NeuNett - Neural Networks for Low-Rate Predistortion

In upcoming 4G and future 5G systems the bandwidth of the transmitted signals is significantly increased compared to current standards. Using carrier aggregation, multi-band, and multi-standard transmission the power amplifier becomes a nonlinear system with substantial memory effects and time-varying behavior. In order to compensate for the nonlinear impairments of the power amplifier, the transmitter has to be linearized adaptively. To perform the linearization, the output signal to the antenna has to be evaluated in some way. Usually, the signal is down converted and sampled by an AD converter and evaluated digitally. However, using traditional approaches, the sampling requirements on the AD converter are immense. A transmission bandwidth of 100 MHz requires sampling rates of up to 800MHz and more to satisfactorily cover the bandwidth due to the nonlinear effects. In this way, the AD converter becomes a showstopper regarding power consumption of the system. From a perspective of system identification and learning, it seems to be ridiculous to use this high amount of data per second for identifying just 10-30 parameters of a system. Therefore, we will investigate new learning paradigms to significantly decrease the requirements on the AD converter.