Digital predistortion (DPD) has been the primary linearization technique in wireless communication systems for over two decades due to its flexibility and precision in digital implementation. Conventional DPD models, such as memory polynomial, generalized memory polynomial, and orthogonal memory polynomial, have been widely adopted to mitigate power amplifier (PA) nonlinearities and memory effects. However, these approaches often exhibit high computational complexity and limited adaptability, making them less suitable for real-time software-defined radio platforms. Furthermore, most existing solutions primarily address PA nonlinearity while neglecting hardware-induced impairments such as DC offset and in-phase/quadrature imbalance, which significantly degrade overall system performance. To overcome these limitations, this paper proposes a neural network-based DPD technique that integrates a piecewise Swish (P-Swish) activation function within a real-valued time-delayed neural network architecture. The proposed method jointly compensates for PA nonlinearity and hardware imperfections while leveraging the physical characteristics of nonlinear distortions to enhance modelling accuracy. By focusing on both signal de-noising and joint compensation, the approach achieves superior linearization performance with reduced complexity compared to state-of-the-art ANN-based and polynomial-based models. Extensive simulations and experimental evaluations demonstrate significant improvements in adjacent channel power ratio, normalized mean square error, and computational efficiency, making the proposed technique a strong candidate for next-generation SDR systems and advanced wireless communication infrastructure.
Dubey et al. (Fri,) studied this question.