Partial discharges in transformer insulation are of major concern to utilities which cause the catastrophic failure of insulation. One of the major challenges is the identification of discharges from multiple sources when it occurs concurrently. Hence it is imperative to devise methods for identifying and separating those signals for corrective measures. In this study, an unsupervised learning approach is proposed for clustering of individual partial discharge signals and then using that information for separating the multi-source signals. Our clustering approach works by constructing a dendrogram by measuring the cosine similarity between the feature vectors and then computing a threshold, to group the individual source signals into different clusters. The feature vectors include the relative energies from the wavelet packet decomposed tree and the Higuchi fractal dimension of the wavelet coefficients at the terminal nodes. The generated clusters are trained using a classifier model to separate the individual and multi-source signals. The proposed approach is a simple and robust technique for individual cluster groupings and individual to multiclass separations and could be used for multiclass cluster groupings. © 2020 IEEE.