Fault Prognostication is a 2017 new start funded by MBARI. The 3-year project will examine ways of improving persistence by improving the vehicle’s fault diagnostic abilities and by implementing ways of prognosticating system failure before it happens. The project has three components:
- development of fault diagnostic algorithms using unsupervised learning
- development of incipient fault prognostication algorithms
- development and testing of a hardware module.
Initial testing will take place on a benchtop version of the LRAUV augmented with additional sensing capabilities and signal processing algorithms:
- every subsystem has electrical, mechanical, and + software elements; to take advantage of this we’ll be tracking patterns of power consumption and patterns of vibration – eventually we’ll all track patterns of software performance.
- engineered systems are designed and integrated subsystem by subsystem, which should translate into weak coupling between the individual components’ behavioral patterns. To take advantage of this we’ll train each subsystem independently and use this independent representation as a starting point for learning the combined system.
Large-scale multi-scale persistent observation will require AUVs that can run failure-free for thousands of hours. Achieving this will require advanced health monitoring and prognostication technology. We plan to field a first prototype in 2019.[/caption]