The application of partial discharge (PD) pattern recognition in gas-insulated metal-enclosed switchgear (GIS) based on machine learning (ML) is gradually increasing, but poor reliability limits its availability and effectiveness on site. In this context, the enhanced robustness and interpretability can obtain an intrinsically reliable algorithm, achieving to establish practical PD pattern recognition applications. In this paper, a series of phase-independent features with clear physical meaning are proposed, through analyzing the PD mechanism of typical insulation defects by gas discharge theory. The decision trees algorithm is selected as the interpretable classifier, and it is trained by the training data from the real 550 kV GIS. As a result, the proposed PD pattern recognition is intrinsically reliable.