Detecting and Grading Valve-Plate Wear in an Axial Piston Pump from Pressure and Vibration Signals: Within-Campaign Separability Versus Leave-One-Condition-Out Generalisation

Valve-plate wear in axial piston pumps reduces volumetric efficiency, and data-driven classifiers often report near-perfect accuracy in grading it. This paper asks how much of such accuracy reflects the fault rather than the recording. Outlet and precharge pressure, leakage flow and housing vibration at four points were recorded on a test bench for five induced wear states (estimated volumetric efficiency 94.9–69.5%) at an unloaded and a loaded (≈700 psi) operating point, in a single campaign. Statistical features were classified with support vector machines, random forests and shallow neural networks under protocols of increasing difficulty. Record-level cross-validation separates the five states with 94–100% accuracy from pressure or from any single accelerometer. Leave-one-condition-out validation shows that only a coarse decision generalises to a wear state absent from training: significant degradation (efficiency below about 80%, where the pump loses its precharge pressure) is detected by the pressure channels at both operating points and by three accelerometers at the loaded point, whereas the fine grading and a continuous efficiency estimate (mean absolute error 3.6–7.0 points) do not transfer, nor do models between operating points. The results are delimited to one campaign with one mounting per wear state; multi-session validation on naturally worn components is required.