A Deep Learning Method for Open-Set Transmission Mode Recognition Based on Multiresolution Time-Frequency Spectrograms

Abstract

Transmission mode recognition (TMR) is a wireless signal recognition technology that plays an essential role in high frequency (HF) communication reconnaissance systems. However, in noncooperative communications, TMR remains a challenging task due to the complexity of HF environments, and the presence of unknown transmission modes may seriously affect recognition performance. To address these challenges, we focus on the open-set TMR problem and propose a deep learning-based open-set TMR framework that integrates signal preprocessing, feature extraction, and feature template matching for recognition. The core of this framework is a novel neural network called multiresolution feature fusion network (MFFN), which takes multiresolution time-frequency spectrograms (MTFS) of the received signals as input. Specifically, the MFFN takes full advantage of multiresolution feature fusion (MFF) and deep metric learning (DML) technologies to learn discriminative feature representations of different transmission modes for open-set recognition. In addition, two benchmark TMR datasets, the active transmission mode (ATM) dataset and the civilian transmission mode (CTM) dataset, are established in this paper to evaluate the effectiveness of various methods. The experimental results indicate that our proposed method outperforms several baseline methods in open-set recognition performance and exhibits strong generalization ability for unknown transmission modes.