution for real-time perimeter security by transforming optical fibers into dense arrays of vibration sensors. However, its high-dimensional, noisy, and spatio-temporally complex data make accurate event recognition challenging. Conventional deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, often struggle to capture long-range dependencies and rely heavily on handcrafted or single-domain features, limiting their effectiveness in complex intrusion scenarios. Designing compact and discriminative feature representations remains a critical challenge. To address these issues, we propose DASViT1D, a novel intrusion detection framework that integrates a one-dimensional Vision Transformer (ViT1D) with fused multi-domain features. Specifically, we extract and combine Mel-Frequency Cepstral Coefficients (MFCC), Redundant Discrete Fourier Transform (RDFT), and Discrete Wavelet Transform (DWT) features to capture complementary spectral and temporal characteristics. ViT1D then leverages self-attention to model long-range dependencies across time. Evaluated on a public ϕ-OTDR DAS dataset with nine event classes, our approach achieves 93.5% accuracy, outperforming CNN-based baselines by nearly 8 points, with a false negative rate (FNR) of 0.009, a nuisance alarm rate (NAR) of 0.005, and fast inference (0.13s/sample), demonstrating its suitability for real-time DAS-based intrusion detection.