Prognostic and Health Management for Suspended Time-Series

Prognostic systems are expected to provide predictive information about the Remaining Useful Life (RUL) for equipment and components. During the last ten years, numerous RUL prediction models have been developed. These methods usually treat completed time-series only, i.e. full statistics before the item fails. Under actual operating conditions occasionally number of failed items is too small, and therefore application of uncompleted (suspended) time-series is necessary, and using Semi-Supervised methods instead of Supervised is required. In this paper, we propose an approach based on regression and classification models we have introduced in the past. These models consider monitoring data (time-series) as inputs and RUL estimation as output. Significant difference of this model is using suspended time-series to estimate optimal RUL for each suspended time-series, so they can be used for initial model training. This article describes the procedures that have been developed and applied successfully for Suspended Time-Series using. Several models based on modification of the SVR and SVC methods (Support Vector Regression and Support Vector Classification) are proposed for consideration. Number of uncompleted time-series used for training and cross-validation is proposed as additional control parameter. Suggested methodology and algorithms were verified on the NASA Aircraft Engine database. Numerical examples based on this database have been also considered. Experimental result shows that the proposed model performs significantly better estimations than pure supervised learning based model.