The state-of-the-art diagnosis and estimation algorithms rely on hybrid continuous and discrete system models, but defining these models is a significant challenge. Furthermore, system dynamics may change; hence acquiring hybrid models from observed data is an appealing approach. While there is a rich body of research on system identification for continuous dynamic systems, hybrid model learning remains an open problem. My research has developed a solution to this problem for a restricted class of stochastic hybrid systems, described here.
"Model Learning for Switching Linear Systems with Autonomous Mode Transitions." L. Blackmore, S. Gil, S. Chung and B. C. Williams. In the proceedings of the Control and Decision Conference, 2007.