Replay-based continual learning for Φ-OTDR event classification

Phase-sensitive Optical Time Domain Reflectometry (Φ-OTDR) enables precise sensing and localization of disturbance events along optical fibers, with applications in perimeter security, pipeline monitoring, and other fields. Accurate classification of disturbance events is essential, but current methods rely on static learning strategies limited to predefined disturbance categories. Real-world environments are dynamic, with new disturbance categories continually emerging, causing these methods to fail when classifying novel events. Updating models solely with new-class data induces catastrophic forgetting, degrading recognition of previously learned categories. To adapt to open dynamic environments, enabling models to continually learn new disturbance categories while preserving recognition of previously learned categories, we propose a replay-based continual learning method integrating Gradient Coreset Replay (GCR) and the OverLoCK network. GCR maintains a replay buffer that stores the gradient information of all observed data. During the learning of new disturbance classes, it replays old-class samples from this buffer, effectively mitigating forgetting. OverLoCK enhances feature robustness through top-down attention and Context-Mixing Dynamic Convolution (ContMix). With a buffer capacity of 600 after three tasks, the model achieves 96.38% accuracy across six disturbance types with 4.07% forgetting. The proposed method demonstrates high effectiveness and stability, providing reliable support for practical optical sensing applications.