Training data generation and validation for a neural network-based equalizer
Tao Liao, Lei Xue, Luyao Huang, Weisheng Hu, Lilin Yi
Abstract
The neural network (NN) has been widely used as a promising technique in fiber optical communication owing to its powerful learning capabilities. The NN-based equalizer is qualified to mitigate mixed linear and nonlinear impairments, providing better performance than conventional algorithms. Many demonstrations employ a traditional pseudo-random bit sequence (PRBS) as the training and test data. However, it has been revealed that the NN can learn the generation rules of the PRBS during training, degrading the equalization performance. In this work, to address this problem, we propose a combination strategy to construct a strong random sequence that will not be learned by the NN or other advanced algorithms. The simulation and experimental results based on data over an additive white Gaussian noise channel and a real intensity modulation and direct detection system validate the effectiveness of the proposed scheme.