An Entropy-Optimized Wavelet Synchrosqueezing and Gaussian Clustering Approach for Separation of Partial Discharge Signals

This paper proposes an enhanced source separation technique for multiple overlapping partial discharge (PD) acoustic emission signals in a power transformer. The proposed method utilises third-order Rényi entropy-based optimisation to achieve highly concentrated time-frequency (TF) representations. Four statistical TF features, namely, temporal centroid, temporal spread, frequency spread, and temporal duration, are extracted and clustered using a Gaussian mixture model for probabilistic source assignment. Experimental results on mixed PD signals demonstrate improved TF concentration, accurate source separation, and robust reconstruction performance, confirming the effectiveness of the proposed approach for high-resolution PD signal measurement and analysis. The performance of the proposed method is evaluated with laboratory-acquired and onsite field-acquired PD signals. The proposed method is validated by comparison with other established source separation techniques.