Data-based process fault detection using Active Cost-sensitive Learning

Abstract

Fault detection for industrial process is important to improve the qualities and quantities of products. Since it is difficult to obtain exact mathematic model for fault detection, data-based model for fault detection is popular in many industrial process applications. There are two class-imbalanced problems in fault detection for industrial process: the problem between unlabeled instances and labeled instances, the one between normal instances and fault instances. Active Cost-sensitive Learning (ACL) is proposed to solve the above two class-imbalanced problems in this paper. Margin sampling, one step of ACL algorithm, is used to select an informative instance from unlabeled instances. Cost-sensitive support vector machine (CSVM), the other step of ACL algorithm, is trained to predict the class label with minimum expected cost for an unlabeled instance. The effectiveness of the proposed algorithm is demonstrated by the benchmark Tennessee Eastman problem. Experiments show that the proposed algorithm can effectively reduce average cost and increase fault sensitivity.