Nonlinear behavioural models for RF devices using an artificial neural network technique

Tian, Mengyue 2024. Nonlinear behavioural models for RF devices using an artificial neural network technique. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (11MB) | Preview
	PDF (Cardiff University Electronic Publication Form) Restricted to Repository staff only Download (330kB) | Request a copy

Abstract

This work focuses on exploring the development of the Cardiff Model (CM) and the Artificial Neural Network (ANN) based behavioural model for reasonable model extrapolation performance while maintaining the interpolation accuracy. As an opening segment following the theories explained in the literature review, both the CM and the ANN model accuracy are verified. The potential extrapolation problems of the two modelling techniques are investigated in detail. The experiments demonstrate that the determination of the internal model parameters, such as the model order of the CM, the initial values of hidden neuron numbers, weights and biases of the ANN, are the key correlated elements to reasonable/accurate model performance. One solution to the extrapolation problem with the user-defined CM order is provided. A modified Levenberg-Marquart (LM) backpropagation training algorithm that can be utilised as part of the ANN based behavioural model is introduced. The modified LM algorithm allows the implementation of an ANN technique-based CM coefficient extractor. The invented coefficient extractor is verified with datasets acquired under ideal simulation and practical load-pull measurement scenarios from theWolfspeed 10Wpackaged device. Results prove that the novel coefficient extractor is able to provide interpolation predictions at the same level of accuracy as the conventional CM, at a Normalised Mean Square Error (NMSE) level below - 50 dB. Then, a method that combines the conventional A-B wave-based ANN behavioural model and the invented coefficient extractor for extracting high user-defined order i CM coefficients is also presented. Two sets of practical load-pull measurement data, acquired with the Wolfspeed 10W packaged device and the WIN NP12 4x25um on-wafer device, are used for method verification. The coefficients, extracted using the proposed combined method, have been proven with reasonable extrapolation ability and robustness under different measurement scenarios, maintaining the interpolation accuracy at a NMSE level below - 50 dB for the predicted output power and - 40 dB for the predicted efficiency. Another solution to the configuration of the ANN with the values of internal model parameters is also given. A direct link between the CM and the A-B wave-based ANN behavioural model is established. The established equation set enables an alternative ANN determining method. The ANN models proposed using the determination method can provide accurate prediction for the behaviour acquired from load-pull characterizations of the Wolfspeed 10W packaged GaN device simulation and a dense load-pull measurement of WIN NP12 4x75um GaN HEMT at 20 GHz, with NMSE levels lower than - 40 dB, and also guarantee a reasonable extrapolation ability for both device output power and efficiency.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1).Cardiff Model 2).Artificial Neural Network (ANN) 3).behavioural model 4). gallium nitride (GaN)5).load–pull measurement 6). power amplifier
Date of First Compliant Deposit:	7 March 2025
Last Modified:	07 Mar 2025 15:30
URI:	https://orca.cardiff.ac.uk/id/eprint/176704

Actions (repository staff only)

Edit Item