Soft Sensor Development Based on Hybrid Modeling With Ensemble Learning for Multimode Batch Processes

Batch processes are important manufacturing process. Data-driven soft sensor has become an essential means to measure key process variables of batch processes online. However, due to the complex characteristics of batch processes, data-driven soft sensor models do not always perform well and these models with large measurement errors at sometimes are risky. In addition, purely data-driven soft sensor models lack interpretability, which limits the application of soft sensor technology in batch processes. To address these issues, an effective soft sensor based on hybrid modeling with ensemble learning for multimode batch processes is proposed. First, the proposed soft sensor uses the warped K-means (WKM) algorithm to partition the batch process into multiple modes, with bootstrap sampling and different secondary variables to construct training data for each mode; Then data-driven soft sensor base models and mechanism-based soft sensor models are established for each mode. Considering the influence of sample size, model parameter quantity on soft sensor models, Bayesian information criterion (BIC), and structural risk minimization principle (SRMP) are introduced to assess model risk of multimode soft sensor base models. According to the model risk assessment results, each mode base model is pruned; finally, the stacking ensemble strategy is used to fuse mechanism-based models and data-driven models. The experimental results of the penicillin fermentation process and the industrial yeast fermentation process demonstrate the effectiveness of the proposed soft sensor.