A spatio-temporal data decoupling convolution network model for specific surface area prediction in cement grind process. (April 2023)

• Record Type: Journal Article
• Authors: Hao, Xiaochen; Huang, Gaolu; Li, Ze; Zheng, Lizhao; Zhao, Yantao
• Is Part Of: ISA transactions. Volume 135(2023)
• Page Start: 380
• Page End: 397
• Publication Date: 2023-04
• Subjects: Convolutional neural network -- Industrial map -- Prediction of cement quality -- Feature extraction -- Spatio-temporal decoupling; Engineering instruments -- Periodicals; Engineering instruments; Periodicals; Electronic journals; 629.805
• DOI: 10.1016/j.isatra.2022.10.006
• Languages: English
• ISSNs: 0019-0578

Abstract

Abstract: The specific surface area is one of the important indicators for measuring the quality of cement products. Realizing accurate prediction for specific surface area is very important for the production scheduling of the cement industry, energy conservation and consumption reduction and improvement of cement performance. However, due to the non-linearity, uncertainty, multiple interference, dynamic time-varying delay and multi scales in cement grinding process, it is difficult to establish an accurate soft-sensing model for cement quality prediction. Aiming at the above problems, we proposed a spatio-temporal decoupling convolution neural network model (STG-DCNN) to predict specific surface area by extracting and fusing data features in temporal and spatial dimension. To complete the prediction of specific surface area, we established the temporal series map and spatial series map by the production variables data according to the mechanism of cement grinding process. Then, sliding window technique was utilized to match the time scale in temporal series map and construct variable coupling relationship in spatial series map. The prediction accuracy, robustness and superiority of the proposed method were demonstrated by experiments results implemented on the actual cement grinding quality management database in a cement production enterprise. Highlights: A method of spatio-temporal map is proposed to make industrial data graphical. The multi-scale sliding window method is used to reconstruct the input layer of a dual-channel convolutional neural network. The dual-channel model structure is built to extract and fuse the features of spatio-temporal sequences. The prediction of the specific surface area of cement provides reference for energy scheduling and quality inspection.