Development of Bearing Fault Detection Models using Multibody Simulation Training Data

This study evaluates the performance of simulation-trained fault detection models on large spherical roller bearings vibration data. A high-fidelity multibody (MB) model of a SKF 22240 CCK/W33 bearing is developed through Simscape Multibody to reproduce the coupled dynamics of inner and outer rings, cage, and 38 rolling elements. Localized defects on raceways are represented through a pointcloud contact formulation, where selected nodes are radially displaced to emulate faults. The model outputs triaxial accelerations at the outer ring under realistic loading and speed conditions that mirror an experimental test campaign. Simulation signals are processed through bandpass filtering, envelope analysis, and segmentation. A set of 23 time and frequency domain features is extracted from each segment, then each feature vector is normalized. The same processing chain is applied to experimental data acquired on a medium-to-large bearing test rig at Politecnico di Torino, mounting SKF 22240 CCK/W33 bearings with machined inner race, outer race, and rolling element defects. A supervised Artificial Neural Network (ANN) classifier is trained only on the simulated feature dataset and then directly evaluated on the independent experimental dataset, in a process free of any data transfer. The network addresses a two-class problem (healthy and damaged), and its performance is assessed through standard classification metrics computed over multiple bootstraps of both training and test sets. Despite the intrinsic differences between simulated and experimental signals, the ANN trained purely on simulations provides reliable and selective fault detection on real measurements. Most residual classification errors are concentrated in low-speed inner race damage conditions, where fault signatures are weak and partially overlap with healthy observations, while high-speed and outer race damage conditions are recognized more robustly. These results show that MB simulation can generate sufficiently realistic vibration data to train ANN-based fault detection models that generalize experimental measurements for large spherical roller bearings. The proposed framework introduces an alternative to costly fault campaigns and offers a flexible way to expand training datasets across loads, speeds, and defect sizes in industrial condition monitoring applications.