TriFusion-Lite: A Temporal-Frequency-Phase Fusion Lightweight Network for Modulation Recognition

Abstract

Automatic Modulation Recognition (AMR) is an essential technology for modern wireless communication systems. However, existing deep learning models are often computationally complex and frequently overlook the phase relationships within in-phase/quadrature (I/Q) signals, thereby hindering their deployment on resource-constrained devices. This paper introduces TriFusion-Lite, a lightweight multi-stream deep learning architecture designed to optimize both efficiency and accuracy in AMR. The proposed framework begins with a tailored preprocessing pipeline that compresses the input signal while enriching its representation with robust statistical features. A novel four-stream parallel network then processes the enhanced signal: a complex-valued convolutional stream to preserve phase integrity, two parallel 1D convolutional streams for independent I/Q channel analysis, and a Short-Time Fourier Transform (STFT) stream to capture spectral characteristics. A hierarchical fusion mechanism progressively integrates these multi-domain features for final classification. Comprehensive evaluations on benchmark datasets demonstrate the effectiveness and competitive performance of the proposed approach. The experiments confirm the effectiveness of the compression stage and analyze its performance across various compression levels. Furthermore, the proposed method achieves competitive results compared with state-of-the-art approaches while maintaining a favorable balance between classification performance and computational efficiency, making it a promising solution for AMR applications on edge devices. The source code of the proposed framework is publicly available at https://github.com/sansi34jun/TriFusion-Lite