A multimodal soft sensor method for industrial processes based on manifold learning
Download as PDF
DOI: 10.23977/iset2021.019
Author(s)
Guohong Qin, Zilu Zhu, Qingxi Zheng
Corresponding Author
Guohong Qin
ABSTRACT
Data-driven soft sensor technology is essential for modern industrial processes. However, traditional soft sensor techniques usually assume that the data is in a single mode. Due to the switching of working conditions, industrial data often presents multimodal characteristics, which cannot be covered by the single model method. In this work, a multimodal partial least squares method based on manifold learning is proposed to measure industrial processes' key variables. First, considering that process variables' dimension is higher, and the complex process data is distributed in a manifold space, the original information is reduced to a smaller size by t-distributed stochastic neighbor embedding (t-SNE). Then the data belonging to different modes are clustered. Finally, in each mode, the partial least squares model is established to obtain measured variables' values. A real industrial case demonstrates the effectiveness and superiority of the algorithm.
KEYWORDS
Multimode, soft sensor, industrial process, Manifold learning, Partial least squares