Near-Infrared Spectrum Calibration by Deep Dictionary Learning Regression With Data Augmentation for In-Situ Moisture Measurement of Fluidized Bed Drying Process

For in-situ measurement of moisture content (MC) during fluidized bed drying (FBD) processes via the near-infrared (NIR) spectroscopy, a novel spectrum calibration method is proposed by developing a deep dictionary learning regression (DDLR) modeling approach based on sampled data augmentation (DA), to address the troublesome issues of very limited and uneven distributed samples along with spectral absorption nonlinearity in practice. The robust principal component analysis (RPCA) is combined with the synthetic minority oversampling technique (SMOTE) for DA, which could balance the distribution of all samples, eliminate the influence of random noise associated with SMOTE when introducing the augmented data, and sort out outliers in the synthesized virtual spectra and MC labels. The dictionary learning model is subtly extended to a deep dictionary learning (DDL) form in order to describe the latent features of measured spectra for model calibration. Experimental results on predicting the MC of silica gel granules during an FBD process demonstrate that the root-mean-square error (RMSE) could be significantly reduced by the proposed NIR calibration model, over 45% compared to traditional calibration modeling methods.