Machine condition change detection based on data segmentation using a three-regime, [formula omitted]-stable Hidden Markov Model

Abstract

An adaptation of HMM for signal segmentation is proposed. It is developed for Health Index (HI) data, which behave in different ways depending on machine condition. HI can be constant, then linearly or exponentially increasing. Knowledge of the current regime is critical for the early stage, as well as developed fault detection and RUL prognosis. The analysed data have time-varying variance/scale and noise changes from Gaussian to non-Gaussian. We initially propose an HMM with α-stable distribution and a three-parameter model. We analyse three versions of HMM: for untransformed data (model ‘G’), for detrended data (‘D’) and additionally rescaled data (‘DN’). We show that model ’D’ is sufficient to find the first change point (detection of early stage damage), while model ’G’ is the most effective for the identification of the second change point. The proposed and verified approach allows us to effectively find the borders between the segments.