Data-Driven Serial Testing of Axial Piston Units: Integrating Signal Processing and Machine Learning for Roller Bearing Fault Detection

Data-driven fault detection is crucial for hydrostatic drives and axial piston units (APUs) due to their central role in both conventional and electrified powertrains. The reliability of their internal roller bearings is pivotal to prevent machine downtime and guarantee high efficiency. However, detecting bearing faults in End-of-Line (EoL) serial testing for quality control is particularly challenging. Strong hydraulically induced noise and, most notably, significant manufacturing-related serial dispersion across units often mask the subtle fault signatures. To address this, this paper introduces a novel experimental methodology designed to simulate a realistic EoL scenario. By systematically interchanging drive shafts and housings among a set of seven units to create 30 unique component combinations, the inherent serial dispersion found in production is effectively replicated. This innovative approach allows for the efficient testing of distinct manipulted bearings against a realistic backdrop of component variability. Using vibroacoustic data from this setup, the present study integrates semi-supervised machine learning (ML) with a comparative analysis of different sensors and signal processing techniques. It is demonstrated that a fixed accelerometer on the test bench, when combined with a knowledge-based bandpass filter for envelope spectrum analysis, provides the most robust fault detection. This optimized configuration consistently achieves an Area Under the Curve (AUC) exceeding 0.95, effectively separating faulty from healthy units despite the challenging conditions. The findings provide a clear framework for implementing a reliable and automated fault detection system in industrial manufacturing, proving that data-driven quality control can succeed even in high-variance, noisy production environments.