Locally spatiotemporal soft sensor for key indicator prediction in cement production process

Abstract

Cement production exemplifies a complex chemical engineering process where chemical reactions and material transformations critically affect product quality. Noise and outliers pose significant challenges in developing reliable soft sensors for cement quality key indicator prediction. To this end, a just-in-time two-dimensional long short-term memory (JTLSTM) soft sensor with correntropy is proposed for predicting free calcium oxide (f-CaO), as a key indicator of cement production. First, convolution operations are employed to extract spatial information and describe process non-linearity. LSTM is then set up to model the process dynamics by capturing temporal information. The avoidance of outlier influence is achieved by adopting a correntropy-based objective function that automatically identifies and assigns small weights to the outlier. Moreover, to accommodate dynamism, JTLSTM is operated in a just-in-time learning manner and updated parameters. The soft-sensing task about f-CaO in a practical cement process demonstrates the superiority of JTLSTM compared to several existing soft sensors.