The authors address the issue of robustness in artificial neural networks subject to small input perturbations. The robustness in artificial neural networks is studied using the concept of input-output sensitivity analysis applied to an incipient fault detector artificial neural network (IFDANN). The IFDANN was designed to detect winding insulation fault and bearing wear in single-phase squirrel-cage induction motors. Modification of the IFDANN, with the intention of increasing its robustness to input noise during real-time applications, is discussed. Analytical and simulation results are presented to show the significant improvement in robustness of the modified IFDANN for operation with noisy measurements.