A Data-Driven Soft Sensor Based on CVA-DTR Model for Tennessee-Eastman Process

Abstract

Soft sensor is widely used in the industrial process to capture variables that are difficult to measure. Multivariate statistical analysis has wildly applied to the industrial soft sensor. However, most of the modeling methods have disadvantages in coping with the process that simultaneously contains nonlinear and dynamic characteristics. Considering that the industrial process's dynamic and nonlinear characteristics are widely contained, we proposed an algorithm that combines canonical variate analysis (CVA) and decision tree (DT), named the CVA-DTR method, to simultaneously capture those characteristics. In this method, features related to the time the most closely will be extracted by CVA first. Then we develop the DT model to associate the temporal features and the output, which handles the nonlinear characteristics. The Tennessee Eastman process (TEP) case is applied to illustrate to prove its correctness and practicability. The results show that the proposed method gives excellent performance in time-varying information analysis and data prediction.