A Study of Subspace Generation Method in the Anomaly Detection of Time Series Data

Abstract

Fast and accurate anomaly detection is becoming essential in maintenance services. In this paper, learning data is respectively selected based on the distance between observed data and learning data, to form an optimum subspace. Through the experiments of projection distance method and mutual subspace method, it is confirmed that generating the constrained pair-wise subspaces for both observed data and learning data can effectively detect weak anomaly, absorbing the transient variation of data.