Support Vector Machines for the disruption prediction and novelty detection at JET

Disruptions are an endemic and likely unavoidable aspect of tokamak operation that poses serious problems to the integrity and the machine lifetime. For that reason in the last 15 years, there have been several studies for disruptions prediction, most of which using neural networks. One of the major drawbacks of this approach is that the network performances normally deteriorate when new plasma configurations are presented to the network, in particular the number of false alarms increases. A possible improvement can be achieved using Novelty Detection techniques, reducing in particular false alarms. Recently, such a Novelty Detection system has been implemented to integrate the disruption predictor proposed for JET. One of the issues to be investigated is that the two systems worked separately and the novelty detector was only able to capture the information present in the database, but no information on what the network learnt about that database. In this work, both the prediction and the novelty detector tasks are performed by the same system using a Support Vector Machine (SVM) instead of a Multi-layer perceptron neural network. A SVM, which is able to find an optimal separating hyper-plane that maximizes the margin between itself and closest data points, is trained to detect the presence or the absence of disruption symptoms, in a certain time instant, by means of selected measured plasma diagnostic signals. In a SVM classifier/predictor the analysis of the distance between input data and support vectors in the feature space gives useful information about the novelty of an input. Moreover, the analysis of the predictor output, corresponding to new input data, gives useful information about the reliability of the predictor output, during on-line applications. The aim of this new approach is to demonstrate how the predictor-novelty detector is able to automatically update itself and if it is possible to improve its performances during lifetime operations. This work will show the performance of this new technique and the effects of the novelty detection on false and missed alarms using a database starting from 27/10/1999(pulse n. 49141) to nowadays. Not only are considered all classes of disruptions but also we do not look for any particular behaviour already reported in literature. (author)