Transformer Fault Warning Based on Spectral Clustering and Decision Tree

Abstract

The insufficient amount of sample data and the uneven distribution of the collected data across faults are key factors limiting the application of machine learning in power transformer fault warning, as demonstrated by the poor adaptability of the established data-driven models under actual operating conditions. In this paper, an unsupervised and supervised learning method is designed for power transformer fault early warning based on electrical quantities and vibration signals. The method is based on the Fourier levels of transformer vibration signals under different electrical conditions measured in the field, and the vibration features are clustered according to their intrinsic properties by means of a spectral clustering algorithm. A decision tree model of the vibration characteristics under each cluster is then constructed to calculate early warning values for the transformer vibration spectrum under different electrical conditions, enabling the assessment of transformer production variability. The above process, which is based on field measurement data and data mining analysis methods, is cheaper than the existing transformer fault warning techniques at home and abroad and makes better use of information and training models.