Robust Classification of High-Dimensional Spectroscopy Data Using Deep Learning and Data Synthesis.

Area of Science:

• Spectroscopy
• Machine Learning
• Chemometrics

Background:

• High-dimensional spectroscopy data classification is challenging.
• Current methods struggle with outlier detection and accuracy.
• Identifying chlorinated solvents in samples requires robust classification.

Purpose of the Study:

• Introduce a novel locally connected neural network (NN) for spectroscopy data classification.
• Enhance classification accuracy using synthetic training spectra.
• Develop robust outlier detection for unrepresented samples.

Main Methods:

• Applied a locally connected neural network (NN) for binary classification of Raman spectra.
• Utilized synthetic training spectra to augment the dataset.
• Investigated autoencoder-based one-class classifiers and outlier detectors.
• Developed a two-step classification process combining NN, synthetic data, and outlier detection.

Main Results:

• The locally connected NN demonstrated superior accuracy compared to traditional algorithms.
• Synthetic training spectra further improved the NN's classification performance.
• The two-step approach achieved high accuracy and robustness against negative outliers.

Conclusions:

• Locally connected neural networks offer a powerful new approach for spectroscopy data analysis.
• Synthetic data generation and autoencoder-based outlier detection enhance model reliability.
• The proposed two-step classification method provides an accurate and robust solution for complex spectroscopy tasks.