A convolutional regression-based approach for short-circuit fault diagnosis in turbine generator excitation windings

Excitation winding short-circuit faults rank among the most prevalent failures in turbine generators. While their initial symptoms are often subtle, prolonged progression can induce severe secondary damages such as rotor grounding, exacerbating equipment deterioration. To address the challenges of high-cost physical fault replication and scarce experimental data, we developed a finite-element model based on operational power plant units. This model enables simulation of winding shortcircuit faults and facilitates fault characterization through radial search-coil monitoring. To overcome the limited sensitivity of search-coil signals during symmetrical short-circuits, we propose a CNN-based regression method that predicts healthy operational baselines. Comparative analysis between predicted and measured waveforms enables precise fault detection and localization. The algorithm’s diagnostic accuracy has been rigorously validated through comprehensive simulation studies. Furthermore, magnetic asymmetry analysis derived from the predicted curves provides critical indicators for early-stage and symmetrical fault identification, offering vital support for operational safety maintenance.