An improved deep learning-based algorithm for 3D reconstruction of vacuum arcs

Extensive attempts have been made to enable the application of deep learning to 3D plasma reconstruction. However, due to the limitation on the number of available training samples, deep learning-based methods have insufficient generalization ability compared to the traditional iterative methods. This paper proposes an improved algorithm named convolutional neural network-maximum likelihood expectation maximization-split-Bergman (CNN-MLEM-SB) based on the combination of the deep learning CNN and an iterative algorithm known as MLEM-SB. This method uses the prediction result of a CNN as the initial value and then corrects it using the MLEM-SB to obtain the final results. The proposed method is verified experimentally by reconstructing two types of vacuum arcs with and without transverse magnetic field (TMF) control. In addition, the CNN and the proposed algorithm are compared with respect to accuracy and generalization ability. The results show that the CNN can effectively reconstruct the arcs between a pair of disk contacts, which has specific distribution patterns: its structural similarity index measurement (SSIM) can reach 0.952. However, the SSIM decreases to 0.868 for the arc between a pair of TMF contacts, which is controlled by the TMF and has complex distribution patterns. Compared with the CNN reconstruction method, the proposed algorithm can achieve a higher reconstruction accuracy for any arc shape. Compared with the iterative algorithm, the proposed algorithm's reconstruction efficiency is higher by 38.24% and 35.36% for the vacuum arc between the disk and the TMF contacts, respectively.