Raman spectral feature extraction and analysis methods for olefin polymerization and cracking based on machine...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy offers fast, sensitive, and cost-effective real-time gas monitoring.
• High dimensionality, spectral overlap, and noise in Raman data challenge traditional mixture composition analysis.
• Accurate gas composition determination is crucial for complex chemical process monitoring.

Purpose of the Study:

• To optimize the XGBoost machine learning model for enhanced gas composition prediction using Raman spectral data.
• To evaluate the effectiveness of different feature extraction and selection methods in improving predictive accuracy.
• To compare the performance of XGBoost against other machine learning models for Raman spectral analysis.

Main Methods:

• Utilized Raman spectral data from a gas mixture including hydrogen, ethylene, propylene, and butene.
• Implemented and compared three distinct feature extraction and selection techniques.
• Trained and evaluated XGBoost, Decision Trees, Random Forests, Support Vector Machines, and Neural Networks.

Main Results:

• The optimized XGBoost model demonstrated superior accuracy and generalization ability in predicting gas composition from Raman spectra.
• XGBoost outperformed Decision Trees, Random Forests, Support Vector Machines, and Neural Networks in quantitative analysis.
• Feature extraction and selection methods significantly enhanced the predictive performance of the XGBoost model.

Conclusions:

• XGBoost is a highly effective machine learning model for quantitative analysis of complex Raman spectral data.
• The optimized XGBoost approach provides a robust solution for real-time gas composition monitoring in chemical processes.
• This work highlights the potential of advanced machine learning techniques to overcome limitations in spectroscopic data analysis.