Threshold lines identification for non-Gaussian distributed diagnostic features

Machine condition monitoring systems are frequently used in the industry, especially for critical infrastructure. Decision making is still challenging due to the lack of limit values (called also thresholds). Identification of thresholds is in particular difficult for unique machines with specific Health Index (HI) data properties. Our contribution is a procedure for threshold values identification for HI data with time-varying characteristics and non-Gaussian behaviour. The proposed methodology consists of few crucial steps, such as data segmentation, data modelling, and Monte Carlo simulations for quantile lines identification based on the fitted models. The proposed methods are based on robust statistics dedicated to non-Gaussian distributed data. Novelty of the paper is related to the extension of simple Trend+Noise model for HI data and model-based framework for threshold lines estimation. Thus, the method provides more accurate results for non-Gaussian, time varying data. We demonstrate the superiority of the proposed procedure over the classical ones. The presented simulation study clearly indicates the efficiency of the approach for non-Gaussian HI data. The methodology is applied to two real datasets with different properties (level of non-Gaussianity and presence of interdependence in the random part). We believe that proposed procedure could be considered as a guideline for future research and may be implemented in commercial monitoring systems.

Automated summary

This study proposes a method for identifying thresholds for health index (HI) data with time-varying characteristics and non-Gaussian behavior. The methodology includes data segmentation, data modeling, and Monte Carlo simulations for quantile lines identification based on fitted models. The proposed method provides more accurate results for non-Gaussian, time varying data.