Fault Diagnosis of Power Converter based on Multi-Channel 1D-BNCNN for Tokamak Magnetic Field Power Supply

Fault diagnosis has been of greater importance as a performance assessment of system degradation by accurately classifying the current health state from various signals. Rapid, reliable, and intelligent short-circuit fault (SCF) diagnosis in converter, as a supplement to existing fault diagnostic methods, is necessary for the sake of the safe and reliable operation of magnetic field power supply (MFPS) in complex external environments. The critical issue for converter fault diagnosis of MFPS in a Tokamak is to extract fault features effectively. In this paper, the proposed end-to-end multi-channel 1D-BNCNN is automatically learn the most essential high-level fault features from raw 1D simulation data. In order to learn and fuse abundant and complementary fault information, 5 channels of simulation data are recorded from the simulation model of converter units established by MATLAB/SIMULINK. The overlap sampling is considered to obtain a large number of samples to prevent overfitting. From the perspective of effective feature extraction related to the fault components and RF, wide first-layer kernels are introduced. Dropout is also added to enhance the model's generalization performance. With the BN layer, 1D-BNCNN can adaptively recalibrate features to offset internal covariant shift and speed up the training process. The experiment results demonstrate that the proposed method is more effective and the diagnostic accuracy is up to 0.95 while the loss is 0.40. This indicates that the proposed model is promising for converter fault diagnosis in MFPS.