Convolutional Neural Network-Based PUSCH DMRS Pattern Optimization for 5G New Radio

A convolutional neural network (CNN)-based architecture for the optimization of demodulation reference signal (DMRS) patterns is proposed for 5G new radio systems. The proposed architecture employs a structured 4-step CNN classification, with each step designed to determine the optimal number of DMRS symbols, starting symbol, configuration, and sub-configuration types. The proposed approach achieves a data rate performance highly comparable to the optimal pattern, showing a 4.2% improvement over other network architectures and a 12.9% gain over the conventional scheme with a fixed DMRS pattern.