In order to reach satisfactory performance, fault diagnosis methods require the tuning of internal parameters, usually called hyperparameters. This is generally achieved by optimizing a performance criterion, typically a trade-off between false-alarm and non-detection rates. Perturbations should also be taken into account, for instance by considering the worst possible case. A new method to achieve such a tuning is described, which is especially interesting when the simulations required are so costly that their number is severely limited. It achieves min-max optimization of the tuning parameters via a relaxation procedure and Kriging-based optimization. This approach is applied to the worst-case optimal tuning of a fault diagnosis method consisting of an observer-based residual generator followed by a statistical test. It readily extends to the tuning of hyperparameters in other contexts.