The quality of the failure diagnosis procedure has a critical impact on the safety of a human machine system - especially the decision taking part. Based on the actual knowledge, a decision must be taken quickly, even when the available data is suspicious, conflictive or incomplete. The wanted diagnosis methodology should take into account such uncertainty in a clever way so as to be able to draw a conclusion about the cause of failure. Two major approaches are used to deal uncertainties: The Bayesian approach and the Dempster-Shafer theory. The Bayesian modeling of the uncertainty is based on strong assumptions of a supposed normal behavior of the outputs, and gives good results in cases where the assumptions hold using Bayesian Networks to converge to a decision. The Dempster-Shafer theory approach deals natively with the uncertainty modeling and uses the generalized bayesian theorem and the disjunctive rule of combination to propagate beliefs (forward and backward) to the connected causal components forming the decision graph. The assumption part is shifted to the decision stage reducing the impact of a wrong belief and keeping track of the effects of uncertainty and conflict between information. A decision process based on the amount of uncertain data is actually an interesting problem to solve methodologically, since most of the time, human operator based diagnosis is performed using experience motivated interpretation of small hints that may yield uncertain knowledge about the state of the system. The extension of the Dempster-Shafer theory is a promising way to manage uncertainty in a reliable way in decision-taking applications.