Enhancing one-Class classifiers performance in multivariate time series through dynamic clustering: A case study on hydraulic system fault detection

Abstract

Datasets composed of multivariate time series arising from real applications are usually affected by many factors such as noise and disturbances. Any modeling procedure benefits from having its training data carefully selected. This paper presents a methodology designed to enhance the performance of one-class classifiers in time series by incorporating dynamic time series clustering. The clustering process leverages DTW Barycenter Averaging (DBA) and k-means to group multivariate time series based on similarity. The Apriori algorithm is used to generate subsets of instances, which are then used to train multiple one-class classifiers for the same class. Three distinct strategies are applied to combine the outputs of these classifiers for each class. The proposed method is evaluated on a hydraulic system dataset to investigate typical faults that occur simultaneously and with varying intensities. The results show that improving the similarity of the training subsets and combining the outputs of the classifiers led to a performance improvement of more than 89 %. In addition, the methodology successfully reduced a hydraulic system dataset from 17 variables to as few as 3 or even 1, while still achieving better classification performance compared with recent findings in the literature.