Research On QoT Estimation Strategy For Optical Networks With Few Samples Based On Weakly Supervised Learning

With the rapid development of communications technology,network requirements are becoming increasingly complex.In order to improve the reliability of network transmission,it is necessary to quickly and accurately predict the Quality of Transmission(QoT)of a lightpath before it is deployed,thus assisting in the design planning of the network.Machine learning is widely used in QoT estimation due to its flexibility and accuracy,but its performance is often limited by the number and quality of labeled samples.In real network scenarios,a large number of transmission parameters for characterising lightpaths have been accumulated,but obtaining their QoT labels is limited by factors such as the lack of lightpath monitoring equipment.In real network scenarios,a large number of transmission parameters have been accumulated for characterising lightpaths,but obtaining their QoT labels is limited by factors such as the lack of lightpath monitoring equipment.As a result,there are more unlabeled lightpath samples and fewer labeled samples that can be collected in the network,making it difficult to support the effective training of QoT estimation models and affecting their prediction accuracy.Therefore,it is essential to investigate how to achieve accurate QoT estimation when the number of labelled samples is limited.Facing this problem,this thesis adopts active and semi-supervised learning methods in weakly supervised learning to improve QoT estimation accuracy by actively constructing high-quality datasets and using unlabelled samples to achieve regularisation.The main research work and innovation points of this thesis are as follows:(1)To address the problem that randomly deploying probe lightpaths to obtain QoT labels can easily lead to data redundancy and waste of resources,this thesis proposed an active learning-based QoT estimation method.First,the unlabelled lightpath samples that are more valuable for training are actively selected by sensing the amount of information contained in the lightpath samples.Then,the corresponding probe lightpaths are deployed to obtain their QoT labels,and the training set is expanded with these high-quality samples to optimise model performance.Simulation results show that our scheme achieves more accurate QoT estimation with as few training samples as possible compared to traditional random sampling schemes.Simulation results show that our scheme achieves more accurate QoT estimation with as few training samples as possible compared to traditional random sampling schemes.(2)To address the problem that small training samples easily lead to overfitting of the model,resulting in poor prediction accuracy of the model for the test lightpaths,this thesis proposed a QoT estimation method based on semi-supervised learning.Lightpaths without QoT labels are considered as possible candidates for future deployment,i.e.lightpaths to be predicted.Prediction variance is a commonly used metric to measure the model’s ability to predict unknown data.A larger variance indicates greater dispersion in predicted values and lower confidence in the model’s predictions.Therefore,this thesis uses "minimizing the prediction variance of the model for the candidate lightpaths" as the regularization term of the classical objective function.It allows the model to fit the labeled samples while focusing on its predictive performance for the "lightpaths to be predicted",avoiding overfitting the model and enhancing its generalisation performance.Simulation results show that our scheme can significantly improve the QoT prediction accuracy of the model for candidate lightpaths.