Fusion reactors will need to be operated in a stable regime without transient events such as plasma disruptions and (large) ELMs. From the point of view of machine learning, such events can be characterized as anomalous and specialized techniques exist to detect them by monitoring various diagnostic signals. If this can be done sufficiently in advance of their predicted occurrence, the plasma control system may have sufficient time to react and avoid or mitigate these events. Particularly dangerous are plasma disruptions, which are major instabilities causing excessive thermal loads and large transient forces that are potentially harmful to the operation of a tokamak and that will have to be avoided or mitigated in ITER. Disruptions can be caused by various complex mechanisms and it is very important to be able to predict them in real time and to recognize the disruption cause. We have recently developed a disruption classifier based on the discrimination of the distribution of wavelet coefficients corresponding to several predictor signals. We are also working on a reliable discrimination between different disruption types. In addition, research is ongoing in collaboration with ITER to establish a new set of predictor quantities (features) with a clear link to disruption physics and minimal dependence on machine characteristics. Finally, similar machine learning techniques are being used for predictive maintenance of subsystems and components in fusion devices.
