Comparison of model order reduction techniques for digital predistortion of power amplifiers

Gilabert, Pere L.; Montoro, Gabriel; Wang, Teng; Ruiz, M. Nieves; Garcia, Jose

doi:10.1109/eumc.2016.7824308

Cited by 11 publications

(4 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The running time depends on both the number of coefficients considered in the search and on the digital signal processor used for running the algorithm. However, by comparing the SW-OMP algorithm with the Less Relevant Basis Removal (LRBR) brute-force technique presented in [10] in terms of computational time, the OMP is 21 times faster, while the accuracy of the search is similar to that obtained with OMP. The computational complexity of the SW-OMP in comparison to the OMP is slightly higher due to the FFT transformations and filtering operations.…”

Section: E Other Considerationsmentioning

confidence: 97%

See 1 more Smart Citation

Spectral Weighting Orthogonal Matching Pursuit Algorithm for Enhanced Out-of-Band Digital Predistortion Linearization

Gilabert

López‐Bueno

Montoro

2019

IEEE Trans. Circuits Syst. II

Self Cite

View full text Add to dashboard Cite

This paper presents a new variant of the orthogonal matching pursuit (OMP) algorithm for reducing the computational complexity of the digital predistortion (DPD) behavioral model in the forward path. The proposed spectral weighting OMP (SW-OMP) algorithm focuses on selecting the most relevant basis functions to compensate for the out-of-band residual distortion which may eventually be masked by the dominant in-band residual error. This basis selection is carried out in an off-line process that does not affect the computational complexity of the real-time closed-loop DPD but, on the contrary, reduces its complexity while enhancing the robustness. Experimental results show that by selecting the DPD coefficients with the SW-OMP, the inherent ACLR and NMSE degradation suffered when reducing the number of coefficients is mitigated under strong nonlinear operation, when compared to using the basis functions selected by the classical OMP algorithm.

show abstract

Section: E Other Considerationsmentioning

confidence: 97%

“…This paper is not aimed at providing an overview or comparison on the existing model order reduction techniques as found in [10]. Instead, in this paper an alternative approach to the classical OMP algorithm which is based on a spectral weighting strategy is proposed.…”

Section: Introductionmentioning

confidence: 99%

Spectral Weighting Orthogonal Matching Pursuit Algorithm for Enhanced Out-of-Band Digital Predistortion Linearization

Gilabert

López‐Bueno

Montoro

2019

IEEE Trans. Circuits Syst. II

Self Cite

View full text Add to dashboard Cite

show abstract

“…Polynomial-based DPD exhibits structural multicollinearity between predictors, enabling researchers to intelligently prune DPD coefficients that do not contribute to the efficacy of the DPD linearisation [3][4]. Authors of [5] allow for function reduction and a change of basis by employing Principal Component Analysis (PCA).…”

Section: Introductionmentioning

confidence: 99%

Acceleration of Digital Pre- Distortion Training Using Selective Partitioning

Loughman

Byrne

Farrell

et al. 2022

2022 IEEE Topical Conference on RF/Microwave Power Amplifiers for Radio and Wireless Applications (PAWR)

View full text Add to dashboard Cite

In recent years model and Digital Pre-Distortion dimension reduction has been widely researched. The operations involved when running DPD are often far less than those needed during the training of the DPD coefficients. The proposed partitioned Least Squares (LS) adaptation allows a selected subset of DPD coefficients to be updated while the remaining coefficients are held constant. This technique allows a more adaptive training procedure, improved interpretability of the important DPD coefficient's during training and the ability to partition the DPD function into specific groups. The Frisch-Waugh-Lovell (FWL) theorem is exploited to partition the coefficients of a DPD basis function trained using LS regression. The proposed methodology was experimentally validated with a Generalized Memory Polynomial (GMP) DPD function, used to linearize a 5W power amplifier (PA) driven by a 40MHz 5G-NR signal.

show abstract

“…• Feature extraction techniques: creating a reduced set of new variables that are linear or nonlinear combinations of the original variables. Some examples of this group are the principal component analysis (PCA) [78], partial least squares (PLS) [79] and canonical correlation analysis (CCA) [80]. Feature extraction techniques (PCA, PLS and CCA) will be discussed in Subsection 4.1.2.…”

Section: Dimensionality Reduction Techniquesmentioning

confidence: 99%

Contribution to dimensionality reduction of digital predistorter behavioral models for RF power amplifier linearization

Pham

View full text Add to dashboard Cite

The power efficiency and linearity of radio frequency (RF) power amplifiers (PAs) are critical in wireless communication systems. The main scope of PA designers is to build the RF PAs capable to maintain high efficiency and linearity figures simultaneously. However, these figures are inherently conflicted to each other and system-level solutions based on linearization techniques are required. Digital predistortion (DPD) linearization has become the most widely used solution to mitigate the efficiency versus linearity trade-off. The dimensionality of the DPD model depends on the complexity of the system. It increases significantly in high efficient amplification architectures when considering current wideband and spectrally efficient technologies. Overparametrization may lead to an ill-conditioned least squares (LS) estimation of the DPD coefficients, which is usually solved by employing regularization techniques. However, in order to both reduce the computational complexity and avoid ill-conditioning problems derived from overparametrization, several efforts have been dedicated to investigate dimensionality reduction techniques to reduce the order of the DPD model. This dissertation contributes to the dimensionality reduction of DPD linearizers for RF PAs with emphasis on the identification and adaptation subsystem. In particular, several dynamic model order reduction approaches based on feature extraction techniques are proposed. Thus, the minimum number of relevant DPD coefficients are dynamically selected and estimated in the DPD adaptation subsystem. The number of DPD coefficients is reduced, ensuring a well-conditioned LS estimation while demanding minimum hardware resources. The presented dynamic linearization approaches are evaluated and compared through experimental validation with an envelope tracking PA and a class-J PA The experimental results show similar linearization performance than the conventional LS solution but at lower computational cost. La eficiencia energetica y la linealidad de los amplificadores de potencia (PA) de radiofrecuencia (RF) son fundamentales en los sistemas de comunicacion inalambrica. El principal objetivo a alcanzar en el diserio de amplificadores de radiofrecuencia es lograr simultaneamente elevadas cifras de eficiencia y de linealidad. Sin embargo, estas cifras estan inherentemente en conflicto entre si, y se requieren soluciones a nivel de sistema basadas en tecnicas de linealizacion. La linealizacion mediante predistorsion digital (DPD) se ha convertido en la solucion mas utilizada para mitigar el compromise entre eficiencia y linealidad. La dimension del modelo del predistorsionador DPD depende de la complejidad del sistema, y aumenta significativamente en las arquitecturas de amplificacion de alta eficiencia cuando se consideran los actuales anchos de banda y las tecnologfas espectralmente eficientes. El exceso de parametrizacion puede conducir a una estimacion de los coeficientes DPD, mediante minimos cuadrados (LS), mal condicionada, lo cual generalmente se resuelve empleando tecnicas de regularizacion. Sin embargo, con el fin de reducir la complejidad computacional y evitar dichos problemas de mal acondicionamiento derivados de la sobreparametrizacion, se han dedicado varies esfuerzos para investigar tecnicas de reduccion de dimensionalidad que permitan reducir el orden del modelo del DPD. Esta tesis doctoral contribuye a aportar soluciones para la reduccion de la dimension de los linealizadores DPD para RF PA, centrandose en el subsistema de identificacion y adaptacion. En concrete, se proponen varies enfoques de reduccion de orden del modelo dinamico, basados en tecnicas de extraccion de caracteristicas. El numero minimo de coeficientes DPD relevantes se seleccionan y estiman dinamicamente en el subsistema de adaptacion del DPD, y de este modo la cantidad de coeficientes DPD se reduce, lo cual ademas garantiza una estimacion de LS bien condicionada al tiempo que exige menos recursos de hardware. Las propuestas de linealizacion dinamica presentados en esta tesis se evaluan y comparan mediante validacion experimental con un PA de seguimiento de envolvente y un PA tipo clase J. Los resultados experimentales muestran unos resultados de linealizacion de los PA similares a los obtenidos cuando se em plea la solucion LS convencional, pero con un coste computacional mas reducido.

show abstract

Comparison of model order reduction techniques for digital predistortion of power amplifiers

Cited by 11 publications

References 7 publications

Spectral Weighting Orthogonal Matching Pursuit Algorithm for Enhanced Out-of-Band Digital Predistortion Linearization

Spectral Weighting Orthogonal Matching Pursuit Algorithm for Enhanced Out-of-Band Digital Predistortion Linearization

Acceleration of Digital Pre- Distortion Training Using Selective Partitioning

Contribution to dimensionality reduction of digital predistorter behavioral models for RF power amplifier linearization

Contact Info

Product

Resources

About