Mechanical Anomaly Detection and Early Warning for Ultrahigh-Voltage Shunt Reactors via Adaptive Thresholds and WGAN-GP

Ultrahigh-voltage shunt reactors (UVSRs) are the main equipment of ultrahigh-voltage (UHV) electricity transmission systems. Timely detection and early warning (D&W) of mechanical anomalies are essential for the safe and economical operation of UVSRs. However, due to the electromagnetic-mechanical–oil-structural coupling vibration mechanism, the strong noise, and variable working conditions, developing an excellent mechanical anomaly D&W method based on vibration analysis for UVSRs is more difficult than traditional equipment. Motivated by this, a novel methodology for mechanical anomaly D&W for UVSRs is proposed. First, following the analysis of the special structure and complex vibration mechanism of UVSRs, a vibration monitoring system was developed to detect the internal mechanical vibrations of UVSRs from the surface of their tanks. Thirteen vibration-related features were initially extracted from the vibration signals, while a random forest algorithm (RF) was used to rank their importance in detecting anomalies. Then, the mechanical health index (HI) is constructed based on the health status-sensitive features. To address omitted detection or false detection problems of traditional fixed-threshold methods, an adaptive threshold method based on sliding windows, Gaussian mixture model (GMM), and “
$3\sigma $
” criteria is proposed to detect mechanical anomalies on the online calculated HIs. Furthermore, a Wasserstein generative adversarial nets optimized by gradient penalty model (WGAN-GP) was used to predict the HIs, on which the proposed adaptive thresholds were applied to warn of potential mechanical anomalies. Finally, the proposed framework was applied to a UVSR in China. Field operational records and multiple sets of comparison experiments verify the feasibility and effectiveness of the proposed framework.