The paper presents some results of research work in the field of artificial neural networks (ANN) applied to nuclear safety. It shows how a priori knowledge in the form of qualitative physical reasoning can provide a powerful basis for designing a set of ANN-based detection subsystems. In particular, it explains how each ANN is in charge of modelling a physical relationship between a set of state variables (thermal balance, mass balance, etc.) by trying to predict one particular variable from other ones; then, the residual signal, defined by the difference between the predicted value and the real one is used to decide whether abnormalities are present. As far as the decision logic is concerned, the paper describes how robustness can be improved by adequate filters on the residuals. The proposed approach is then validated on data coming from a fullscope simulator of one of the Belgian nuclear power units: the neural-based detection system is trained on “normal” scenarios and is able, after learning, to detect reliably and rapidly most of the incidental situations chosen as tests.