Physics informed neural networks for fault severity identification of axial piston pumps

Abstract

Artificial intelligence (AI) has shown great potential in the maintenance stage of industrial manufacturing. However, the existing data-driven methods often lack integration with physics knowledge, leading to the implicit decision-making processes. To overcome this limitation, this study proposes a physics informed neural network for identifying fault severity in axial piston pumps. The method begins by establishing a discharge pressure model that incorporates physics information to describe instantaneous pressure changes. Subsequently, a neural network model is employed to estimate the temporal function of discharge pressure based on observed signals. Moreover, by allowing the learnability of fault-related physics parameters, the discharge pressure model is transformed into a physics informed loss term that guides the learning process. By incorporating the physics informed loss, the proposed method effectively combines observations and physics knowledge, providing explicit physical interpretation to the identified parameters. Additionally, the introduction of volume efficiency establishes a relationship between the physical degradation indicator and fault severity, offering a physical explanation of the identified results. Experimental results on an axial piston pump demonstrate the effectiveness and robustness of the proposed method, particularly in terms of interpretability for fault severity identification.