Spatio-temporal bidirectional redundancy removal for quality prediction of dynamic processes: A novel dynamic feature extraction strategy

Data-driven modeling has been widely recognized as an effective and essential approach for predicting key quality variables in industrial processes. However, due to the long production process and the dynamic fluctuation of operating conditions, it is necessary to integrate data from various spatial and temporal points to comprehensively describe the production process. Furthermore, data is characterized by massive redundancies and high dimensionalities, which pose challenges for modeling dynamic processes. To address these issues, this paper proposes a strategy of spatio-temporal bidirectional redundancy removal to extract dynamic features and applies it to quality prediction. Firstly, it constructs input samples with the spatio-temporal two-dimensional structure. Then, information redundancy is removed by implementing feature extraction in both the spatial and temporal directions, respectively. Additionally, a spatio-temporal weight matrix is used in this process to capture dynamic information which is more related to the output quality variable. Its effectiveness was verified by applying it to the debutanizer column process and the leaching process in zinc smelting. Results show that the proposed strategy generates efficient dynamic features with no information redundancy and low dimensionalities. Moreover, it outperforms other common types of dynamic process modeling methods in terms of quality prediction performance.