Deep learning algorithms for temperature field reconstruction of nonlinear tomographic absorption spectroscopy
Andong Deng, Jianqing Huang, Hecong Liu, Weiwei Cai
Abstract
Nonlinear tomography can be combined with line-of-sight measurement techniques and take the full advantage of the nonlinear relationship between the target fields to be reconstructed and the measured projections. For example, it has been integrated with absorption spectroscopy for simultaneous imaging of temperature and species concentration, and the technique is referred to as nonlinear tomographic absorption spectroscopy (NTAS) which is especially suitable to applications where the optical access is extremely limited. However, the major drawback of nonlinear tomography is its high computational costs for the inversion process, which involves the solution of a large-scale nonlinear equation system. The situation becomes more severe when thousands of tomographic frames need to be processed. This limitation can be potentially overcome by applying a deep learning algorithm, which can build efficient mapping between the projections and the target fields for rapid reconstruction. Nevertheless, only preliminary study of convolutional neural networks was reported so far, and there are also other well-established methods which have not been investigated yet. In this work, we cite NTAS as an example and demonstrate the reconstruction of temperature field using deep brief network (DBN) and recurrent neural network (RNN) for the first time. This work also aims to provide systematic comparative studies between these representative algorithms. As expected, all the deep learning algorithms are very efficient in solving NTAS problems, and a typical reconstruction time is on the order of milliseconds. In addition, the results suggest that, in NTAS problem, RNN is superior to the other methods in terms of both accuracy and noise immunity, and it is more favorable for practical applications.