Vibration-Based Fault Diagnosis of Power-Transformer On-Load Tap Changers Using SGMD and CMPA-SMAGNet

Fault diagnosis of power-transformer on-load tap changers (OLTCs) is difficult because their vibration signals are nonlinear and nonstationary. This study evaluates CMPA-SMAGNet, a diagnostic framework that combines symplectic geometry mode decomposition (SGMD), a dual-branch multi-scale network, and a chaotic marine predators algorithm (CMPA). SGMD converts each vibration record into a symplectic-component tensor and a vector of engineered descriptors, including modal energy ratios, kurtosis, and symplectic geometric spectral entropy. SMAGNet processes temporal components and global descriptors in parallel, while CMPA tunes the learning rate and gate sensitivity. Experiments used 1,200 samples from four mechanical conditions on a KM-type OLTC platform. On the reported test split, the method achieved 95.0% accuracy; across ten runs, mean precision and recall were 94.98% and 95.05%, respectively. Under the current protocol, these results were higher than those of the tested ResNet-18, TCN, and 1D-CNN baselines. The findings support the feasibility of the framework on this experimental platform, while robustness across different devices, operating points, sensor arrangements, and measured field-noise conditions remains to be established.