Unsupervised learning to detect wear faults in axial piston pumps by the similarity of periodic discharge pressure ripples

Abstract

The anomaly detection of axial piston pumps is essential to ensure the reliable and safe operation of hydraulic systems. Discharge pressure ripple is an inherent property of axial piston pumps, which exhibits a periodic waveform over each cycle under normal conditions. This work proposes an unsupervised learning method to detect wear faults in axial piston pumps by the similarity of pressure ripple signals. Specifically, the healthy discharge pressure ripples are transformed into similarity matrices to train an unsupervised network comprising a convolutional encoder and a deconvolutional decoder. The reconstruction error of similarity matrix under normal conditions determines the anomaly threshold. The anomaly detection of axial piston pumps is performed by comparing the reconstruction errors of testing samples with the predefined anomaly threshold. A test rig was built to carry out experiments of an actual axial piston pump under eight operating conditions in both normal and abnormal scenarios. The experimental results show that the proposed method can effectively detect the anomaly behavior due to wear faults in the axial piston pump with an average accuracy rate of more than 90%.