An online video-based fault diagnosis method for oil-immersed transformers using GWO and BP

Conventional condition assessment methods applied to oil-immersed transformers rely on offline detection and exhibit low efficiency in video feature extraction. A video-based online fault diagnosis model is developed in this study by integrating the Grey Wolf Optimizer and a Back Propagation neural network. This model blends synthetic minority over-sampling technique imbalance processing, Gated Recurrent Unit sequential feature extraction, and Random Forest feature selection, adaptive boosting ensemble learning. This combination forms a multi-stage collaborative diagnosis process and achieves accurate identification of multiple fault types in oil-immersed transformers. The proposed model achieved an online diagnosis accuracy of 96.02%, a recall rate of 91.02%, and a fault classification rate of 94.23% in the experiments. Under harsh conditions such as strong Gaussian noise, 10% video frame loss, and 20 °C oil temperature fluctuation, the model still maintains a high F1 score. Meanwhile, this model contains only 2.89 M parameters, and the inference delay is as low as 178 ms, so the comprehensive performance far exceeds the comparison models. In conclusion, the proposed model makes use of the strengths of high accuracy, good robustness, and a compact structure. The proposed approach can be used for online detection of faults in oil-immersed transformers.