Abstract
Assessment of system health and prediction of remaining useful life can be performed effectively through the evaluation of degradation levels configured by multiple states. Commonly, degradation progression is modeled according to the specific configuration using existing algorithms with assuming numbers and state conditions. However, due to the complexity, especially in the case of a system with multiple hidden states, the proper configuration is hard to assign and to identify. The need for unsupervised state estimation process to assist degradation modeling preventing under or over assumption becomes obvious. Among the existing approaches is the application of clustering methods to classify data and to estimate the number of degradation states might exist. However, integration into the lifetime prediction framework is still infancy and often not considered. Therefore, in this work, a previously developed state machine lifetime model is extended to allow flexibility in configuring state topology based on K-means clustering algorithm and cluster validity index for the optimal number of states identification. Combining unsupervised state estimation process with a new state machine lifetime model has transformed it into a semi-supervised prognostic approach. For validation, hydraulic pressure data from tribology experiment are deployed for training and test the algorithm. Based on the evaluation, this approach demonstrates the ability to improve health state assessment and lifetime prediction in a more flexible way to address the variability in the system.