Inverse physics-informed neural networks for fault diagnosis of friction pair in axial piston pump under imbalanced samples

Abstract

Fault diagnosis of friction pairs in axial piston pumps is essential for ensuring the safe operation of hydraulic systems. However, deep learning-based diagnostic methods are often limited by scarce and imbalanced fault samples, while conventional dynamics simulations rely on empirical parameters, resulting in poor data fidelity. To overcome these challenges, this paper proposes a novel fault diagnosis method that integrates an Inverse Physics-Informed Neural Network (IPINN) with a Bidirectional Temporal Convolutional Network (BiTCN). A dynamic model of an axial piston pump with swash plate defects is developed. The network structure and hyperparameters are determined through preliminary experiments and grid search. The IPINN is then employed to optimize key dynamic parameters, generating high-fidelity simulation data that closely align with experimental measurements, thereby significantly improving data quality and alleviating class imbalance. Subsequently, the augmented dataset is fed into the BiTCN, which utilizes bidirectional residual units and an attention mechanism to extract complex fault features from vibration signals. Experimental results demonstrate that the simulation data optimized by IPINN exhibit significantly better statistical metrics than those based on empirical parameters. The BiTCN achieves a diagnostic accuracy of 0.98, outperforming traditional algorithms such as TCN and LSTM by more than 8%. Moreover, noise robustness analysis confirms that the BiTCN maintains an accuracy of 0.91 under strong noise conditions of 5 dB, highlighting its excellent environmental adaptability. This study provides an effective solution for the intelligent maintenance of axial piston pumps.