RiSi: Spectro-temporal RAN-agnostic Modulation Identification for OFDMA Signals

Blind modulation identification is essential for 6G's RAN-agnostic communications, which identifies the modulation type of an incompatible wireless signal without any prior knowledge. Nowadays, research on blind modulation identification relies on deep convolutional networks that deal with a received signal's raw I/Q samples, but they mostly are limited to single-carrier signal recognition thus not pragmatic for identifying spectro-temporal OFDM/OFDMA signals whose modulation varies with time and frequency. Therefore, this paper proposes RiSi, a semantic segmentation neural network designed to work on OFDMA's spectrograms, by replacing vanilla DeepLabV3+'s 2D convolutions with 'flattened' convolutions to enforce the time-frequency orthogonality constraint and to achieve the grid-like pattern of OFDMA's resource blocks, and by introducing three-channel inputs consisting of I/Q/amplitude. Then, we synthesized a realistic and effective dataset consisting of OFDMA signals with various channel impairments to train the proposed network. Moreover, we treated varying communication parameters as different domains to apply domain generalization methods, to enhance our model's adaptability to diverse communication environments. Extensive evaluation shows that RiSi's modulation identification accuracy reaches 86% averaged over four modulation types (BPSK, QPSK, 16-QAM, 64-QAM), while its domain generalization performance for unseen data has been also shown to be reliable.