A Spatial-information-based Semi-supervised Soft Sensor for f-CaO Content Prediction in Cement Industry

f-CaO is a key factor affecting the quality of cement in production. In this paper, the cement clink production process is introduced and discussed in detail. The time delay between the variables leads to an inaccurate matching relationship with each other and defects the performance of traditional soft sensors. To this end, a semi-supervised spatial-information-based soft sensor for f-CaO content is proposed. First, we analyzed the relationship between process variables and quality variable and then reconstruct the input of samples into data matrix by stitching unlabeled process data together. The semi-supervised structure helps retain process information in the data. Then, an end-to-end soft sensor based on CNN is established: convolution and pooling operations are used to extract the features of two-dimensional data containing spatial information; a multi-layer perceptron models the extracted features regressively. Further, in order to solve the defect of insufficient generalization ability of the CNN-based model, a framework for spatial feature extracting and transferring is proposed. Compared with the multilayer perceptron, strong regression models with spatial features get better prediction accuracy. An actual cement production case is used to verify the effectiveness of the proposed method.