本文对机器学习技术在光传输网络传输质量（QoT）估计领域的进展进行了全面的研究。由于不断增长的数据流量需求和新技术的出现，光网络的复杂性日益增加。高精度估计QoT对于优化网络资源和确保可靠通信至关重要。传统QoT估计方法通常无法捕获光网络的高度动态和非平稳行为，导致性能和资源利用率不理想。为了应对这些挑战，这项工作引入了两种创新的机器学习框架：不变卷积神经网络预测器（ICNNP）和在线持续学习（OCL）框架。ICNNP旨在通过考虑变化的链路跨度和时间特征分布来预测QoT，从而处理时变特征，而不需要大量的重新训练。 该模型适应不同的网络配置和条件，展示了光网络系统中初始部署的实用方法。ICNNP通过跨不同传输设置的广泛实验数据收集进行了测试。 结果表明ICNNP明显优于基准替代方案。 值得注意的是，即使跨度数量在9到12之间，评估时间范围从12小时后到72小时后，我们的模型仍保持信噪比预测误差，标准偏差始终分别低于0.4 dB 和0.25 dB。本文OCL框架下提出了正则化在线持续学习(OCL-REG)算法，旨在在不断变化的网络环境中保持QoT估计的高精度。OCL-REG能够通过数据流适应不断学习，并使用正则化技术来避免灾难性遗忘，确保模型的性能随着时间的推移保持一致，即使网络不断发展。本文在各种场景（包括渐进式、循环式和随机分布变化的环境）中进行了一系列测试，以评估所提出模型的性能。与传统的OCL方法和静态模型相比，OCL-REG模型具有更短的适应窗口尺寸，表现出优越的适应性和稳定性。与平均需要253个样本的传统再训练策略相比，OCL-REG模型的适应窗口大小更短，平均约为107个样本。OCL-REG表现出卓越的适应性和稳定性，在包含1000个批次的数据分布偏移序列的测试基准上实现了0.19的平均累积均方误差(C-MSE)。此外，该指标在1000个批次中的标准偏差仅为0.001。论文中的讨论概括了这些机器学习框架对现实世界部署的影响。ICNNP提供了一种解决方案，满足了对能够从有限的训练数据中很好地泛化的模型的迫切需求，而OCL-REG框架解决了模型能够实时适应网络变化而无需大量再训练或人工干预的必要性。它们与网络控制平面的集成可以显着改善资源分配、预测性维护和整体网络性能，为下一代智能光网络铺平道路。
The thesis presents a comprehensive study on the advancements of machine learning (ML) techniques in the domain of quality of transmission (QoT) estimation within optical transport networks. Optical networks are experiencing increasing complexity due to escalating data traffic demands and the advent of new data rich technologies. Estimating QoT with high precision is paramount for optimizing network resources and ensuring reliable communication. Traditional QoT estimation methods often fail to capture the highly dynamic and non-stationary behavior of optical networks, leading to suboptimal performance and resources utilization. To address these challenges, this work introduces two innovative ML frameworks: the invariant convolutional neural network predictor (ICNNP) and the online continual learning (OCL) framework. The ICNNP is designed to predict QoT by accounting for variant link spans and temporal features distribution shifts, thereby handling time-variant features without requiring extensive retraining. This model adapts to different network configurations and conditions, demonstrating a practical approach for initial deployment in optical network systems. ICNNP was put to the test through extensive experimental data collection across varying transmission setups. The results indicate that ICNNP significantly outperforms the benchmark alternatives. Notably, even as the number of spans ranged between 9 and 12 and the evaluation timeframe extended from 12 hours to 72 hours, our model maintained a signal-to-noise ratio prediction error, and the standard deviation is consistently under 0.4 dB and 0.25 dB, respectively.On the other hand, the OCL framework, particularly focusing on the regularized online continual learning (OCL-REG) algorithm, is developed to maintain high precision in QoT estimation amidst the ever-changing network environment. OCL-REG‘s capability to learn continually through data stream adaptation and its use of regularization techniques to avoid catastrophic forgetting ensures the model‘s performance remains consistent over time, even as the network evolves. The thesis conducts a series of tests across various scenarios, including environments with gradual, recurring, and random distribution shifts, to evaluate the performance of the proposed models. The OCL-REG model, with its shorter adaptation window size, displays superior adaptability and stability compared to traditional OCL methods and static models. The OCL-REG model has a shorter adaptation window size, averaging around 107 samples, compared to the conventional retraining strategy, which requires an average of 253 samples. OCL-REG demonstrates superior adaptability and stability, achieving an average cumulative mean squared error (C-MSE) of 0.19 on a testbench with a data distribution shift sequence containing 1000 batches. Furthermore, the standard deviation of this metric over the 1000 batches is only 0.001.The discussion within the thesis encapsulates the implications of these ML frameworks for real-world deployment. The ICNNP offers a solution to the pressing need for models that generalize well from limited training data, while the OCL-REG framework addresses the necessity for models that can adapt in real-time to network dynamics without extensive retraining or human intervention. Their integration with the network‘s service layer can lead to significant improvements in resource allocation, predictive maintenance, and overall network performance, paving the way for the next generation of intelligent optical networks.
