Toward effective utilization of similarity based residual life prediction methods: Weight allocation, prediction robustness, and prediction uncertainty

Abstract

Similarity based residual life prediction method is an emerging method for component residual life prediction. Studies on (a) the effect of weight function on prediction accuracy; (b) prediction robustness; and (c) prediction uncertainty of such method are rare. However, the abovementioned factors are essential concerns for wide application of a similarity based residual life prediction method. In this article, the essential elements of a similarity based residual life prediction method is outlined first with an extended weight function introduced. Afterward, an evaluation framework for investigating the prediction robustness of a similarity based residual life prediction method is established. In addition, a prediction uncertainty estimation method is proposed based on historical samples, inspired by cross-validation technique. In an extensive numerical investigation, a comparative study on the effect of weight function on prediction accuracy is conducted by tuning the parameters in the weight function. The prediction robustness of the similarity based residual life prediction method is evaluated in comparison with a time-series forecasting based residual life prediction method. Finally, the proposed prediction uncertainty estimation method is illustrated, which may facilitate further application of the similarity based residual life prediction method.