Design of Data-Driven Soft Sensor for Quality Prediction in Industrial Polyester Resin Production Process

In the present study, a data-driven soft sensor is designed based on a state-dependent parameter modeling method using the Local Instrumental Variable (LIV) approach for a polyester resin production batch process. Data from an industrial process has been used for soft sensor modeling. To design an accurate soft sensor, the non-stationary characteristic of the process is considered in the calculations by adopting the output variable of the previous moment to the set of input variables. The number of input variables of the final model was reduced from 23 variables determined by process knowledge to only 4 variables for viscosity and 3 for acidity number in this study. The final model of the soft sensor was trained with the data of one batch, as a result, the time and amount of calculations were significantly reduced. The performance results of the LIV method by MAE, RMSE, and R2 indicators were obtained as 0.0015, 0.0019, and 0.9999 for viscosity and 0.0030, 0/0094, and 0/9995 for acidity number, respectively for the batch process of polyester resin production. Compared to other soft sensor modeling methods, the LIV model predicted the quality index variables (QIV) of the product more accurately using less number of batches and input variables for model training.