An effective procedure exploiting unlabeled data to build monitoring system

Currently, condition-based maintenance becomes increasingly important with additions of factory automation through the development of new technologies. For many complicated machines, it is difficult to use mathematical models to describe their conditions. Intelligent maintenance makes it possible to perform maintenance similar to that of a human being. However, conventional artificial intelligent methods such as neural network and SVM use only labeled data (feature/label pairs) for training. Labeled instances are often difficult, expensive, or time consuming to obtain. Active learning and semi-supervised learning address this problem by using a large amount of unlabeled data together with labeled data to build better models. In this paper, a new active semi-supervised procedure was proposed to perform fault classification for machine condition monitoring. The effectiveness of the procedure was verified by its application to bearing diagnosis and gear fault detection.