SIMULTANEOUS DATA SCALING AND TRAINING OF DATA DRIVEN REGRESSION MODELS FOR QUALITY CONTROL OF BATCH PROCESSES

摘要

This contribution presents a novel approach for data scaling and training of Partial Least Squares (PLS) models, for use in applications of multivariate statistical process monitoring and control of batch processes. PLS models are intended for training soft sensors and for quality prediction. Additionally, aspects of model inversion are investigated to explore the use of the trained models for quality control. Batch processes are widely used in food industry as the preferred platform for the production of added value products due to the characteristics of the raw materials, i.e., natural elements with complex flow properties and compositions. This is also due to the nature of the most common transformations which involve bio-processes (e.g., fermentation) and the unit operations (e.g., drying) applied in the food industry. However, due to the complexity of their process' dynamics, quality control in the batch platform has traditionally been limited to the measurement of the end-product properties and the process control to the univariate monitoring in proven ranges. Data-driven models used in applications of fault identification, soft sensors, and quality prediction have been proven and exploited in the industry as they provide a solution for online monitoring and control. However, there are still many limitations on the existing data-driven methods that restrict their performance and application. In this contribution, it is demonstrated how the novel approach for simultaneous data scaling and model training results in significant improvements in the performance of the data-driven models. Advantages of the proposed algorithm regarding rank identification, the accuracy of the quality predictions and PLS model inversion are described.