Neural network based adaptive predistortion for the linearization of nonlinear RF amplifiers

The good spectral efficiency of linear modulation techniques makes them attractive for use in high data rate digital radio systems. Unfortunately, the fluctuating envelopes of such systems combined with the nonlinear nature of the high power RF amplifiers commonly used gives rise to spectral spreading, adjacent channel interference, I-Q crosstalk, and degraded bit error rates. A possible solution to these problems is the linearization of the transmitter system by predistortion of the baseband digital signal using a nonlinear filter which is the inverse of the amplifier response. We realize the inverse filter using a backpropagation neural network. A coupler and demodulator provide a feedback path which provides the input to the neural network while the desired signal is the original waveform. The shaped waveform is passed through the network which adaptively corrects for the changing response of the high power RF amplifier. This technique allows us to correct for quite general nonlinearities, significantly reduce spectral spreading, and to improve the bit error rate.