Heterogeneous data-driven measurement method for FeO content of sinter based on deep learning and tensor decomposition

Abstract

FeO content is an important index reflecting the reducibility and drum strength of sinter, which is of great significance to the subsequent iron-making process. However, due to the time delay of chemical analysis method, FeO content cannot be obtained in real-time, which promotes the research on online measurement methods. Based on deep learning and information fusion, this paper proposes an intelligent heterogeneous data-driven method to accurately measure the level of FeO content in real time. Firstly, in view of the interference information such as background and furnace contour, the region proposal network is adopted to detect the burning ore region in the cross-section image of sintering machine. Secondly, a split attention-based deep convolutional neural network is constructed to extract deep image features. In particular, a fuzzy labeling strategy is proposed aiming at the problem of dividing images at the junction of different levels. In the end, the image features and the sintering process variables are fused through a bilinear model, and tensor decomposition is introduced to obtain the final fusion feature. Experimental results on actual industrial data demonstrate the effectiveness of the proposed method for measuring the level of FeO content, achieving an excellent accuracy of 98.35%.