Transformer-based Modulation Recognition Algorithm with Multi-domain Feature Fusion

Abstract

In low signal-to-noise ratio environments, the performance of modulation recognition is severely affected by noise, making it hard to boost the overall recognition rate. What’s more, existing modulation signal recognition algorithms based on a single feature can’t adapt to the different signal characteristics under varying signal-to-noise ratios. To solve these problems, this paper proposes a Transformer-based modulation recognition algorithm with multi-domain feature fusion. This algorithm extracts three types of signal features at the same time: I/Q time-domain features, constellation diagram features and instantaneous features. It uses the Transformer self-attention mechanism to automatically learn the contribution of different features under different signal quality levels and generate weight scores for each feature. Each score represents how important the corresponding feature is to the final classification decision. The original generated scores are normalized into a probability distribution via Softmax to get the final attention weights, which are then used to perform a weighted summation of the three feature sets—each feature set is multiplied by its corresponding attention weight. The fused final feature set is then used for classification. Experimental results show that the method proposed in this paper has obvious advantages in the low signal-to-noise ratio range of -10dB to -4dB. It outperforms the LSTM model (with a close overall recognition rate) by 6.62%, and its overall recognition rate is 12.13%, 6.81%, 1.45% and 5.19% higher than that of CNN2, ResNet, LSTM and MCformer respectively.