| Along with the fast development of wireless technology,the modulation types and the channel environment of signals are increasingly complicated.In military and civil fields,accurate recognition of modulation signals is becoming more crucial.Traditional methods of modulation recognition based on decision theory and feature extraction require the assistance of prior information,artificial screening features,etc.With the development of deep learning,the neural network-based signal modulation recognition methods are widely adopted.Currently,most deep learning methods are highly dependent on data,and they have a poor ability to extract interpretable information and knowledge from data.Therefore,when designing neural networks,it is necessary to embed professional information and physical laws related to a given prediction task into the architecture to improve recognition stability.In addition,noise and other interference have a serious impact on the recognition of modulation signals.How to reduce the effect of noise and increase the precision of signal recognition has become an important research topic in the field of modulation recognition.To solve the above problems,this paper combines wavelet signal processing with deep learning to carry out research on signal modulation pattern recognition based on invariant scattering convolution network.First of all,this paper summarizes the basic knowledge of signal modulation principle and neural network theories,analyzes the structure of the invariant scattering convolution network built based on wavelet theory to provide a theoretical basis for the combination of invariant scattering convolution network and deep learning algorithm.Then,a signal modulation recognition algorithm based on the invariant scattering convolution network is designed to improve the performance of signal recognition.In this algorithm,the scattering coefficient of the output of the invariant scattering convolution network is taken as the characteristic representation of signals,and a neural network classifier is used to classify different signals.Given the characteristics of the modulation signals and the feature representation provided by the invariant scattering convolution network,the applicability of invariant scattering convolution network in modulation recognition is analyzed.According to the characteristics of the scattering coefficient of the convolution network of the input-modulation signal,a feature fusion parallel neural network classifier is designed.The classifier includes long short-term memory,sub-spectrum normalization,and residual modules,and combines the advantages of convolutional neural network and recurrent neural network in spatial and temporal feature extraction.To reduce noise interference,wavelet threshold denoising is integrated into the invariant scattering convolution network.Finally,a multi-task signal modulation recognition algorithm based on signal-to-noise ratio interval division is designed to improve the generalization ability of the modulation recognition model under different signal-to-noise ratio conditions.The algorithm treats signal recognition under different signal-to-noise ratio intervals as independent tasks,reducing the variation range of signal-to-noise ratio and minimizing the influence of noise.Then,a recognition method based on a parallel residual neural network identifies the signal-to-noise ratio interval to determine the task of the signal.Next,the signal is input into the invariant scattering convolution network to obtain the scattering coefficient to extract features of the data,and according to the recognition results of the signal-to-noise ratio interval,classifiers with corresponding task parameters are used independently to classify and recognize the signal.Experimental results indicate that the proposed algorithm can effectively reduce noise disturbance and increase the precision of signal recognition,which holds great reference value for the research and development of modulation pattern recognition methods. |
