GPON Physical-Layer Fault Dataset

This dataset supports the research article "Performance Evaluation of Lightweight Machine Learning Classifiers for Real-Time Physical-Layer Fault Diagnosis in GPON Infrastructures" and provides all resources necessary to reproduce, extend, and benchmark the synthetic fault classification dataset described in the Methods section of that work.
The repository contains two primary files: (1) the pre-generated dataset in comma-separated values format (gpon_dataset_v3.csv), ready for immediate use in classification experiments; and (2) the Python generation script (gpon_dataset_v3.py), which allows complete reproduction and controlled modification of the dataset from first principles.
Dataset characteristics. The dataset comprises 6,600 samples distributed across five mutually exclusive physical-layer fault classes representative of ITU-T G.984 GPON operational conditions: Healthy (n = 3,300), Loss Event (n = 1,800), OSNR Degradation (n = 780), Laser Drift (n = 420), and Hard Failure (n = 300). Each sample is described by five features directly observable from standard OLT telemetry interfaces: extra optical insertion loss (dB), optical signal-to-noise ratio (dB), Q-factor (dB), base-10 logarithm of pre-FEC bit error rate, and received optical power (dBm). The class distribution deliberately reflects the imbalanced prior probabilities observed in operational GPON networks, with the Healthy class accounting for 50% of samples and Hard Failure for 4.5%. An additional 600 composite fault samples — generated by combining Loss Event and OSNR Degradation conditions — are included to model compound degradation scenarios frequently observed in field deployments.