Experimental Evaluation of Multiantenna Spectrum Sensing: Unsupervised Learning Approach

Cognitive radio (CR) networks rely on spectrum sensing algorithms to determine the availability of unused frequency bands. In this regard, several statistical tests have been formulated while considering presumed models. Nevertheless, their detection performance depend on the accuracy of these models, and learning-based detectors have shown to overcome these drawbacks by adapting themselves to the uncertain radio environment. In this work, we aim to evaluate an unsupervised spectrum sensing employing a multi-antenna testbed. It consists of a training stage by resorting to clustering algorithms to determine the label of the data, and eventually the detection performance is assessed during a testing stage employing an artificial neural network (ANN). Exhaustive experimental eval-uations on a software-defined radio (SDR) platform are carried out to assess its capacity to adapt itself to a real environment facing real aspects of the sensed signal, multipath channel, and user mobility. The experimental results show significant gains in comparison to model-based detectors, and reveals the feasibility to avoid significant amount of labeled data for training, thus being suitable for practical CR applications.