Anomaly Detection Method Based on Fast Local Subspace Classifier

Abstract

Anomaly detection method from multi-dimensional time-series sensor data has been developed which detects anomalies based on normal state models. LSC method was employed to deal with various normal states and fast LSC method was proposed which reduces a computational time. Clustering is utilized to reduce the data for searching in FLSC method. Availability of FLSC method was confirmed using data of real equipment. FLSC method was 1 to 10 times faster than LSC method.