Cross target attributes and sample types quantitative analysis modeling of near-infrared spectroscopy based on instance transfer learning

Abstract

During the near‐infrared (NIR) spectroscopy analysis process, most existing methods can carry out calibration transfer only between the same samples. In the machine learning area, transfer learning has the potential to achieve calibration transfer across different kinds of samples. This ability raises the following questions: Is this transfer process feasible in the field of NIR spectroscopy? How can this transfer process be realized? To solve these problems, on the basics of boosting extreme learning machine (ELM), the instance transfer learning method was applied. The TrAdaBoost for classification problems was improved to the TrAdaBoost for regression. Simulation verification of ten datasets (fuels and foods) from different instruments was performed. The results demonstrated that by applying this instance transfer model after principal component analysis (PCA) dimension reduction, the conditions of NIR spectroscopy analysis could be relaxed; in other words, the target attributes and sample types need not be the same.