Quantification of on-line Partial Discharge (PD) measurements is a challenge in the industry for several reasons, amongst them: instrumental characteristics, type of sensors, location of the sensors in the machine (line terminals, parallel circuits, neutral point) and the measurement procedure used. PD can be measured in picocoulombs, in millivolts or in dB and is most commonly displayed in 2D or 3D representation. All of these variations make comparison difficult and partially explains why no acceptable PD level has yet been defined for generator diagnostics. Another problem is to select the best parameter for quantification: maximum PD amplitude, discharge current, repetition rate, number of pulses … The former is one of the most common one, but it often neglects the identification of the discharge source causing the PD signal. Differentiation between PD sources is not straightforward and cannot only rely on simple quantification rules. In the present work, a methodology to automatically recognize individual PD sources from 2D PDA files was implemented using deep learning techniques. A Deep Convolutional Variational Autoencoder (DCVAE) was used to help PD experts through an iterative process in separating PDA files in different classes representing each type of PD sources (symmetric, positive asymmetry, negative asymmetry, gap type discharges …). The approach was tested on the entire Hydro-Quebec database of about 33 000 files and each group of files associated with each PD source was then selected to carry out independent discharge rate as a function of amplitude analysis. The statistics of each group were thereafter compared between themselves, but also with statistical analysis of data including the global PD activity when no PD source separation is done.