College

Frank Batten College of Engineering and Technology

Department

Electrical/Computer Engineering

Program

PhD Electrical & Computer Engineering

Publication Date

4-2021

Abstract

Clustering is widely used in unsupervised machine learning to partition a given set of data into non-overlapping groups. Many real-world applications require processing more complex multivariate time series data characterized by more than one dependent variables. A few works in literature reported multivariate classification using Shapelet learning. However, the clustering of multivariate time series signals using Shapelet learning has not explored yet. Shapelet learning is a process of discovering those Shapelets which contain the most informative features of the time series signal. Discovering suitable Shapelets from many candidates Shapelet has been broadly studied for classification and clustering of univariate time series signals. Shapelet learning has shown promising results in the case of univariate time series analysis. The analysis of multivariate time series signals is not widely explored because of the dimensionality issue. This work proposes a generalized Shapelet learning method for unsupervised multivariate time series clustering. The proposed method utilizes spectral clustering and Shapelet similarity minimization with least square regularization to obtain the optimal Shapelets for unsupervised clustering. The proposed method is evaluated using an in-house multivariate time series dataset on detection of radio frequency (RF) faults in the Jefferson Labs Continuous Beam Accelerator Facility (CEBAF). The dataset constitutes of three-dimensional time series recordings of three RF fault types. The proposed method shows successful clustering performance with average value of a precision of 0.732, recall of 0.717, F-score of 0.732, a rand index (RI) score of 0.812 and normalize mutual information (NMI) of 0.56 with overall less than 3% standard deviation in a five-fold cross validation evaluation.

Keywords

Shapelet, Multivariate

Disciplines

Electrical and Computer Engineering | Longitudinal Data Analysis and Time Series

Files

Download

Download Full Text (372 KB)

Unsupervised Multivariate Time Series Clustering


Share

COinS