Unified self-training framework for open-set semi-supervised event recognition in distributed optical fiber sensing

Distributed optical fiber sensing (DOFS) has become a prominent solution for real-time and wide-area monitoring owing to its long sensing range and high sensitivity. However, existing event recognition approaches rely heavily on large-scale annotated datasets, making them difficult to deploy in practical applications where labeled data are scarce and unknown events frequently occur. Moreover, most current models are designed for closed environments and struggle to identify unseen disturbances under real-world conditions. To address these challenges, we propose an open-set semi-supervised event recognition method for DOFS systems based on a unified self-training framework. The proposed method selects reliable unknown samples using a buffer-based selection strategy and integrates closed-set and open-set self-training into a unified optimization process, enabling simultaneous recognition of known and unseen events. Experimental results show that our method delivers strong and consistent performance in both close-set and open-set event recognition scenarios, achieving 95.14% open-set accuracy and substantially outperforming representative baselines such as IOMatch and OpenMatch under low-supervision conditions (1% labeled data). This demonstrates that the proposed method markedly improves the recognition performance and robustness of DOFS systems, providing a reliable foundation for intelligent optical fiber sensing in open and dynamic environments.