USVDD: A one-class outlier detection method with imprecisely- observed data

Abstract

Detecting outliers is essential to recognizing potential faults in high-value equipment. Since such equipment typically undergoes careful maintenance to prevent breakdowns, outlier detection often involves a one-class classification problem. Current advanced methods frequently face difficulties in managing the complexities of one-class observations, particularly when dealing with uncertain and imprecise monitoring data. This study proposes a novel approach to fault detection, called uncertain support vector data description (USVDD), in contexts involving one-class uncertain data. Drawing on uncertainty theory, USVDD conceptualizes imprecise observations as uncertain variables. By integrating novel hyperparameters, particularly belief degrees ${\alpha }_k$, the method effectively addresses and quantifies data uncertainty, enhancing its ability to handle complex, uncertain datasets. Testing on 12 real-world datasets highlights that the proposed USVDD method outperforms alternative methods by achieving the highest balanced F1-scores on 10 datasets. Additionally, it demonstrates remarkable efficiency, completing computations on high-dimensional datasets up to 60 times faster than competing methods on the same hardware setup. USVDD excels in managing observational data influenced by aleatory uncertainty, making it a reliable solution for one-class diagnostic modeling in situations with scarce fault samples.