Answers are grounded in this paper and cite their sources.
To address the performance degradation of existing modulation recognition methods in low signal-to-noise ratio (SNR) environments and the inefficiency of feature fusion, this paper proposes a dynamic SNR-adaptive multi-feature neural network architecture. Specifically, an EnhancedSNRBlock module is designed to achieve environment-aware feature enhancement. By dynamically adjusting feature weights according to real-time SNR conditions, this module enables the model to adaptively emphasize informative features, thereby improving recognition robustness in noisy environments. Furthermore, a dual-branch heterogeneous convolutional structure extracts and fuses time-frequency features with statistical characteristics, significantly improving representation capability for overlapping signals. Experiments show that on the RadioML2016.10a dataset, the recognition accuracy of this method reaches 53.41% at -8dB and 78.45% at -4dB, achieving good recognition results in low SNR environments.
No takes yet. Share an insight, caveat, or question.
Xing et al. (2025) studied this question.