Weight-Variable Scattering Convolution Networks and Its Application in Electromagnetic Signal Classification

Deep learning is an important support for the development of cognitive communication in the cognitive Internet of Things (IoT). Deep convolution neural networks have powerful functional expression and feature extraction capabilities. Mallat et al. proposed a convolutional network model with strict mathematical theory support and excellent feature extraction ability in 2012, i.e., wavelet scattering convolution networks, which is widely used in audio and image classification. In this paper, we improve the wavelet scattering convolution networks and propose construct weight-variable scattering convolution networks by combining the scattering network and the deep convolution neural network organically. We select a variety of wavelet filters for filtering operations and add 1*1 convolution layers to enable the features extracted by different types of wavelet filters to be combined optimally. In order to verify the effectiveness of the proposed network structure, we conduct experiments on classification of typical electromagnetic signals. The experimental results show that the signal classification accuracy obtained by the weight-variable scattering convolution networks is significantly better than that of the traditional wavelet scattering convolution network, and the computational complexity is greatly reduced while the classification accuracy is close to the deep convolution neural network.