Research On Kernel Initialization Of Convolutional Neural Network In Modulation Recognition Application

Traditional modulation recognition algorithms are mostly feature based classifiers or likelihood based classifiers.With the rapid development of deep learning in fields such as computer vision,researchers began seeking its application in modulation recognition.However,up till now,most of the modulation recognition algorithms based on deep learning are only mimicking the algorithm architecture developed for computer vision applications.No fundamental innovation has been brought to the design of core components for the modulation recognition.Therefore,with the knowledge of the distribution characteristics of modulated signals,deep learning algorithms based on custom convolution kernel initialization are proposed.Compared with the random convolution kernel initialization,the custom initialization of the convolution kernel could guides the optimization of the convolution kernel in a direction that is more in accordance with the general distribution of the modulated signal,which may shorten the training time,avoid overfitting and improve the recognition accuracy.In this thesis,the proposed convolutional neural network adopts a double convolutional layer structure.When customizing the convolution kernel,The first convolutional layer is used to extract the local distribution character around the modulation symbols,The second convolutional layer is used to extract the symbol mapping character for different modulations.In this thesis,the initialization method of the convolution kernel is developed under the additive white Gaussian noise and fading channel.In the additive white Gaussian noise channel,the initial value of the convolution kernel in the first convolution layer is calculated according to the distribution of the in-phase quadrature components of the signal,thereby identify the symbol location.In order to make the model more suitable for practical application scenarios,this paper considers the situation that the signal has carrier phase or carrier frequency offset in fading channel.When the signal has a carrier phase offset,the convolution kernel is redesigned in the first convolution layer according to the distribution of each symbol in each modulation,which solves the shortage of designing the convolution kernel in the additive white Gaussian noise channel.Different degrees of phase offset are adapted through convolution kernel diversification in the second convolution layer.When the signal has a carrier frequency offset,the problem is solved by representing the signal in phase and amplitude and as well as designing the matching convolution kernels in the first convolutional layer.In this paper,the simulated test results show that the proposed scheme is able to achieve superior classification accuracy,while maintaining less computational resources.In the additive white Gaussian noise channel,using custom initialization for the model’s convolution kernel reduces overfitting and training time at some signal-to-noise ratios compared to using random initialization.In the fading channel,when there are carrier phase or frequency offset,the model using the custom initialization convolution kernel has strong robustness.