Neural Network Signal Processing for LTE Receiver Application
Josh Robinson, Joseph M. Carmack, Amit Bhatia, John Majewski, Scott Kuzdeba, Joe Farkas, Brandon Hombs, Thomas Koch · 2021
Radio Digital Signal Processing (DSP) elements work well in the conditions for which they are designed, but degrade in edge cases outside of these conditions. In previous work we built individual physics-driven neural network (NN) models to replace standard DSP function models in an LTE-like intelligent receiver. This prevents performance degradation in the edge conditions while giving the same or better performance than the DSP models. In this work we extend these individual NN models to a more realistic LTE receiver with multiple antenna support and stitch together the individual NN models to form a complete end-to-end pipeline. We report results for these NN models as well as the complete pipeline, comparing with standard DSP element analogues using both channel bit error rate (CBER) and bit error rate (BER). The individual NN models perform equal or better than DSP blocks in the more realistic and difficult LTE receiver conditions and the end-to-end pipeline provides a much better performance improvement over a DSP pipeline, providing a modular architecture that can be used in other applications.