Stable Soft Sensor Modeling for Industrial Systems

Abstract

Learning trustworthy models is essential for machine learning tasks, as many researchers have revealed the vulnerability of machine learning models, especially when the fundamental independent and identically distributed (IID) assumption is not satisfied. Building a trustworthy model is promising when training on big representative data but fails to work with limited data. In this paper, we focus on solving small sample problems and unstable prediction problems in machine learning. First, to deal with small sample problems, we propose using a uniform manifold approximation and projection (UMAP) algorithm to generate high-quality virtual samples. Then, with the generated big data and original small data, we use the stable learning method to achieve stable predictions. In addition to a detailed description of the UMAP algorithm and the stable learning algorithm, we also discuss the corresponding theoretical explanations and implementation details. Finally, several comparison studies are implemented on the Tennessee Eastman benchmark process to validate the effectiveness of the proposed method.

Read or Download: PDF