Bayesian based methodology for the extraction and validation of time bound failure signatures for online failure prediction

Increasing demand volume and diversity have led the emergence of high-mix low-volume production lines where success requires sustainable production capacities. However, equipment breakdowns significantly reduce and disrupt these capacities. This give rise of interest in developing methodologies to avoid failures by treating their respective causes, prior to the failure occurrences. In this paper, we present a methodology to extract and validate rules (and patterns) as time bound failure signatures for real time failure predictions, using Bayesian approach. In comparison to existing approaches to learn and extract failure signatures, the presented methodology offers extraction, selection and validation of rules/patterns which is linked to sufficient time to execute corrective and proactive measures to avoid failures (the time bound). Moreover, proposed methodology uses event driven contextual information from product, process, equipment and maintenance data sources, instead of relying only on sensor data. This is to avoid sensor biases, as decision support equipment/module levels and the fact that failure source is not necessarily the equipment which could result in over engineering. This methodology is tested and extracted rules are validated using data collected from a world reputed semiconductor manufacturer.