Circuit fault diagnosis for 12-pulse power converter in HL-3 based on Meta Pseudo Labels

Abstract

The application of deep learning in the field of power electronics fault diagnosis has garnered considerable attention and research interest. In the coil power supply of the HL-3 tokamak device, there exist practical challenges in dealing with fault data from 12-pulse thyristor power converters, including large data volume, variable frequency operation, and sensor noise, among others. To solve these problems, this paper proposes a method that combines wavelet threshold denoising and Meta Pseudo Labels. Firstly, discrete wavelet transform is used to transform the original voltage signals of the 12-pulse thyristor rectifier operating across a wide frequency range into the wavelet domain. Subsequently, a reasonable threshold is employed to filter out high-frequency noise, resulting in the output of wavelet threshold denoising. Then, Meta Pseudo Labels algorithm, an idiographic semi-supervised learning approach, is applied to learn the features of the denoised data, yielding classification models and experimental results. The model identifies three types of states in single thyristor branches of the rectifier circuit, which are normal, short-circuit fault, and open-circuit fault. The experimental results indicate that the accuracy and robustness of the model trained using the Meta Pseudo Labels algorithm outperform those of a convolutional neural network with the same network architecture. Especially, one of the experiments accomplishes a diagnosis accuracy exceeding 90% while training with the whole unlabeled training dataset and only 10% of stratified subsamples from the labeled training dataset.