On the use of non-negative matrix factorisation to characterise wastewater from dairy processing plants

A portable field spectrometer was used to record the time-resolved ultraviolet/visible (UV/Vis) spectrum of the effluent stream at two different dairy processing plants (a Drier Plant and Cheese Factory). The spectra exhibited significant variability. As an alternative to the partial least squares regression methods usually used in the online UV/Vis field a non-negative matrix factorisation technique was employed to compress the spectral data. One of the extracted basis vectors had a physical shape associated with protein absorption. The weightings associated with the basis vector explained 80% of the variability in protein concentration as measured with traditional grab sampling techniques (increasing to 94% with a further three vectors), allowing the spectrometer to be retrospectively calibrated to continuously measure protein.