Machine Learning for PAPR Distortion Reduction in OFDM Systems

Abstract: The purpose of the project is to investigate the possibility of using modern machine learning to model nonlinear analog devices like the Power Amplifier (PA), and study the feasibility of using such models in wireless systems design. Orthogonal frequency division multiplexing (OFDM) is one of the most prominent modulation technique used in several standards like 802.11a,802.11n, 802.11ac and more. Telecommunication systems like LTE, LTE/Aand WiMAX are also based on OFDM. Nevertheless, OFDM system shows high peak to average power (PAPR) in time domain because it comprises of many subcarriers added via inverse fast Fourier transform(IFFT). HighPAPR results in an increased symbol error rate, while degrading the efficiency of the PA. Digital predistortion (DPD) still suffers from high symbol error rate (SER) and reduced PA efficiency, when there is an increase in peak back off(PBO). A receiver based nonlinearity distortion reduction approach can be justified by the fact that base stations have high computation power. A iterative-decision-feedback mitigation technique can be implemented as a receiver side compensation assuming memoryless PA nonlinearities. For successful distortion reduction the iterative-decision based technique required the knowledge of the transmitter PA. The author proposes to identify the nonlinear PA model using machine learning techniques like nonlinear regression and deep learning. The results show promising improvement in SER reduction with small PA model learning time.