Joint Supervised and Unsupervised Machine Learning for Spectrum Sensing

As the Internet of Things (IoT) applications continue to experience exponential growth, the resulting congestion in frequency spectrum highlights the necessity for innovative spectrum monitoring techniques. Spectrum sensing plays a crucial role in cognitive radio systems by facilitating the efficient utilization of spectrum through the detection and utilization of unused frequency bands. This work presents a novel methodology for detecting radio frames from a noisy spectrum. We implement joint machine learning comprising a supervised Semantic Segmentation neural network and unsupervised Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to detect IoT packets from the spectrogram images. The proposed method achieves 95.4% mean accuracy and 93% mean intersection-overunion (IoU) at low Signal-to-Noise Ratio (SNR) level. Moreover, the proposed method outperforms the traditional threshold detection methodologies as demonstrated by the Receiver Operating Characteristics (ROC).