Adaptive soft sensor ensemble for selecting both process variables and dynamics for multiple process states

To improve the predictive ability of soft sensors in chemical and industrial plants, the selection of process variables and consideration of dynamics in the processes have been studied. When multiple process states exist in plants, process variables that are related to an objective variable (y) and their corresponding time delays can be different in each process state. In this paper, we propose a method to optimize process variables and dynamics for each process state, and predict the objective variable values using multiple adaptive soft sensors according to process states. First, a dataset is clustered using Gaussian mixture models; then, time-delayed process variables are selected as explanatory variables (x) for each cluster using genetic algorithm-based process variable and dynamics selection. For each set of explanatory variables, a nonlinear adaptive soft sensor is constructed. An ensemble prediction is performed by assigning a weight to each locally weighted partial least squares (LWPLS) model according to its predictive ability. The novelty of the proposed method is the construction of adaptive soft sensors that are optimized for each process state in a plant where multiple process states exist, and the accurate prediction of objective variable values through integrating the constructed adaptive soft sensors as the process state changes. The effectiveness of the proposed method was tested using two datasets from actual plants, and we confirmed that it could accurately predict the objective variable values in each process state.