RF Power Amplifier Behavioral Modeling using a Globally Recurrent Neural Network

In this paper it is shown that a globally recurrent time delay neural network can accurately model a nonlinear RF power amplifier having significant memory. The recognized difficulty of training a recurrent neural network is overcome by reducing it initially to a feed forward network, training that network, and then using the weights established by this training sequence in a restructured recurrent network. The training of the recurrent network thus reduces to the training of a feed forward network and a simple restructuring. The required maximum input delay is established by examination of the temporal profile of the energy contained in the amplifier impulse response. The model was successfully trained with an RF pass band time domain multi-sine signal and subsequently validated with another multi-sine signal composed of different sine components at different amplitudes. A second model trained with a wider bandwidth multi-sine was successfully validated with a W-CDMA signal