Millimetric Diagnosis: Machine Learning Based Network Analysis for mm-Wave Communication

Abstract

Troubleshooting millimeter-wave (mm-wave) wireless networks is complex due to the directionality of the communication. Issues such as deafness, misaligned antennas, or blockage may severely impact network performance, and identifying them is crucial to improve network deployments. To this end, access to lower-layer information is important. However, commercial off-the-shelf mm-wave wireless devices typically do not provide such information. Even if they would, detecting effects such as deafness based on information of a single node that forms part of the network is typically hard. In this paper, we present the design and evaluation of an external sniffing device that can infer the aforementioned performance issues only using narrowband physical layer energy traces. Our sniffer does not need to decode any data, resulting in a simple but effective approach which also preserves privacy and works on encrypted networks. Our key contribution is a machine learning framework which enables automated energy trace analysis while coping with the non-stationarity of the traces. We evaluate its performance in practice using off-the-shelf wireless devices operating in the 60 GHz band. Our results show that the above framework correctly infers physical layer events in virtually all cases, thus providing valuable information to troubleshoot issues in mm-wave networks.