Cross-domain knowledge transfer in industrial process monitoring: A survey

Abstract

The last decades have witnessed rapid progress in machine learning and data analytics-based industrial process monitoring. However, the underlying assumption that the training and test data should have the same feature space and the same distribution is generally challenged in practical industrial applications due to varying working conditions, mechanical wear, feed changes, etc. To this end, knowledge transfer, which reduces the discrepancy between different data and facilitates the target model learning, has given rise to tremendous advances for mitigating this trap. Motivated by the success, in this survey, the state-of-the-art techniques are investigated and a review from a broad perspective in the field of cross-domain industrial process monitoring applications is provided, including fault detection and diagnosis, fault prognosis, and soft sensors. Owing to the extensive developments, the cross-domain knowledge transfer in process monitoring can be divided into three branches in this survey, i.e., the multivariate statistical analysis-based, the shallow neural networks-based, and the deep neural networks-based methods. Benefiting from the theoretical development and elaborately developed approaches, current challenges and instructive perspectives are further conceived for inspiring new directions in this exciting research field. The aim of this paper is to sketch the basic principles and frameworks for cross-domain knowledge transfer in process monitoring and provide inspiration for future studies in the process industry.