Deep Learning Based Sparse Array Adaptive Beamformer Design Under Class Imbalance

Deep learning has demonstrated significant potential in the field of array processing. Sparse array reconfigurability can play a vital role in cognitive sensing within dynamic radio frequency (RF) environments. For practical implementation in such environments, it is essential to minimizes excessive switching and sparse array reconfiguration. In this paper, we design sparse arrays optimized for broader angular sectors instead of targeting a single specific angle achieving optimal beamforming that maximizes the signal-to-interference-plus-noise ratio (SINR). To further improve sparse array classification rates and SINR, we incorporate class imbalance handling techniques into a convolutional neural network framework. To address this, we employ a class imbalance strategy of oversampling less frequent configurations to improve classification accuracy. We evaluate the impact of this approach, along with the angular sectorbased design, on classification accuracy and achievable SINR. Our results demonstrate that the proposed strategies achieve higher SINR performance and improved classification accuracy compared to existing designs.