Quantitative Evaluation of DP of Insulating Paper by NIRS Based on Negative Correlation Learning

The introduction of NIR spectroscopy for accurate and nondestructive aging condition assessment of insulating paper has important engineering application values. At present, neural network is the mainstream method to establish the correlation model between NIR spectra and the degree of polymerization (DP), but it has the problems of random weights and susceptible to local optimum, and its accuracy and generalization ability are poor in field applications. This paper proposes a neural network integration strategy based on negative correlation learning (NCL) to achieve weak correlation coupling of multiple sub-neural networks, which adding correlation penalty in the cost function to enhance the feature information diversity of neural network integration and improve the model accuracy and generalization ability. The results show that: the appropriate number of sub-neural networks and correlation penalty coefficients can utilize the maximum advantages of negative correlation integration when modeling; compared with a single neural network, negative correlation neural network integration (NCNN) has higher accuracy (RMSE =74.1, MRE=8.0%), it can be used as an alternative method to neural networks for a more precise model.