On Optimal Threshold Selection for Condition Monitoring

Abstract

Well designed features and properly selected detection thresholds are important prerequisites for reliable performance of the condition monitoring systems. Ideally, thresholds should be selected in a way that the diagnostic system keeps alert to the appearance of fault with minimal delay while under normal conditions false alarms should be avoided. If the thresholds are set too high, missed alarms may occur while too low values implicate false alarms. In practice thresholds are often set heuristically by a skilled person for each component of the machine. The task is nontrivial as usually many thresholds need to be defined. Moreover, a feature may be related to diverse faults with different sensitivity levels. Motivated by this issue, the intention of this paper is to lay the basis for rigorous threshold selection that implies the need for minimal design parameters. The idea is to first elaborate the probabilistic model of the feature. In order to check the relative change in the probabilistic pattern, the statistical hypothesis tests are employed. The only required priors, needed to tune the diagnostic algorithm, are data records collected under nominal condition and the probability of false alarm (PFA) as a sole “tuning knob”. Technically, the approach converts into the problem of statistical hypothesis testing. The performance of the algorithms is preliminarily confirmed via simulations and a real case study with a motor drive subjected to imbalance.