[Study on the gasoline classification methods based on near infrared spectroscopy].

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Context:

• Gasoline classification is crucial for quality control and regulatory compliance.
• Traditional methods may lack efficiency and accuracy in distinguishing similar fuel grades.
• Near-infrared (NIR) spectroscopy offers a rapid, non-destructive analytical technique.

Purpose:

• To investigate and compare classification methods for gasoline (90#, 93#, 97#).
• To evaluate the efficacy of discriminant cluster analysis in specific spectral regions (700-1100 nm and 1100-1700 nm).
• To develop and validate a novel classification model using principal component analysis (PCA) and self-organizing competitive neural networks (SoCNN).

Summary:

• Discriminant cluster analysis was initially tested, with the 1100-1700 nm region showing higher accuracy.
• A new model was developed by first condensing spectral data using PCA, selecting three principal components (97% cumulative credibility).
• A three-layer SoCNN was trained using 32 PCA-condensed wavelengths, demonstrating feasibility and superior performance over discriminant cluster analysis.

Impact:

• Establishes the viability of NIR spectroscopy combined with PCA and SoCNN for precise gasoline grade classification.
• Provides a more accurate and potentially cost-effective alternative to existing classification techniques.
• Highlights the advantages of PCA for data dimensionality reduction and SoCNN for robust pattern recognition in spectral analysis.