To solve such a challenging problem, a framework of a novel fault detection method robust against uncertainty
has to be developed. In the proposed method the model of the nominal product can be obtained with the application
of the non-linear product identification method e.g. the Extended Kalman Filter, Artificial Neural Networks (ANNs)
or Fuzzy model [33]. It should be underlined that for such kind of models the uncertainty description can be
obtained, however, it is not a trivial task. For example, in the case of application of the ANNs the parameters of
neural model obtained during training procedure are not uniquely obtained but they are approximated by a so-called
feasible parameter set which represent the neural model uncertainty. The size of such parameters set depends on the
inaccuracy of parameters estimates resulting from the values of noise contained in the training data and neural
architecture inaccuracy. The mathematical description of the model uncertainty enables to calculate the output
adaptive thresholds [33] which allow performing the robust fault detection according to the scheme presented in Fig.
8. The adaptive threshold, contrary to the constant one, bounds the residual at a level that is dependent on the model
uncertainty, and hence it provides a more reliable fault detection.