Comparison and ensemble of temperature-based and vibration-based methods for machinery prognostics

This paper presents a comparison of a number of prognostic methods with regard to algorithm complexity and performance based on prognostic metrics. This information serves as a guide for selection and design of prognostic systems for real-time condition monitoring of technical systems. The methods are evaluated on ability to estimate the remaining useful life of rolling element bearing. Run-to failure vibration and temperature data is used in the analysis. The sampled prognostic methods include wear-temperature correlation method, health state estimation using temperature measurement, a multi-model particle filter approach with model parameter adaptation utilizing temperature measurements, prognostics through health state estimation and mapping extracted features to the remaining useful life through regression approach. Although the performance of the methods utilizing the vibration measurements is much better than the methods using temperature measurements, the methods using temperature measurements are quite promising in terms of reducing the overall cost of the condition monitoring system as well as the computational time. An ensemble of the presented methods through weighted average is also introduced. The results show that the methods are able to estimate the remaining useful life within error bounds of +-15\%, which can be further reduced to +-5\% with the ensemble approach.