A Semi-supervised Convolutional Autoencoder for Multiple Physical Parameters Measurement in Laser Absorption Tomography

Tunable Diode Laser Absorption Spectroscopy (TDLAS) tomography is a high-precision flow field imaging technique that reconstructs the spatial distribution of flow field parameters by scanning absorption data from different angles. It is particularly useful for quantitative analysis in complex flow fields, such as combustion, gas flow, and pollution monitoring. Recently, deep learning has been applied to TDLAS tomography, improving the accuracy of 2D or 3D reconstructions of gas parameters and enhancing robustness against noise. However, traditional supervised learning methods in TDLAS require large amounts of labeled data, which are costly. To overcome this, we propose a convolutional autoencoder-based semi-supervised learning network. This network combines unsupervised autoencoder features with supervised learning, effectively utilizing both limited labeled and abundant unlabeled data, thereby improving model generalization and reducing reliance on labeled data. Simulation results demonstrate that the reconstruction error of temperature was significantly reduced from 6.7% to 1.8%. The method shows strong potential for efficient and accurate flow field reconstruction in TDLAS applications.