Deep neural-network laser absorption spectroscopy tomography for combustion diagnosis

Abstract

Reactive flows are present in many industrial applications, particularly in energy-generation systems, where they directly impact combustion efficiency and environmental pollution. Therefore, effective diagnosis of reactive flows is crucial, involving the monitoring of physical and chemical parameters such as flow velocity, temperature, and species concentration. Specifically, measuring temperature and species concentration can greatly reduce harmful emissions and improve fuel efficiency. Given the highly dynamic and complex nature of reactive flows, measurement techniques are required to be accurate, robust, and have a fast response.

In most industrial applications, probe-based sensing methods, such as thermocouples and gas samplers, have gained significant attention due to their simplicity in setup and measurement. However, these point-wise detection methods suffer from limited spatial and temporal resolutions. The advent of Laser Absorption Spectroscopy (LAS) provides a superior approach for diagnosing reactive flows. Due to its non-intrusive nature, fast response and robustness, LAS has become increasingly popular in industrial settings. Utilizing an emitter-receiver configuration, LAS measures the Line-of-Sight (LoS) laser intensity, which is partially absorbed by the target gas, and infers temperature and species concentration from it. This technique exhibits a temporal resolution of Kilo Hertz and above, benefiting from the quick scanning ability of tunable diode lasers for specific absorption spectra. Despite the above advantages, achieving rapid and accurate measurements remains challenging due to the time-consuming laser signal post-processing. Therefore, accelerating this process is essential.

In addition, LAS can produce spatially resolved images of gas properties when combined with tomography. This is achieved by employing multiple LoS measurements from different projection angles, known as LAS tomography. However, physical constraints of optical access in industrial combustors often result in insufficient laser beams for LAS tomography, leading to rank deficiency in tomographic inverse problems and thus inaccurate diagnostic results. Recent advancements in deep learning techniques demonstrate great potential for addressing these complex tasks. Previous studies have shown that applying deep learning techniques to LAS and LAS tomography can enhance both the measurement speed and accuracy. Therefore, the objective of this thesis is to design applicable and effective deep learning algorithms for LAS signal post-processing and tomography in reactive flow diagnosis.

In this thesis, a novel deep learning-based hybrid neural network model has been developed for signal post-processing to predict path-average temperature of a Gas Turbine Engine (GTE). This model integrates Wavelength Modulation Spectroscopy (WMS) with fundamental physical absorption principles. By effectively extracting both temporal and spatial features from the spectral lineshape, the model provides reliable temperature predictions and ensures high industrial applicability.

For LAS tomography, two training sets are established to simulate the reactive flows of an annular combustor with ten injectors using customized multiple Gaussian profiles and a modeled circular burner simulated by the Fire Dynamics Simulator (FDS), respectively, rather than the typical Gaussian-shaped phantoms. Based on the comprehensive training sets, a Convolutional Neural Network (CNN)-based algorithm has been designed to reconstruct temperature distributions. This newly introduced method has been validated through numerical simulations, demonstrating good accuracy and sensitivity in monitoring the dynamic combustion process.

Although the simulation can mimic the dynamics of reactive flows to some extent, the reconstruction results may be suboptimal due to unpredictable discrepancies between real-world data and simulations. To address this challenge, this thesis proposes an innovative untrained neural network for LAS tomography called the Model-Informed Double Image Prior (MI-DIP). For the first time, this network introduces a dualpath architecture for dataset-free, joint reconstruction of two-dimensional temperature and species concentration. The developed network closely imitates the problem formulation of LAS tomography and regularizes the inverse problem using its physical model and inherent priors, thereby stabilizing the image reconstruction process. This approach has been validated through both numerical simulations and lab-scale experiments, demonstrating its ability to reconstruct the temperature and water vapor concentration profiles of the flames. The results highlight improved imaging accuracy and enhanced resistance to noise, showcasing the potential of the MI-DIP to provide reliable diagnostics in complex, dynamic environments.