Anomaly Detection and Unsupervised Classification of Plasma Events

In nuclear fusion experiments, massive databases of time series signals store all relevant information about the plasma evolution. However, most of the physics knowledge remains hidden due to two factors. On the one hand, most measurements are taken by indirect methods and the inversion techniques are quite complex. On the other hand, the plasma shows highly non-linear interactions that are difficult to model. The ITER database is expected to store more than 1 Tbyte of data per discharge (about 1 million signals, mainly time series and video-movies) and the extraction of hidden knowledge will be essential. One of the most important aspects in ITER will be the automatic recognition of off-normal events as the plasma evolves. A crucial objective is the real-time determination of this kind of events. However, this is an extremely difficult task not tackled so far that requires off-line analysis of the databases to develop proper methods to be applied under real-time requirements. This work presents a three step off-line method to perform the temporal location of anomalies, the unsupervised grouping of events and the potential analysis of causal relationships. An example of the latter can be the sequence of events that can produce different types of plasma disruptions. The first step of the method is the recognition and temporal location of anomalies by analyzing multi-dimensional parameter spaces made up of plasma quantities. The second step consists of determining characteristic time lengths of such anomalies. The third step is the unsupervised classification of the events in order to assign labels to each class of physics event. The unsupervised classification into different classes is used to filter out those groups without statistical relevance.