Electromagnetic (EM) telemetry is a cost-competitive method for communication between downhole tools and oilfield surface equipment. However, EM telemetry use is limited to wellsite deployments that result in favorable signal-to-noise ratio (SNR) levels. Increasingly, the EM attenuation brought by deeper wells, conductive service fluids, and conductive formation layers has made it more important to increase EM telemetry SNR. This study explored the ability of various deep learning (DL) methods to improve EM telemetry SNR with data sets composed of mixed synthetic downhole signals and real field noise. The DL models explored were previously proven to achieve state-of-the-art performance with seismic random noise attenuation and speech separation use cases.
Furthermore, architectural adjustments that improved model performance with EM telemetry data were targeted. Then, the SNR improvement performance of different design changes was evaluated. Finally, the study targeted improvements that reduced the requirements for deployment hardware, including number of stakes, without sacrificing SNR improvement performance. Optimized models were found to successfully improve SNR in EM telemetry data, outperforming traditional bandpass filter methods and enabling the demodulation of mixtures that were otherwise too noisy to be demodulated.