THz spectral data analysis and components unmixing based on non-negative matrix factorization methods.

Area of Science:

• Spectroscopy
• Data Analysis
• Matrix Factorization

Background:

• Terahertz time-domain spectroscopy (THz-TDS) data from complex samples often contain overlapping information, necessitating advanced data unmixing techniques.
• Traditional Nonnegative Matrix Factorization (NMF) can be unstable due to initialization sensitivity and noise, limiting its application in THz-TDS data analysis.

Purpose of the Study:

• To develop and evaluate smooth constraint Nonnegative Matrix Factorization (CNMF) algorithms for improved feature component extraction and identification in THz-TDS data.
• To compare the performance and robustness of CNMF against standard NMF for data decomposition tasks.

Main Methods:

• Implementation of low-rank approximate NMF and CNMF algorithms incorporating sparseness, independence, and smoothness constraints.
• Evaluation of NMF and CNMF algorithms using simulated THz-TDS data from binary and ternary systems, including applications like medicine tablet inspection.

Main Results:

• CNMF demonstrates superior performance in finding optimal solutions and exhibits greater robustness to random initialization compared to standard NMF.
• The CNMF method effectively extracts and identifies feature components from complex THz-TDS data, even with overlapping information.

Conclusions:

• Smooth constraint NMF (CNMF) is a promising advancement for THz-TDS data resolution and unmixing.
• The developed CNMF approach enhances the reliability and accuracy of identifying unknown mixtures in spectroscopic data.