A nonnegative matrix factorization algorithm for the detection of chemicals from an incomplete Raman library

摘要

Raman spectroscopy has proven to be a powerful technique for the standoff identification of surface-deposited chemical agents. In the supervised detection framework, the measured Raman spectrum is compared to a reference library of known spectra. A well-known shortcoming of the supervised approach is that no comprehensive library exists, and when chemicals are present that are not contained in the reference library, the supervised algorithms may confuse those chemicals with library members. One way to deal with this problem is to use an unsupervised method such as nonnegative matrix factorization (NMF) to estimate both the constituent spectra and their relative quantities directly from a block of measured spectra. Chemical identification may then be performedby associating the extracted spectra with the reference library spectra. This two-stage NMF approach often fails because knowledge of the reference library was not used in extracting the spectra. We present anovel modification of NMF in which a subset of the extracted spectra are constrained to be equal to the known reference library. This method is shown to outperform the standard NMF approach and the common supervised identification algorithms when there are chemicals present that are not in the library. This algorithm is applicable to any problem in which a target is identified by comparing a block of measured data to a library of known constituent signatures.