A Noise-Robust Modulation Signal Classification Method Based on Continuous Wavelet Transform

The feature extraction of Automatic Modulation Classification (AMC) is difficult, and the classification performance is poor, particularly for low SNRs and fading channels. To solve these problems, a combinatorial model based on Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM) is proposed in this paper. This method transforms the time-domain signals to the two-dimensional time-frequency domain samples by Continuous Wavelet Transform (CWT). It then extracts in-depth features through the CNN-LSTM model to conduct classification on modulation signals. Analyses show that the proposed model yields an classification accuracy higher than 90% at varying SNR conditions ranging from -6dB to 12dB. Compared with traditional algorithms, this method effectively improves the problem of poor classification performance for low SNRs and possesses better robustness.