Ph.D. received on: 8/7/1997

E-mail: magnino@istel.ing.unipg.it

Tutor: Prof.  Michele La Cava, Università degli Studi di Perugia
 
 

 ___________________________________________________________________________________________________________

Application of neural networks in control system fault diagnosis   ___________________________________________________________________________________________________________

Advisor:

Prof. Michele La Cava, Università degli Studi di Perugia

Summary:

Fault diagnosis has become a topic of primary importance in modern process automation as it gives the basis for the realization of fault tolerant systems. In this work, the chances of application of artificial neural networks in control system fault detection and diagnosis are examined. In this analysis different kinds of system are taken into account going from more or less complex stationary systems to linear and non linear dynamic systems. Different diagnosis architectures are proposed according to the considered system and they are designed to employ:
- neural networks as configuration classifiers;
- neural networks for identification of the “normal model”, fault free, of the system under monitoring (particularly for non linear dynamic systems) to allow residual generation to perform detection and diagnosis of system failures.
For this purpose, some neural network architectures, like backpropagation, radial basis function networks and Kohonen maps, are examined in detail along with different training and testing algorithms. Advantages and drawbacks of the examined approaches are shown by means of examples of application of the proposed diagnosis architectures to different kinds of system.
 

 _______________________________________