Abstract: Signal modulation recognition is the core of cognitive radio and spectrum sensing. With the rapid development and application of deep learning technology in recent years, multi-modal deep learning has become the mainstream of multi-modal machine learning. However, its usage in communication systems has not been well explored. This paper proposes a signal contour stellar images domain recognition method based on deep learning (DL) to achieve the problem of low recognition accuracy under low signal-to-noise ratio. A signal I/Q waveform domain recognition method based on deep complex-valued neural network is proposed to extract the amplitude and phase features of signals to achieve high-precision and high-robustness recognition of multiple signals. A multi-modal deep learning method is proposed to fuse image features, amplitudes, and phase features extracted by complex-valued neural networks to further improve the recognition accuracy and robustness of signals. Finally, the simulation results show the superiority of the scheme and prove that the scheme utilizes the complementarity between signal multi-modalities, removes the redundancy between the modes, and realizes the deep intelligent extraction of signal features, which can lead to a better signal recognition effect.

Key words: signal modulation recognition, multi-modal fusion, contour stellar images, complex-valued networks