Modulation Signal Recognition Based on Feature Fusion and Residual CNN in Fractal Noise Background

Abstract

In this paper, a modulated signal recognition method based on feature fusion and ResCNN (FF-ResCNN) under generalized fractal noise background is proposed. Firstly, in the feature extraction and fusion part, considering the robustness, anti-noise and interpretability of the features, the fusion feature vectors of generalized fractal spectrum, instantaneous feature and high-order cumulant feature are constructed. Secondly, in order to suppress the possible gradient and degradation problems, the ResCNN model is used to train and classify the fusion features when the CNN model is used to further collect the deep-level abstract features. Finally, feature extraction, feature fusion and classification experiments are carried out for nine modulated signals under two noise types (GWN and fractal noise). Experimental results show that in the case of generalized fractal noise, when SNR=0dB, the recognition accuracy of signal modulation type reaches 95.42%, which verifies the superiority of the proposed method. At the same time, the simulation analysis is carried out for different feature fusion schemes, which provides a valuable reference for feature combination optimization under different noise environment.