An Interpretable Acoustic Partial Discharge Detection Framework for Oil Filled Power Transformers Using Multiphysics Modeling and Machine Learning

This paper presents an interpretable framework for acoustic partial discharge (PD) detection in oil-filled power transformers by integrating experimental field measurements, physics-based signal modeling, multi-domain feature extraction, and machine learning-based classification. Partial discharge is a critical indicator of insulation degradation, and reliable detection under noisy operating conditions remains a major challenge in transformer condition monitoring. In this study, a hybrid dataset is constructed by synergizing authentic field measurements from high-voltage transformer testbeds with multiphysics finite element simulations. To address the limitations of experimental testing, diverse environmental and operational disturbances are systematically synthesized to emulate realistic, complex noise scenarios. The resulting comprehensive dataset enables robust evaluation of the proposed detection approach. After wavelet-based denoising, each signal is represented using a set of statistical, temporal, spectral, and wavelet domain features. Two complementary descriptors, the Wavelet Scale Centroid (WSC) and the Wavelet Domain Cross Correlation Index (WD CCI), are incorporated to better characterize the energy distribution and structural similarity of transient signals. Classification is performed using a support vector machine (SVM) with probabilistic output, and the decision reliability is further assessed using distance-based validation metrics in the feature space. The results demonstrate that the proposed framework effectively distinguishes PD signals from non-PD disturbances, providing a robust and interpretable solution for transformer condition monitoring. This approach contributes to improving the reliability of acoustic PD diagnostics and supports the development of intelligent monitoring systems for power equipment.