Composition analysis of ceramic raw materials using laser-induced breakdown spectroscopy and autoencoder neural...

Ceramic manufacturing relies on precise elemental composition in raw materials. Variations in Si, Al, Mg, Fe, and Ti content affect product quality. Traditional methods are slow and costly. Rapid analysis techniques are needed. LIBS is promising but faces matrix effects. Neural networks can help but risk overfitting. This paper addresses these challenges with a novel approach. The study focuses on reducing overfitting in spectral data analysis. The goal is to improve accuracy and efficiency in ceramic raw material testing.

The aim is to develop a reliable method for analyzing ceramic raw materials using LIBS and machine learning. Matrix effects complicate accurate elemental quantification. High spectral data dimensions increase overfitting risks. The study introduces a new feature extraction method. Combining LR and SUAC reduces dimensionality. This approach improves model training efficiency. The method is tested for performance against alternatives. The goal is to enhance accuracy while minimizing overfitting.

Laser-induced breakdown spectroscopy was used to collect spectral data. The dataset included 8188 spectral features per sample. Linear regression was applied for initial dimensionality reduction. A sparse and under-complete autoencoder followed for nonlinear feature extraction. The data dimension was reduced from 8188 to 100, then to 32. Back Propagation Neural Network was used for final analysis. The LR + SUAC combination was tested for performance. Cross-validation compared this method with others.

The LR + SUAC + BPNN model outperformed other methods in cross-validation. Dimensionality reduction improved model training efficiency. The spectral data was reduced from 8188 to 32 features. Overfitting was significantly reduced in the final model. The method achieved the highest accuracy in elemental quantification. The results suggest this approach is effective for ceramic raw materials. The model's performance was consistently better than alternatives. The study confirms the effectiveness of combining LR and SUAC for this purpose.

The LR + SUAC + BPNN model effectively reduces overfitting in LIBS data analysis. This method improves accuracy in ceramic raw material composition analysis. The combination of linear and nonlinear dimensionality reduction is key. The approach is suitable for high-dimensional spectral data. The study supports the use of this method in industrial settings. The results align with the authors' hypothesis about model performance. The method provides a practical solution for matrix effect challenges. The findings suggest this technique can be applied to other material analyses.