Soft sensor of processes based on dual attention spatio‐temporal interaction network

Aiming at the problems of layer information loss as well as the effectiveness of process data spatio‐temporal feature fusion in stacked network‐based soft sensor methods, this paper proposes a dual‐attention spatio‐temporal interaction network (DA‐TSINET) method. Firstly, the dual‐attention stacked network is constructed to overcome the layer information loss. Self‐attention is added to different layers of stacked denoising autoencoder (SDAE) to enhance local denoising features, and global enhanced features are obtained by self‐attention fusion. A gated recurrent unit (GRU) and a convolutional neural network (CNN) are used in parallel to further extract spatio‐temporal relations, and an interactive gating module is designed for fusion to obtain globally enhanced spatio‐temporal features for constructing a soft sensor model. Simulation experiments are carried out by debutanizer column and thermal power plant and compared with stacked autoencoder (SAE), SDAE, stacked isomorphic autoencoder (SIAE), variable‐wise weighted SAE (VW‐SAE), and gated stacked target‐related autoencoder (GSTAE); the results show that the proposed method has high prediction accuracy, which verifies the effectiveness of the proposed method.