Abnormal state prediction of flotation process based on dual attention mechanism and multivariate information fusion

Abstract

AbstractPredicting abnormal conditions in flotation processes is vital for safety, efficiency, and product quality. However, existing studies lack predictions of abnormal working conditions in flotation processes and neglect temporal information in data. To address this, this paper proposes a novel approach for predicting abnormal work conditions in flotation processes. It utilizes a dual attention mechanism and multivariate information fusion. Features are extracted from froth images using the Xception model, a pre‐trained convolutional neural network. These features are combined with flotation process monitoring variables, creating fused data. An encoder and decoder time feature seq2seq (EDTF‐seq2seq) model with time and feature attention modules enables end‐to‐end information fusion and work condition prediction. The attention modules assign weights to each feature point, capturing the time–feature relationship and improving prediction accuracy. Four sets of experiments using real flotation process data validate the effectiveness of the proposed method, achieving favourable prediction accuracy.