IoT Technology Recognition Using Deep Clustering

Hyeongyun Kim, Adnan Shahid, Jaron Fontaine, Eli De Poorter, Ingrid Moerman, Haewoon Nam

Abstract: This article proposes a fully unsupervised technology recognition (TR) method using deep clustering for identifying wireless technologies to learn from raw data without requiring manual label annotations. Identifying and recognizing wireless technologies is important to realize effective spectrum management. Recent scientific publications utilized supervised deep learning with great success to train AI models to recognize different wireless technologies based on labeled RF signal data sets. However, assigning labels to wireless technology signal data sets for supervised deep learning is time consuming and may not always be practical. To remedy this issue, the proposed method combines an autoencoder and clustering layers, extended with a gaussian mixture model, to learn low-dimensional salient features from raw in-phase and quadrature (IQ) data and cluster them into distinct wireless technologies. The optimal input dimensions and performance of the method in different signal-to-noise ratio conditions are analyzed for a set of several different low-power wide area network (LPWAN) technologies: Sigfox, LoRa, IEEE 802.11ah, and IEEE 802.15.4g. Extensive simulations show that the proposed method for wireless TR exhibits a recognition accuracy exceeding 90%, surpassing the performance of the state-of-the-art deep clustering algorithms by 4% to 5% in terms of accuracy.