Nonlinear Activation Function-Free Artificial Neural Network Enabled Optical Performance Monitoring

Optical performance monitoring (OPM) by leveraging artificial neural network (ANN) has become a promising data-driven solution for stability, reconfiguration, and efficient utilization of heterogeneous optical networks. The implementation complexity as well as the involved dataset of conventional ANN, however, remains a common concern, preventing the efficient OPM implementation. Nonlinear activation functions (NAFs) are commonly used in conventional ANN for feature attraction and nonlinear mapping, which largely reduce the convenience of the OPM, posing the question on the feasibility of removing NAF in ANN-enabled OPM. Here, by investigating the necessity of NAF for the ANN-based OPM, we find that, when either amplitude histogram (AH) or asynchronous delay-tap sampling portrait (ADTP) is used as the ANN input, the commonly-used multi-layer perceptron (MLP) and convolutional neural network (CNN) can be implemented without NAFs. Acceptable variations of mean absolute error (MAE) for the OSNR monitoring are observed by 0.03 dB and 0.23 dB, respectively. Moreover, the removal of NAF does not affect the identification of both the baud-rate and the modulation format information. Our results reveal that, the linear mapping is sufficient to realize the OPM with stable statistical images, thus leading to a simplification prospect for the on-chip ANN-enabled OPM.