Contaminated Multiband Signal Identification Via Deep Learning

Multiband signals, whose active frequencies lie within continuous intervals, arise in a wide range of applications like radar imaging. In this paper, given limited and varying-length time-domain samples of a contaminated multiband signal, we propose novel deep networks to estimate the number of bands and locate the bands’ centers. A multiband signal representation model, which combines the long short-term memory (LSTM) and convolutional neural network, is trained to map varying-length observed samples to a frequency spectrum representation. A counting model then counts the number of bands based on the estimated spectrum. Combining the spectrum representation and estimated number of bands, the bands’ centers can be recovered efficiently and automatically. Numerical experiments demonstrate that the proposed method is very effective and can leverage extended samples for better performance. Moreover, it outperforms other deep architectures for line spectral estimation at different noise levels and is much faster than an atomic norm-based method.