Simulations of nuclear fusion diagnostics based on projections with Venn predictors and context drift detection

Tomography is used to reconstruct the inside cross section of an inaccesible object. To this end, a high number of cross section projections are required. In nuclear fusion devices, the plasma follows a toroidal shape and tomography can be used to estimate the two-dimensional spatial distribution of the plasma emission through a cross section. One of the objectives of tomography in plasma physics is to determine the number of localized emission peaks within the cross-section. Due to space restrictions in fusion devices, only a very limited number of projections can be experimentally measured and this introduces limitations to detect the number of emission peaks in the cross-section. This article describes a set of simulations to show that a multi-class classification system can be a valid alternative to tomography in plasma physics, even with the data of a single projection. Each class of the classifier represents the number of perturbations that can appear at any time in the plasma emissivity. To ensure the accuracy and reliability of the predictions, probabilistic classifiers (Venn predictors and Bayesian classifiers) have been used, with more reliable results from the Venn predictors. The probability intervals of Venn predictors with a single projection and intense/low respectively perturbations are [0.85, 0.93] and [0.62, 0.71] (to be compared with a probability 0.2 of a class random assignation). However, when the perturbation peaks are not strong enough and are lost into the emissivity background, the determination of the number of emission peaks can produce unreliable results. In these cases, to detect the plasma transition from an unperturbed state to a perturbed one (regardless of the number of perturbations) a martingale framework has been used. The changes have been detected by testing exchangeability with two different martingales: randomized power martingale and simple mixture martingale. Neither false alarms nor missed alarms have been observed with the former one and the average delay in the change recognition is between 16 and 72 samples (depending on cases). This means between 16 and 72 ms if the sampling period is 1 ms (it should be noted that the discharge length in ITER will be 50 min).