Hybrid semi-supervised fault detection framework under limited high-confidence data scenarios

Abstract

In the broad field of fault detection, approaches utilizing process conditions are established for systems with adequate datasets. However, systems including recently commissioned and novel processes have limited datasets available for model-based fault detection. Moreover, these systems have far greater proportions of normal operating data than adequate fault examples due to the time for which they have been operated. In this work, a combined hybrid framework for fault detection is developed that enables augmentation of the limited dataset available with HAZOP data, allowing for the utilization of both human expert knowledge and generated pseudo-process data. Additionally, the generation of artificial data is performed for reducing false positives in adversarial training. A semi-supervised distance variant of center loss is used to improve the consistency of deep feature activations from paired and unpaired data. A comparison between the proposed approach and an approach utilizing only process data in the limited data case is presented. Overall, the proposed approach shows 4.1 % and 8.8 % improvements in average detection rate when compared to the state-of-the-art supervised method for the Tennessee Eastman process and subcritical coal-fired power plant case studies, respectively, enabling the use of unlabeled data to supplement labeled process data for fault detection.