Online Updating Soft Sensor Modeling and Industrial Application Based on Selectively Integrated Moving Window Approach

In this paper, the moving window (MW) approach is introduced to update the soft sensor model with the latest process information, which provides powerful efficiency of tracking the shifting feature of the process. The weighted supervised latent factor analysis is utilized to build predictive model in the MW. However, information on previous important process variations is lost when the models are updated by the MWs. To fully take advantage of the former windows, a set of most recent window data sets are stored to be integrated by the Bayes' rule for the current quality estimation. Nevertheless, these windows may contain similar information about the process, and the redundancy would increase the calculation with a low-efficient accuracy improvement. To this end, a selecting method is proposed through a statistical hypothesis testing to determine whether the local model should be incorporate to the model set or not. In this way, the most informative models are remained to integrate an efficient predictive model. A real industrial case study demonstrates the feasibility and efficiency of the proposed online updating soft sensor.