The Plasma Disruption Prediction in Tokamak Reactors: Design of an Alarm System by Fuzzy Neural Approach

sciencediscussionAI and Robotics

Oct 20, 2013 (3 years and 9 months ago)

62 views

The Plasma Disruption Prediction in Tokamak Reactors:

Design of an Alarm System by Fuzzy Neural Approach


Francesco Carlo Morabito, Mario Versaci

Associazione Create
-
Enea
-
Euratom,University of Reggio Calabria, Faculty of Engineering, DIMET

Via Graziella, L
oc. Feo di Vito, Reggio Calabria, Italy


Aim of the work

Disruption in a tokamak reactor is a sudden loss of confinement and transfer of plasma energy to the surrounding structure.
During a disruption, the plasma current and the thermal energy content of a

tokamak plasma discharge collapse in an
uncontrollable way, thereby generating mechanical forces and heat loads which threaten the structural integrity of surroundin
g
structures and vacuum vessel components. The objective of the paper is to design an alar
m system for detecting the onset of a
disruption in tokamak plasma discharges in due time for control actions to be taken. In order to overcome the real time
operational requirements and with the aim of using a very large number of potentially predicting v
ariables, Neural Network
(NN) models seem to be the most viable tool. NN tool is also more amenable to hardware implementation of the alarm system.
The paper also focuses on analysing the impact of the measurement noise in the model.


Method (Fuzzy Neural

Approach)

A group of available diagnostic signals of different physical nature is used in order to determine if the plasma is approachi
ng
the disruption limit. The output of the procedure is the estimated time to disruption. A theoretic limit based on phy
sical
approximations can be used for performance comparison. We firstly use a static NN trained on a suitable database of real
experimental data based on both non
-
disruptive and disruptive discharges. A dynamic NN model aiming to overcome the
limitations o
f the static models coming from the possibility of facing different types of disruptions not included in the main
database was also studied. Finally, since some physical knowledge could advantageously be included in the network
framework by using the fuzzy

system approach, the original model was accordingly modified by obtaining a fuzzy neural
system. The paper will also present the way in which the two aspects of the model were combined. The use of fuzzy logic
concept is suggest by the consideration that p
revious techniques make use of expert knowledge for deciding about the onset of
a disruption.


Results

The proposed NN based model far exceeds the performance of standard techniques. The NN approach has been successfully
applied for solving different probl
ems to be faced during work, namely, the detection of redundancy in the input data; the
ranking of the input variables, and finally, the automatic extraction of rules from the database to be proposed to experts fo
r
further interpretation. The most interest
ing result achieved was that the NN approach allows to automatically classify the
different kinds of possible disruptions. The figures reported below show the results achieved in ranking the sensors by usin
g a
technique based on fuzzy curve concept and th
e qualitative plot used for classifying different disruptive shots.

The work is completely new for the part regarding the combination of fuzzy techniques and NNs, and also for the use of
unsupervised techniques for generating classification. In this sens
e, the paper extends the work carried out in USA, in Orincon
Corporation, which uses multilayer perceptron to foresee prediction.


Conclusions

This paper proposes a really working NN
-
based tool for prediction of incipient disruption in tokamak plasma physi
cs
applications.




Fig. 1


R慮king of 瑨攠inpu琠s敮sors

䙩F. 2


P汯琠tf 瑨攠楮puts for qu慬楴慴av攠捬慳s楦楣慴楯n