CTabDDPM-based data augmentation for fault diagnosis of optical networks

Abstract

Data imbalance constitutes a primary bottleneck that constrains the performance of intelligent fault-diagnosis models for optical networks. To address this challenge, a data augmentation method based on a conditional tabular denoising diffusion probabilistic model (CTabDDPM) is proposed. Accurate conditional generation is enabled by the CTabDDPM, such that samples can be generated in a fault-class-directed manner. In addition, the diffusion mechanism allows the joint data distribution to be learned progressively, thereby effectively ensuring the physical consistency of the generated data. To further enable a systematic evaluation of generative performance, a multidimensional assessment framework is established that encompasses distributional similarity, structural fidelity, and task utility. Experiments conducted on a real optical module dataset demonstrate that, under an extremely imbalanced setting (0.84%), the proposed method achieves a diagnostic accuracy of 0.9987 and an F1 score of 0.9788, while the false negative rate and false positive rate are reduced to 0.0107 and 0.0304, respectively. These results substantially outperform the baseline generative models, thereby validating both the effectiveness and practical value of the proposed method for improving fault diagnosis under severe class imbalance.