Machine-Learning Assisted Digital Predistortion Using Feedback via Dual-Polarized Antenna Arrays

This paper proposes a digital predistortion (DPD) method that is enhanced by machine learning. Instead of relying on feedback data from over-the-air (OTA) measurements or a dedicated embedded feedback path, we capture a set of feedback data via a secondary transceiver connected to a dual-polarized antenna array. An artificial neural network is employed to generate the OTA-approximated data from the captured feedback data, which may contain additional noise or extra distortion due to the limited dynamic range of the feedback path. Experimental measurements are conducted using a 28-GHz 256-element dual-polarized phased-array test board. The proposed DPD method achieves a 7.6 dB improvement in adjacent-channel-leakage ratio. That is comparable to the 8.0 dB improvement obtained using OTA data.