Research On Closed Set Recognition Of Convolutional Codes

With the wide application of modern digital communication systems,the importance of blind identification of channel codes parameters has become increasingly prominent.Blind identification of channel codes parameters is to identify the type and the encoding parameters of the received or intercepted data without any prior knowledge,so as to lay a foundation for subsequent channel decoding and further signal and information processing.Convolutional code is a kind of channel codes.Compared with linear block codes,convolutional code has better performance at the same code rate and is widely used.In the fields of adaptive intelligent communication,military reconnaissance and electronic countermeasure,the blind identification of convolutional codes parameters plays a vital role.Deep learning is a popular research direction at this stage and has been successfully implemented in many application fields.It is necessary to explore the application of deep learning in communication systems.In this thesis,the blind recognition of convolutional codes in Gaussian noise channel is studied,and the recognition of convolutional codes is studied from two aspects: hard decision and soft decision.Based on hard-decision convolutional codes identification,a combination of convolutional codes feature map and deep learning algorithm is proposed to realize codes parameters identification.The experiment analyzes the codeword processing method based on matrix IP transformation,and summarizes the advantages and disadvantages of the algorithm.Based on this,we propose the feature extraction method based on Gaussian column transform,and compare the optimal algorithm under two feature extraction methods,the AlexNet algorithm based on Gaussian column transform has the best recognition performance,and still has 80% recognition accuracy at the BER of 0.02.On the other hand,this thesis uses the received soft-decision information combined with deep learning technology to solve the convolutional codes parameters recognition problem.The methods of convolutional codes recognition based on TextCNN and ResNet models are mainly studied.To address the shortcomings that the recognition algorithm based on TextCNN model is easily affected by the length of word construction and the recognition algorithm based on ResNet model has too much computational complexity,a convolutional codes recognition model based on improved one-dimensional AlexNet（DAN）is proposed.The effects of input sample length,number of convolutional layers and other factors on the recognition performance of the model are analyzed.Comparing and analyzing the recognition performance of different algorithms,the recognition performance of all algorithms reaches more than 90% when the signal-to-noise ratio is greater than 18 d B.At the signal-to-noise ratio of 2 d B,the recognition accuracy of the DAN algorithm proposed in this thesis is 70% higher than that of the traditional algorithm and 16% higher than that of the TextCNN algorithm.When the signal-to-noise ratio is less than 6 d B,the recognition performance of DAN algorithm is better than ResNet algorithm.By comparing the training time and test time of the model,it is found that the computational complexity of DAN algorithm is nearly five times lower than ResNet algorithm.Therefore,the DAN algorithm proposed in this thesis has higher recognition accuracy,lower computational complexity,and achieves fully automated and integrated parameter recognition compared to traditional algorithms.