A Framework for Bearing Prognostics in Changing Operating Conditions

In this paper, a framework for conducting data-driven prognostics presence of a domainshift is introduced. Domain shift is brought about by factors such as changing operations which alterthe distribution of data thus degrading the performance of learned prognostic models. The use ofWeibull-based hazard functions of the kurtosis and shape factor time domain features are exploredand affirmed as not only being trendable and monotonic but robust across operating conditions; thus,desirable for prognosis. In the learning stage, the usual procedure is to use the designated training datawhich is full lifetime data. Unfortunately, a key characteristic of test data is that its truncated whichmay offer a serious impediment to the predicting model’s performance due to it being trained with fulllifetime data only. In this research, the learning data is extended by adding its truncated versions thussignificantly improving the model’s prediction accuracy for test data. In this work, it is demonstratedthat the proposed method results in over 95% accuracy of remaining useful life prediction for bearingsoperated in different operating conditions.