Health Assessment for Crane Pumps based on Vehicle Tests using Deep Autoencoder and Metric Learning

Abstract

Pump is one of the key components in a crane, which once fails will severely hurt the reliability of the hydraulic system and cause great loss. Therefore, accurate, reliable and effective crane pump health assessment must be performed. However, the research about pump health assessment still stays at the stage of bench tests, which have the limited help for the real-world pump health prognosis. In this paper, to evaluate crane pump health status and avoid the issue above, the real-world vehicle tests of several cranes with different service years are performed to acquire the pump signals during the cranes’ actual operations. Deep Autoencoder (DAE), a kind of unsupervised learning approach, which possesses the capacity to learn meaningful representations from raw signal, is used reduce the data dimension before they are sent to Mahalanobis-Taguchi System in metric learning. Mahalanobis distance (MD) is utilized to reveal the performance degradation and assess the health condition. Performances of other feature learning methods such as statistical features, EMD, MLP, CNN are tested and contrasted. Results show that the proposed approach achieves the best performance.