Transformer acoustic fault diagnosis method based on selective kernel Siamese network modeling

Aiming at the problems of low recognition accuracy and lack of diagnostic reliability faced by the existing transformer non-outage fault diagnosis techniques when dealing with small samples and complex fault scenarios, this paper proposes a transformer acoustic fault diagnosis method based on the selective kernel Siamese network model. An adaptive synchronous S-transform denoising method is proposed to optimize the acoustic signal processing through the S-transform with adaptive windowing to generate the acoustic pattern with high time-frequency resolution and no energy divergence. Further, a time-frequency selective kernel enhancement (TFSK) module is designed for dynamic convolution kernel size selection and expansion rate adjustment, which effectively enhances the ability to capture the time-frequency features of the atlas and its adaptability. Finally, the Siamese network is constructed to evaluate the transformer state, and the dynamic metric function is introduced to enhance the sensitivity to the small-sample scale difference, so as to realize the high-precision acoustic diagnosis under the conditions of few samples and information interference. Experimental results show that the method not only maintains high diagnostic accuracy under different data sample sizes and noise interference conditions, but also proves its excellent generalization ability in fault diagnosis of multi-model transformers.