Analysis of artificial neural network applications for assessing the technical condition of traction transformers in electric locomotives

Abstract

This paper presents a critical and systematic analysis of the scientific literature on assessing the technical condition of traction transformers of electric locomotives using artificial neural networks (ANN). Unlike earlier descriptive reviews, the analysis was carried out in three stages: chronological evolution of the directly relevant studies, classification of existing approaches according to Dasarathy's data fusion taxonomy, and critical evaluation of each study's methodological strengths and limitations, including constraints acknowledged by the original authors themselves. Nine studies dedicated directly to traction transformers of electric locomotives were identified. All of them operate within a single physical-domain parameter group and a single data-fusion level; their datasets are limited to one institution or one train type and are not publicly available. It is shown, both empirically and theoretically, that no existing approach fuses heterogeneous thermal, insulation, mechanical and electrical parameters at a level higher than single-domain feature fusion. The results provide a methodological foundation for developing a multi-parameter hybrid neural network model for traction transformer diagnostics.