Application of Gaussian Process Regression on the Quantitative Analysis of the Aging Condition of Insulating Paper by Near-Infrared Spectroscopy
Li Yuan,Zhang Wen-bo,Chen Xiao-lin,Li Han,Zhang Guan-jun
DOI: https://doi.org/10.3964/j.issn.1000-0593(2022)10-3073-06
2022-01-01
Abstract:As the aging condition of the insulating papers determines the remaining lifetime of the oil-immersed transformers, a fast and effective aging assessment method for insulating paper is of great significance. As it is known, the degree of polymerization (DP) is the most direct parameter to characterize the aging condition of insulating papers. However, the traditional detection method or so-called viscometry is time-consuming and destructive. Near-infrared spectroscopy (NIRS) technology, as a non-destructive detection method can rapidly determine the samples' components and contents. Until now, it has been successfully applied in many fields and will hopefully be employed as an alternative method to viscometry. However, the current spectral quantitative analysis method is still not accurate enough to predict the DP of insulating paper samples. In this paper, we introduce Gaussian process regression ( GPR) to predict DP of insulating papers accurately. Firstly, the NIRS database of insulating papers under different aging conditions is established, and in this procedure, the raw spectra are preprocessed by the Savitzky-Golay method to improve the signal ratio to noise. Then GPR models with various kernels are established, and the prediction accuracy and stability of the different models are comparatively studied. The results show that the GPR model with Exp kernel is of poor generalization performance, and the models with Matern32, Matern52 and RQ kernels are highly sensitive to the model parameters. Finally, the SE kernel is selected as the optimal kernel function of the GPR model. The DP prediction results of the SE kernel GPR model are compared with traditional PLS, SVR and BPNN models, and the results show that our established GPR model has the lowest RMSE (65. 5 and 70. 6) and highest correlation coefficient r (0. 94 and 0. 93), both for the training set and testing set. The RMSE of the GPR model is lower than PLS, SVR and BPNN models by 54. 1% , 58. 8% and 12. 9% respectively. It is indicated that the established GPR model can be a powerful tool for the aging assessment of insulating papers by the NIRS technique.