Modeling of probability distribution functions for automatic threshold calculation in condition monitoring systems

Abstract

A suitable setting of threshold levels has been a dilemma for engineers in a number of science and industrial fields. It is a common problem in monitoring, where deviation from a correct state needs to be detected. As the number of monitored values in modern systems reaches thousands, threshold calculations became a significant yet frequently underestimated concern. Due to a prohibitive cost of manual threshold setting, many systems generate thousands of false alarms consequent upon running on default threshold levels. In the paper, the authors illustrate a methodology for automatic threshold calculations in a large monitoring system. The paper is mainly addressed to engineers and machine monitoring systems developers, therefore selected statistical topics were treated briefly with main focus on practical solutions. Two fundamental data types are considered; namely, vibration signal measures, which can be extended to any nonnegative data, and symmetrical process values. As it is shown, these data types have significantly different probability distributions. Since real data seldom fits the Gaussian model, an investigation of several distributions and their comparison is presented. The proposed approach is validated on four datasets including process values from a gas compressor and vibration signal measures from a wind turbine.