Linear and nonlinear hierarchical modeling strategy for dynamic soft sensor

In real industrial process, linearity and nonlinearity often exist at the same time, which brings difficulty to the modeling of soft sensor in industrial process. In this paper, a linear and nonlinear hierarchical strategy is proposed for soft sensing of dynamic processes. First, a linear identification coefficient (LIC) is designed to measure the degree of linear correlation between input variables and output variables. Process variables are divided into linear variable group and nonlinear variable group. Then, we use dynamic partial least squares (DPLS) to build a linear model. In view of the prediction residuals of linear models, a long short-term memory (LSTM) model is established to fit them, so as to compensate for the failure of linear methods to capture nonlinear relationships. The validity of the method is proved by the experiment of three-phase flow. Compared with other linear and nonlinear models, the proposed method has better accuracy and clearer structure.