Improving Autoencoder Training with novel Goal Functions based on Multivariable Control Concepts

Abstract Autoencoders are becoming more representative in all fields of knowledge, due to their ability to classify, compress, and identify data patterns. This study objective was to propose entirely new objective functions using multivariable process control concepts as the gain matrix and Relative Gain Array to improve the quality of prediction and classification of an autoencoder. The advantages of the proposed approach are illustrated through a pulp-and-paper industry. The new function results show an improvement in the detection, leading to savings of up to 22 to 38 thousand dollars per month compared to a model using only MSE.