CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Machine learning techniques applied to dimensionality reduction for digital predistortion linearizers

Estudiants que han llegit aquest projecte:

Director/a: GILABERT PINAL, PERE LLUÍS

Departament: TSC

Títol: Machine learning techniques applied to dimensionality reduction for digital predistortion linearizers

Data inici oferta: 06-11-2020     Data finalització oferta: 06-06-2021


Estudis d'assignació del projecte:
    DG ENG AERO/SIS TEL
    DG ENG AERO/TELEMÀT
    DG ENG SISTE/TELEMÀT
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
machine learning, digital predistortion linearization
 
Descripció del contingut i pla d'activitats:
Over the past half century, the improvements in spacecraft technology have been primarily in the areas of microelectronics for on-board processing, high frequency electronic devices, and integrated circuits for communications and navigation, solar cells and batteries for on-board power generation and storage among many others.

Despite the fact that energy-storage technologies have advanced dramatically over the past years, the power consumption of on-board communications, sensors and digital signal processing systems is of paramount importance in battery or solar powered systems such as small satellites, HAPs or UAVs (drones). There is panoply of applications that involves the use of these systems, e.g., Earth observation applications, surveillance, broadcast communications, scientific research, etc.

In wireless communications, the power amplifier is a critical subsystem in the transmitter chain. Not only because it is one of the most power hungry devices that accounts for most of the direct current power consumption, but also because it is the main source of nonlinear distortion in the transmitter. Amplitude and phase modulated communications signals presenting high peak-to-average power ratio have a negative impact in the transmitter's power efficiency, because the PA has to be operated at high power back-off levels to avoid introducing nonlinear distortion.
Digital predistortion (DPD) linearization is the most common and spread solution to cope with power amplifiers (PA) inherent linearity versus efficiency trade-off. When considering wide bandwidth signals, carrier aggregation or multi-band configurations in high efficient transmitter architectures, such as Doherty PAs, envelope tracking PAs or outphasing transmitters, the number of parameters required in the DPD model to compensate for both nonlinearities and memory effects can be unacceptably high. This has a negative impact in the DPD model extraction/adaptation, because increases the computational complexity and drives to over-fitting and uncertainty. However, by applying dimensionality reduction techniques we can both avoid the numerical ill-conditioning of the estimation and reduce the number of coefficients of the DPD function, which ultimately impacts the baseband processing computational complexity and power consumption.

In this Project, several dimensionality reduction techniques will be described and compared in terms of model order reduction capabilities and evaluation performance. In particular, some of the machine learning techniques for dimensionality reduction under study will be:
- Genetic Algorithms.
- Hill climbing.
- Random Forest.
- Decision Trees.
- High Correlation using Variance Inflation Factor.
- Orthogonal Matching Pursuit (OMP).
- Double OMP (DOMP).
- Sparse Bayesian learning.
- LASSO regularization.
- Ridge regularization.

In particular, the performance of these dimensionality reduction techniques will be evaluated and compared when considering the nonlinear behavioral modeling and linearization of a high efficient load-modulated balanced amplifier (LMBA).
 
Overview (resum en anglès):

Over the past half century, the improvements in spacecraft technology have been primarily in the areas of microelectronics for on-board processing, high frequency electronic devices, and integrated circuits for communications and navigation, solar cells and batteries for on-board power generation and storage among many others.

Despite the fact that energy-storage technologies have advanced dramatically over the past years, the power consumption of on-board communications, sensors and digital signal processing systems is of paramount importance in battery or solar powered systems such as small satellites, HAPs or UAVs (drones). There is multiple applications that involves the use of these systems, e.g., Earth observation applications, surveillance, broadcast communications, scientific research, etc.

In wireless communications, the power amplifier is a critical subsystem in the transmitter chain. Not only because it is one of the most power hungry devices that accounts for most of the direct current power consumption, but also because it is the main source of nonlinear distortion in the transmitter. Amplitude and phase modulated communications signals presenting high peak-to-average power ratio have a negative impact in the transmitter's power efficiency, because the PA has to be operated at high power back-off levels to avoid introducing nonlinear distortion.

Digital predistortion (DPD) linearization is the most common and spread solution to cope with power amplifiers (PA) inherent linearity versus efficiency trade-off. When considering wide bandwidth signals, such as Doherty PAs, envelope tracking PAs or outphasing transmitters, the number of parameters required in the DPD model to compensate for both nonlinearities and memory effects can be very high. This has a negative impact in the DPD ceofficients extraction, because increases the computational complexity and drives to over-fitting and uncertainty. However, by applying dimensionality reduction techniques we can both avoid the numerical ill-conditioning of the estimation and reduce the number of coefficients of the DPD function, which ultimately impacts the baseband processing computational complexity and power consumption.

In this Project, several dimensionality reduction techniques will be described and compared in terms of model order reduction capabilities and evaluation performance. In particular, some of the machine learning techniques for dimensionality reduction will be studied.

© CBLTIC Campus del Baix Llobregat - UPC