Active learning with cross-dataset validation in event-based non-intrusive load monitoring
Florian Liebgott, Bin Yang · 2017
Supervised event-based NILM systems usually require a large set of labeled training data to achieve high classification accuracies. To minimize the cost of labeling a sufficient amount of events, active learning can be employed. By using only a small set of labeled samples for initial training followed by selecting only the most informative samples to be labeled, the total number of labeled training samples can be reduced significantly. The performance of an active learning system strongly depends on the choice of the initial training set and the used query strategy. We thus investigated the impact of different methods to select the dataset for initial training as well as various query strategies on the resulting classification accuracy in an event-based NILM framework. For evaluation we used two datasets, BLUED and ISS kitchen, on which we were able to achieve high classification accuracies with significantly less training samples compared to conventional training without active learning.