Noise-Tolerant Co-Trained Semisupervised Soft Sensor Model for Industrial Process

This is the authors' abstract. We don't add key points for this paper.

Soft sensors are often used to obtain variables that are difficult to directly measure in an industrial process. In this article, a co-training-based semisupervised soft sensor model is presented to address noise removal and labeled sample acquisition in an industrial process. To fully utilize unlabeled samples, a semisupervised method based on co-training is proposed to build a semisupervised model based on <tex-math notation="LaTeX">ε </tex-math> -SVR (SSCo- <tex-math notation="LaTeX">ε </tex-math> SVR) soft sensor model. At the same time, to improve the noise tolerance of the soft sensor model, <tex-math notation="LaTeX">ε </tex-math> -insensitive loss is introduced in the training process. To improve the nonlinear data estimation ability of the soft sensor model, the gradient-boost ensemble learning method is introduced and an ensemble SSCo- <tex-math notation="LaTeX">ε </tex-math> SVR (ESSCo- <tex-math notation="LaTeX">ε </tex-math> SVR) soft sensor model is established. The noise tolerance and generalization ability of the soft sensor model are discussed using numerical simulations. Furthermore, the debutanizer column data and actual data from a coking process are used to establish a soft sensor model. The results demonstrate that the method is effective and practical.