Classification-Based Transfer Learning for Blind Adaptive Receiver Beamforming

Abstract

Adaptive receiver beamforming processors typically require expert design and can be limited by their convergence rate in data-starved applications. In this paper, we present a new type of machine learning beamformer using classification-based transfer learning (CBTL) to alleviate these limitations. The architecture consists of a pre-trained signal classifier, in our case a convolutional neural network, prepended by a beamforming layer. Narrowband beamforming weights are optimized by minimizing the classification loss, in turn nulling interference and amplifying a signal of interest (SOI). There are no requirements for calibration of the array, synchronization to the SOI, or training data modulated by the SOI. We describe the CBTL beamformer and demonstrate its effectiveness using several modulated signals. Simulated performance was compared to two well-established methods for blind source separation, and we achieved average signal-to-interference-plus-noise ratio gains of 3 to 9 dB when fewer than 100 samples were available from a 4-element array. The technique shows promise for applications where there is limited prior knowledge and few samples are available for beamformer estimation.