Spatiotemporal-Mamba: A Lightweight Dual-Branch Spatiotemporal Network for Intelligent Classification of Vibration Events in $\bm {\Phi }$-OTDR Systems

In recent years, distributed acoustic sensing (DAS) systems based on phase-sensitive optical time-domain reflectometry ($\bm {\Phi }$-OTDR) have demonstrated significant potential for power communication monitoring. However, the inherent long-range transmission and high spatial resolution of such systems often lead to challenges in achieving high classification accuracy and maintaining low nuisance alarm rates. To address these issues, we propose Spatiotemporal-Mamba, a lightweight dual-branch framework for joint spatiotemporal modeling. Unlike conventional approaches that rely on Transformer architectures or 2D signal transformations, the proposed method directly processes raw one-dimensional time-series data. The framework integrates an enhanced ResNet module for multi-scale spatial feature extraction and a bidirectional Mamba2 module for efficient temporal modeling. A gated fusion mechanism is introduced to adaptively combine spatial and temporal features with minimal computational overhead. Experimental results on a public $\bm {\Phi }$-OTDR dataset from Beijing Jiaotong University show that our model achieves a classification accuracy of 99.8% and a false alarm rate as low as 0.001, with only 0.67M parameters-outperforming state-of-the-art methods. Furthermore, the model attains 94.1% accuracy on a private dataset collected in complex outdoor environments, successfully validating its robustness against multi-source overlapping events. Despite the increased environmental complexity, these results confirm the model's strong generalization capability. These results highlight the effectiveness and deployment potential of the proposed method for real-time vibration event recognition in $\bm {\Phi }$-OTDR systems.