A Novel Interpretable Ensemble Learning Method for NIR-Based Rapid Characterization of Petroleum Products

Near-infrared (NIR) spectroscopy has become an important analytical tool to perform rapid characterization of petroleum products due to its effectiveness and efficiency. However, NIR analysis is a typical small sample problem, and a unitary regression model under specific assumption often suffers from poor generalization ability during online implementation. Besides, the interpretability of complex machine learning models is relative incompetent to reveal the variable contribution of wavelength variables, which is significant to guarantee the model reliability in practical application. Thus, this article proposes a novel ensemble learning algorithm to improve the poor model generalization ability caused by insufficient sample size and complex unexpected online application situations. Furthermore, the constructed model is interpreted by combining model-agnostic explanation method with spectral structure knowledge, and a novel reliability index (RI) is proposed to ensure model reliability in practice. The prediction performance of the proposed algorithm is verified using two industrial NIR datasets. Experimental results confirm the effectiveness of the proposed ensemble learning algorithm. The model interpretability and model reliability are guaranteed by the proposed interpretation method.