Deep Learning Signal Waveform Characterization of Partial Discharge for Underground Power Cable Conditions

In this study, deep learning models were applied to an extensive partial discharge (PD) dataset to provide a foundation for automatizing the detection of underground cable defects. Traditional signal detection methods require explicit prescription of PD features that characterize the state of the cable system. When the sample size is sufficiently large, on the other hand, deep learning models allow complex interrelations of autogenerated features. A distinct challenge is the characterization of the waveform signal, which depends on cable length. The deep learning models presented in this paper outperformed predictions with traditional methods. In addition to classifying PD signals, the models identified source locations of the defects within a cable system through recurrent neural networks. Additional assessments included advanced data augmentation strategies and interpretability. The success of the results demonstrates the potential use of the model for predictive maintenance.