Studies On High Performance And Low Complexity Digital Predistortion For Wireless Communication System

Due to the inherent nonlinear property of RF power amplifier,the non-constant envelope communication signals amplified by RF PAs will present obvious nonlinear distortion.To mitigate the nonlinearity of RF PAs,digital predistortion has become one of the indispensable modules in wireless communication system because of its high flex-ibility,high precision and low cost properties.With the continuous evolution of wire-less communication system to the fifth generation mobile communication technology,wireless communication base stations have the tendencies towards large bandwidth,miniaturization and high efficiency,which has brought new challenges for digital pre-distortion modules.On the one hand,the application of micro base stations imposes more strict restrictions on the cost and power consumption of the communication sys-tem,how to reduce the complexity of DPD has become the current urgent problem accordingly.On the other hand,RF PAs excited by wider bandwidth and higher peak-to-average power ratios(PAPR)signals can exhibit much more complex and stronger nonlinearities,high correlation between different polynomial basis functions of the tra-ditional Volterra-based model limits its performance improvement even when the num-ber of polynomial terms increases dramatically,which leads to the limited linearization performance for RF PAs with high nonlinearities.This dissertation focuses on how to reduce DPD implementation complexity and design high performance DPD models.In traditional DPD modules,each feedback path needs two high-precision and high-resolution analog-to-digital converters(ADCs)to collect the output of RF PAs,and in order to collect PA output’s out-of-band information,the sampling rate of ADCs should not be less than 3-5 times of the bandwidth of input signal,which will greatly increase the power consumption and cost of DPD modules,especially in broadband communication scenarios.In this paper,starting from the statistic properties of com-munication signals,a novel close-loop in-phase observation DPD method is designed.Then combine the 1-bit observation method,a close-loop 1-bit DPD system with a sin-gle feedback channel has been presented.Compared with the existing DPD systems,the power consumption and the cost of the feedback path of the proposed method are reduced significantly as only one 1-bit comparator is needed in the feedback path.After the completion of the above close-loop(1-bit)in-phase observation DPD architecture design,considering that open-loop indirect learning architecture(ILA)presents faster convergence,lower computational complexity and simple structure com-pared with the close-loop direct learning architecture(DLA),a novel in-phase observa-tion DPD scheme based on open-loop ILA has been further proposed.Compared with the classical ILA method,both the theoretical and experimental results reveal that the proposed scheme could effectually mitigate the measurement noise to reach better per-formance,and with lower cost of the feedback path.However,when there is obvious I/Q imbalance effect in the system,the classical joint PA nonlinearity and I/Q modu-lator predistortion method is no longer suitable for the in-phase observation scheme.Accordingly,this thesis discussed the I/Q modulator impairments and presented a time-division method to correct the I/Q imbalance for the above mentioned in-phase obser-vation methods.Due to its excellent capability of nonlinear fitting,neural network(NN)has been considered as a promising method for behavioral modeling for RF PAs.In this pa-per,three novel vector-decomposition based multilayer preceptron neural network DPD models are proposed.First,the irrationality of widely used splitting input and output into in-phase andquadrature parts to construct models was analyzed.Then,we propose three new neural network models and they conform more with the nonlinear physi-cal mechanisms of RF PAs,where only the magnitudes of input signals are conducted nonlinear operations,and the phase information is then recovered with linear weight-ing operations.The theoretical analysis and various experimental results show that the proposed models can achieve better performance and with significantly lower compu-tational complexity,compared with the existing neural network DPD models.