Active Learning for On-Line Partial Discharge Monitoring in Noisy Environments

Partial discharge (PD) results from deterioration to the insulation of high voltage equipment in power grids, such as transformers, switch gears, and cable terminals. PD monitoring is a promising approach that ensures the reliable performance of electrical assets through condition based maintenance. Machine learning techniques have been successfully used to discover features and specific patterns that can differentiate between partial discharges and noise. On-line PD monitoring systems have been recently deployed for continuously assessing the equipment health, so that its maintenance requires minimal cost and shorter disruption to operational services. However, labelled data is required in order to build predictive models for on-line PD monitoring. Labelled data is expensive to obtain since it requires expert input. In addition, in real scenarios the ratio of noise pulses to real PDs can be high, making the learning task harder. This paper investigates an active learning (AL) approach for streaming data that aims to maintain an on-line model that is robust to class imbalances. The empirical evaluation, using real HFCT sensor data, shows that the proposed AL approach is able to achieve over 80% accuracy in noisy scenarios with minimal expert dependence (only 2% of the labelled instances).