Early detection of transformer internal arcs using transient signal processing and machine learning

In power transformers, phenomena such as lightning strikes, switching operations, and short-circuit faults can lead to deformation of the winding insulation, causing electrical arc formation between windings. Over time, arc formation may result in a rigid short circuit, leading to excessive heating and, under pressure, transformer explosions. This thesis investigates the detection of arc formation in transformer windings by analyzing transient in the current and voltage signals at the transformer terminals. In this study, a 15 MVA power transformer is modelled in three dimensions in ANSYS@Maxwell environment and the arc model created in MATLAB@Simulink is adapted to the ANSYS@Maxwell magnetic model by performing the arc analysis in the transformer windings via synchronised operation. Finite Element Method (FEM) is used to analyse the designs. In order to ensure the accuracy of the transformer model, normal, short circuit and harmonic analyses are performed and the results obtained are compared with the factory and design data and it is determined that the results are compatible. The data obtained from the transformer terminals are transformed into frequency spectrum using Fast Fourier Transform (FFT) and high frequency harmonics generated during arc fault conditions in the transformer are detected from these signals. The models are trained with Machine Learning (ML) algorithms such as K-NN, DT, SVM, ANN and ESM using harmonic data at 361 different points obtained from frequency spectra. High accuracy and performance values are achieved in fault detection models of faulty winding and faulty disc. Thus, with the developed method, it is possible to detect arc fault formation in transformer windings in a short time without need of any sensor.