Ghost-ResNet-Based Fault Diagnosis Technique Detecting Weak AC Arc Failures in a Three-Phase PWM Inverter

This paper proposes an artificial-intelligence-based diagnostic algorithm that combines residual feature learning with a Ghost module to reliably identify weak AC series arc faults at the output of a three-phase pulse-width-modulated (PWM) inverter. The current waveform and spectral content of an output-side AC series arc vary with the output fundamental frequency, switching frequency, load inductance, and arc intensity. A weak arc produces substantially smaller waveform distortions and spectral changes than a strong arc, making robust detection difficult using conventional signal-processing methods based on a single fixed threshold. In this study, strong and weak arcs are distinguished according to arc power and the electrode-gap opening condition, and the model is evaluated using output AC-current data acquired over a wide range of operating conditions. With fast Fourier transform (FFT) magnitude spectra as the input, the proposed Ghost-ResNet achieved a diagnostic accuracy of 96.89% on the test set, outperforming all comparison models. It also achieved high precision and recall for the AC-arc class, including weak-arc samples. These results demonstrate that the proposed model is effective for detecting output-side AC series arcs in three-phase PWM inverters, including weak arcs and arcs occurring under varying operating conditions.