Design of batch process with machine learning, feature extraction, and direct inverse analysis

1In the context of process design, control, and management, it has become common to utilize time-series data measured in industrial plants to construct mathematical models based on machine learning and to design and control processes based on these models. In a batch process y, it is difficult to design time-series data for each process variable x to meet the target value of the endpoint. Therefore, in this paper, we propose a method based on machine learning to design a batch process through the direct inverse analysis of a Gaussian mixture regression (GMR) model. The GMR model is constructed after the time-series data are transformed into latent variables. Through case studies, it is confirmed that the GMR model can predict y by inputting x into the model and design x for a new batch by inputting y into the model and inverse-transforming latent variables to x. The proposed method can successfully perform the direct inverse analysis of time-series data and design new profiles in the batch process, even when multiple y variables exist.