Abstract
The significance of the distribution and power transformers as the key components of the electric power grid is not overrated. Therefore, a thorough diagnosis and quality maintenance procedures are matters of the primary interest for power companies. Among most accurate and well-established transformer mechanical condition monitoring methods is the Frequency Response Analysis (FRA). FRA has a quite solid background in terms of research, industry application and international standardization. Nevertheless, the interpretation of the FRA still suffers from the ambiguity of the test results. In this paper, a modified measurement setup for FRA is proposed in order to collect the comprehensive data sets for statistical analysis and machine learning application. Based on the collected data a winding short-circuit detection and fault severity prediction model is developed. The estimated accuracy of the proposed interpretation and classification method demonstrated the validity of the novelty approach.