When commissioning or restarting a system after a maintenance action there is a need to properly tune the decision thresholds of the diagnostic system. Too low or too high thresholds may implicate either missed alarms or false alarm rates. This paper suggests an efficient data-driven approach to optimal setting of decision thresholds for a PEM fuel cell system based solely on data acquired from the system in reference state of health (i.e. under fault free operation). The only design parameter is the desired false alarm rate. Technically, the problem reduces to analytically determining the probability distribution of the fuel cell's complex impedance and its particular components. Employing pseudo-random binary sequence perturbation signals, the distribution of the impedance is estimated through the complex wavelet coefficients of the fuel cell voltage and current. The approach is validated on a PEM fuel cell system subjected to various faults. (C) 2014 Elsevier B.V. All rights reserved.