Date of Award
Master of Science
Deep learning consists of various machine learning algorithms that aim to learn multiple levels of abstraction from data in a hierarchical manner. It is a tool to construct models using the data that mimics a real world process without an exceedingly tedious modelling of the actual process. We show that deep learning is a viable solution to decision making in mechanical engineering problems and complex physical systems.
In this work, we demonstrated the application of this data-driven method in the design of microfluidic devices to serve as a map between the user-defined cross-sectional shape of the flow and the corresponding arrangement of micropillars in the flow channel that contributed to the flow deformation. We also present how deep learning can be used in the early detection of combustion instability for prognostics and health monitoring of a combustion engine, such that appropriate measures can be taken to prevent detrimental effects as a result of unstable combustion.
One of the applications in complex systems concerns robotic path planning via the systematic learning of policies and associated rewards. In this context, a deep architecture is implemented to infer the expected value of information gained by performing an action based on the states of the environment. We also applied deep learning-based methods to enhance natural low-light images in the context of a surveillance framework and autonomous robots. Further, we looked at how machine learning methods can be used to perform root-cause analysis in cyber-physical systems subjected to a wide variety of operation anomalies. In all studies, the proposed frameworks have been shown to demonstrate promising feasibility and provided credible results for large-scale implementation in the industry.
Kin Gwn Lore
Lore, Kin Gwn, "Deep Learning for Decision Making and Autonomous Complex Systems" (2016). Graduate Theses and Dissertations. 15965.