Geometrical Kernel Machine for Prediction and Novelty Detection of Disruptive Events in TOKAMAK Machines

Abstract

This paper presents a recently introduced Kernel\ud Machine, called Geometrical Kernel Machine, used to\ud predict disruptive events in nuclear fusion reactors. The\ud algorithm proposed to construct the Kernel Machine is able\ud to automatically determine both the number of neurons and\ud the synaptic weights of a Multilayer Perceptron neural\ud network with a single hidden layer. It has been demonstrated\ud that the resulting network is able to classify any finite set\ud of patterns defined in a real domain. The prediction\ud problem has been here modeled as a two classes classification\ud problem. The geometrical interpretation of the\ud network equations allows us both to develop the disruption\ud predictor and to manage the so called ageing of the kernel\ud machine. In fact, using the same kernel machine, a novelty\ud detection system has been integrated in the predictor,\ud increasing the overall system performance, and the reliability\ud of the predictor