A distributed PCA-TSS based soft sensor for raw meal fineness in VRM system

Abstract

Vertical roller mill (VRM) is an increasingly popular comminution equipment in cement plants. Raw meal fineness at the outlet of VRM is one of the most important indicators to measure product quality. A soft sensing model developed for powder fineness in real-time could assist operators to monitor the comminution process online. However, due to frequent fluctuation of raw material properties, intrinsic nonlinearity of the process and changeable operation conditions, the data present a multimodal characteristic. Therefore, this paper proposes an indicator to measure the similarity between variables, that is, the shortest distance between nodes in the constructed weighted network. By combining the Girvan Newman (GN) algorithm, the nodes in the variable network are divided into multiple groups, and based on this, the distributed PCA (DPCA) similarity is adopted for time series segmentation (TSS). Compared with traditional similarities between samples (distance, density, etc.), the similarity between time series focuses more on the dynamic characteristics of variables. And the implementation of DPCA similarity is equivalent to increasing the sparse characteristics of principal components, which is beneficial for enhancing the generalization of the model. Support vector regression (SVR) models along with a support vector machine (SVM) based classifier are built to obtain final predictions of the powder fineness. Effectiveness of the proposed method is verified by actual industrial data. It has a root mean square error (RMSE) index of 0.3451 on the test set, which is much smaller than that of the multi-modal soft sensor based on either clustering approaches (0.6524) or PCA similarity based TSS method (0.4282).