基于稀疏滤波神经网络的智能调制识别

针对传统调制识别中特征提取依赖人工经验的问题，该文提出了一种基于抗噪预处理及稀疏滤波卷积神经网络的智能通信调制识别算法。该算法将调制信号的循环谱作为卷积神经网络的输入图像，并引入低秩表示算法去除循环谱图中的噪声及干扰。在有监督训练卷积神经网络之前，该文设计了一种新型的稀疏滤波准则对网络进行无监督的逐层预训练，从而提升了泛化性能。仿真表明算法在信噪比为0 dB时仍可达94.2%的识别准确率，优于传统方法及相关深度学习方法。

Intelligent Modulation Recognition Based on Neural Networks with Sparse Filtering

To overcome the disadvantage that the feature extraction in traditional automated modulation recognition is heavily dependent on manual experience, this paper proposes an intelligent modulation recognition algorithm for communication signals, which is based on antinoise processing and sparse filtering convolutional neural network (AN-SF-CNN). The cyclic spectra of modulated signals are calculated, then low-rank representation is performed on cyclic spectra to reduce noises and disturbances existed in signals. After that, before fine-tuning the convolutional neural network, we propose a sparse filtering criterion to conduct the unsupervised pre-train of the network layer-by-layer, so as to improve the generalization performance effectively. Simulation results demonstrate that the average correct classification rate of proposed method can even reach to 94.2% when the signal to noise ratio is 0 dB, it is superior to traditional methods and some popular deep learning methods.