A TCN Based Data-Driven Method Combined With TL for Modeling in Experimental IM-DD System

Accurate modeling of fiber channel is crucial for the design and optimization in optical fiber communication systems. This study proposes a data-driven approach to model the end-to-end channel in an intensity modulation-direct detection (IM-DD) system. The method utilizes temporal convolutional network (TCN) to capture the attenuation, chromatic dispersion (CD), and nonlinear characteristics of the end-to-end channel under six different transmission distances and four launch powers for pulse amplitude modulation 4 (PAM4) signals. We experimentally evaluate the generalization capability of trained model in another two launch powers, and obtain an average normalized mean square error (MSE) of 0.0131. Furthermore, we apply transfer learning (TL) technology with our model for PAM4 signals in strong nonlinearity and high-level formats of PAM8 signals, acquiring the average normalized MSEs of 0.0133 and 0.016 with 85% and 70% epochs of initial training reduction, respectively. The experiments demonstrate the accuracy and efficiency of the proposed method.