During a discharge at the TCV tokamak, the plasma be classified as varying between Low (L), High (H) and, in some cases, a temporary (intermediate) mode, called Dithering (D). In addition, while the plasma is in H mode, Edge Localized Modes (ELMs) can occur. The ability to accurately, and automatically, detect changes between these states, and ELMs, is considered important for future tokamak operation. However, it is difficult to design a traditional rulebased system that can accurately account for all the possible reasons behind these phenomena.The alternative is to use an approach whereby data generated in fusion experiments is used by algorithms which can, by themselves, learn the underlying rules that explain these events. Deep learning algorithms are exactly suited for this task. By feeding them with enough data, these models can automatically extract any existing correlations that allow for accurately detecting plasma confinement states and ELMs. In this work, we present a series of different deep learning algorithms for this task, namely, convolutional neural networks, recurrent neural networks, and sequence to sequence encoder-decoder models. The algorithms presented differ from each other with regards to the assumptions made regarding the data that they process, their architectures, and their capacity to accurately carry out the classification task. We will show, in particular, that a sequence to sequence model can achieve the best results, while also allowing for explicit incorporation of domain knowledge into the classification task.
[mehr]