Dualformer Enhances Blind Communication Signal Analysis with Dual-Channel Networks

This research introduces Dualformer, a new Transformer-based architecture built upon the concept of a dual-channel neural network (DualNN). DualNN is designed to efficiently process complex-valued signals, which are common in communication systems, by sharing network parameters between the in-phase (I) and quadrature (Q) channels. This parameter sharing is theoretically shown to reduce generalization error while maintaining the model's expressive capacity.Dualformer specifically segments input signals into patch-level tokens, allowing it to capture multi-granularity features. This design contributes to its robust performance across various blind signal analysis tasks. The paper highlights its application in automatic modulation recognition (AMR), signal scheme recognition (SSR), and signal structure parsing (SSP).Extensive experiments comparing Dualformer against three Transformer-based baselines and four conventional deep learning approaches consistently demonstrate its superior performance across these tasks. Furthermore, the modular nature of DualNN suggests its applicability extends to other unsupervised and weakly supervised complex-valued signal processing challenges, such as blind source separation and low-SNR spectrum sensing.