Enhancing Autoencoder-Based Machine Learning through the Use of Process Control Gain and Relative Gain Arrays as Cost Functions

Abstract

Autoencoders are neural networks utilized for unsupervised learning and reconstructing input data, making them helpful in for analyzing industrial process data. To enhance their effectiveness, we introduce two cost functions based on the Gain Matrix and Relative Gain Array (RGA) concepts, referred to in this paper as Gain Autoencoder (GAE) and Relative Gain Autoencoder (RGAE). These cost functions aid in reducing dimensionality and improving the model's performance in industrial settings. This article delves into applying of these functions in machine learning, particularly in autoencoders, to predict Mooney viscosity in styrene butadiene rubber (SBR) production. The findings indicate that the proposed GAE and RGAE models outperform traditional linear (linear regression) and nonlinear models (SVR), as evidenced by an increased explained variance of up to 10% and a decrease in mean square error of up to 13%. The successful integration of advanced data analysis techniques with domain knowledge in process control systems opens up new avenues for optimizing industrial processes and resource utilization.