An Operation Condition-Matched Similarity Method for Remaining Useful Life Estimation with Dynamic Sample Fusion

Abstract

Remaining useful life (RUL) estimation is a key technology in prognostics and health management (PHM). Considering the problem that operating condition (OC) is easily overlooked and sample fusion is mainly determined by experience, this paper presents an OC-matched similarity method with dynamic sample fusion. The method contains two main stages, including training stage for obtaining the library of OC-based degradation models and the trained parameter for dynamic sample fusion, and testing stage for RUL prediction. In the training stage, we extract the sensor data on account of the linear correlation coefficient and expand the library of degradation models through adding OC information. Then, cross-validation is implemented to train the parameter for dynamic sample fusion and parameter is optimally selected by minimizing the target function. When estimating RUL of test data, OC-matched similarity is measured by calculating the distance between test data and OC-matched model. Eventually, RUL is estimated by the weighted average of each sample based on the similarity measurement. This method is validated by the 2008 PHM Conference Challenge Data, which contains both sensor measurements and operating settings. The results have suggested significant improvement comparing with traditional similarity method.