ABSTRACT
Spectral measurement techniques, such as the near-infrared (NIR) and Raman spectroscopy, have been intensively researched. Nevertheless, even today, these techniques are still sparsely applied in industry due to their unpredictable and unstable measurements. This paper put forward two data fusion strategies (low-level and mid-level fusion) for combining the NIR and Raman spectra to generate fusion spectra or fusion characteristics in order to improve the in-line measurement precision of component content of molten polymer blends. Subsequently, the fusion value was applied to modeling. For evaluating the response of different models to data fusion strategy, partial least squares (PLS) regression, artificial neural network (ANN), and extreme learning machine (ELM) were applied to the modeling of four kinds of spectral data (NIR, Raman, low-level fused data, and mid-level fused data). A system simultaneously acquiring in-line NIR and Raman spectra was built, and the polypropylene/polystyrene (PP/PS) blends, which had different grades and covered different compounding percentages of PP, were prepared for use as a case study. The results show that data fusion strategies improve the ANN and ELM model. In particular, mid-level fusion enables the in-line measurement of component content of molten polymer blends to become more accurate and robust.
