Fault diagnosis of hydraulic pump based on stacked autoencoders

Abstract

The fault diagnosis of hydraulic pump is always a challenging issue in the field of machinery fault diagnosis. Not only is manual feature extraction time-consuming and laborious, but also the diagnostic result is easily affected by subjective experience. The stacked autoencoders (SAE) which has powerful learning and representation ability is applied in hydraulic pump fault diagnosis, and it is directly used for training and identification based on vibration signals, so that we don't need to extract features manually. To avoid gradient vanishing and to improve the performance for small training set, ReLU activation function and Dropout strategy are both introduced into SAE. Validated by experiments, proposed SAE is superior to the BP, SVM and traditional SAE. And it can recognize hydraulic pump condition accurately even if the training set is small, which satisfies the engineering application.