DBSCAN-Based Clustering for Nonlinearity Induced Penalty Reduction in Wavelength Conversion Systems

We experimentally demonstrate, for the first time, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to improve the receiver performance of wavelength converting 40-Gb/s 16-quadrature amplitude modulation (16-QAM) and 30-Gb/s 64-QAM using four-wave mixing in semiconductor optical amplifier (SOA). Two novel modified DBSCAN methods are proposed, in which “un-clustered” received constellation points that are considered to be noisy points are further processed using method-(1) K-means and method-(2) the minimum distance between an unlabeled point and the clustered points. DBSCAN is more effective for distorted signals due to optical nonlinearities at higher than 25 and 30 dB of optical signal-to-noise ratios (OSNR) for 16- and 64-QAM, respectively. We find that DBSCAN can result in improvements in Q-factor by ~0.8 dB at optimum power when compared to linear equalization. The soft-clustering ability of DBSCAN method-(2) is particularly useful at high OSNR (30 dB) and high input signal powers in the SOA where excessive nonlinearity causes rotation of the outer-constellation points. Method-(1)'s hard clustering is valuable at lower input signal powers, where the clusters maintain a Gaussian-circular shape.