Machine learning techniques for adaptive polynomial and neural network digital predistorters
(English) The power amplifier (PA) is a core element in the radio transmitters to support the required mobile and fixed broadband communication link ranges. However, the PA is a power-hungry and nonlinear by nature device. Under spectrally efficient wideband modulated waveforms with high peak-to-ave...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/690181 |
| Acceso en línea: | http://hdl.handle.net/10803/690181 https://dx.doi.org/10.5821/dissertation-2117-403561 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Enginyeria de la telecomunicació Àrees temàtiques de la UPC::Informàtica 004 621.3 |
| Sumario: | (English) The power amplifier (PA) is a core element in the radio transmitters to support the required mobile and fixed broadband communication link ranges. However, the PA is a power-hungry and nonlinear by nature device. Under spectrally efficient wideband modulated waveforms with high peak-to-average power ratio, the PA energy efficiency is significantly decreased since back-off operation is needed to meet the transmission quality requirements. Moreover, when employing highly efficient amplification architectures, like Doherty, load modulated balanced amplifier (LMBA) or envelope tracking (ET) PAs, the added distortion is left as an issue to be addressed at system level by a linearizer. In this context, the closed-loop adaptive digital predistorter (DPD) is a key component of the digital front-end (DFE) to counteract the PA nonlinear response under varying conditions and to cope with the inherent trade-off between linearity and efficiency. According to the fifth generation (5G) and beyond communication technologies and the proposed radio transmitter and PA architectures, the DPD may have to deal with strong nonlinearities and memory effects, in-phase and quadrature (IQ) modulator imbalances and DC offsets, additional PA supply or load modulation distortion, and multi-antenna PA input and output cross talk and beam-dependent effects. Such impairments degrade the radio access network (RAN) energy efficiency, capacity, and the number of potential RAN users due to the increased in-band and out-of-band distortion. The adaptive DPD can overcome such undesired effects but faces relevant obstacles. At every new generation of mobile communication systems, the signal bandwidth is increased and the DPD needs to be operated at higher speed. The DPD challenges are twofold. On the one hand, combining massive bandwidth operation together with handling complex multi-dimensional effects may increase exponentially the complexity of the DPD and make it both commercially unaffordable and energy inefficient due to the increased cost and power consumption at the DFE and data conversion stages. On the other hand, the adaptive DPDs need significantly larger training periods to compensate for all the undesired effects. In line with the above-mentioned challenges, the research presented in this dissertation aims at guaranteeing best DPD linearity versus efficiency trade-off in complex nonlinear scenarios, by leveraging on efficiently deployed machine learning (ML) and artificial intelligence (AI) techniques to reduce the computational complexity of DPD modeling and identification at the DFE, guaranteeing well-conditioned and robust DPD estimation, and drastically reducing the DPD training times while meeting performance requirements. To accomplish that, several newly applied and customized ML feature selection and feature extraction dimensionality reduction techniques are combined with new training data length reduction schemes, to ensure both reduced DPD behavioral modeling matrices and input datasets in single-antenna and multi-antenna adaptive polynomial and neural network DPD architectures, respectively. To validate the benefits of these contributions in accordance with the previous goals, all these techniques have been deployed and thoroughly benchmarked under adverse conditions in realistic laboratory test benches. |
|---|