The dynamical response of a train rolling on a real track depends on several parameters. Most of them cannot be accurately identified and have to be considered as uncertain. The aim of this thesis is the construction of a probabilistic model of the rail fatigue life considering the track geometry and the rail wear as random fields modelled with the Karhunen-Loève expansion. This latter requires the modal decomposition of the covariance operator. This step can be very expensive if the domain if much larger than the correlation length. To deal with this issue, an adaptation of the KLE, consisting in splitting the domain in sub-domains where this modal decomposition and the sample generation can be comfortably computed, is proposed. A correlat...