An machine learning-enhanced reconfigurable software defined radio architecture for adaptive 5G wireless systems

Abstract

This paper presents a machine learning (ML)-enhanced software defined radio (SDR) architecture optimized for adaptive 5G wireless communication. The system incorporates predictive ML algorithms to enable real-time modulation selection, finite impulse response (FIR) filter reconfiguration, and spectrum adaptation based on dynamic channel parameters such as bit error rate (BER), received signal strength indicator (RSSI) and signal-to-noise ratio (SNR). A decision tree classifier and a deep Q-learning agent dynamically determine optimal modulation schemes (BPSK, QPSK, 16-QAM, OQAM) and filter tap configurations (4/8/16 taps), ensuring efficient communication under varying network conditions. Implemented on a Xilinx Zynq SoC using Verilog for datapath design and Python for ML control via AXI4-Lite, the architecture achieves a maximum operating frequency of 182.4 MHz, 40.7% logic utilization, and only 122.3 mW power consumption. Compared to existing SDR implementations, the system demonstrates a 17% frequency improvement, 28% power reduction, and 21% area savings. Real-time electrocardiogram (ECG) transmission confirms the system’s adaptability, achieving BER < 10⁻³ at 22 dB SNR and < 10⁻⁵ at 26 dB. These results affirm the viability of the proposed ML-SDR for edge-based biomedical and ultra-reliable low-latency communications (URLLC) applications in 5G networks.