To identify individual partial discharges(PD)signals produced by multiple insulation defects in gas insolated switchgear (GIS),this paper analyses the mechanism of propagation and mixing of multiple PD electromagnetic wave signals in GIS cylinder and proposes the convolutive mixing model to describe it for a separation algorithm to acquire individual PD signals. With the non-stationary property,mixing PD signals are changed at time domain into a set consisting of short-time stationary PD signals and then the Molgedey-Schuster decorrelation approach is employed to separate these stationary PD signals at frequency domain. The correlation of the envelope of separated PD signals in this set is used to reconstruct PD signals to realize the separation of non-stationary UHF PD mixtures. The effective separation of actual UHF PD mixing signals validates the assumption of convolutive mixing process in GIS and also offers a new approach to the identification of mixing PD signals by external ultra-high frequency detection scheme from multiple insulation defects in GIS.