Disruptions are sudden and unavoidable losses of confinement that may put at risk the integrity of a tokamak. However, the physical phenomena leading to disruptions are very complex and non-linear and therefore no satisfactory model has been devised so far either for their avoidance or their prediction. For this reason, machine learning techniques have been extensively pursued in the last years. In this paper a real-time predictor specifically developed for JET and based on support vector machines is presented. The main aim of the present investigation is to obtain high recognition rates in a real-time simulated environment. To this end the predictor has been tested on the time slices of entire discharges exactly as in real world operation. Since the year 2000, the experiments at JET have been organized in campaigns named sequentially beginning with campaign C1. In this paper results from campaign C1 (year 2000) and up to C19 (year 2007) are reported. The predictor has been trained with data from JET's campaigns up to C7 with particular attention to reducing the number of missed alarms, which are less than 1%, for a test set of discharges from the same campaigns used for the training. The false alarms plus premature alarms are of the order of 6.4%, for a total success rate of more than 92%. The robustness of the predictor has been proven by testing it with a wide subset of shots of more recent campaigns (from C8 to C19) without any retraining. The success rate over the period between C8 and C14 is on average 88% and never falls below 82%, confirming the good generalization capabilities of the developed technique. After C14, significant modifications were implemented on JET and its diagnostics and consequently the success rates of the predictor between C15 and C19 decays to an average of 79%. Finally, the performance of the developed detection system has been compared with the predictions of the JET protection system (JPS). The new predictor clearly outperforms JPS up to about 180 ms before the disruptions.

Disruptions remain one of the most hazardous events in the operation of a tokamak device, since they can cause damage to the vacuum vessel and surrounding structures. Their potential danger increases with the plasma volume and energy content and therefore they will constitute an even more serious issue for the next generation of machines. For these reasons, in the recent years a lot of attention has been devoted to devise predictors, capable of foreseeing the imminence of a disruption sufficiently in advance, to allow time for undertaking remedial actions. In this paper, the results of applying fuzzy logic and classification and regression trees (CART) to the problem of predicting disruptions at JET are reported. The conceptual tools of fuzzy logic, in addition to being well suited to accommodate the opinion of experts even if not formulated in mathematical but linguistic terms, are also fully transparent, since their governing rules are human defined. They can therefore help not only in forecasting disruptions but also in studying their behaviour. The analysis leading to the rules of the fuzzy predictor has been complemented with a systematic investigation of the correlation between the various experimental signals and the imminence of a disruption. This has been performed with an exhaustive, non-linear and unbiased method based on decision trees. This investigation has confirmed that the relative importance of various signals can change significantly depending on the plasma conditions. On the basis of the results provided by CART on the information content of the various quantities, the prototype of an adaptive fuzzy logic predictor was trained and tested on JET database. Its performance is significantly better than the previous static one, proving that more flexible prediction strategies, not uniform over the whole discharge but tuned to the operational region of the plasma at any given time, can be very competitive and should be investigated systematically.

The importance of predicting the occurrence of disruptions is going to increase significantly in the next generation of tokamak devices. The expected energy content of ITER plasmas, for example, is such that disruptions could have a significant detrimental impact on various parts of the device, ranging from erosion of plasma facing components to structural damage. Early detection of disruptions is therefore needed with evermore increasing urgency. In this paper, the results of a series of methods to predict disruptions at JET are reported. The main objective of the investigation consists of trying to determine how early before a disruption it is possible to perform acceptable predictions on the basis of the raw data, keeping to a minimum the number of ‘ad hoc’ hypotheses. Therefore, the chosen learning techniques have the common characteristic of requiring a minimum number of assumptions. Classification and Regression Trees (CART) is a supervised but, on the other hand, a completely unbiased and nonlinear method, since it simply constructs the best classification tree by working directly on the input data. A series of unsupervised techniques, mainly K-means and hierarchical, have also been tested, to investigate to what extent they can autonomously distinguish between disruptive and non-disruptive groups of discharges. All these independent methods indicate that, in general, prediction with a success rate above 80% can be achieved not earlier than 180 ms before the disruption. The agreement between various completely independent methods increases the confidence in the results, which are also confirmed by a visual inspection of the data performed with pseudo Grand Tour algorithms.

No abstract

This paper describes a pattern recognition method for off-line estimation of both L/H and H/L transition times in JET. The technique is based on a combined classifier to identify the confinement regime (L or H) at any time instant during a discharge. The classifier is a combination of two different classification systems: a Bayesian classifier whose likelihood is computed by means of a non-parametric statistical classifier (Parzen window) and a support vector machine classifier. They are combined through a fuzzy aggregation operator, in particular the Einstein sum. The success rate achieved exceeds 99% for the L to H transition and 96% for the H to L transition. The estimation of transition times is accomplished by following the temporal evolution of the confinement regimes.