Interpretable prediction of sinter particle size distribution based on multi-task learning and bidirectional gated recurrent unit

Abstract

The particle size distribution of sintered ore is the key indicator reflecting the quality of sintered ore products. Aiming at the problems of low accuracy and lack of interpretability of the existing sinter ore particle size distribution prediction models, this paper proposes a multi-task learning prediction model that uses the bidirectional gated recurrent unit as an information sharing layer to predict the particle size distribution of sinter ore. Firstly, features with high correlation with different particle size ranges (<5 mm; <10 mm) are selected as model inputs by the feature selection method. Secondly, a bidirectional gated recurrent unit is used as the information-sharing layer to construct a multi-task learning model, and the information sharing between two tasks is used to obtain their coupling information to consider the correlation between different particle sizes and improve the prediction accuracy of the model. Additionally, the Shapely interpretation method is introduced to analyse the interpretability of the important features of the model to enhance the predictive model interpretability. Finally, the model proposed in this article is compared with the rest of the combined models, and the results show that the proposed model has high prediction accuracy. This method provides a new direction for the high-quality production of sintered ore.