Research On Quantization Of Learned Digital Back-Propagation Nonlinear Compensation Algorithm In Optical Fiber Communication System

With the rapid development of Internet,mobile Internet and highdefinition video services,the continuous and rapid growth of data traffic has brought huge challenges to the optical fiber communication system,and it is urgent to further improve the transmission capacity of the optical fiber communication system.Compared with the intensity modulation/direct detection system,the coherent optical fiber communication system has the advantages of high receiving sensitivity,large channel capacity and can use digital signal processing technology to equalize the damage,and has become the main solution in the field of long-distance optical fiber communication.In the long-distance coherent optical fiber communication system,the linear damage equalization has been fully studied,and the nonlinear effect of the optical fiber has become the main factor restricting the further improvement of the system capacity.In the existing research on nonlinear compensation algorithms,the Digital Back-Propagation(DBP)algorithm based on the Split-Step Fourier method is a very promising research scheme,but it has disadvantages such as high computational complexity and relatively poor compensation effect in frequency domain and time domain compensation respectively.In recent years,with the development of artificial intelligence technology,machine learning has provided new ideas for solving the problems in the field of optical fiber communication.The Learned Digital Back-Propagation(LDBP)proposed based on the neural network structure reduces the complexity and improves the compensation performance at the same time.At present,the research based on LDBP nonlinear compensation algorithm basically does not consider the influence of quantization noise on the performance of the algorithm when the algorithm is implemented in hardware.Therefore,this paper conducts in-depth research on the quantization scheme of nonlinear compensation algorithm based on LDBP for long-distance coherent optical fiber communication system.The main research contents and innovations are as follows:(1)The uniform quantization scheme of the LDBP nonlinear compensation algorithm is studied,and the performance of the uniform quantization algorithm is simulated and analyzed.The nonlinear compensation algorithm based on LDBP needs to quantize the neural network weight and the signal after each calculation in hardware implementation.The uniform quantization scheme evenly divides the signal or neural network weight amplitude range into 2n-1 equal parts according to the number of quantization bit n,and take the nearest endpoint value as the quantized value.This paper firstly studies the quantization scheme that uniformly quantizes the signal and LDBP weight parameters directly after floating-point training;Secondly,studies the quantization scheme that uniformly quantizes the weight parameters and the signal in each training process of the LDBP algorithm.The study found that the performance of the algorithm quantized during the LDBP training process is better than that of the algorithm directly quantized after floating-point training,and the weights trained have a certain tolerance for the quantization error of the algorithm.By building a 20 GBaud dualpolarization 16QAM long-distance coherent optical fiber communication simulation system of 3200 km,the performance of the uniform quantization scheme based on LDBP nonlinear compensation algorithm is studied.The research results show that the quantization training improves the tolerance of the LDBP nonlinear compensation algorithm to the quantization error,and the effective SNR is 15.66 dB after 8 bit uniform quantization training.Compared with the direct uniform quantization of LDBP nonlinear compensation algorithm,the effective SNR is increased by 0.82 dB.(2)The non-uniform quantization scheme of LDBP nonlinear compensation algorithm is studied.According to the amplitude probability distribution of the signal and the weights of the neural network,an improved non-uniform quantization scheme based on Alaw quantization and broken line approximation is proposed,and the performance of the improved non-uniform quantization algorithm is simulated and studied.According to the amplitude probability distribution of the signal and the weights of the neural network,the quantization range is divided into 5 intervals using different quantization units,and uniform quantization is performed in each interval.By building a 20 GBaud dual-polarization 16QAM coherent optical fiber communication simulation system with a transmission distance of 3200 km,the performance of the non-uniform quantization scheme based on LDBP nonlinear compensation algorithm is studied.The research results show that when the signal and weights are quantized at 8 bits,the effective SNR of using the improved non-uniform quantization scheme in LDBP quantization training is improved by 0.41 dB compared with the use of the uniform quantization scheme;at the same time,under the same performance of the algorithm,the requirements of the improved nonuniform quantization scheme for the quantization bit of weights are 2 bits lower than that of uniform quantization.To sum up,this paper studies the quantization problem of nonlinear compensation algorithm of learned digital back-propagation in the longdistance coherent optical fiber communication system,and analyzes the algorithm performance of the uniform quantization scheme and the nonuniform quantization scheme.On this basis,an improved non-uniform quantization scheme based on A-law compression and broken line approximation is proposed,and the performance of the improved nonuniform quantization scheme based on LDBP nonlinear compensation algorithm is verified by simulation,which provides a new idea for the quantization scheme of nonlinear compensation algorithm based on LDBP in long-distance coherent optical communication system.