Visualization and accuracy improvement of soil classification using laser-induced breakdown spectroscopy with deep...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Traditional methods for ore classification often require extensive feature engineering.
• Laser-induced breakdown spectroscopy (LIBS) generates complex spectral data.
• Deep learning offers a powerful approach for analyzing intricate datasets without manual feature extraction.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN) model for data mining of LIBS spectra for ore classification.
• To investigate the use of heat diffusion for an affinity-based transition embedding model within the DNN architecture.
• To introduce and validate a novel training set update method to enhance deep learning model performance.

Main Methods:

• Implementation of a deep neural network (DNN) model for processing LIBS spectral data.
• Application of heat diffusion for affinity-based transition embedding to map fully connected layer data non-linearly.
• Development of a DNN output-based training set update strategy for iterative model improvement.

Main Results:

• The DNN model achieved a recognition accuracy of 75.92%, outperforming traditional methods.
• The proposed training set update method significantly improved the model's accuracy to 85.54%.
• The combined LIBS and DNN approach demonstrated high accuracy for ore classification.

Conclusions:

• Deep learning, particularly DNNs, is highly effective for spectral detection and ore classification using LIBS data.
• The novel training set update method accelerates sample labeling and boosts deep learning model accuracy.
• This integrated approach represents a valuable and accurate tool for the geological and mining industries.