Physics-Informed Neural Networks for Monitoring Dynamic Systems: Wind Turbine Study Case

This work addresses the use of Physics-Informed Neural Ordinary Differential Equations (PINODEs) for the identification and monitoring of dynamic systems. By combining prior knowledge regarding the physics that governs the behavior of a dynamic system with the flexibility and learning capabilities of Artificial Neural Networks (ANNs), it is possible employ data-driven methods that result in a robust representation of the systems, even without full knowledge of the underlying physical phenomena, while retaining a degree of interpretability of the model's outputs. A description of the overall framework for modeling the system identification problem and training the ANNs under the is presented, along with an application case study for condition monitoring of a wind turbine's gearbox using vibration data. The results demonstrate the ability of the identified model to generalize with great accuracy to scenarios not accounted for during training, a property attributed to the inclusion of information regarding the physics of the problem during the training procedure. It is also shown that when exposed to data collected during a fault condition, the model's output significantly deviate from the actual measurements, hence its potential use as a tool for condition monitoring and fault detection is also successfully demonstrated.