Automatic Modulation Recognition Method for Adaptive Noise Reduction

Experiments on modulation automatic recognition combined with previous deep convolutional neural network models show that most feature extraction models have low recognition accuracy under a low signal-to-noise ratio, and the problems of more model parameters and high computational complexity are ignored while pursuing high recognition accuracy, so the research focuses on adaptive noise reduction and balancing model accuracy and computational complexity. A lightweight convolutional neural network model based on adaptive noise reduction is proposed, and a phase accumulation method based on forward propagation is proposed to enhance the feature representation of signal samples in the training concentration and an adaptive noise reduction module based on attention mechanism to reduce the noise of signal samples, and time-frequency features are extracted serially by a convolutional neural network (CNN) and gated recurrent unit (GRU), and the recognition accuracy of the method is maintained at 0dB signal-to-noise ratio on the benchmark data set, and the recognition rate is more than 87.82%, and the overall recognition accuracy is high, the model parameters are small-scale.