Comprehensive prediction and grade evaluation of sinter chemical composition by maximise sintering process information

Abstract

Aiming at the problem that the low frequency of sinter chemical composition detection leads to the problem that part of the sintering process data cannot be utilised by machine learning, a sinter chemical composition prediction model that maximise sintering process information is proposed. Firstly, by combining the Gaussian mixture model and the K-nearest neighbour algorithm, useful information on unlabelled samples that cannot be utilised by the model is extracted. Then, combined with long short-term memory, it is used to predict the mass fractions of the four sinter chemical components of TFe, FeO, CaO and SiO₂. Compared with the three models: backpropagation neural network, deep neural network and Elman neural network, the results show that the constructed model has lower prediction error and higher fitting performance, which can fully extract useful information from sensor data. In addition, a method for evaluating the chemical composition grade of sinter has been proposed, which can effectively classify the sinter chemical composition grade of different qualities based on the model prediction results, and can comprehensively evaluate the iron content, reducibility, and alkalinity of sinter.