Neuro-fuzzy modeling in model based fault detection, fault isolation and controller reconfiguration (DEL.4506)

The demand for increased efficiency, in various branches of industry, such as the manufacturing, aerospace, process and energy production industry leads to an increased degree of automation of the production process. In order to meet this demand while at the same time maintaining or improving the quality and safety, it is necessary that the (complex) control system is fault tolerant. This means that the effects of failures, such as component failures, actuator or sensor failures, on the quality and safety of the production process are minimized. The latter criterion is of course context dependent and may e.g. result in a safe shut down of the process unit avoiding casualties or in modifying the performance specifications to safely continue operation under minimal economical losses.

An important class of approaches to design fault tolerant control systems, consists of a so-called Model Based Fault Detection and Isolation (FDI) part whose task is to detect whether a failure has occurred and to localize it, and a Controller Reconfiguration (CR) part whose task is to modify the control structure or parameters to maximize production quality and safety. The two key elements in designing these two parts are the development of a mathematical model and a suitable decision mechanism to localize the failure and to select a new controller configuration. Despite the fact that Neural Networks have been widely investigated for this purpose, they have demonstrated a number of shortcomings, such as the lack of analytical tools to analyze their performances (stability, robustness, etc), that hamper their practical acceptance in fault diagnosis advisory systems.

In this project a neuro-fuzzy modeling methodology will be developed based on the accredited expertise of the applicants of this project. The methodology represents the mathematical model and derived observers as a composition of local models, each describing the system in a particular operation regime or failure mode. The switching between models is determined by fuzzy set membership functions, which may be modeled as neural networks and tuned by means of neuro-fuzzy learning algorithms. This approach exhibits a number of advantages that will be fully explored in the research project. One advantage is that the set of rules by which the model is represented can be interpreted linguistically by the operators. This allows to extend and update these models with their expert knowledge improving both the reliability of the models and their acceptance and validation. The i-node-switching model will be embedded in the framework of hybrid system theory and heterogeneous control theory, paving the way to put FDI and CR in a fundamental framework to analyze the robustness, stability and performance by means of analytical tools rather then by exhaustive simulation runs only. The advantages pursued in this research will enhance the acceptability of FDI and CR in industrial applications.

Because of the scale of the problem, the research will be performed by two Ph.D. students each concentrating on one of the two main research tasks in the project, one on the topic of neuro-fuzzy mathematical modeling for fault detection and one on the topic of decision making and controller reconfiguration. The integration of both activities is assured by tuning the time schedules of the workprogrammes proposed for each of the two main tasks. Finally, in order to speed up the transfer from the obtained research results to industry and to help focusing the project on relevant long term industrial objectives a close cooperation is proposed between the groups of the university and the applied research laboratories (ECN, TNO, NLR), who all have active research programs on the topic of Fault tolerant control.

An algorithm for controller reconfiguration for non-linear systems has been developed. It combines a multiple model estimator and a Generalized Predictive Controller (GPC). The set of models is built up, each corresponding to a different operating condition of the system. The algorithm is illustrated for two different case studies - one with a linear model of one joint of a space robot manipulator, subjected to failures, and one with a non-linear model of the inverted pendulum on a cart [1].

[1] S.K. Kanev and M. Verhaegen, "Controller Reconfiguration of Non-linear Systems", to be submitted to Automatica CEP, 1999.

Voor meer informatie kunt u de website van de vakgroep raadplegen.

Er zijn 4 bedrijven en 2 onderzoeksinstellingen bij dit project betrokken.

Dr.ir. M.H. Verhaegen
Dr. C. Witteveen
Technische Universiteit Delft
Faculteit ITS
Subfaculteit Eletrotechniek
Vakgroep Regeltechniek
Postbus 5031
2600 GA DELFT.

Titel van het onderzoek

Neuro-fuzzy modeling in model based fault detection, fault isolation and controller reconfiguration (DEL.4506).