Bayesian Feature Discovery for Predictive Maintenance

Amir Dib, Charles Truong, Laurent Oudre, Mathilde Mougeot, Nicolas Vayatis, Heloïse Nonne

Abstract: This paper considers predictive maintenance, which is the task of predicting rare and anomalous events (typically, system failures) using event logs data, which are series of time-stamped symbolic codes emitted at regular or irregular intervals by a monitored system. Our objective is to find small sets of codes (called itemsets or patterns) that occur shortly before failures. Current prediction methods either produce patterns at a high computational cost or resort to kernel approaches which are often difficult to interpret. We introduce Bayesian Pattern Feature Discovery (BPFD), a new generic algorithm for pattern discovery. Our method, based on a pattern mining technique, produces informative and explainable features and is computationally efficient. The performance of BPFD is highlighted on real-world data sets, showing that enriching the feature space with the discovered patterns improves significantly the prediction power of a broad range of predictors and offers useful insight on the predictive maintenance task.