Prediction of abnormal conditions for fused magnesium furnace based on improved LSTM model and rule-based reasoning

Abstract

There are some abnormal conditions during the fused magnesia smelting process (FMSP), making it crucial for the operators to predict their occurrence to ensure proper guidance of the operation. In this paper, a prediction method is proposed for current and sound signals in the FMSP based on the Temporal convolutional network-Bidirectional-Long short-term memory-Attention (TCN-Bi-LSTM-Attention) model. Through analysis of abnormal conditions in the FMSP, according to the characteristics of abnormal conditions and the change of current change rate before and after prediction, an abnormal condition identification scheme based on rule-based reasoning (RBR) that utilizes current and sound signals is constructed. The results of the experiment showed that when considering the same prediction steps, the proposed model had higher prediction accuracy compared to the TCN-Attention model and the Bi-LSTM-Attention model. The proposed abnormal condition identification scheme can predict the type of abnormal condition that will occur in the next 60 s, with recall, precision, and F1 scores of over 90% for four different abnormal conditions.