Aiming at the diagnostic challenges of faults such as winding deformation
inter-turn short circuit
and clamping pressure loosening in oil-immersed transformers
a fault diagnosis method based on multi-physics field simulation and an improved convolutional neural network is proposed. Firstly
an electromagnetic-structural force field coupled simulation model is established based on transformer parameters. Experimental verification shows an error of only 4.53% between simulated and measured vibration signals
proving the model's reliability. Building on this
by setting winding faults with different severity levels
magnetic field and vibration signals are extracted. The Gramian angular field (GAF) is employed to convert the one-dimensional signals into two-dimensional feature maps to suppress noise interference. Subsequently
the AlexNet architecture is improved by introducing a multi-head self-attention mechanism to quantify the contribution of different features
combined with transfer learning for fault classification. Test results indicate that the diagnostic accuracy of this method reaches 96.36%
significantly outperforming traditional models and single-signal diagnosis methods. This research provides an effective solution for the precise detection and identification of latent internal faults in transformers.