Comparative analysis of data-driven models for spatially resolved thermometry using emission spectroscopy.

Area of Science:

• Spectroscopy
• Data Science
• Thermodynamics

Background:

• Line-of-sight emission spectroscopy has limitations in measuring temperature in non-homogeneous fields.
• Spatially resolved temperature measurements are crucial for understanding complex systems.

Purpose of the Study:

• To develop and evaluate data-driven models for spatially resolved temperature measurements using emission spectroscopy.
• To compare the performance of feature engineering with classical machine learning against end-to-end convolutional neural networks (CNNs).

Main Methods:

• Investigated two categories of data-driven methods: feature engineering with classical machine learning and CNNs.
• Evaluated fifteen feature groups combined with fifteen classical machine learning models.
• Assessed eleven CNN models for temperature distribution measurement.

Main Results:

• Feature engineering combined with machine learning outperformed direct CNN application.
• Physics-guided transformation, signal representation, and Principal Component Analysis proved most effective for feature extraction.
• The light blender ensemble model, using extracted features, achieved the best performance (RMSE: 64.3, RE: 0.017, RRMSE: 0.025, R: 0.994).

Conclusions:

• The proposed methodology, leveraging feature engineering and the light blender model, accurately measures non-uniform temperature distributions from low-resolution spectra.
• This approach is effective even when species concentration distributions are unknown, overcoming limitations of traditional spectroscopy.