The potential of Fourier-transform near-infrared (FT-NIR) spectroscopy for qualitative and quantitative analysis in solid-state fermentation (SSF) of protein feed was verified based on FT-NIR spectroscopy combined with multivariate data analysis. The raw spectra were processed and analyzed by multivariate analyses, which integrated the approaches of discrete wavelet transform (DWT), principal component analysis and extreme learning machine (ELM) modeling. The noise of raw spectra was filtered and latent information was extracted by DWT, and then the characteristic information obtained by DWT was visualized in principal component space, in which the structures with the time course of the SSF were explored. Thereafter, some parameters of the calibration models were optimized by cross-validation. The results of the final models were achieved as follows: root mean square error of prediction (RMSEP) = 0.0987/Rp2 = 0.9322 for pH model, RMSEP = 0.0092 w/w/Rp2 = 0.8991 for moisture content model, and an identification rate of 91.43% for the discrimination model of the fermentation phase in the validation set. Finally, compared with partial least squares (PLS)/PLS-discriminant analysis and back propagation artificial neural network models, the ELM model showed excellent performance for prediction and generalization. This study demonstrates that FT-NIR spectroscopy coupled with appropriate chemometrics approaches could be utilized to monitor the SSF, and ELM reveals its superiority in model calibration.
You have access to this article
Please wait while we load your content...
Something went wrong. Try again?