Vibration response of rotating machines is typically mixed and corrupted by a variety of interfering sources and noise, leading to the necessity for the isolation of the useful signal components. A relevant frequently encountered industrial case is the need for the separation of the vibration responses of the same type of bearings inside the same machine. For this purpose, a Blind Source Separation procedure has been successfully applied, based on the maximization of the information transferred in a neural network structure. Thus, a key element for the success of the proposed procedure is the non-linear function used in this single layer Neural Network structure. However, since the vibration response of defective rolling element bearings is characterized by signals with super-Gaussian distributions, a sensitivity analysis of this non-linear function is necessary. First, this analysis is performed in a set of numerical experiments, based on dynamic models of defective bearings. Finally, the same analysis is applied in an experimental test rig.