Data-Driven Damage Assessment of Axial Piston Pumps

Axial piston pumps, vital for industrial applications, demand proactive maintenance strategies to avert failures and ensure system availability. Precise prediction of the current health state is crucial for selecting the most appropriate countermeasure, optimizing both time and resource allocation. While previous studies predominantly focus on classifying damage types, this paper explores regression models for precise and continuous damage severity assessment. It addresses key questions regarding prediction accuracy across the whole operating range, optimal sensor positions, and interpolation capability. A vibration data set comprising 96 operating points and eight distinct pump health states measured at three sensor positions was utilized. Two funda-mentally different methods, a random forest and a convolutional neural network, were employed to predict severity for the damage types axial clearance and cavitation damage. Results reveal that the convolutional neural network outperforms random forest in terms of accuracy and variance, while the random forest excels in assessing undamaged pumps. Notably, convolutional neural networks exhibited interpolating capabilities, crucial for real-world applications.