Data-driven methods for power fault diagnosis heavily rely on the signal data quality of power sensors. The power systems in Tokamak fusion devices often operate in environments with complex electromagnetic field coupling, leading to the mixing of physical characteristic signals with a significant amount of inseparable noise in the collected data.

This study aims to mitigate the impact of noise on the final diagnostic results by proposing a multi-branch denoising network, termed Hierarchy Branch Denoising Convolutional Neural Network (HBD-CNN) that utilizes noise-resistant wavelet enhancement in conjunction with one-dimensional convolutional neural networks to accomplish power system fault diagnosis tasks under the influence of noise interference.

Firstly, the signal decomposition function of discrete wavelet transform (DWT) was incorporated into the network layer of the convolutional neural network (CNN), and the optimization of the traditional 1D-CNN network structure was deepened alongside the more robust exponentially linear unit (ELU) for noise. Then, a data multi-level structure was constructed based on prior knowledge to leverage and couple it with hierarchical classification modules within the network, hence the generalization capability of HBD-CNN was enhanced. Finally, preliminary validation of the architecture of this model was conducted based on the simulated power supply dataset.

Validation results show that the fault diagnosis accuracy for the power converter reaches 98.31% when the signal-to-noise ratio (SNR) is 10 dB. Even at an SNR of 2 dB, the accuracy remains above 92%.

The results of this study indicate that HBD-CNN demonstrates excellent fault diagnosis performance and potential under noisy conditions.