Citation

Zhang Y, Gong ZW, Zheng Y, Xie LB, Zhang Z, Liu ZX. Sci. Sin.-Tech. 2021; 52(1): 180-197.

Copyright

(Copyright © 2021, Chinese Academy of Sciences and the National Natural Science Foundation of China, Publisher Science in China Press)

DOI

10.1360/SST-2021-0337

PMID

unavailable

Abstract

As a basic component of industrial internet and industrial internet platform, the health state of industrial equipment (IE) is related to the stable operation of industrial systems and the quality level of industrial products. Therefore, both health status assessment (HSA) and degradation trend prediction (DTP) of IE have important theoretical value and engineering significance. This study proposes constructing a dual-task framework to realize the HSA and DTP of IE using deep learning techniques. First, monitoring signals of IE are extracted in the time and frequency domains so that strongly correlated features are selected by a light gradient boosting machine. After principal component analysis-based dimensionality reduction, both the health index and the category label of health state are constructed. Especially, an adaptive transfer learning algorithm, namely, manifold spatial distribution alignment, is used to migrate the source domain to the target domain at the feature level to complete distribution alignment. Finally, by integrating causal expansion convolution, a bidirectional gate recurrent unit, and an attention mechanism, a dual-task depth network framework is designed to realize the dual functions of HSA and DTP of IE. Experimental results show that the proposed method effectively addresses the limitation of existing methods, which is to investigate HSA and DTP alone. In addition, compared with typical deep learning methods, the presented methods significantly improve DTP accuracy and reduce the frequency of negative migration.

Language: en