A Hybrid-Structure Soft Sensor Based on Model Risk Assessment and Weight Fusion for Batch Processes

In batch processes, some quality variables, which are known as primary variables, are hard to be measured such as product concentration, biomass concentration, and substrate concentration. The online prediction of primary variables is essential and significant for process control and fault diagnosis, which directly affects the safety and stability of manufacturing processes. In the prediction of primary variables online, soft sensors play an important role. Due to the limitation of the number of samples and the generalization ability of modeling methods, it is a challenge to develop a soft sensor with high accuracy and strong robustness. The out-of-sample deviance will increase significantly, especially lack of model risk assessment. Therefore, a soft sensor based on model risk assessment and weight fusion is developed and a hybrid model is established to realize the online prediction for batch processes. The weight-fusion-based soft sensor first constructs a weight-fusion-based structure with undetermined weights to introduce the model risk into hybrid modeling. Sequentially, a model risk assessment method is proposed and applied to the calculation of the weights in the hybrid model. As a result, the weight-fusion-based hybrid model can be determined and the online prediction of primary variables can be realized. The proposed soft sensor considers the influence of the model risk and improves the prediction accuracy and robustness. A fermentation process of penicillin was applied to illustrate the accuracy and robustness of the soft sensor based on model risk assessment and weight fusion.