Improvement of One-Class Classifier Performance via Time Series Clustering: Application to a Hydraulic System

Abstract

This work proposes a methodology aimed at improving the performance of one-class classifiers through time series clustering. The clustering process involves utilizing DTW Barycenter Averaging (DBA) and k-means to group multivariate time series based on their similarity. The Apriori algorithm is employed to generate subsets of instances, which are used to train and select multiple one-class classifiers for the same class. The effectiveness of the proposed method is evaluated on a dataset from a hydraulic system to study typical failures happening simultaneously and with several intensities. One of the faults was selected and the results showed that increasing similarity in the subset training data led to an improvement of the classifier performance of more than 34%.