An approach to model-based fault detection is described which is suitable for nonlinear plants that are information poor. Such plants may have a bare minimum of sensors available with which to operate the process, the sensors may output at frequencies which are low relative to the dynamics of the plant, a complete set of possible failure modes may not be identified and there may be considerable uncertainty surrounding any models that are available. The approach is based on two principles, a principle of redistribution and a principle of a minimum number of explanations. By arguing that faults and modelling inaccuracies manifest themselves as an erroneous redistribution of mass, energy and so on throughout a plant, the former enables the diagnostician to relate to a plant simulation at a more qualitative level whilst maintaining laws of conservation. The latter argues that the diagnostician’s main task is that of identifying suitable plausible candidate sets of a few faults and model inaccuracies. This leads to the construction and analysis of simple regression models. It is demonstrated how this might be applied to one particular application, that of near-real time materials accountancy in nuclear fuel reprocessing plants.