This thesis investigates the application of the Transformer Neural Network for digital pre-distortion (DPD) to counteract the nonlinearities of power amplifiers in multi-carrier signal scenarios. Building upon previous work, which tested the Transformer architecture on a single-carrier signal and demonstrated its superiority over other state-of-the-art models, this study extends the exploration to more complex and realistic multi-carrier environments. The research investigates and proposes improvements and modifications to the Transformer architecture and to the input features used to train the network, and assesses its capabilities across diverse signal conditions. The proposed model is compared against other neural network models, specifically a Long Short-Term Memory Neural Network (LSTMNN), a Bidirectional LSTM Neural Network (BiLSTMNN) and a Bidirectional Gated Recurrent Unit Neural Network (BiGRUNN), known for their effectiveness in DPD modeling. Additionally, the model has been evaluated alongside Ericsson’s proprietary DPD models. Evaluative metrics include normalized mean square error (NMSE) and adjacent channel leakage ratio (ACLR). The Transformer-based model shows superior performance when compared to the other neural networks, and also comparable performance to Ericsson's DPD. However, the model's complexity is still too high for practical FPGA implementation. This aspect is identified as a significant area for future research and optimization, although it falls outside the scope of this thesis.
La presente tesi analizza l’applicazione delle reti neurali Transformer alla pre-distorsione digitale (DPD) per compensare le non-linearità degli amplificatori di potenza in scenari a segnali multi-portante. Partendo dai risultati ottenuti in letteratura, che hanno evidenziato l’efficacia dell’architettura Transformer per segnali a singola portante rispetto ad altri modelli all’avanguardia, questo studio estende l’analisi a contesti più complessi e realistici. L’obiettivo è valutare e migliorare le prestazioni del modello apportando modifiche sia all’architettura della rete che alle caratteristiche delle variabili usate durante l’addestramento, esaminandone il comportamento in diverse condizioni operative. Il modello proposto viene confrontato con altre reti neurali impiegate nella modellazione DPD, tra cui una rete Long Short-Term Memory (LSTMNN), una Bidirectional LSTM (BiLSTMNN) e una Bidirectional Gated Recurrent Unit (BiGRUNN), oltre che con i modelli proprietari di Ericsson. Le prestazioni sono valutate attraverso metriche standard quali l’errore quadratico medio normalizzato (NMSE) e il rapporto di dispersione sul canale adiacente (ACLR). I risultati dimostrano che il modello basato su Transformer offre prestazioni superiori rispetto alle altre reti analizzate e risultano comparabili a quelle dei modelli DPD di Ericsson. Tuttavia, la complessità computazionale del modello risulta ancora eccessiva per un’implementazione efficiente su FPGA, individuando così un rilevante ambito di ricerca futura.
Transformer DPD for non-linear power amplifiers in multi-carrier signals
BARONI, TOMMASO
2024/2025
Abstract
This thesis investigates the application of the Transformer Neural Network for digital pre-distortion (DPD) to counteract the nonlinearities of power amplifiers in multi-carrier signal scenarios. Building upon previous work, which tested the Transformer architecture on a single-carrier signal and demonstrated its superiority over other state-of-the-art models, this study extends the exploration to more complex and realistic multi-carrier environments. The research investigates and proposes improvements and modifications to the Transformer architecture and to the input features used to train the network, and assesses its capabilities across diverse signal conditions. The proposed model is compared against other neural network models, specifically a Long Short-Term Memory Neural Network (LSTMNN), a Bidirectional LSTM Neural Network (BiLSTMNN) and a Bidirectional Gated Recurrent Unit Neural Network (BiGRUNN), known for their effectiveness in DPD modeling. Additionally, the model has been evaluated alongside Ericsson’s proprietary DPD models. Evaluative metrics include normalized mean square error (NMSE) and adjacent channel leakage ratio (ACLR). The Transformer-based model shows superior performance when compared to the other neural networks, and also comparable performance to Ericsson's DPD. However, the model's complexity is still too high for practical FPGA implementation. This aspect is identified as a significant area for future research and optimization, although it falls outside the scope of this thesis.File | Dimensione | Formato | |
---|---|---|---|
2025_03_Baroni_Tesi.pdf
accessibile in internet per tutti
Dimensione
19.96 MB
Formato
Adobe PDF
|
19.96 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/234590