(476g) Soft Sensor Development of Batch Bioprocess Using Bayesian Inference Based Support Vector Regression

The chemical and biochemical industry usually faces the challenge of unavailable on-line measurements of critical process and product variables, which are necessary to ensure successful control, monitoring and optimization of process operation. As an alternative solution, soft sensor can provide accurate estimates of unmeasured variables in a real-time fashion and has attracted increasing attention in advanced process control field. The commonly used data-driven models in building soft sensors require accurately recorded historical data and lab measurements. In industrial practice, however, the process input variables often have missing values and the time records of lab measurements may be inaccurate as well. Due to such uncertainty of modeling data, the traditional multivariate statistical or artificial intelligence methods can easily fail to construct reliable soft sensors. In this study, the Bayesian inference strategy is employed to handle the historical data uncertainty and estimate the posterior probabilities of input process variables and output lab measurements. Then the missing process values and misaligned lab data can be estimated and corrected within the probabilistic framework. Further, the support vector regression (SVR) is used to fit the predictive model between the process measurements and the lab based product quality variable. The proposed approach is tested on a fed-batch penicillin fermentation process and found to be effective in developing accurate soft sensors for key product variable predictions.