Alongside the progress in technology related to the Internet of Things, the Mahalanobis–Taguchi (MT) system, which is an anomaly detection technique suitable for monitoring the condition of production equipment, has attracted attention. However, with the conventional MT method, historical data acquired and accumulated from sensors and smart devices cannot be analysed appropriately. This is because very often the accumulated historical information is data not labelled as either ‘normal’ or ‘anomaly’. Therefore, in this research, we propose a procedure that enables to detect anomalies with the MT method even when learning data are categorised as ‘unlabelled’. Specifically, we introduce a process for estimating the Mahalanobis distance on the population by applying a robust method based on γ divergence in the MT method. Through numerical experiments, we show that the proposed procedure is a useful anomaly detection technique for unlabelled data. Since this implies that labelling is redundant in anomaly detection, we conclude that the practicality of the MT method can be improved.
ASJC Scopus subject areas