This paper deals with digital acquisition, classification and analysis of the stochastic features of random pulse signals generated by partial discharge (PD) phenomena. Focus is made on a new measuring system for the digital acquisition of PD-pulse signals, which operates at a sampling rate high enough to avoid the frequency aliasing, but that provides an amount of PD pulses which enables PD stochastic analysis. A separation and classification method, based on a fuzzy classifier, is developed for the analysis of the acquired PD-pulse shape signals. The result of the fuzzy classification is a cluster of signals homogeneous in terms of stochastic features of PD pulses. The classification efficiency is evaluated resorting to the PD-pulse height and phase distributions analysis. The instrumentation, and the associated classification methodology, are applied to measure and analyze PD data recorded for mica-insulated stator bars and coils, where typical defects, occurring during normal operations, were simulated. It is shown that the proposed procedure enables PD-source identification to solve the identification problems which arise, in particular, when different sources of PD are simultaneously active. In addition fuzzy classification provides an efficient noise-rejection tool.