Blind multicarrier waveform recognition based on spatial-temporal learning neural networks

Abstract

Blind multicarrier waveform recognition has become a more daunting task and open problem for the current and future radio surveillance and signals interception, with the advent of new multicarrier technologies such as the state-of-the-art F-OFDM, UFMC, FBMC, OTFS, GFDM and CP-OFDM techniques. Therefore, the practical recognition scheme for multicarrier waveforms is necessary to keep up with the pace. To tackle this challenge, we propose a novel multicarrier waveform recognition framework based on Spatial Temporal-Convolutional Long Short-Term Deep Neural Network (ST-CLDNN) in the entirely blind context. The complementary information of the raw in-phase, quadrature and amplitude samples are first extracted to provide more distinguishable features. Then ST-CLDNN collects the advantages of one-dimensional convolutional and long short-term memory (LSTM) to extract high-level spatial and temporal features for the recognition task. Later, we introduce the transfer learning strategy to put the computational resource to good use and obviate the retraining from scratch for a time-varying channel. Experimental results indicate that the proposed ST-CLDNN can perform better than the traditional feature-based classifiers and existing deep learning (DL)-based neural networks, and deliver a substantial recognition performance in a time-varying multipath fading channel.