A survey of deep networks-based data imputation and soft sensing for industrial processes : from small models to large models

Abstract

Process data, characterized by strong nonlinearity, dynamics, and complex coupling, are ubiquitous in real-world industrial production. With the rapid development of increasingly complex modern industries, traditional shallow models struggle to capture the wealth of implicit information in massive industrial data. The robust feature extraction capabilities of deep neural networks have inspired the development of a large body of deep networks-based methods in the field of process modeling. However, there remains a notable absence of an up-to-date and systematic review on data imputation and soft sensing techniques, ranging from small models to large models. To address this gap, this work conducts a comprehensive review of deep learning methodologies for data imputation and soft sensing, spanning five classic architectures and the emerging technical frameworks of Large Language Models (LLMs). We primarily give the motivation of jointly handling data imputation and soft sensing, and present a general framework of deep network for process modeling. Finally, we propose a prospective research framework that utilizes LLMs for imputation and sensing tasks. This survey brings together the latest strides in both small and large models, offering researchers an up-to-date perspective on current breakthroughs and future research opportunities.