Time-Attention Graph Convolutional Network Soft Sensor in Biochemical Processes

Most data-driven soft sensor methods can model nonlinear time-varying characteristics of biochemical processes. However, the intrinsic relationship between variables, which is helpful for understanding model behavior, has rarely been investigated in existing data-driven methods. In this work, a novel soft sensor model of time-attention graph convolutional network (TA-GCN) is proposed, which jointly leverages variable relationships and long-term temporal dependencies to improve interpretability and prediction accuracy. This model first uses the maximum information coefficient to construct a topology graph and trains edge strengths end-to-end. The data are then encoded in the spatial-temporal dimension based on GCN and attention mechanism. Finally, the empirical knowledge that analyzes the operating state of the process and graph are combined to explain the model behavior. In comparison to existing soft sensors, TA-GCN enables efficient and scalable training for long-term spatial-temporal dependencies. Experimental results on InPenSim dataset demonstrate that TA-GCN is competitive with state-of-the-art methods.