【作者】 杨光；

【导师】 刘发林；

【作者基本信息】 中国科学技术大学 ， 电磁场与微波技术， 2014， 博士

【摘要】 凭借其成本适中、编程灵活以及性能卓越的优势,数字预失真(Digital Predistortion,DPD)成为现代通信发射机中主流的线性化技术。在高速宽带数据业务需求以及多标准多频带通信体制发展推动下,宽带/双频DPD技术成为当前的研究热点。本文围绕宽带/双频DPD的行为建模、模型参数辨识以及预失真系统学习结构展开研究。在宽带DPD系统中,为了降低反馈回路中模数转换器的采样速率,对反馈回路进行带限处理,利用带限建模技术进行预失真建模。本文对带限建模方法进行分析,理论上证明带限模型参数辨识归结为广义最小二乘问题。从频域角度解释带限模型参数辨识的物理意义,并提出一种基于频域数据的带限模型参数辨识算法。所提出的频域模型辨识算法能够取得与传统的时域算法相当的性能,然而计算复杂度大大降低。将传统的DPD系统学习结构推广到带限DPD情况,利用所提出的频域模型辨识算法研究了不同学习结构能够达到的线性化性能。为了同时补偿交调失真和互调失真,双频DPD系统中每个频段的预失真器被建模为双输入单输出系统。二维记忆多项式(Two Dimensional Memory Polynomial,2D-MP)能够有效地建模并发双频功放的失真特性,但数值稳定差,且模型复杂度高。提出二维正交多项式缓解数值不稳定问题,在有限精度的数字处理系统中鲁棒性更好。提出一种包含同阶包络交叉项的简化2D-MP模型,在模型复杂度和建模精度之间取得良好的折中,同时便于利用一维查找表实现。提出基于乘法单元的双频模型,能够以较低的模型复杂度取得良好的线性化性能。提出一种直接学习DPD模型参数的一步辨识算法。利用基带失真分量迭代注入技术得到期望的预失真信号,然后通过最小二乘估计对DPD模型参数进行一步辨识。该算法能够取得与传统直接学习DPD参数辨识算法相当的线性化性能,然而计算复杂度更低。将该算法用于直接学习带限/双频DPD系统参数辨识,得到了良好的线性化效果,进一步证明所提算法的有效性。更多还原

【Abstract】 Owing to its advantages of moderate cost, high flexibility and superior performance, digital predistortion (DPD) has become the mainstream linearization technique in modern communication transmitters. Spurred by the demands of high-speed wideband data service and multi-standard/multi-band communication systems, wideband and dual-band DPD are the highlights in DPD research at present. This dissertation focuses on behavioral modeling, model identification and DPD learning architectures for wideband and dual-band DPD.To reduce the sampling rate of analog to digital converters in the distortion acquisition path, the feedback loop is band-limited in a wideband DPD system, which requires a band-limited modeling technique for satisfactory performance. The author proves that the band-limited modeling technique can be formulated as a generalized least squares problem, with clear physical meaning from a frequency domain perspective. A frequency domain data based model extraction algorithm is proposed, which provides as good linearization performance as the conventional time domain data based algorithm with greatly reduced computational complexity. Conventional DPD learning architectures are extended for band-limited DPD, and their linearization performances are studied with the proposed algorithm.To compensate for cross-band modulation distortion as well as intra-band intermodulation distortion in a concurrent dual-band power amplifier (PA), a dual-band behavioral model of either band takes the form of a dual-input-single-output system. The reported two dimensional memory polynomial (2D-MP) shows high modeling accuracy for characterizing distortion behaviors in a concurrent dual-band PA, but it exhibits numerical instability and suffers from high complexity. To alleviate the numerical instability problem,2D orthogonal polynomials are proposed, which show robust performance in finite precision processing. To reduce the model complexity of2D-MP, a simplified2D-MP model with envelop cross terms of equal order is proposed, which proves to be a good tradeoff between complexity and performance, and facilates implementation with1D lookup tables. A multiplicative cell based dual-band model is also proposed, which yields good linearization performance with low model complexity. A one-step model identification algorithm is proposed for direct learning DPD. The expected predistortion signal is firstly derived by iterative injection of the baseband distortion components, which facilitates one-step model extraction of the DPD model by least squares estimation afterwards. The algorithm yields comparable linearization performance to the conventional algorithm, but bears much lower computational complexity. It is applied to model extraction of direct learning based wideband and dual-band DPD systems, and good linearization performance is achieved, which validates the effectiveness of the proposed algorithm.更多还原