Quality prediction for multi-grade batch process using sparse flexible clustered multi-task learning

Abstract

Data-driven quality prediction methods are widely used in industrial chemical plants. However, it is often difficult to develop prediction models for multi-grade batch processes. Two major issues need to be considered when developing high-accuracy models. The first is the unavailability of sufficient data to create models for each grade of these processes. The other is that each batch cycle typically has an excessive number of explanatory variables. This paper proposes two methods to predict the quality of products manufactured in these multi-batch processes in chemical plants. These methods combine the features of two techniques: the first is a flexible clustered multi-task learning method, which utilizes data from other grades effectively to create high-performance quality prediction models with a small amount of data. This is useful when more data are available for the other grades. The other is a sparsity technique to overcome the high-dimensionality problem of input features. The effectiveness of the proposed methods is demonstrated on a numerical dataset, and finally applied to data generated during an actual industrial blow molding process.