TF-Transformer: Temporal-Frequency Transformer for OFDM Signal Recognition

The automatic signal modulation recognition of Orthogonal Frequency Division Multiplexing (OFDM) multi-carrier signals holds significant research importance and application value in areas such as Wi-Fi, 4G, and 5G communications. However, traditional automatic modulation recognition methods perform poorly in multi-carrier communication systems, especially in complex real-world channel environments. In this study, we first create an OFDM multi-carrier signal dataset based on real channel data, named OFDM2024, which includes data on various modulation types under multiple signal-to-noise (SNR) ratio conditions. Secondly, we propose a modulation recognition method for OFDM signals based on a temporal-frequency attention mechanism using the Transformer model, called TF-Transformer. First, we design temporal feature extraction and frequency feature extraction modules to extract temporal and frequency domain features from the signal data. Next, we develop a feature fusion module to integrate the obtained features. Finally, we different the modulation types of the signals based on the fused features. Experimental results indicate that TF-Transformer significantly outperforms traditional signal recognition algorithms such as ViT, CNN, and ResNet in terms of performance.