Application of physics-informed machine learning in performance degradation and RUL prediction of hydraulic piston pumps

Abstract

Hydraulic pumps have been widely used in various application domains, especially in aerospace and industrial machinery. Therefore, the accurate prediction of the performance degradation and remaining useful life (RUL) is crucial for ensuring the high reliability and safety of hydraulic pump pressure supply systems. However, the hydraulic pump performance degradation is a complex multi-factor coupling process, which is affected by the wear of internal key friction pairs and external operating conditions. This paper proposes a method that integrates failure mechanisms with data-driven approaches to forecast the degradation trajectory of hydraulic pumps, considering both the wear and operating conditions. Based on the friction and wear failure mechanism, wear evaluation models for three key friction pairs are first developed, and the wear of the hydraulic pump is evaluated by the output of the model. A Long Short-Term Memory (LSTM) network is then used to construct the mapping relationship between the wear level and return oil flow of the hydraulic pump. Afterwards, the dynamic model is iterated by updating the degradation sensitive parameters of three pairs of key friction pairs. Finally, the effectiveness of the proposed physics-informed LSTM (PI-LSTM) network framework is verified on four collected hydraulic pump degradation datasets, and compared with those of state-of-the-art methods.