Self-supervised pre-training for distributed acoustic sensing data feature learning with transformers

Distributed acoustic sensing (DAS) is a vibration measurement technology, achieving high spatial resolution while maintaining cost-effectiveness and robustness in harsh environments. The sensing data, i.e., waterfall plots, are two-dimensional (2D) spatial-temporal, which is abstract for humans to analyze. Consequently, it is essential to present data with more understandable features for various DAS applications. This paper introduces a new Transformer model and extends the self-supervised masked autoencoder to learn the features of waterfall plots, referred to as DAS-MAE. The DAS-MAE can understand the semantic meaning of waterfall plots without human intervention or data labels. After pre-training on the open dataset obtained in the lab, DAS-MAE demonstrates high capability and generalization in practical applications.