Resolution of overlapping terahertz spectra using non-negative matrix factorization base on pure variables initialization

In many cases, especially for multi-component mixture, the overlap of spectra will greatly affect the identification accuracy. It is not a reliable method to identify components only by analyzing absorption peaks of mixture directly. Resolution of overlapping THz spectra is a useful way for identification. In this paper, we used the non-negative matrix factorization (NMF) base on pure variables initialization for resolution THz spectra of multi-component mixture. Instead of random non-negative initialization, we introduced the pure variables of which makes the initialization process more intuitive and more interpretable. To verify the proposed method, binary and ternary mixture was preparation. The results show that NMF base on pure variables initialization can resolve the THz spectra of mixture into the spectra of components. Moreover, the initialization method base on pure variables needs fewer iterations than the initialization method base on random non-negative values. It also means that NMF base on pure variables initialization is a potential new approach to identify each pure component in a multi-component mixture.