Digital twin-driven zero-shot fault diagnosis of axial piston pumps using fluid-borne noise signals

Axial piston pumps are crucial components in fluid power systems, where reliable fault diagnosis is essential for ensuring operational safety and efficiency. Traditional data-driven methods require extensive labeled fault data, which is often impractical to obtain for high-value pumps, while model-based diagnosis is affected by simulation fidelity and parameter uncertainties. To address this limitation, this study develops a calibrated-source zero-shot fault diagnosis framework using fluid-borne noise signals. The source generator is an inverse-transient-analysis-calibrated digital twin derived from our prior fluid-borne-noise modeling work, which uses healthy pressure measurements to synthesize pressure-ripple and outlet-flow-ripple fault signals for classifier training; experimental fault samples are excluded from training and used only for target-domain testing. Directly measured pressure ripples and outlet-flow ripples estimated by a physics-informed neural network are evaluated as separate diagnostic branches in both one-dimensional and synchrosqueezed time–frequency forms. The implemented learning model is a convolutional neural network classifier for zero-shot fault diagnosis of axial piston pumps. Among the tested temporal, residual, and transformer-style backbones, a compact wide-kernel convolutional neural network with a subsequence-length first-layer kernel gives the best one-dimensional pressure-ripple performance. Gradient-weighted class activation mapping visualizations indicate that this kernel scale helps the model focus on complete pressure-ripple events rather than isolated local details. The flow-ripple branch provides a more mechanism-oriented view of leakage-related faults, although its accuracy is limited by virtual-sensing uncertainty. Low-fidelity domain-adaptation comparisons further show that distribution alignment can improve clean-condition performance but degrades under strong noise. Overall, the results indicate that reliable zero-shot pump diagnosis requires synthetic fault data with sufficient source-domain fidelity and a classifier whose receptive field is consistent with fluid-borne-ripple physics.