Deep Learning for Depth Registration of DAS Channels in Vertical Seismic Profiling
Arvin Karpiah, Daniel O. Trad, Marcelo Guarido
Depth registration is a persistent issue with Distributed Acoustic Sensing (DAS) data.While the depth of DAS channels can be determined along the optical fiber, the depth of theDAS channels with respect to the formation is unknown due to several factors. This makesit difficult for DAS data to be calibrated and synthesized with other datasets, despite it beinga powerful sensing technology. In this study, we investigate two different deep learningarchitectures: a Long Short-Term Memory (LSTM) model with an attention mechanism,and a transformer model, to map DAS channels to the correct formation depths. We usedDAS and accelerometer data collected from a Vertical Seismic Profile (VSP) survey in acarbon dioxide storage monitoring facility. At common depths, DAS and accelerometertraces should show some correlation in pattern between them, as they sample the samegeology. Originally developed for language processing, we modified these deep learningmodels to understand time-series data and predict depths based on these time series. Themodels were trained using accelerometer data with known depths, and the resulting modelwas applied to predict depths for DAS channels.