An ensemble active learning for a fluidized bed granulation in the pharmaceutical industry. (October 2022)

• Record Type: Journal Article
• Authors: Chen, Zhongxin; Tang, Yongwei; Gao, Zenglin; Zhou, Jun; Huang, Panling
• Is Part Of: Journal of process control. Volume 118(2022)
• Page Start: 16
• Page End: 25
• Publication Date: 2022-10
• Subjects: Active learning -- Pharmaceutical technology -- Selective ensemble -- Conditional variational auto-encoder; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.08.007
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: Active learning (AL), allowing to query the labels by an oracle, can dwindle the generalization error from a few labeled samples, thus well-motivating in the scenario where there are fewer labeled samples and abundant unlabeled samples. However, the query samples cannot be exploited efficiently in the existing methods based on AL for regression tasks, and AL is typically applied for classification problems. In this paper, an active learning framework (CALF) is proposed to improve the efficiency with fewer query samples, aiming at the moisture content prediction in fluidized bed granulation. The proposed method, based on conditional variational auto-encoder (CVAE) and selective ensemble algorithm, can efficiently incorporate the information in query samples into the labeled sample space. The game of CVAE and selective ensemble algorithm improves the performance of the framework effectively, and its effectiveness is verified by the results obtained from the large batch of fluidized bed granulating experiments. Highlights: We propose an AL-based regression method (CALF) for the case of few labeled samples and abundant unlabeled samples. The game of CVAE and selective ensemble algorithm optimizes and improves the performance of the framework. The query criteria over the committee is the ambiguity on input which can reduce the computation complexity. The evaluation results show that the learner based on CALF can obtain better prediction performance through fewer samples.