On-line vacuum degree monitoring of vacuum circuit breakers based on laser-induced breakdown spectroscopy combined with random forest algorithm

Abstract

The present study introduces a novel method for online vacuum monitoring of vacuum circuit breakers, which overcomes the limitations of conventional offline monitoring methods in engineering applications. Specifically, the proposed method uses laser-induced breakdown spectroscopy (LIBS) in combination with the random forest (RF) algorithm to monitor the vacuum level of the breaker. The experimental setup included the selection spectral lines from four elements (one in the target material and three in the ambient gas). Spectral data were collected using the LIBS platform under various pressure conditions ranging from 10⁻³ Pa to 10⁵ Pa. An RF model was established to classify the pressure level using spectral data. The classification performance is tested with a confusion matrix. We employed various spectral preprocessing techniques to enhance the distinctiveness of spectral features. The choice of the optimal preprocessing method was determined by assessing the correlation coefficient (R²) and root mean square error (RMSE), resulting in improved accuracy in vacuum level classification. The variable importance random forest (VIRF) algorithm was also employed for feature selection on the raw spectra to eliminate redundant information, thus enhancing the model's efficiency and accuracy. Utilizing out-of-bag (OOB) error as the evaluation metric, we systematically explored the impact of decision tree quantity (n_tree) and the number of selected features (m_try) on the model's performance, optimizing it to its highest performance. The results indicate that the proposed method can achieve an accuracy of over 99% of vacuum classification, comparable to traditional offline vacuum monitoring technology. This demonstrates that LIBS technology combined with the RF algorithm represents a promising new approach for online vacuum monitoring of vacuum circuit breakers.