Physics-informed recurrent neural network with cross-condition transfer for hydraulic oil filter degradation prognostics

The hydraulic oil filter plays a crucial role in oil filtering, and its performance degradation can significantly affect the reliability and safety of the entire hydraulic system. However, accurately predicting its degradation remains challenging due to the complexity of its failure mechanism. In addition, the failures of the filter highly depend on its working conditions (e.g., pressure and temperature), which requires a dynamic model adaptation to various environmental parameters. To solve this problem, this study proposes a novel physics-informed recurrent neural network (PI-RNN) that systematically incorporates a hydraulic filter physics-of-failure (PoF) model into the recurrent neural network (RNN) architecture. Firstly, a nonlinear and coupled hydraulic oil filter pressure drop degradation model is developed based on Darcy's law. Secondly, a novel PI-RNN model is proposed by embedding PoF into the structure of RNN. The loss function is also reconstructed considering monotonicity and concavity constraints. The proposed PI-RNN also achieves multi-condition migration applications through a combination of pre-training and parameter calibration. Thirdly, two real case studies are conducted to validate the proposed degradation model. The obtained results show that the proposed PI-RNN model has high prediction accuracy. Moreover, the proposed PI-RNN model allows to rapidly adapt the parameters of the model to deal with different working conditions.