Inter-sequence-attention Transformer network for distributed fiber-optic sensing signal recognition

Abstract

Distributed fiber-optic acoustic sensing (DAS) is an emerging technology which can capture acoustic waves everywhere along the sensing fiber. The recognition of such spatial-temporal sensing data is a crucial step in some applications of DAS. In this work, we extend the Transformer neural network well known for its state-of-the-art performance in sequential tasks into an inter-sequence form, mainly by rearranging the query matrices, key matrices, and value matrices separately from each temporal sequence, into piles of matrices that each matrix contains information from all sequences, consequently getting inter-sequence attention weights and attention matrices. The network is named inter-sequence-attention Transformer (ISAT), which simultaneously extracts the temporal and spatial joint features of DAS data, while explicitly preferring temporal dimension to spatial dimension rather than treating them equally like in 2D convolution neural networks (CNNs). According to evaluation results based on the open dataset, the ISAT network shows an improvement in aspects of both recognition accuracy and parameter efficiency compared with previously proposed networks.