Fault Diagnosis of On-Load Power Transformers via Weak Sideband Vibration Identification

Abstract

The winding structure is a critical component of a power transformer. Assessing its structural condition during operation poses significant challenges for grid maintenance. Unlike static windings, the on-load windings undergo periodic stiffness changes under electromagnetic forces (EMFs). In this study, a novel parametric vibration model based on the time-varying stiffness is investigated. Next, the Mathieu equation is employed to analyze the sideband vibration response, and the spectral correlation (SC) method is applied to detect weak signals. This study presents a transfer learning framework employing a knowledge separator and a domain-adversarial network to address the challenge of limited fault samples. In laboratory experiments, the structural degradation of windings was simulated by adjusting the clamping force. Both numerical simulations and experimental results show that reducing the clamping pressure from 4.0 to 2.0 MPa causes the characteristic frequency to shift from approximately 46–20 Hz. The performance of the transfer learning is assessed based on the prediction accuracy of the unlabeled samples in the target domain following training in the source domain. Finally, vibration monitoring was performed on a field transformer, which served to verify the correlation between the observed 16.1 Hz low-frequency sideband vibration and the defects in the winding structure.