Feature Extraction and Fusion for Anomaly Detection Using Hybrid Multi-Scale CNN-Transformer Model

This paper explores the challenge of effectively detecting and analyzing human intrusion events using vibration signals from Distributed Fiber Optic Sensors (DFOS) in high speed train security monitoring. A novel Hybrid Multi-Scale CNN-Transformer Model is proposed to capture both frequency-domain and time-domain features from vibration signals. The model employs parallel networks—a multi-scale Convolutional Neural Network (CNN) for frequency-domain analysis and a Transformer network for time-domain analysis. The multi-scale CNN uses different convolution kernels to learn features at various time scales, enhancing its ability to capture diverse signal characteristics. Concurrently, the Transformer effectively models temporal dependencies and long-range relationships in the time domain, capturing sequential patterns and dynamics in the vibration signals. This dual approach aims to enhance signal understanding and analysis by integrating comprehensive information from both domains. It effectively handles the diversity and complexity of signals by capturing latent features even in inconsistent signal conditions. The experimental results show that the proposed method achieves superior classification performance.