Real-Time Downhole Geosteering Data Processing Using Deep Neural Networks On FPGA

Abstract

The success of machine learning has spread the deployment of Deep neural Networks (DNNs) in numerous industrial applications. As an essential technique in today’s oilfield industry, geosteering requires performing DNN inference on the hardware devices that operates under the severe down-hole environments. However, it can produce massive power dissipation and cause long delays to execute the computation-intensive DNN inference on the current hardware platforms, e.g., CPU and GPU. In this paper, we propose an FPGA-based hardware design to efficiently conduct the DNN inference for geosteering tasks in downhole environments. At first, a comprehensive analysis is presented to choose the optimal computation mapping method for the target DNN model. A detailed description of the customized hardware implementation is then proposed to accomplish a complete DNN inference on the FPGA board. The experimental results shows that the proposed design achieves 7× (1.4×) improvement on performance and 82× (1.3×) reduction on power consumption compared with CPU(GPU).