Abstract

The lack of real-time measurement of certain critical product and process characteristics is a major problem in the manufacturing industry, and it can lead to an out of specification production. A soft sensor is a predictive model that uses readily available process measurements to infer variables that are impossible or difficult to obtain in real-time. In this work, historical process data related to the black liquor recovery circuit from a Canadian kraft pulp and paper mill is used to develop soft sensor models for the black liquor solid content at the concentrator feed. Prior to modeling, irrelevant variables and observations not representative of a normal operating regime are eliminated from the dataset. For practical reasons related to modeling restrictions and soft sensor industrial implementation, is proposed that a limited number of variables be used as model inputs. Two Partial Least Squares-based selection criteria are used to select the most relevant predictors. Two different sets of ten variables are obtained and used to develop Sugeno-type fuzzy logic, neural network and Partial Least Regression models. Their predictive performance is compared in order to determine the best model configuration and input selection method. iii Currently, the black liquor solid content at the concentrator feed is measured once every eight hours, by performing a laboratory analysis. The proposed soft sensor model can be used to provide a real-time value of the solid content, allowing operators to monitor the process and act timely if corrective actions are required.