人工智能
计算机科学
模式识别(心理学)
规范化(社会学)
自动目标识别
计算机视觉
语音识别
噪音(视频)
信号处理
特征提取
作者
Wendi Xiao,Yi Gong,Yuan Zeng
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-05-01
卷期号:: 67-72
标识
DOI:10.1109/acc6g70208.2026.11605561
摘要
Recent deep learning-based automatic modulation recognition (AMR) methods have achieved remarkable recognition results when training and test radio signals are collected under the same communication conditions. However, the received signals encounter unseen impairments in complex communication environments, leading to distribution shifts between the training and test signals. Pre-trained AMR models typically exhibit poor generalization on such shifted signals. To address this, we propose an entropy-aided test-time normalization adaptation (TNA) method for AMR. TNA adapts normalization estimates using limited unlabeled signals and requires no backpropagation process. Specifically, since the entropy of test signals is strongly correlated with recognition confidence and signal distortion, an entropy-aided sample selection mechanism is designed to retain low-entropy signals for stable adaptation at test time. Then, the proposed method adapts the normalization estimates within the pre-trained AMR models via an exponential moving average that fuses the pre-trained estimates with the statistics of the selected signals. Experimental results demonstrate that TNA effectively improves the generalization performance of the pre-trained AMR model under various wireless propagation environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI