State monitoring method of hydraulic system based on  attention machine multi-task network

HUANG Peng-cheng, LI Hai-yan, LIN Jing-liang, LIANG Gui-ming
(School of Mechanical and Electrical Engineering, Guangdong University of Technology, Guangzhou 510006, China)

Abstract: Aiming at the poor monitoring performance of the current hydraulic system state monitoring methods under complex multi-state conditions, the multi-task learning and attention mechanism methods were studied. Combining with the multi-task and attention mechanism, a multi-task network-based monitoring method for hydraulic system was proposed. Firstly, the attention mechanism was used to assign different weights to each sensor according to the degree of contribution of the sensor signals to the task. Secondly, a convolutional network (CNN) was used to construct an adaptive feature extractor to extract depth features from multiple sensor signals with weights. Finally, a multi-task feature sharing diagnostic network was established to realize simultaneous monitoring of multiple states. The results indicate that the proposed method is superior to the previous method, and can effectively monitor the various states of the hydraulic system, with an average accuracy of 99.3%.
Key words: state monitoring of hydraulic system；monitoring performance；attention machine; multi-tasking network