Abnormal Detection of Power Transformer Based on Generative Adversarial Network and Stacked Auto Encoder

An operated power transformer continuously emits vibration signals which can reflect its mechanical condition. Although some researches have succeeded in detecting the abnormality of power transformer by monitoring and analyzing the vibration signal, most of them requires a large number of abnormal samples, which is not obtainable for transformer in operation. To solve this problem, a transformer abnormal detection method which only need the vibration signal of normal operating condition is proposed. The Stacked Auto Encoder (SAE) is adopted to extract latent features as well as reduce dimension of vibration signals. And the Generative Adversarial Network (GAN) is trained to learn the data distribution of normal signals. Then if the discriminator of GAN encounters an abnormal sample, the output will be distinguishable. The proposed method can detect the abnormal operating condition of transformer with an accuracy of 94.4%, which demonstrate the feasibility and effectiveness of the proposed method.