A review of semi-supervised learning for industrial process regression modeling

It is difficult to directly measure the parameters such as product quality and environmental protection index in complex industrial process by testing instrument. This is mainly caused by the restriction of site environment, measurement technology and economic cost. The establishment of data-driven soft sensor model is one of the effective methods for online estimation of these difficult parameters. However, the truth values of these difficult-to-detect parameters are usually obtained after off-line laboratory analysis. This method has the disadvantage of high time lag and high cost, which leads to the lack of labeled samples used to construct soft sensor. In contrast, a large number of auxiliary variables can be acquired and stored in real time through industrial control systems, which are used as input of soft sensor model. This makes Semi-supervised Learning (SSL) become a research hotspot. The reason is that it can make full use of a small number of labeled samples and a large number of unlabeled samples to improve the generalization performance of soft sensor model. With this in mind, this article aims to provide an overview of existing SSL methods for industrial process regression modeling. At first, a brief description of the basic idea, assumptions and applications of SSL is given. Secondly, the semi-supervised regression (SSR) method for industrial process is described from several perspectives. Then, the research difficulties of SSR are pointed out. Finally, the research and outlook is carried out, and the future research direction is pointed out, that is, combining the characteristics of specific industrial process to build the SSR intelligent soft sensor model based on multi-mode data.