Mixed Intrusion Events Recognition Based on Group Convolutional Neural Networks in DAS System

Abstract

The mixed intrusion events recognition is still a challenging problem in optical fiber perimeter security with distributed acoustic sensing (DAS), because the vibration signals will be mixed when multiple events occur at the same time and in close proximity. In order to identify mixed events consisting of two single events, we propose a recognition scheme based on deep group neural networks algorithm. The 100 groups convolutional neural networks (100G-Net) model is designed to make the best of vibration information of samples for feature extraction and classification. The experiments show that not only average training recognition accuracy of the proposed algorithm can reach 99.6%, but also the generalization ability of the proposed algorithm is better than typical CNN models.