Development of Deep Learning Based Prognostics

Rotating components such as bearings and gears play a very crucial role in virtually all mechanical equipment. Over time these parts degrade and eventually need to be replaced. The role of prognostics is to predict the remaining useful life of these components. Having an accurate method for doing so improves both safety and efficiency. The majority of current methodologies require much expertise in both signal processing and prognostic modeling. Due to the somewhat recent success of many deep learning based approaches, this dissertation explores how deep learning can be incorporated into prognostics in order to automatically extract important and useful features that can be used for prognostics. These deep learning methods avoid the need of much prior prognostic experience. A restricted Boltzman machine and a deep belief network are utilized in order to estimate the remaining useful life of rotating components. In addition, the deep learning based approaches are combined with a particle filtering approach to capture the uncertainty of the remaining useful life estimations. A Mixture Density Network and an ensemble of deep belief networks are used to model the different probability density functions required for use in a particle filter. These methods are tested and validated using real bearing and gear run-to-fail tests. The test results show state of the art results when compared to traditional prognostic approaches and show the promising potential for deep learning based prognostic models.