Soft sensors for industrial processes based on hierarchical spatio-temporal enhance quality-related network

Abstract Aiming at the problem that process data have spatio-temporal characteristics and the accumulation of quality-independent information layer by layer leads to the reduction of the effectiveness of soft sensor models, a method based on hierarchical spatio-temporal enhancement quality-related network (HSTE-QN) is proposed. By stacking the HSTE-encoder, deep local and global spatio-temporal features are captured. Specifically, based on the construction of data graph, the graph attention network was used to dynamically capture the spatial features. The cross-convolutional network is proposed to polymeric local temporal features from spatial features, and a multi-head hierarchy time enhancement module is further designed to reinforce their global features. Finally, adaptive strategies are used to integrate local and global multi-level spatio-temporal features. The loss function of the quality regularization mechanism is constructed to overcome the accumulation of quality-independent information when stacking layer by layer. The maximum information coefficient and Jacquard similarity coefficient are introduced to calculate the nonlinearity and spatial correlation between inputs and qualities at each layer of the network, so as to obtain the hierarchical spatio-temporal features of the quality correlation and prevent overfitting. After that, the soft sensor is modeled through the fully connected layer. The experimental results demonstrate that, in the case of the debutanizer tower, the proposed HSTE-QN model achieves an root mean square error (RMSE) of 0.0145 and an R 2 value of 0.9865. In the sulfur recovery case, it attains an RMSE of 0.0251 and an R 2 value of 0.8072. These results significantly outperform the seven compared models, validating the prediction accuracy and stability of the proposed method.