UA-ConstNet: Uncertainty-Aware Dual-View Fusion for Automatic Modulation Classification

Abstract

Automatic modulation classification (AMC) is difficult when the receiver can only observe a short and noisy signal burst. In such cases, a single signal representation may not provide enough information for reliable recognition. This paper presents UA-ConstNet, an uncertainty-aware dual-view network for short-burst modulation classification. The model uses both the raw in-phase/quadrature (I/Q) sequence and a constellation-density map. The raw I/Q branch extracts temporal features through residual one-dimensional convolution and squeeze-and-excitation attention, while the constellation branch learns spatial symbol-cluster patterns from the density map. The two feature vectors are projected into a shared space and combined through a feature-wise adaptive gate, which dynamically balances temporal I/Q information and spatial constellation information. An evidential output head further estimates class evidence, class probabilities, and predictive uncertainty in a single forward pass. Experiments on 5,760 synthetic bursts show that UA-ConstNet achieves 41.09% accuracy and 40.66% macro-F1. However, its ECE remains 0.141, indicating imperfect probability calibration, and high-order QAM classes, especially cross-32QAM and 64QAM, remain difficult to distinguish under short-burst and noisy conditions.