Ensemble-learning for pressure prediction in vacuum circuit breaker using feature fusion of laser-induced plasma spectra and images

Abstract

The detection of vacuum level using laser-induced plasma shows a great potential for safe and reliable online monitoring of vacuum level on vacuum interrupter. However, traditional methods often rely on a single type of plasma signal (spectra or image), which leads to relatively low accuracy. To tackle it, this paper employs ensemble learning to fuse plasma spectra and images to construct a regression model using Gradient Boosting Regression (GBR), Multilayer Perceptron Regression (MLP), Support Vector Regression with Radial Basis Function (SVR-RBF), and Linear Regression (LR). The model consists of Image prediction module, Spectra prediction module, and Ensemble learning module. In the first two modules, the base models predict pressure based on images and spectra, respectively. The predictions are then stacked in Ensemble learning module and fed into LR for final pressure prediction. The results indicate that the base models within the Image prediction module and Spectra prediction modules can predict pressure; however, their prediction errors are significantly greater than those of Ensemble learning module. The prediction errors and evaluation parameters reveal that ensemble learning, in combination with feature fusion, effectively reduce the prediction errors, and significantly enhance predictive performance of model. More importantly, the Limit of Quantification (LOQ) of this method reaches 10⁻⁵–10⁻⁴ Pa, meeting the requirements of the electric power industry standards. A beneficial exploration has been conducted on the application of vacuum level detection based on laser-induced plasma.