The methods available for partial discharge diagnosis use single classifier for multi-type division. These methods are sensitive to cross-overlapped regions and greatly affected by inherent defects of the single classifier.This paper proposes a depth stratification diagnosis method of partial discharge

achieves multi-class division by layer-by-layer dichotomy decision. Benign samples are recognized preferentially in hierarchical decision making

and classify crossover regions into deep nodes. Different classifiers may be chosen at each dichotomous node. Five typical GIS discharge models are designed

PRPD spectra and U-Δt sequence spectra are formed

and 16 characteristic parameters are constructed. The diagnostic classification accuracy under different depth is explored

and compared with traditional direct classification methods. The results show that the accuracy of the depth stratification recognition is improved by 20% compared with the direct recognition diagnosis

especially for the surface and particle defects（improved by 30%）.