Robust Multi-task Learning for Calibration Transfer in DP Detection by NIRS of Insulating Paper

Abstract

In recent years, researchers have proposed the use of Near-Infrared Spectroscopy (NIRS) to detect the Degree of Polymerization (DP) of insulating paper, thus obtaining the aging of transformer insulation in a convenient, fast and nondestructive way. To cope with the problem that the previously established quantitative analysis models are no longer applicable to newly produced spectrometers due to the differences between spectrometers, also called calibration transfer problem, we proposed an robust multi-task learning (RMTL) method, which unites the multi-task learning model of trace norm regularization and l<inf>2,1</inf> norm regularization to obtain the correlation relationships between tasks, improving the generalization ability of each task and reducing the risk of overfitting. Therefore, RMTL can use the large amount of data accumulated by the host spectrometer (HS) and the small amount of data from the slave spectrometer (SS) to train at the same time to obtain a relatively high-quality quantitative analysis model of the slave machine. In addition, we compare the RMTL method with the classical DS, PDS, MU-PLS, PLS with direct slave modeling, and three other multi-task learning methods with different norm regularization, and the results show that the proposed method has the best performance in terms of root mean square error (RMSE) and correlation coefficient(R) on the dataset.