In the context of numerical software verification, a major difficulty is to handle correctly and precisely the inputs of the program. These inputs are given by sensors and suffer from two kinds of uncertainty: their value lie within a non deterministic range and they are affected by a probabilistic noise. Thus, if one wants to rigorously and precisely model such sensors, it is necessary to mix, in the static analysis of programs, non-deterministic and probabilistic behaviors. In this talk, I will present the first results we obtained in this direction. More precisely, I will present how P-Boxes can be seen as an abstract domain in the sense of the abstract interpretation theory and why they suffer from the same drawbacks as intervals. Then I will present an evolution of affine sets that include both probabilistic and non-deterministic behaviors, while keeping a good precision by considering linear relations between variables.