This paper deals with a fault detection method taking into account model uncertainties described by bounded variables. A parity space approach is used for generating testable redundancy relations in which each uncertain parameter is defined by an interval containing all its feasible values. Consistency tests consist in evaluating these set-membership relations and lead to convex sets containing the feasible free-fault behaviours of the supervised system. The objective is to improve fault detection performance by taking into account constraints on variations of uncertain parameters, which do not randomly vary.