Mahalanobis-Taguchi Method for Anomaly Detection and Classification

Abstract

The Mahalanobis-Taguchi (MT) method helps perform anomaly detection or pattern recognition. This method creates a unit space wherein the normal group forms a homogeneous population, such that whether an individual belongs to the unit space is detected as an anomaly. However, when anticipating the homogeneity of abnormal individuals, which we refer to as “known anomaly,” we can define the unit space for them as well as normal. Any individual that does not belong to any unit space is likely to be an “unknown anomaly” after multiple unit spaces are set accordingly. In this study, we propose two novel analyses to detect and classify any “unknown anomaly” and “known anomaly” within the MT framework. Through Monte Carlo simulation and data analysis, we demonstrate that the proposed procedures can appropriately detect and classify two types of abnormal individuals. We focus not only on “Supervised Learning” using the training data labeled “normal” and “known anomaly,” but also on “Semi-Supervised Learning” using labeled and unlabeled data. We introduce “Semi-Supervised Learning” as a parameter estimation method in the proposed procedures, which confirms its contribution to reducing the labeling cost by engineers.