Intelligent Blockage Recognition using Cellular mmWave Beamforming Data: Feasibility Study

TL;DR

This study explores the feasibility of using cellular mmWave beamforming data for blockage recognition in 6G networks, highlighting challenges like domain shift and data volume, and suggesting strategies for effective deep learning-based sensing.

Contribution

It demonstrates the potential of using Doppler Frequency Spectrum data for blockage recognition and provides insights into data handling and model training for JCAS applications.

Findings

Domain shift affects model performance and must be addressed.

Diluting DFS data reduces inference accuracy.

Small datasets are insufficient for reliable inference.

Abstract

Joint Communication and Sensing (JCAS) is envisioned for 6G cellular networks, where sensing the operation environment, especially in presence of humans, is as important as the high-speed wireless connectivity. Sensing, and subsequently recognizing blockage types, is an initial step towards signal blockage avoidance. In this context, we investigate the feasibility of using human motion recognition as a surrogate task for blockage type recognition through a set of hypothesis validation experiments using both qualitative and quantitative analysis (visual inspection and hyperparameter tuning of deep learning (DL) models, respectively). A surrogate task is useful for DL model testing and/or pre-training, thereby requiring a low amount of data to be collected from the eventual JCAS environment. Therefore, we collect and use a small dataset from a 26 GHz cellular multi-user communication device with hybrid beamforming. The data is converted into Doppler Frequency Spectrum (DFS) and used for hypothesis validations. Our research shows that (i) the presence of domain shift between data used for learning and inference requires use of DL models that can successfully handle it, (ii) DFS input data dilution to increase dataset volume should be avoided, (iii) a small volume of input data is not enough for reasonable inference performance, (iv) higher sensing resolution, causing lower sensitivity, should be handled by doing more activities/gestures per frame and lowering sampling rate, and (v) a higher reported sampling rate to STFT during pre-processing may increase performance, but should always be tested on a per learning task basis.