DASFormer: A Long Sensing Sequence Classification and Recognition Model for Phase-Sensitive Optical Time Domain Reflectometers

In recent years, the application of distributed fiber optic sensing technology in road traffic monitoring has gained significant attention, greatly enhancing the efficiency and safety of transportation systems. To comprehensively understand the dynamic changes in road traffic, it is essential to collect long monitoring sequences and analyze them to gain deeper insights. With current methods, however, it is still difficult to precisely process the long sequences collected by phase-sensitive optical time domain reflectometry (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\Phi $ </tex-math></inline-formula>-OTDR) in real-time. To address this issue, we propose DASFormer, a model designed with temporal and spatial reduction mechanisms, making it highly suitable for processing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\Phi $ </tex-math></inline-formula>-OTDR long sequences. DASFormer model excels in processing such long sensing signal with remarkable accuracy, providing a significant improvement over previous method. The model adopts a dual-tower architecture to separately extract temporal and spatial features, with a multilevel and multiscale feature fusion structure, with adaptive multihead attention to dynamically adjust the importance of each attention head, and with DeepNorm normalization functions to enhance the network’s stability. We conducted comprehensive recognition and classification on both laboratory-acquired and field-tested datasets. The model achieved an impressive 99.6% accuracy in identifying six types of events in the laboratory data and a 98.7% accuracy for five types of events in the field data. Additionally, DASFormer demonstrates the ability to identify radial distances, further enhancing its capabilities. These experimental results demonstrate that the model has outstanding adaptability and practicality for handling a broader range of data in complex road traffic monitoring environments and beyond.