Volumetric reconstruction for combustion diagnostics via transfer learning and semi-supervised learning with limited labels

Abstract

Volumetric tomography (VT) is a powerful tool for combustion diagnostics due to its capacity in resolving flame structures in three-dimensional (3D). Recently, convolutional neural network (CNN) has been applied to solve the inversion problems of VT, which features an overwhelming advantage over classical iterative methods in terms of computational efficiency. However, a large number of labels have to be prepared for the supervised learning of CNN using iterative methods, compromising its efficiency advantage. Moreover, previous studies were limited to a single dataset and the generalization performance of CNN has not yet been tested. In this work, both transfer learning and semi-supervised learning were employed to construct the CNN networks with limited labels. The comparative studies between them and supervised learning confirmed that a significant improvement in reconstruction accuracy can be achieved even with limited labels. The correlation coefficient between the reconstruction and ground truth is larger than 0.98 for three commonly encountered application scenarios. The training strategies developed in this work are expected to be valuable for all VT modalities as applied to flow/combustion diagnostics.