Enhancing Fault Detection in Optical Networks with Conditional Denoising Diffusion Probabilistic Models

The scarcity of high-quality anomalous data often poses a challenge in establishing effective automated fault detection schemes. This study addresses the issue in the context of fault detection in optical fibers using reflectometry data, where noise can obscure the detection of certain known anomalies. We specifically investigate whether classes containing samples of low quality can be boosted with synthetically generated examples characterized by high signal-to-noise ratio (SNR). Specifically, we employ a conditional Denoising Diffusion Probabilistic Model (cDDPM) to generate synthetic data for such classes. It works by learning the characteristics of high SNRs from anomaly classes that are less frequently affected by significant noise. The boosted dataset is compared with a baseline dataset (without the augmented data) by training an anomaly classifier and measuring the performances on a hold-out dataset populated only with high quality traces for all classes. We observe a significant improved performance (Precision, Recall, and F1 Scores) for the noise affected training classes proving the success of our methods.