Multi-Signal Classification Using Deep Learning and Sparse Arrays

In uncoordinated spectrum scenarios, such as shared spectrum, accurate spectral analysis is challenging if devices are interfering with one another. This paper examines wireless standard classification for spatially separated emitters that are interfering in time and frequency. The proposed approach involves estimating the number of interfering signals and their directions of arrival, spatial isolation, and then wireless standard classification using a deep learning-based convolutional neural network (CNN). The CNN is trained to classify commercial wireless standards, including 4G LTE, 5G, IEEE 802.11ax, Bluetooth Low Energy 5.0, and Narrowband Internet-of-Things, using a synthetic dataset impaired with typical channel and receiver impairments, including AWGN and Rayleigh/Rician fading. When a single emitter is transmitting, spatial filtering using a four element uniform linear array (ULA) increases the classification accuracy of low SNR signals by up to 26% compared to a single antenna system. When multiple emitters are transmitting, especially when there are three or more, a four-element ULA cannot spatially filter well enough to yield high classification accuracies. In these cases, we show that switching to a sparse array can improve classification accuracy without adding more array elements. Sparse arrays can improve classification accuracy by up to 33% compared to ULAs in scenarios with multiple emitters or closely spaced emitters.