Mining Rare Patterns in Wind Tunnel by a Sequential Outlier Ensemble

Abstract

Data-driven techniques have been extensively leveraged in modeling and control of wind tunnel systems. Under this situation, mining rare patterns (also referred to as outliers or anomalies) in wind tunnel databases is becoming increasingly critical, since these samples may imply interesting data patterns. Recently, some ensemble outlier detection methods have been developed to improve the performance of single detectors. However, these ensemble models are often parallel structure and designed for variance reduction. This paper proposes a sequential ensemble model, by which both variance and bias can be reduced simultaneously. Experimental results on real world wind tunnel databases have shown that our sequential ensemble model could outperform most existing parallel ensemble models and single models.