Physics informed neural networks for detecting the wear of friction pairs in axial piston pumps

Abstract

The wear of friction pairs is one of the most common failure mechanisms for axial piston pumps and its accurate detection is essential for ensuring safety and reliability of hydraulic systems. The existing studies on the wear detection of friction pairs in axial piston pumps is focused on data-driven fault diagnosis models, but these black-box data-driven models are limited by poor interpretability and physical inconsistency. To overcome this limitation, this paper proposes an interpretable wear detection method for the friction pairs of axial piston pumps based on physics informed neural networks. First, we establish an ordinary differential equation (ODE) for the time derivative of discharge pressure to relate the instantaneous discharge pressure with the fluid film thicknesses in friction pairs that represent the wear condition of axial piston pumps. Second, we develop a physics informed neural network and a multi-parameter dynamic identification method to identify the fluid film thickness in each friction pair by inversely solving the ODE based on observed discharge pressure signals. Finally, we propose an interpretable wear detection method based on the pump's volumetric efficiency and effect size of fluid film thickness. Experimental results suggest that the identification results of fluid film thickness in the friction pairs have a good physical consistency, and the proposed method can locate the worn friction pair with a high model interpretability.