[Near infrared spectroscopy quantitative analysis model based on incremental neural network with partial least...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Context:

• Quantitative analysis of flue gas composition is crucial for environmental monitoring and combustion process optimization.
• Traditional methods like Partial Least Squares (PLS) and Back-Propagation Neural Networks (BPNN) have limitations in handling complex spectral data and incremental learning.
• Near-infrared spectroscopy (NIRS) offers a non-destructive and rapid method for chemical analysis.

Purpose:

• To develop an improved quantitative analysis model for near-infrared spectroscopy using an incremental neural network integrated with Partial Least Squares (PLS).
• To enhance the prediction accuracy and robustness of flue gas component concentration measurements.
• To enable efficient model updates with new data without requiring access to historical samples.

Summary:

• A novel three-layer Back-Propagation Neural Network (BPNN) model is proposed, incorporating Partial Least Squares (PLS) for optimized initial weight determination and incremental updates.
• PLS regression is used to establish initial weights based on historical data and to compute updated weights when new samples become available.
• The model demonstrates superior performance in predicting carbon dioxide, carbon monoxide, and methane concentrations in flue gas compared to PLS, BPNN, PLS-BPNN, and Recursive PLS (RPLS).

Impact:

• The proposed model significantly reduces prediction errors (RMSEP) for key flue gas components, indicating higher prediction effectiveness.
• The incremental update capability allows the model to adapt to new data without retraining on the entire historical dataset, improving efficiency and reducing computational load.
• The enhanced robustness and generalization of the model make it suitable for real-time monitoring and adaptive control in combustion processes.