A Spectroscopy-Constraint Network for Fast Thermochemical Process Monitoring Using Wavelength Modulation Spectroscopy

Performance optimization of various combustion-based power generation systems requires fast and accurate online monitoring of their thermochemical parameters. As an in situ sensing technology, laser absorption spectroscopy (LAS) with modulated wavelength, has been widely adopted. However, rapid parameter retrieval from modulated LAS signal can be challenging due to the underlying complex and nonlinear spectroscopy model. Most existing acceleration algorithms utilize supervised neural networks in an end-to-end manner ignored constraints on the spectroscopic model constraint. In addition, most state-of-the-art neural networks exhibit complicated structures with low computation efficiency. In this work, we developed a spectroscopy-constraint neural network for rapid thermochemical parameters extraction. The laser spectroscopic model is integrated in the proposed network through an encoder-decoder structure, offering independency on synthetic labeled dataset and hence enhancing its performance on measurement thermochemical parameters in industrial scenarios. Furthermore, the developed network has a simple structure and lightweight parameter size. A case study of aircraft engine exhaust monitoring is presented. The proposed model effectively reveals the dynamic behaviors of the engine. Compared with two representative supervised models, the new model exhibits better performance on spectral recovery as well as higher computational efficiency.