Abnormality Detection of Blast Furnace Ironmaking Process Based on an Improved Diffusion Convolutional Gated Recurrent Unit Network

Accurate abnormality detection is of critical importance to the blast furnace (BF) ironmaking process monitoring. However, the multivariate, nonlinear, dynamic characteristics of the BF ironmaking process pose great challenges for this task. In this article, we propose a novel graph neural network (GNN) consisting of a sparse relationship learning (SRL) module and an adaptive diffusion convolutional gated recurrent unit (ADCGRU)-based forecasting module, and further apply it in abnormality detection of the BF ironmaking process. Herein, SRL aims to formulate implicit nonlinear relationships among multiple process variables into a sparse association graph. Distinct from most existing graph construction methods based on prior knowledge or similarity metrics, the sparse association graph can be learned by SRL from historical observational data and used to describe the complex variable relationships in a reasonable and explainable way. In addition, a novel ADCGRU is proposed by introducing a self-attention mechanism into the traditional DCGRU to adjust the node adaptability. In this way, the nodes that benefit to improve the dynamic modeling performance will receive more attention, which helps to efficiently capture the dynamic nature of the BF ironmaking process. Furthermore, we propose an SRL-ADCGRU-based abnormality detection scheme for the BF ironmaking process. Experiments on practical data collected from a large-scale BF located in Liuzhou Iron & Steel Company, Ltd., China demonstrate that the proposed method compares favorably against the state-of-the-art approaches for abnormality detection of the BF ironmaking process.