MmWave beam path blockage prevention through codebook value prediction under domain shift

Abstract

Use of millimeter and terahertz spectra for communication is very sensitive to obstacles blocking signal beam paths. Beam angle codebook values can be adapted to control beam operation angles for blockage prevention, but this requires prediction of beam paths that are blocked. The performance of the prediction pipeline may be affected by domain factors such as physical characteristics of an operation environment and a specific blocker. This can be illustrated by artificially introducing domain factor shifts between training and test data subsets where a specific domain factor is left out of the training subset. Our experiments reveal significant performance drops in the blockage prediction performance on left-out test subset folds that contain all the samples of a specific domain factor. Thus, the prediction pipeline must employ effective domain shift mitigation techniques to attain consistent prediction performance in different domains. Pipeline performance should be supported by logical input data to prediction causation. We quantify causation by means of Shapley importance values with input regions attributable to signal aspects such as linear array antennas. Shapley importance results show high neural network prediction confidence value affection for amplitude variance and a limited set of subsequent fast-time blocks. Random inductively biased convolutions affection differs in a limited number of spatially separated antennas causing affection. Equally high prediction confidence value affection is assumed for iterative component search due to internal extraction mechanics and time complexity increases when zero-masking input data regions. We link equally high assumed prediction confidence value affection for iterative component search to highly logical IF signal to prediction causation. The affection of amplitude variance and a limited set of subsequent fast-time blocks shows weaker causation, still considered logical if the neural network can separate observations in representation distributions for varying distance and angle combination sets. The random inductively biased convolutions show illogical causation. They rely on direct IF signal features. Affection by a limited number of antennas indicates reliance on features with inadequate separation ability along angles at appropriate resolution.