Structural break detection method based on the Adaptive Regression Splines technique

TL;DR

This paper introduces a new time series segmentation method based on Adaptive Regression Splines to detect structural breaks, tested on simulated data and real vibration signals from industrial machinery.

Contribution

It presents a novel segmentation technique utilizing Adaptive Regression Splines and a statistical test for regime change detection in time series data.

Findings

Effective in simulated signals with various distributions

Successfully applied to industrial vibration data

Provides a statistical hypothesis testing framework

Abstract

For many real data, long term observation consists of different processes that coexist or occur one after the other. Those processes very often exhibit different statistical properties and thus before the further analysis the observed data should be segmented. This problem one can find in different applications and therefore new segmentation techniques have been appeared in the literature during last years. In this paper we propose a new method of time series segmentation, i.e. extraction from the analysed vector of observations homogeneous parts with similar behaviour. This method is based on the absolute deviation about the median of the signal and is an extension of the previously proposed techniques also based on the simple statistics. In this paper we introduce the method of structural break point detection which is based on the Adaptive Regression Splines technique, one of the form of regression analysis. Moreover we propose also the statistical test which allows testing hypothesis of behaviour related to different regimes. First, the methodology we apply to the simulated signals with different distributions in order to show the effectiveness of the new technique. Next, in the application part we analyse the real data set that represents the vibration signal from a heavy duty crusher used in a mineral processing plant.