Gary G. Yen
Oklahoma State University, USA

Syllabus
As technology advances and complexity of the systems grows, so does the required degree of system availability. At the same time, faults increase in chance of occurrence, diversity and severity of consequences. One of the three goals of this tutorial is to provide a comprehensive literature review on Fault Tolerant Control (FTC), a field of intelligent control research that emerges to address this dichotomy by specifically designing control algorithms capable of maintaining stability and performance despite the occurrence of faults.
　In order to achieve the required degrees of reconfiguration and stability, the adaptive controller can benefit greatly if more than the simple instantaneous difference between desired and actual states is available to be used as performance index. Due to the continuous interaction between the controller and the plant, the quality of a certain control strategy can only be fully measured after analyzing all future effects it has on the control mission. This kind of problem is the main focus of approximate dynamic programming, a machine learning technique that solves it through a backward search from the final step. To make the problem tractable to an on-line learning approach, adaptive critic designs (ACDs) introduce a critic block responsible for approximating the cost-to-go or strategic utility function. Such a function is defined by the Hamilton-Jacobi-Bellman equation and represents the core of dynamic programming.
　Heuristic dynamic programming (HDP) is the most straightforward application of ACD. In HDP, the critic block is trained forward in time trying to minimize error measure. The controller, usually referred to as action network in ACD, is then adapted in the direction of the minimization of cost function by using the critic to analyze the impact of its actions over the cost-to-go. Although ACD can be implemented with any differentiable structure, neural networks have been the best known candidate due to their generalization and nonlinear mapping capabilities as well as having suitable methods for on-line learning. Given the system of interest, dynamic or recurrent neural networks were chosen due to their more efficient handling of dynamic nonlinear mapping.
　Dual Heuristic Programming (DHP) reevaluates the purpose of the critic network and redesigns it with a different adaptation path, capable of generating smoother derivatives and has shown improved performance. The addition of a third neural network and a change in the training paradigm leads to the state-of-the-art adaptive critic architecture, Globalized Dual Heuristic Programming (GDHP).
In this tutorial, the second goal is to review five different neural control architectures in order of growing sophistication, starting from two basic neural control architectures and leading to three adaptive critic approaches. The merits of each will be thoroughly discussed and their shortcomings will be elaborated.
　Final goal of this tutorial is to apply the ACD to the proposed fault tolerant architecture capable of increasing the availability of complex nonlinear systems potentially subject to a wide range of fault scenarios. Motivated by an encompassing understanding in the areas of fault information extraction and FTC itself, the proposed hierarchical architecture is composed of three levels.
　The lowest level is composed of a baseline nonlinear reconfigurable controller that generates identification models and new control solutions for previously unknown faults. To implement such a controller as well as an identifier for fault modeling, a GDHP manages a set of three recurrent neural networks. The use of GDHP grants the architecture the power to preserve system stability and as much performance as possible in the presence of faults that may extend the order or add crucial nonlinearities to the dynamics of the system.
　Operating on a middle level, a novel supervisor increases the reconfiguration speed of the GDHP controller for abrupt faults known at design time as well as faults autonomously modeled and addressed online during a previous occurrence. Moreover, the supervisor also increases the stability of the online GDHP reconfigurable controller by preventing malfunctions within its training algorithm (that would lead to divergence or local minima convergence) from building up to the point of degrading the tracking performance of the plant. At the core of the supervisor, two innovative decision logics based on three quality indexes perform fault detection and diagnosis as well as controller malfunction detection.
　Overviewing the entire architecture, a fault development rule extraction algorithm is positioned at the highest level. Through information gathered from the GDHP controller, identifier and from the supervisor, this final component’s goal is to use all historical data from the system to build linguistic rules that inform the human mission planner (e.g., user, operator or pilot) of the probability that different fault scenarios have of taking place at particular future time frames. Once implemented, the fault development rule set will present crucial information to the mission planner when deciding if the desired trajectory of a mission should be altered after the occurrence of a minor fault to reduce the chance of a major, possibly irremediable, fault occurring.
　To substantiate the presented architecture, extensive simulation results will be presented, covering both the workings of specific components and the integration of the overall architecture. The power of the algorithm can be observed in the series of proof-of-the-concept simulated systems, ranging from SISO linear systems to MIMO nonlinear systems with unobserved states.

Speaker
Gary G. Yen received the Ph.D. degree in electrical and computer engineering from the University of Notre Dame, Notre Dame, IN, USA in 1992 under Professor Anthony N. Michel. He is now a Professor in the School of Electrical and Computer Engineering, Oklahoma State University, Stillwater, OK, USA. Before he joined OSU in 1997, he was with the Structure Control Division, U.S. Air Force Research Laboratory in Albuquerque, NM, where he was a technical lead and program manager for several Information Technology programs in deployable space vehicles. His technical contribution and leadership in “predictive maintenance,” “information assurance” and “condition based monitoring” based solely upon vibration and acoustic signatures, have been highly regarded in several DoD publications. His current research is supported by the DoD, DoE, EPA, NASA, NSF, and a Process consortium, amounted to over $2.5M. His research interest includes intelligent control, computational intelligence, conditional health monitoring, signal processing and their industrial/defense applications.

Dr. Yen was an associate editor of the IEEE Transactions on Neural Networks and the IEEE Control Systems Magazine during 1994-1999 and was an associate editor of Automatica during 2001-2004. He is currently serving as an associate editor for the IEEE Transactions on Control Systems Technology, IEEE Transactions on Systems, Man and Cybernetics and IFAC Journal on Mechatronics. Additionally, he serves as the editor for the IEEE conneCtIonS, a publication of the IEEE Computational Intelligence Society. Under the IEEE Control Systems Society, he served as Program Chair for the 2000 IEEE Conference on Control Applications held in Anchorage, AK and is the General Chair for the 2003 IEEE International Symposium on Intelligent Control held in Houston, TX (a conference of 250+ attendees). He will be the General Chair for the 2006 IEEE World Congress on Computational Intelligence sponsored by the IEEE Computational Intelligence Society to be held in Vancouver, Canada. WCCI, held every fours years, has been a conference of 1,200-1,800 attendees. On behalf of IEEE Robotic and Automation Society and later IEEE Control Systems Society, he was a Representative to the IEEE Neural Network Council Administrating Committee during 1995-2000. Dr. Yen has served as Chair for IEEE Control Systems Society Student Activities Standing Committee and Chair for IEEE Neural Network Council Neural Network Technical Committee from 2000-2003. In 2004-2005, he serves as the Vice President for Technical Activities within the IEEE Computational Intelligence Society. Dr. Yen has published over 50 peer-reviewed journal articles in this field of his research. He was recipient of Halliburton Outstanding Young Faculty and Outstanding Faculty Awards from Halliburton Foundation in 2001 and 2004, respectively.