【摘要】 针对复杂信道环境下调制识别对标注数据高度依赖的问题，提出一种基于双置信度融合机制的半监督信号调制识别方法。该方法采用基于自训练半监督学习的多通道卷积长短期深度神经网络（MCLDNN）框架，利用高阶累积量、瞬时参数与小波系数构建特征向量，通过高斯混合模型（GMM）聚类并评估聚类置信度，动态融合基于自训练的半监督学习模型置信度形成双置信度机制。在公开数据集RML2016.10b上的实验表明，该方法在0～18 dB信噪比下对7种数字调制信号的平均识别准确率达97.78%，较传统半监督方法提升4.88%，接近全监督性能；模型收敛速度提升约20%，有效解决了传统基于自训练的半监督学习中模型认知闭环导致的过拟合问题，为低标注成本场景下的调制识别提供了高效解决方案。更多还原

【Abstract】 In response to the problem of modulation identification with high dependence on labeled data in complex communication channel environments, a semi-supervised signal modulation recognition method based on dual-confidence fusion mechanism was proposed. This method adopted the multichannel convolutional long short-term deep neural network(MCLDNN) framework based on self-training semi-supervised learning, used high-order cumulants, instantaneous parameters and wavelet coefficients to construct feature vectors, clusters and evaluated clustering confidence through the Gaussian mixture model(GMM), and dynamically integrated the confidence of the semi-supervised learning model based on self-training to form the dual confidence mechanism. Experiments on the public dataset RML2016.10b shows that under the test scenarios of 0～18 dB SNR and 7 digital modulation signals, the average recognition accuracy of this method reaches 97.78%, which is 4.88% higher than the traditional semi-supervised method and is close to the full supervision performance. At the same time, the model convergence speed has improved by about 20%. This method effectively solves the overfitting problem caused by the model cognitive loop in traditional self-training based on semisupervised learning, and provides an efficient solution for modulation recognition in low-labeling cost scenarios.更多还原