Deep Convolutional Neural Network with Wavelet Decomposition for Automatic Modulation Classification

In cognitive radio, signal recognition is an important technology and modulation recognition plays a key role in it. With the development of artificial intelligence, deep learning algorithms applied in automatic modulation recognition have developed quickly, whereas they usually depend on a large number of labeled samples for training. Few samples directly affect the network convergence, which will lead to network overfitting and cannot achieve good results. The loss of prior information makes feature extraction more difficult. In this paper, we propose a wavelet-decomposition-based algorithm for modulation recognition to solve the small sample size problem. To obtain rich information relatively, we adopt the wavelet function to analyze signals from multiple scales, extract the time domain features of the signals without any prior information by Residual Blocks, and fuse these features by attention blocks. In order to reduce the risk of overfitting, we use the Batch Normalization layer, Global Average Pooling, and Additional Random Noise to improve robustness. The proposed algorithm has a classification accuracy of 95.6% with only 20 samples per category when the SNR is 20dB, which outstrips other classical methods under the same condition.