A multimode soft sensor method combining common feature transfer and special feature correction

Aiming at the problem of multiple conditions for complex industrial processes, in this paper, a multimode soft sensor method combining common feature transfer and special feature correction is introduced. First, a novel public information extraction method is used to extract common features and special features of multi-condition data. After using the geodesic flow kernel (GFK) to achieve distribution alignment, the initial condition is selected as the modeling condition, respectively, the partial least squares regression (PLSR) models with other historical conditions and target conditions are established. Then, the label biases of historical conditions can be obtained by comparing the real label with the pseudo label. The relationship between label biases and special features in historical conditions can be represented by the label correction model. Finally, label correction value of target condition can be obtained by its special feature, and soft sensor value of the dominant variable is obtained. The simulation experiment in the Tennessee Eastman (TE) process proves the validity and practicability of the model.