Predictive maintenance using PCA

Abstract This paper describes a suite of algorithms for a novel automated predictive maintenance system based on Principal Component Analysis (PCA). The system: (a) supports the automatic generation of concise, reliable models which describe the normal mode of operation of a monitored device; (b) identifies alarm conditions; and (c) conjectures the most likely cause of failure and computes a prediction of the device failure time. In on-line monitoring mode, these models will enable immediate alarming of conditions outside the normal mode of operation. Failure prediction will be based on the capability of the models to “fingerprint” potential failure branch types. The main application area is predictive maintenance; the approach is meant to identify the subset of all monitored devices that are most likely to fail within the prediction horizon. This information is important for the optimal scheduling of maintenance work.