Research On Machine Learning Techniques For Optical Fiber Nonlinearity Mitigation

For the long-haul large-capacity optical fiber communication systems,higher-order quadrature amplitude modulation(QAM)signals with increasing spectral efficiency,are competitive candidate.In combination with coherent detection and digital signal processing(DSP)technologies,we can equalize linear impairments and mitigate nonlinearity effects of transmission to improve the performance of coherent optical transmission system.The finite impulse response(FIR)-filter-based adaptive equalizers have found successful use in the digital domain to mitigate the linear transmission distortions caused by the chromatic dispersion and polarization mode dispersion of an optical fiber.However,the fiber is a nonlinear medium and its refractive index is dependent on the optical signal power level.Higher-power signals with higher OSNR always cause the severe nonlinear effects,which degrades the system performance and dominantly limits the maximum transmission distances of the higher-order QAM signals.Therefore,nonlinear equalization DSP techniques are indispensable to mitigate the fiber nonlinearity.In the thesis,we introduced two machine learning algorithms,i.e.,K-means algorithm and KNN algorithm to improve system bit-error-ratio(BER)performance.For the drawbacks of the conventional algorithms,we proposed the modified algorithms,i.e.,blind-K-means algorithm and non-data-aided(ND)-KNN algorithm and verified them in the 16-QAM and 64-QAM coherent optical transmission system.Higher classification accuracy,lower classification complexity and zero data redundancy make the proposed blind-K-means and ND-KNN techniques be suitable candidate in practical M-QAM long-haul coherent optical communication system.