ST-T: A Spatio-Temporal Transformer for Φ-OTDR Multi-Location Time Series Classification

Multilocation time series classification, one of the most fundamental Φ-OTDR vibration signal applications, has not only gained substantial research attention but has also been widely applied in security monitoring. The current prevailing classifiers are almost all built upon convolutional operations. As a consequence, these methods only perform well in modeling local relationships of signal content but are limited to capturing long-range global interactions. Besides, previous research typically emphasized temporal feature exploration. Yet, the fusion of both spatial and temporal perspectives is pivotal for unearthing richer details, aiding models in enhanced classification decisions. In this paper, we propose a novel Transformer-based method, called ST-T, to capture long-term spatiotemporal features simultaneously. Specifically, we first propose to take time series subsequences which stem from different location variables, as the inputs. Second, we employ a two-dimensional variable-position encoding layer to keep the spatiotemporal information of each patch. Third, we introduce a novel cross-shape mask that enables self-attention mechanism to focus on the pertinent information that requires attention. We evaluate our method on a publicly available Φ-OTDR vibration dataset. ST-T achieves the best accuracy rank compared with several competitive methods. Further ablation studies validate the effectiveness of our model.