Domain discrepancy feedback optimization method for few-sample QoT estimation in optical networks

Accurate quality of transmission (QoT) estimation is crucial for optimizing optical network planning and operation. Machine learning has attracted attention as a promising scheme for QoT estimation. Usually, it is difficult to obtain a large amount of optical path data due to environmental and cost constraints. The data collection difficulties can be mitigated by transfer learning (TL) that leverages knowledge across domains. However, the QoT estimation model’s performance will degrade when the data distribution between the source and target domains is significantly different. Therefore, this paper proposes a domain discrepancy feedback optimization (DDFO) method to improve the QoT estimation accuracy under few-sample conditions. Within the proposed method, first, a dual-branch neural network with shared weights is constructed to learn from both the source and target domains jointly. Subsequently, an adaptive alignment layer is introduced to measure the feature distribution discrepancy between domains. Finally, the discrepancy loss is incorporated into the overall loss function to guide model training through the feedback mechanism. Simulation results show that the performance of the DDFO method demonstrates an improvement ranging from 26.38% to 42.65% compared with traditional artificial neural networks, TL, and sample-distribution-matching-based transfer learning in diverse network topology scenarios. DDFO demonstrates strong generalization and superior performance in few-sample QoT estimation, making it a promising solution for practical optical network optimization.