This work presents an on-line diagnosis algorithm for dynamic systems that combines model based diagnosis and machine learning techniques. The Possible Conflicts method is used to perform consistency based diagnosis. Possible conflicts are in charge of fault detection and isolation. Machine learning methods are use to induce time series classifiers, that are applied on line for fault identification. The main contribution of this work is that
Possible Conflicts are used to decompose the physical system. This decomposition allows defining the input-output structure of an ensemble of classifiers. Hence the structural knowledge provided by the Possible Conflicts is exploited by the machine learning methods. Possible Conflict decomposition may be used for class selection or for class and attribute selection. Experimental results on a simulated pilot plant show that class selection has
an important potential to increase the classifier accuracy for several learning algorithms. The effect of an additional attribute selection depends on the kind of machine learning method, although it improves the accuracy of the most precise classifiers, especially at the first stages of the diagnosis processes, just after a fault has been detected.
