Sistemas de clasificación automáticos con confianza y credibilidad en fusión termonuclear

Abstract

In this Thesis, the problem of associating reliability measures to automatic classification tasks in thermonuclear fusion devices has been tackled. To this end, a set of classifiers, whose implementation has been based on the conformal prediction theory, have been analyzed and applied to data generated in a specific fusion device, the TJ � II stellarator. In particular, the focus has been put on the set of images corresponding to the TJ � II�s Thomson scattering diagnostic. Also, it is important to note that the big amount of data stored in the databases of fusion devices makes essential the problem of selecting the most suitable samples to train a classifier. The aforementioned problem has been tackled in this Thesis from an active learning perspective. An active learning algorithm has been implemented to allow, with a reduced amount of training data, reaching both success rates and reliability measures that are better than, or as good as, the ones reached training the classifiers using a much bigger amount of randomly selected data. In fusion devices, more and more needs of real time signal classification are present. Bearing this kind of requirement in mind, methodologies for the reduction of the computational overload in the classification processes have been emphasized in this Thesis. This goal has been reached by two different ways. On the one hand, the implementation of variants of the classification method has been taken into account. On the other hand, the use of reduced but effective training sets has been analyzed.