GPU-Based implementation of pruned artificial neural networks for digital predistortion linearization of wideband power amplifiers
This paper presents a feature selection technique based on l1 regularization to select the most relevant weights of artificial neural networks (ANNs) for digital predistortion (DPD) linearization of wideband radio-frequency (RF) power amplifiers (PAs). The proposed pruning method is applied to the f...
| Autores: | , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/445544 |
| Acceso en línea: | https://hdl.handle.net/2117/445544 https://dx.doi.org/10.1109/JMW.2025.3560420 |
| Access Level: | acceso abierto |
| Palabra clave: | Radio frequency Graphics processing units Linearity Artificial neural networks Peak to average power ratio Throughput Predistortion Real-time systems Wideband Power generation Neural network Power amplifier Digital predistortion GPU-based implementation Predistortion linearization Output power Hidden layer Processing unit Graphics processing unit Power efficiency Relevant weight Mean power output Pruning strategy Peak-to-average power ratio Model performance Behavioral model Parallelization Iterative learning control Normalized mean square error Throughput performance Artificial neural network model Orthogonal frequency division multiplexing Reference signal Digital signal processing Parameter vector Multiple blocks Load-modulated balanced amplifier Model-order reduction |
| Sumario: | This paper presents a feature selection technique based on l1 regularization to select the most relevant weights of artificial neural networks (ANNs) for digital predistortion (DPD) linearization of wideband radio-frequency (RF) power amplifiers (PAs). The proposed pruning method is applied to the first hidden layer of a feed-forward real-valued time-delay neural network, commonly used for DPD purposes. In addition, this paper presents the ANN-based DPD implementation using a graphic processing unit (GPU) with compute unified device architecture (CUDA) units. Thanks to the proposed pruning strategy, it is possible to reduce the ANN complexity significantly, thereby achieving a higher data throughput with the GPU-based implementation. The trade-off among RF performance metrics, number of model parameters and throughput of the GPU implementation is evaluated considering the linearization of a high-efficiency pseudo-Doherty load modulated balanced amplifier (LMBA). The linearized PA operating at an RF frequency of 2 GHz delivers a mean output power of 40 dBm with approximately 50% power efficiency when excited with 5G new radio (NR) signals with up to 200 MHz bandwidth and an 8 dB peak-to-average power ratio (PAPR). The real-time GPU implementation of the ANN-based DPD can meet the linearity specifications with a throughput circa 1 GSa/s. |
|---|