Low Complexity Deep Learning LDPC Decoding

As an efficient channel coding,LDPC code(Low Density Parity Check Code)has good performance in correcting errors,Belief Propagation(BP)decoding has the advantage of low complexity and has been widely used in the field of communications.However,in its Tanner graph,there is a trap set problem,which causes the performance curve of the BP decoding algorithm in the decoding process to appear flat in the area of high signal-to-noise ratio,which is called an error floor phenomenon.The emergence of a phenomenon restricts its application in low-error-rate storage and other fields.This thesis uses the powerful optimization capabilities of deep learning to alleviate the impact of the short-loop structure,achieve high-performance decoding of LDPC codes,and improving the performance of correcting errors.This thesis also studies how to simplify the deep decoding network to achieve low-complexity decoding.First of all,in this thesis,the iterative process of BP decoding is mapped into an equivalent deep neural network,and the weight of soft information propagation in Tanner graph is solved through deep learning,to achieve the purpose of accelerating the speed of iterative convergence and improving the performance of convergence.The neural network decoder is expanded according to the Tanner graph,and weight parameters are added in the message transmission process of variable nodes and check nodes for learning and training,and the optimized solution of message propagation processing parameters is obtained,and good decoding performance is achieved.Channel noise is an important factor that affects decoding performance.This thesis introduces a denoising network,which is cascaded with the neural network decoder as a decoding network.Through joint training and optimization,the effect of noise on the decoding process is reduced.The denoising network is based on the convolutional neural network structure and the residual structure is added to solve the problem of network performance degradation.Through the multi-task learning strategy,the loss function of the denoising network and the decoding network are used as the multi-task loss function of the denoising and decoding network structure,and the scale factor is used to balance.Experiments have proved that in the AWGN channel,the decoding network and the denoising network are trained in parallel,and the learning process of the two strengthens each other,and the denoising part optimizes the decoding performance of the decoding part.In order to further improve the decoding performance,the way to increase the number of weight parameters in the hidden layer of the decoding network and the scale of the neural network has the problem of high complexity.Decoding network has too many learning parameters and high complexity,which brings huge challenges to engineering applications.In response to this problem,the law of the weight parameters in the decoding network structure is analyzed during the experiment.On this basis,two low-complexity neural network decoders are proposed:a low-complexity decoder based on variable node weight sharing and layer weight sharing.The experimental results show that the decoder based on variable node weight sharing reduces the computational complexity by 50%,the two low-complexity decoding networks reduce the complexity of the decoding process,ensure high decoding performance,and the neural network decoders effectively solve the problem that the existing decoders only support the decoding of medium and short code lengths.These two low-complexity neural network decoders can be applied in the field of long codes,and the storage burden is relatively lower and more practical.