Novelty Detection For Conditioned Based Maintenance of Mechanical Systems Using Neural Based Auto-Encoders

We propose to develop inexpensive, self-adapting, anomaly detection algorithms and to apply them to critical monitoring tasks. Specifically in Phase I we focus on mechanical failure of pumps which are used in key Navy systems, for example, for ship board fire pumps. Prior experience with data sets from Helicopter transmissions and induction motors suggest that similar such anomaly detection systems would also be successful with monitoring the mechanical and hydralic behavior of pumps. The key technical approach is taken from earlier work suggesting that Hippocampal function has a similar basis in novelty detection as engineering work which identified Auto-encoders as potentially important technology for representing the "state" of mechanical or non-mechanical systems. The physical properties of such systems once properly captured can provide the Auto-encoder neural networks the opportunity to capture low-dimensional structure which can be used for model-based decisions of anomalous states not yet encountered by the network. Consequently, these algorithms adapt themselves to complex systems, develop an internal model of the signals from the system, and automatically issue an alert when the monitored system enters a heretofore unseen state (i.e. when the signals differ significantly from the norm), potentially signifying imminent failure or some other dangerous or costly state of affairs.