Distributed active learning with application to battery health management

This paper focuses on distributed implementation of active learning with a limited number of queries. In the prognostics and health management domain, the cost to obtain a training sample can be fairly high, especially when studying the aging process for remaining useful life prediction of a mission critical component. Active learning with limited resource is formulated as a reinforcement learning problem, where the sampling strategy has to minimize the expected generalization error within a finite horizon. An importance sampling based method is adopted for active learning and extended to distributed implementation with multiple active learners. The fusion of importance weights from multiple learners is interpreted as a special boosting strategy. The proposed framework is applicable to classification and regression problems as well as semi-supervised learning. The remaining useful life prediction for battery health management is used to compare the proposed method with conventional passive learning methods. Empirical study shows that the fusion of distributed active learners achieves better classification and prediction accuracy with a reduced number of training samples needed to have a complete run-to-failure profile.