A neural network approach to data predistortion with memory in digital radio systems

The authors present an algorithm to determine the coefficients of a general data-predistorter with memory for compensation of high-power amplifier (HPA) nonlinearities in digital microwave radio systems. This technique is based on modeling the predistorter-modulator-HPA system as a neural network with memory. It is shown that, by extending the optimization algorithm of back-propagation to complex signals and with neurons modeled as finite-impulse-response (FIR) filters, the proposed algorithm determines automatically the predistorter with the objective that the overall transmitter behaves as a linear system with a prescribed pulse shape. The novelty with respect to previous techniques is that in the present scheme a control on the spectrum of the signal after the HPA is exercised. This minimizes the interference between adjacent channels. The algorithm has been tested successfully in several radio systems employing quadrature amplitude modulation (QAM) signal formats, and some examples of its application are reported