Integrating Knowledge Distillation and Transfer Learning for Enhanced QoT-Estimation in Optical Networks

A precise assessment of the Quality-of-Transmission (QoT) for a Lightpath (LP) is essential for efficient optical network design and optimal resource utilization. Recent advances in deep neural network (DNN) techniques have yielded promising results for QoT estimation. However, these models typically rely on numerous parameters and require extensive training data and significant processing resources for reliable predictions. In this context, we propose a novel framework integrating knowledge distillation (KD) and transfer learning (TL), offering a synergetic solution to these practical challenges of traditional DNN-based systems. The proposed framework reduces the number of trainable parameters by 93.6%, training time by 48.5%, and achieves a prediction time of 0.09 seconds while maintaining comparable accuracy. Our hybrid model attains 98.4% accuracy, with an MSE of 0.016 dB, demonstrating high-performance efficiency, reduced computational complexity, and enhanced adaptability. The dataset used in this investigation was produced synthetically using the GNPy platform. To the best of our knowledge, this is the first time the hybrid solution (KDTL-QoT), combining both KD and TL, has been used to estimate the QoT of a new LP. The results make this approach a viable solution for real-world applications in optical networks.