A neural network based on-line adaptive predistorter for power amplifier

In this paper, we present a real-time linearizing technique based on Real Valued Tapped Delay Neural Network (RVTDNN) for base band signal predistortion of Power Amplifier (PA). The proposed architecture is suitable for adaptive linearizing of PAs with memory effects. Instead of using indirect learning architecture, we propose a data on-line adaptive predistortion to ensure a continuous adaptation without interrupting the transmitting process. With the proposed architecture, a reliable transmitting process is permanently ensured. The compensation is global including a memory non-linear PA, the modulator-demodulator, and A/D and D/A converter imperfections. The adaptation algorithm minimizes a given cost-function, considered as the mean square error (MSE) between the input Cartesian (I_IN, Q_IN) components and those of the PA output divided by the maximum realizable linear gain. About 30dBc in ACPR improvement is achieved with quick convergence and good stability.