Dynamically scaling spatio‐temporal semi‐supervised adaptive networks based soft sensor for industrial process

Abstract Aiming at the question of information loss between layers when mining spatiotemporal features of process data and whether pseudo‐labels are generated for unlabelled data, this paper proposes the dynamically scaling spatio‐temporal semi‐supervised adaptive networks based soft sensor for industrial process (DSST‐SSAN). In order to extract the local temporal correlation features and decrease the information loss between layers, the dynamic scaled spatio‐temporal feature module is constructed, the local prediction models between the current input and the hidden layer features are built in each hidden layer of the long short‐term memory (LSTM) network respectively, the prediction deviations of multiple local models are calculated and the dynamic scaled factors are constructed to update the corresponding hidden layer features. The spatial features are extracted in parallel using graph attention network (GAT), and the spatio‐temporal features are obtained by fusion to establish a soft sensor model. To address the lack of modelling labelling data, a semi‐supervised thresholding mechanism is proposed to filter the pseudo‐labelled data for adaptive data accumulation. The threshold is constructed using the likelihood root mean square of the root mean square error (RMSE) and mean absolute error (MAE) of the labelled data, which can determine whether the unlabelled data need to generate pseudo‐labels and perform modelling data accumulation and thus update the model. The effectiveness of the proposed method is confirmed by simulation experiments on two industrial cases, debutane tower and sulphur recovery.