Anomaly detection for distributed optical fiber acoustic sensing based on one-class convolutional neural network

Distributed fiber optic sensing has significant research value in the field of anomaly intrusion detection, but it faces the challenge of the scarcity of real sample data. To address this issue, this paper proposes an unsupervised fiber optic sensing feature learning method based on a One-Class Convolutional Neural Network (OC-CNN), which can identify intrusion signals solely relying on sensing data from non-intrusive conditions. The method utilizes a well-pretrained VGG19 network for feature extraction and only trains a custom classifier network. Experimental results validate that the method achieves an average recognition accuracy of 92.59% for multiple intrusion types without prior knowledge.