Deep Graph Network for Process Soft Sensor Development

In the (bio)chemical processes, traditional hardware sensors are difficult to directly measure the quality of critical products due to their time-varying, non-linear, and dynamic characteristics. This makes process soft sensor modeling methods important. Since the process variables can be regarded as natural graph data, this work introduces graphs in the soft sensor modeling area. A soft sensor model based on the graph neural network (GNN) is proposed. The model can learn the topological structure of graph data between each unit variable. Moreover, it characterizes variable relationships from the spatial and temporal dimensions to the output prediction by introducing the spatial-temporal convolutional layer. The effectiveness and advantages of the GNN-based soft sensor model are verified using a simulated fermentation process.