Open-Set Event Recognition Based on Contrastive Prototype Learning for Distributed Optical Fiber Acoustic Sensing System

Distributed optical fiber acoustic sensing (DAS) systems are widely used in vibration monitoring due to their high sensitivity and immunity to electromagnetic interference. Intrusion events recognition is a critical task in DAS. In real-world monitoring, unknown event signals significantly degrade the recognition accuracy of conventional methods. However, most of current research only focuses on the closed-set recognition (CSR), overlooking the open-set recognition (OSR) challenges in practical environments. In order to solve this problem, this paper proposes an inverted residual multi-scale convolutional neural network (IRMS-CNN) combined with contrastive prototype learning, enhancing the recognition accuracy in open environments. The proposed IRMS-CNN captures local and global features through multi-scale feature extraction, leading to more discriminative feature representations. Single feature prototype for each known class is maintained to prevent overfitting and enhance generalization capability. The contrastive prototype learning method forces samples to cluster near their corresponding class prototypes while maximizing the separation between different class prototypes. For the audio event dataset, the scheme achieves 92.97% overall accuracy, with 93.02% acceptance rate (AR) for known events and 92.81% rejection rate (RR) for unknown events. For the field event dataset, we obtain 88.33% overall accuracy, with 88.69% AR for known events and 86.45% RR for unknown events. The proposed method effectively improves recognition performance under the introduction of unknown events and enhance the model robustness.