Transformer Inter-Turn Short-Circuit Fault Recognition Based on Deep Learning

As a core component of the power grid, the operational status of transformers directly or indirectly impacts the safety of the entire power system. Statistical data indicate that among various transformer faults, inter-turn short circuits in the windings occur most frequently, accounting for approximately 60% to 70% of all faults. Current diagnostic techniques for inter-turn short circuits require shutdown and disassembly, which increases maintenance costs and fails to detect potential transformer faults in a timely manner. To address this issue, this paper proposes a transformer inter-turn short circuit fault diagnosis method based on autoencoders. For fault data acquisition, current waveforms of both primary and secondary windings were obtained through MATLAB/Simulink and labeled with six different fault tags. Regarding the diagnostic network, a novel comprehensive loss function was introduced to optimize the network for feature classification. Additionally, an adaptive optimization algorithm was designed to dynamically adjust the structure of the stacked autoencoder, avoiding the arbitrary selection of network structures based on experience. Experimental results demonstrate that this method exhibits excellent performance in terms of diagnostic accuracy.