A Model Integration Strategy for Quantitative Aging Assessment of Insulating Paper by NIRS

In recent years, near-infrared spectroscopy (NIRS) has been used for non-destructive, quantitative assessment of the aging condition of insulating paper. Massive algorithms are proposed to explore the relationship between the NIR spectrum of insulating paper and its degree of polymerization (DP). However, how to evaluate and choose an optimal algorithm of high precision and generalization ability from many algorithms deserve to be further studied. In this paper, we propose a two-layer ensemble learning strategy that takes the output of the base learners as the input of the meta-learner, maximizes the use of valid information mined by each algorithm, and corrects the prediction bias of the base models. Firstly, the selection strategy of base learners and meta-learner is given, and the general structure of the two-layer ensemble learning is described. Then, we test the performance of the base model and the meta-model using a dataset of 509 samples including 8 types of insulating paper. The results show that the two-layer ensemble learning has higher modeling stability (RMSEvar of 68), and the obtained meta-model has higher accuracy (RMSEmean of 72) compared to the best base model (RMSEmean of 79, RMSEvar of 79). The two-layer ensemble learning has a significant advantage compared to the winner-take-all strategy.