Implementation of Tiny Machine Learning (TinyML) as Pre-distorter for High Power Amplifier (HPA)Linearization of SDR-based MIMO-OFDM

This paper implements a new pre-distorter (PD) scheme using Tiny Machine Learning (TinyML) to compensate for the non-linear High-Power Amplifier (HPA) in the Software Defined Radio (SDR)-based Multiple-Input Multiple-Output Orthogonal Frequency Division Multiplexing (MIMO-OFDM). TinyML is used because there are weaknesses in the PD neural network (NN), which has a large model size and high computational costs. Some research uses PDNN because it has flexible advantages and can provide the inverse characteristics of sophisticated PA modelling. Using PDNN in the SDR-based MIMO-OFDM will cause system degradation because the SDR device has low computational memory. The system degraded because the execution time of the transmitter was high, which affected the latency of the communication system. TinyML will reduce the complexity of the NN model, resulting in a smaller model size and a lower computational cost. The implementation of PD TinyML in SDR-based MIMO-OFDM can improve the execution time by 40% compared to the use of PDNN. Even though it has a lower execution time, TinyML can compensate for HPA and perform similarly to PDNN. The maximum SNR value of PD TinyML is only 2.07 dB, lower than the maximum SNR value of PDNN. In addition, the BER graph of PDTinyML’s SNR to achieve a minimum BER is 1.37 dB higher than PDNN. From this, the performance decrease in PDTinyML is not too significant.