Selection of compressed training data for RF power amplifier behavioral modeling

Abstract

Abstract-In this paper, we present an algorithm which uses the probability information of the input signal to inform the selection of a compressed training dataset for RF PA behavioural model extraction. The proposed algorithm can dramatically reduce the number of training samples. The accuracy of this algorithm is validated by extraction of behavioural models using a large dataset of consecutive samples and a reduced training dataset determined using the proposed algorithm. A noticeable reduction in computational complexity and faster execution time is achieved with the new approach.

Digital predistortion based on sample selection with memory effect

“…A solution for sampling based on the probability density distribution of the signal intensity is proposed in Reference 7. First, the classification statistics are performed based on the signal intensity distribution of the complete input signal to obtain the number of occurrences in each amplitude block.…”

“…The samples are first classified according to their amplitudes, and then a certain number of sample points of each amplitude category are selected randomly. In Reference 7, the authors introduced a method based on signal strength probability density distribution, in which the distribution of the target histogram is the same as the probability density distribution of the original signal. Wang et al proposed a histogram sampling method that considers AM/AM characteristics 8 and a mesh‐selecting (MeS) method, 9 which can be regarded as a two‐dimensional histogram‐based method with full consideration of the probability density distribution of the in‐phase and quadrature channels.…”

Early stopping criteria for adaptive training of dynamic nonlinear behavioural models

“…Under certain operating conditions PAs exhibit dynamic and non linear operation. As such they may be closely modelled using non linear filters [4], [5], due to their similarity in behaviour. In order to model a nonlinear PA with memory The operation of a digital filter is to take input signal samples x(n), combine the current sample and previous samples, weight all of them and sum the resulting products.…”

Generalised digital predistortion of RF power amplifiers with low-rate feedback signal

“…Based on a theory of the information carried by training samples used in LS which is introduced in [5], equation (1) indicates the relationship between the determined equation derived from LS and the joint probability information of the training samples,…”

“…However, different number of training samples with respect to different sampling rate are used in the simulation which leads to a degeneration of identification accuracy. This paper analyzes this multi-rate identification method from the view of information carried by training sample [5]. It reveals that by utilizing the probability property of modulated signal, a generalized DPD technique with any time-series based models can be directly performed with any sampling-rate of ADC without additional complexity.…”