One of the main factors limiting the implementation of neural networks in industrial applications is the difficulty of detecting potentially unreliable outputs. This could be the case of the neural disruption predictor installed in JET, where new plasma configurations might present features completely different from the ones observed in the experiments used in the training set. This 'novelty' can lead to incorrect behaviour of the network. A Novelty Detection method, which determines the novelty of the output of the neural network, can be used to assess the network reliability. In this paper, two approaches to Novelty Detection are tested, i.e., Self Organising Maps and Support Vector Machines. Preliminary results are encouraging, in particular when referring to false alarms.