Forecasting Failures in a Data Stream Context Application to Vacuum Pumping System Prognosis

This paper presents a local pattern-based method for forecasting failures in a data stream context. It also details a successful application to complex vacuum pumping system prognosis. More precisely, using historical data, the behavior of a set of pumping systems is first modeled by extracting a given type of episode rules, namely the First Local Maximum episode rules (FLM-rules). Each rule comes along with its proper temporal information: its optimal temporal window width. The most reliable FLM-rules are then selected to further forecast system failures in a data stream context. A forecast time interval is supplied for each forecasted failure by merging the temporal information of FLMrules. The results obtained for production data are very encouraging. Failures are predicted with a good temporal accuracy and precision while very few false alarms are generated. The method presented in this paper is patented and it is being deployed for a customer of the semi-conductor