The increasing use of composite materials on aircraft structures as well as their increasing average age have led to the search and the development of several global nondestructive testing techniques to scan large portions of the aircraft externally. One such technique is Infrared thermography. If rapid inspection can be expected, the size of the data and the complexity of the thermograms make the interpretation difficult. So in order to help the operator in the fulfilment of his job to achieve rapid, reliable and repeatable non destructive evaluation, we have caried out for the last four years a project named SEQUOIA, in which Artificial Intelligence has been integrated. The first approach presented at QNDE 93 was based on spatial analysis which revealed itself to be encouraging but insufficient and with not enough versatility [1]. A complementary approach is presented, it is based on the use of multi-layered Neural Networks. This classification technique is used to correlate temporal thermal signatures with sound and defected regions of an inspected part. As thermal modelling is now well developed and comprehensive, the investigative study relies on the training of the neural network on theoretical thermograms so that we can produce as many examples as one can think of. Different inputs for the neural network have been studied: raw data (temperature curves), derived data (derivative of temperature curves), contrast data (subtraction of reference from raw data). Multi-layer neural networks, as well as related algorithms such as Nearest Neighbour (KNN, Kmeans) and Learning Vector Quantization (L.V.Q) have been tested. The evaluation of the neural network process has mainly been based on its ability to reduce the errors, prior to uncertainties.