This paper deals with detection and classification of intermittent stochastic faults by combining a generalized polynomial chaos (gPC) representation with either Maximum Likelihood or Bayesian estimators. The gPC is used to propagate stochastic changes in an input variable to measured quantities from which the fault is to be inferred. The fault detection and classification problem is formulated as an inverse problem of identifying the unknown input based on the Maximum Likelihood of fit between predicted and measured output variables, or on a Bayesian inference based estimator which recursively updates the gPC coefficients. Simulation studies compare the proposed methods with a particle filter (PF) to estimate the value of an unknown feed mass fraction of a chemical process. The proposed method is shown to be significantly more efficient in terms of computational effort and less sensitive to user defined tuning parameters than a PF.