A novel semi-supervised power transformer defect monitoring technique using unreliable pseudo-labels with highly imbalanced partial discharge signals

In a high-voltage power transformer, one of the most important metrics to measure the effectiveness of an insulating system is partial discharges (PD). However, in real-world settings, several PD sources (single or multiple) with significant interference may hinder the signal’s acquisition, which could result in an erroneous final diagnosis of the equipment’s condition. Convolutional neural networks have outstanding classification performance and robust automatic feature extraction capabilities, making them adept at recognizing patterns. Nevertheless, the collected samples could be more balanced and a significant amount of unlabelled data, which complicates the ability of the existing approaches to provide accurate diagnoses. This study suggests three distinct sections for the transformer defect recognizing technique using unreliable pseudo-labels (TDRUPL), including teacher and student models, to address the abovementioned issues. A unique variational autoencoder first creates synthetic data to equalize the imbalanced samples. Then, external interference immersed in the measured PD signal is removed by employing a bi-directional long short-term memory network and its hyperparameters are tuned by genetic algorithm. Finally, to deal with unlabelled datasets, the instructor model in TDRUPL includes two separate classifiers that help to extract and classify PD characteristics. They also label some unlabelled data as noisy or noise-free samples for convenience. The student model dynamically modifies the conditional and marginal distribution to recognize transformer defects accurately. The proposed algorithm achieves outstanding performance with an accuracy of 98.9%.