Neural Network Signal Processing for LTE Receiver Application

Neural Networks (NN) provide great flexibility in modeling non-linear relationships and have proved to be very valuable in domains such as Computer Vision, Language Processing etc. Under the DARPA AI Exploration initiative of Signal Processing in Neural Networks, we extend the latest NN research to Radio Frequency (RF) domain, by targeting a LTE receiver application in three ways. First, using physics driven Digital Signal Processing (DSP) models to train NN models, we replace the De-noise (Jammer Suppression), Modem (Time/Frequency Synchronization), Equalize (Channel handling) and Decode (Information bit extraction) functions with NN equivalents. These NN based functional equivalents individually give same or better performance than their DSP counterparts and the performance gains increase when the functionality is stitched together in a pipeline. Second, DSP elements in radios work well in the conditions for which they are designed, but degrade in edge cases outside of these conditions. We address this problem by training Generative Adversarial Network (GAN) models that learn novel data distributions and provide a path for generating large amounts of training data. This extended training data is used to update the DSP equivalent NN models, thus making them resilient to novel operating conditions. Finally, with the goal of increasing the TRL level of this research, we push the NN models to FPGA to get the benefits of hardware based processing. We will present the experimental results from these three tasks and discuss our findings.