Data Augmentation Based on Time-Series Generative Adversarial Networks for Failure Prediction in Optical Network

The wide application of machine learning (ML) technology in the field of optical network failure prediction requires sufficient failure samples for model training. However, due to the robustness of optical network system, failures rarely occur and few failure samples are obtained in the collected data, which leads to the problem of data imbalance and brings great challenges to failure prediction. In consequence, it is necessary to find a suitable method to augment the failure data. However, typical data augmentation methods can not make full use of the time-series characteristics of optical network failure data. In this paper, we propose a Time-Series Generative Adversarial Network (TimeGAN) method for learning time-series characteristics of real failure data and augment the failure data, thereby addressing the issue of data imbalance and improving the effect of failure prediction. The TimeGAN method exhibits superior performance in generating failure samples, as evidenced by its lower mean absolute error (MAE), root mean square error (RMSE), and Kullback-Leibler (KL) divergence compared to other algorithms. Qualitatively, highly consistent t-Distributed Stochastic Neighbor Embedding (t-SNE) and Principal Component Analysis (PCA) visualizations of original and augmented data validate the superiority of TimeGAN. Quantitatively, the efficacy of TimeGAN in improving failure prediction accuracy was demonstrated through failure prediction experiments. The accuracy (ACC) of the LSTM model trained with the augmented data increased by 22.41%, and the F1 score increased by 21.68%, which indicates a significant improvement in the model's failure prediction capability.