Multiple Model Fault Tolerant Control Using Globalized Dual Heuristic Programming

An advanced reconfigurable controller improved by a multiple model architecture is proposed as a tool to achieve fault tolerance in complex nonlinear systems. The most complete adaptive critic design, Globalized Dual Heuristic Programming (GDHP), constitutes a highly flexible nonlinear adaptive controller responsible for the generation of new control solutions for novel plant dynamics introduced by unknown faults. Working on a higher hierarchical level, the proposed supervisor makes use of two quality indexes to perform fault detection, identification and isolation based on the knowledge stored in a dynamic model bank. In the event of abrupt known faults, such knowledge is then used to greatly reduce the reconfiguration time of the GDHP controller. The synergy of the proposed supervisor and GDHP goes beyond, as solutions designed by the controller to previously unknown faults are autonomously added to the model bank. The fine interrelations among the algorithm's subsystems are illustrated in a rich numerical simulation of a MIMO nonlinear system subject to different fault scenarios. Index T e r m s fault tolerant control, fault detection and diagnosis, neural networks, adaptive critic, multiple models I. I n t r o d u c t i o n The growing complexity of physical plants and control missions inevitably leads to increasing occurrence, diversity and severity of faults. Take automated production process as an example, availability is now considered to be the single factor with the highest impact on profitability [1]. Therefore, availability, defined as the probability that a system will operate satisfactorily and effectively at any point in time [2], becomes one of the key design criteria of a system. Fault Tolerant Control (FTC) is a field of research that emerges to increase availability by specifically designing control algorithms capable of maintaining stability and performance despite the occurrence of faults [3]. Although some fault scenarios are known at design time and can even be predicted if enough information is available, it is not reasonable in real-world applications to assume knowledge of the dynamics of the plant over all possible fault scenarios. In addition, while an approximate linear model can often be derived for a plant operating close to its nominal point, nonlinearities introduced or augmented by a fault after its occurrence can become of paramount importance to achieve a successful new control solution [4]. Therefore, a complete FTC architecture must present adaptive capabilities for the online generation of new nonlinear control solutions in response to unknown fault scenarios. Neural Network (NN) has been well regarded as an effective tool in function approximation due to its advanced nonlinear mapping capability [5]. The questions that arise when designing a NN adaptive controller are how to perform online learning in an effective manner for the most diversified scenarios, and how to guarantee stability. Given the present state-of-the-art control designs however, adaptation flexibility and stability are still conflicting characteristics and therefore a suitable compromise must be reached. Adaptive critic designs have shown to implement useful approximations of dynamic programming, a method for determining optimal control policies in the context of nonlinear plants [6]. As a candidate solution for the FTC problem, adaptive critic designs are capable of achieving superior performance by combining three different NNs: an identification NN that produces information of the dynamics of faults as they occur, a critic NN to estimate the impact of the underlying control strategy over time; and an action NN to generate real-time control actions to accommodate faults as they emerge. In this paper, Globalized Dual Heuristic Programming (GDHP), appointed as the most complete and powerful adaptive critic design [7], was implemented. Even though a reconfigurable adaptive controller is a key element without which solutions for unknown faults cannot be designed online, if used as a FTC architecture alone, it displays two major limitations. The first involves the fact that a reconfigurable controller makes it impossible for any available fault knowledge to be incorporated during design time. Although an ideal reconfigurable controller will always reach a solution (given its existence) for a given fault scenario, the amount of time it must be allowed to learn the new dynamics and modify itself accordingly could be greatly reduced by direct application of a known solution. The second major limitation is caused by the known tradeoff between adaptability and long-term memory. As a reconfigurable controller is optimized to deal with a broader scope of faults with minimum reconfiguration time, previously configured controllers are forgotten and the reconfiguration process has to be repeated even when returning to the healthy condition from an intermittent fault scenario. In [8] it was shown that implementing a reconfigurable controller in a Multiple Models Architecture (MMA) has the potential to overcome the cited limitations for the tracking of complex nonlinear plants. Since then, MMA has been applied to FTC by combining fault scenarios and their respective control solutions in model banks 0-7803-7891-1/03/$17.00 © 2003 IEEE 523 coordinated by supervisory systems. However, most publications so far are based on fixed model banks built offline and therefore are incapable of improving the controller response in the reoccurrence of faults that were unexpected during design time. In [9] a Dynamic Model Bank (DMB) was used to allow the insertion of new plant dynamics as they were identified online, but the use of a linear controller and the lack of a complete Fault Detection and Diagnosis (FDD) scheme significantly limited its applicability. ~] Supervisor El Dynamic Model Bank ....... i "1 i