Semi-supervised Anomaly Detection on Industrial Process Data Using Self-adaptive Training

Fault detection plays an important role in safe operation and equipment maintenance of process industries. With the boom of online monitoring systems in process industries, the high availability of multi-channel measurement data motivates research into data-driven fault detection methods. Semi-supervised learning is a machine learning framework that trains on a small amount of labeled data and a large amount of unlabeled data, efficiently utilizing unlabeled raw data while leveraging critical label information to achieve detection accuracy. A novel semi-supervised fault detection method using self-adaptive training is proposed in this paper. By constantly calibrating target labels during the training process, most of the unlabeled samples can be correctly labeled. Experiments on the Tennessee Eastman Process (TEP) benchmark dataset is conducted to examine its performance. It achieves high detection accuracy using only partially labeled training data.