A simulation-driven machine learning framework for condition monitoring of rotating machinery

Licentiate thesis, 2026

Rotating machinery is widely used across many industrial fields and functions as the core of operations. Such machines usually work in harsh environments, which accelerates machine degradation. To ensure reliability of rotating machinery, the implementation of condition monitoring (CM) and fault diagnosis strategies is crucial for identifying anomalies at an early stage.

This thesis focuses on the development of a simulation-driven transfer learning framework, with particularly emphasis on bearing fault diagnosis across different rotational speeds. The framework defines a target domain where the actual task takes place, and a well-defined source domain where the fault-related knowledge is acquired. Through transfer learning, knowledge gained from the source domain is effectively applied to real diagnostic tasks in the target domain. To build the target domain, an open-source bearing dataset is first used. After conducting experiments on the test rig at LUT University, an experimental bearing dataset is constructed by collecting vibration signals under different operating conditions.

In parallel, multibody dynamics (MBD) models are developed for generating vibration responses that reproduce representative fault characteristics. To build the source domain, a full-scale MBD model of the test rig is developed, where the geometries of all components and faults are consistent with the physical system. The model is then validated against the measurements, demonstrating its accuracy in reproducing fault signatures. For large-scale data generation, the computational cost of the MBD model is assessed in comparison to an analytical model. This process provides valuable insights into how the system dynamics and computational efficiency evolve with increasing model fidelity.

During the diagnostic model training stage, feature extraction methods are integrated into the transfer learning framework. These methods are designed to extract fault-related signatures across different operating conditions, while preserving consistent physical information. Finally, the diagnostic model, which is trained using the simulated-fault features, successfully distinguishes bearing conditions in the measurement data. The results demonstrate the proposed framework's potential for addressing more complex scenarios in future applications.