Recognition Of Closed Set Turbo Codes Based On Machine Learning

Channel coding technology is very important in the field of communication,and Turbo code is an important coding method that is widely used.Machine learning methods have greatly developed and become its core technology in today's natural language processing and image fields.The closed set recognition of Turbo codes based on machine learning methods has not been studied.For the intelligent development of codeword recognition calculations,it is very important to study Turbo code recognition based on machine learning methods.The generation of Turbo codes can be seen as consisting of an encoder and an interleaver.In this paper,various parts of coding are studied,including m-sequence prediction based on machine learning,closed set recognition of convolutional codes,and closed set recognition of turbo codes.The m-sequence part is expected to be applied to the interleaver part,and the convolutional code part experiment can be widely adapted to the encoder part of the turbo code.The last part is the single frame and multiframe experiments of the closed set identification of the turbo code.The main research results of this paper are mainly divided into three parts:1.The first research on the recognition of pseudo-random sequences by machine learning,theoretically proved the neural network's complete learning ability for msequences,discussed the basic network settings and experimented with sparse different coefficients at 25,50,75,and 100 orders The network performance capability corresponding to the degree.Using the intercepted information sequences of lengths of 2000,5000,and 10000,the sparse coefficients of the m-sequences at the 25 th,50th,and 75 th order can be learned,and the network structure used for reference is given.Sequence experiments explore the reasons why neural networks can tolerate msequences,and summarize the neural network's ability to express binary operations.2.For the first time,the closed-set recognition of convolutional codes by machine learning was studied.The intercepted convolutional code data of unknown priors was used to perform classification experiments through deep convolutional networks and improved convolutional neural networks to achieve unknown alignment positions.The fast closed set identification of the lower convolutional code has strong fault tolerance and high degree of versatility,which provides an important reference for the encoder study of turbo codes for machine learning.The study found that the improved convolutional neural network is superior to short-memory codeword recognition,and its effect is better than a deep residual network with more than three orders of magnitude.3.For the first time,the closed-set recognition of Turbo codes by machine learning was studied.The frame structure of the codeword was obtained through the zeroing structure.Corresponding neural networks were established for single-frame and multiframe data.The single-frame and multi-frame were studied and analyzed.The difference of codeword recognition provides a new multi-frame data set generation scheme for parallel processing.The network under a single frame can complete the identification of the encoder,and the neural network under multiple frames can realize the joint identification of the encoder and the interleaver.In the future,the network with a higher interleaving depth will be identified using a lighter network.