A Robust Random Forest Model for Classifying the Severity of Partial Discharges in Dielectrics

Partial Discharges (PDs) are a common source of degradation in electrical assets. It is essential that the extent of the deterioration level of insulating medium is correctly identified, to optimize maintenance schedules and prevent abrupt power outages. Temporal PD signals received from damaged insulation, collected through the IEC-60270 method is the gold standard for PD detection. Temporal signals may be transformed to the frequency domain, introducing new spectral features that may be beneficial in certain circumstances. Consequently, time delays are introduced, due to the high utilization of computational resources within the signal processing pipeline. Moreover, some microprocessors struggle with the excess computational burden demanded by resource-heavy mathematical transformations. To rectify these issues, an alternative approach is utilized, where Machine learning (ML) algorithms are directly used for the classification of PD severity. Cylindrically-shaped air cavities with lengths ranging from 1mm–6mm are introduced to a resin-based polyethylene terephthalate (PET) insulation material. The cavities are partitioned based on size, to obtain different classes of PD severity. A comparative analysis is performed on various ML algorithms, to determine which algorithm correctly determined the severity of PDs, with highest efficacy. Random Forest was determined to be the most performant, with an accuracy of 98.33%. The high performance illustrates the model’s potential success in accurately determining the hazard level of PDs in real-time, based on merely time-domain signals.