Hybrid digital twin-based data enrichment strategy for reliable machine learning-based diagnosis of ball bearings under multi-regime operating conditions

Reliable fault diagnosis is essential for condition monitoring of rotating machinery, yet the scarcity of experimental data often limits the effectiveness of machine learning algorithms. A promising approach lies in constructing hybrid datasets that combine experimental and numerical data. However, if performed arbitrarily, such hybridization may degrade diagnostic reliability. This study proposes a systematic methodology for building hybrid databases for ball bearing fault diagnosis across multiple fault severities and varying operating conditions. The methodology is evaluated along three complementary axes: (i) single-condition performance, (ii) expandability to a wide range of speeds through the identification of a reliable variation range, and (iii) transferability of calibration parameters to neighboring regime operating conditions. Both homogeneous (quantitative enrichment) and heterogeneous (qualitative enrichment with unseen fault labels) hybridizations are investigated. Results demonstrate that even a small experimental fraction (≈10%) substantially enhances classification accuracy compared with purely numerical models, while stable performance (>90%) is achieved from about 30% of experimental data. In multi-speed scenarios, hybridization maintains generalization provided sufficient experimental anchoring, though the reliable range remains narrow and asymmetric. In heterogeneous configurations, high accuracy is preserved even with a single EC, as numerical enrichment effectively compensates for missing severities, provided the enrichment is balanced with the operating range.