An adaptive active learning method for fault diagnosis

Abstract

In many real world fault diagnosis situations, we only have a limited amount of labeled samples to model supervised leaning methods. It brings two potential problems before we practice the supervised leaning methods: 1) could not find unseen faults. 2) need sufficient fault samples to train classifiers. In order to deal with the issue, this paper proposed an adaptive active learning method for fault diagnosis. The adaptive active learning method could actively discover unseen faults and choose useful samples to increase the performance of the classifier. The adaptive active learning method has two key points over general active learning methods: 1) It could alter the goals of classifier adaptively based on the usefulness of samples. 2) An incremental learning strategy is used to quicken the training procedure. On the benchmark Tennessee Eastman Process data, we evaluated the performance of the proposed method, and the results show that our method was better than the stand-alone supervised fault diagnosis methods on both unseen and seen faults.