ML-Based Classification of Device Environment Using Wi-Fi and Cellular Signal Measurements

• Published in: IEEE access Vol. 10; pp. 29461 - 29472
• Main Authors: Ramamurthy, Arun, Sathya, Vanlin, Rochman, Muhammad Iqbal, Ghosh, Monisha
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 5G mobile communication; Access control; Automatic control; Automatic frequency control; Classification; Datasets; Devices; Frequency control; Global Positioning System; indoor; Indoor environments; Machine learning; outdoor; Performance evaluation; Radio frequency; Satellite broadcasting; Wi-Fi; Wireless communication; Wireless fidelity
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3158056

Future spectrum sharing rules very likely will be based on device environment: indoors or outdoors. For example, the 6 GHz rules created different power regimes for unlicensed devices to protect incumbents: "indoor" devices, subject to lower transmit powers but not required to access an Automatic Frequency Control database to obtain permission to use a channel, and "outdoor" devices, allowed to transmit at higher power but required to do so to determine channel availability. However, since there are no reliable means of determining if a wireless device is indoors or outdoors, other restrictions were mandated: reduced power for client devices and indoor access points that cannot be battery powered, have detachable antennas or be weatherized. These constraints lead to sub-optimal spectrum usage and potential for misuse. Hence, there is a need for robust identification of device environments to enable spectrum sharing. In this paper we study automatic indoor/outdoor classification based on the radio frequency (RF) environment experienced by a device. Using a custom Android app, we first create a labeled data set of a number of parameters of Wi-Fi and cellular signals in various indoor and outdoor environments, and then evaluate the classification performance of various machine learning (ML) models on this data set. We find that tree-based ensemble ML models can achieve greater than 99% test accuracy and F1-Score, thus allowing devices to self-identify their environment and adapt their transmit power accordingly.