A deep learning model enabled multi-event recognition for distributed optical fiber sensing

Fiber optic sensors that utilize backscattered light offer distributed real-time measurements and have been seen tremendous improvements in sensing distance and spatial resolution over the last decades. However, these improvements in sensor capabilities lead to a significant increase in the amount of data that needs to be processed. Traditional processing schemes are no longer adequate, so the development of novel signal processing methods is critical. Phase-sensitive optical time domain reflectometry (Φ-OTDR) is now applied in various applications for multi-event recognition, and it would usually be difficult, sometimes even unrealistic to label all the acquired samples due to its real-time and seamless monitoring nature. To fully take advantage of the information contained within the large number of unlabeled samples, which were formerly not utilized and hence wasted, we propose a semi-supervised model to boost the event classification performance of Φ-OTDR. The model extracts respectively the temporal features and the spatial bidirectional features together with a dual attention mechanism. Its classification accuracy has been improved up to 96.9% with only 1230 labeled samples. In addition, our model shows significant advantages when the number of labeled samples is reduced. Importantly, our method improves the accuracy of multi-event classification without any modification to the optical setup.