A new detection and isolation scheme for unknown parametric faults in non-linear stochastic systems is presented that is particularly suited for small parametric changes. The proposed residual is the moving angle between two stochastic processes: the error derived from the reference model and the expected error in case a certain fault has occurred. Since the fault is unknown, fault classes are defined. Most of the faults that belong to such a fault class can be detected and isolated by just testing one representative of this fault class. As the moving angle itself is a stochastic process, an estimator is designed and characterised. Conditions for the detectability and isolability of fault classes are given for this estimator based on hypotheses tests. Finally, the theoretical results are confirmed by simulations and Monte Carlo tests.