This paper addresses the problem of learning a on-line
classification model that can differentiate between partial
discharges and noise. Nevertheless, labelled data is required
in order to build such PD detection model. Consequently, we
investigate an active learning (AL) approach for streaming
classification that aims to maintain an on-line model that is
robust to class imbalances. To the best of our knowledge this
is the first work that proposes to use AL for PD classification.
The empirical evaluation 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). This is an important result for PD detection since
expert feedback is expensive and not always available. In
addition, we studied the impact of the class distribution on
ability of different AL strategies to learn an accurate PD
classifier. The results show that the the AL strategies with
variable thresholds or that include random selection are good
candidates for this challenging learning scenario. In future
work, we plan to create an alternative AL strategy that deals
explicitly with the class imbalance problem. Moreover, we
want to extend the current results to the muticlass problem,
that can differentiate the type of PD signals (e.g., internal,
surface, corona) as this additional information can support
more effective maintenance.