General-Incremental Neural Network Model for Linearizing the Ultra Broadband Software-Defined Radio Over Fiber System

Software-defined radio (SDR) can offer high flexibility for wireless communication. However, as the operating frequency of wireless communication increases, the coverage of a single active antenna unit (AAU) or remote radio unit (RRU) is reduced. Fortunately, the radio over fiber (RoF) can pull away the radio frequency (RF) signal to extend the cover range for a base station. Therefore, the integration of RoF system and SDR technology, i.e., software-defined radio over fiber (SDRoF) is crucial and has complementary advantages. Nevertheless, a lot of RF and optical components exhibit nonlinearities, which significantly constrain the transmission performance of the SDRoF system. Implementing digital pre-distortion (DPD) to linearize the ultra-broadband frequency band of the SDRoF system requires a significant investment of resources. This study proposes a general-incremental neural network (GINN) model that leverages the generalization ability of neural networks to develop a comprehensive solution for training a single model to linearize the ultra-broadband of the SDRoF system. The proposed model includes three components, i.e., a general module, an adaptive switch, and an incremental module. The general module is trained as a general predistorter, the adaptive switch keeps tracking the linearization status. The incremental module is activated by the adaptive switch to provide additional linearization compensation in linearity degraded frequency bands. To verify the effectiveness of the proposed method, this paper selects four widely used mobile frequency bands as examples: 2.6 GHz, 3.5 GHz, 4.9 GHz, and 5.8 GHz. The experimental results show that the predistorter built by the proposed model can improve the adjacent channel leakage ratio (ACLR) and error vector magnitude (EVM) of the transmitter efficiently. Furthermore, in the case of abnormal output spectrum of the transmitter e.g., at the 3.5 GHz frequency band, the ACLR and EVM of the incremental module can be further improved by 1-2 dB and 0.32%, respectively.