Physics-informed machine learning for high-speed on/off valve performance degradation prediction in water hydraulic manipulator

Water-hydraulic high-speed on/off valves (HSVs) are critical control components, such as water-hydraulic manipulators (WHMs) and underwater flotation regulation systems. These systems are characterized by their numerous applications and demanding dynamic performance requirements. Therefore, ensuring effective data monitoring and performance safeguards for these valves is crucial to the overall functionality of hydraulic control systems. However, due to the complex multi-physics field coupling of HSVs, data-driven machine learning (ML) methods evidently cannot achieve accurate degradation prediction for high-performance HSVs. To tackle this issue, the present study proposes a physical information-based neural networks (PINNs) that incorporates convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM), and the Attention method to accurately predict the degradation process of HSV performance. In these proposed methods, the inclusion of physical knowledge in the form of differential equations into the loss function enhances convergence towards the actual physical failure mechanism during network training. The results demonstrate a significant reduction in RMSE by 30.27% for the physical information-based CNN-BiLSTM-Attention·(PICNN-BiLSTM-Attention) model compared to the traditional CNN-BiLSTM-Attention model, accompanied by an improved R2 value by 2.95%. Comparative analyses with other approaches are conducted to evaluate error and fitting effects, confirming that our proposed method effectively enhances prediction accuracy for HSV performance degradation.