为提升低信噪比条件下雷达/ 通信频率、相位编码信号调制识别性能,降低特征提取复杂度,提出了基于深度信念网络DBN(Deep Belief Network, DBN)以及快速特征提取的调制识别方法。结合快速傅里叶累加算法FAM(FFT Accumulation Method)算法,提出了将循环谱估计图像转化为有效可识别特征向量的提取算法;设计了用于编码信号调制识别的DBN 网络训练与识别框架。仿真结果表明,文中方法较传统方法具有更低的特征提取与预处理复杂度,提取的特征在几种典型编码调制模式信号中具有明显区分,DBN 训练识别框架对雷达/ 通信编码信号调制识别均具有可行性与有效性,在低信噪比条件下对无线电编码信号有更高的识别正确率。

To improve automatic modulation recognition (AMC) performance of radar and communication frequency and phase shift keying signals in low SNR levels and decrease computational complexity of feature extraction, an AMC method combining deep belief network (DBN) and fast feature extraction is proposed. A FFT accumulation method (FAM) based feature extraction algorithm is presented to transform cyclic spectrum images to valid recognizable feature vectors. DBN training and recognition framework is specially designed for AMC of frequency and phase shift keying signals. Simulation results reveal that the proposed feature extraction algorithm has lower computational complexity and extracted features has apparent distinctions among several typical modulation types. DBN training and recognition framework is proved effective to both radar and communication signals. The recognition method we proposed is also verified to achieve better recognition accuracy in low SNR levels.

TN99

国家自然科学基金资助项目(61272333)