An encoder-decoder CNN for DAS-to-geophone transformation

Jorge E. Monsegny, Daniel O. Trad, Donald C. Lawton

Distributed acoustic sensing is a technology that uses optical fibre to record seismic waves. While traditional geophones record the particle velocity created by a passing wave, optical fibre records the strain or strain rate. The conversion between the two kind of signals allows seismic time lapse imaging applications with data from these two different recording systems. Here we use convolutional neural networks to transform fibre to geophone data. Instead of using a supervised model where we provide examples of corresponding fibre and geophone traces, we utilise an encoder decoder scheme that receives fibre traces and produces fibre traces. The important distinction is that the decoder is deterministic and contains the physics of transforming a geophone trace to a fibre trace while the encoder is the convolutional neural network that does the opposite transformation. The whole encoder-decoder is trained to be the identity operator on fibre traces. At the end of the training, the application of the encoder part alone will perform the desired signal conversion from fibre to geophone.