Physics-informed surrogate modeling for a damaged rotating shaft

In online structural health monitoring frameworks, surrogate modeling approaches aim to reproduce the dynamics of an underlying high-fidelity model for facilitating fast simulations. Conventional machine learning techniques based on purely data-driven approaches, i.e., neural networks (NN), often behave as a black box delivering predictions that may lack physical consistency. In an effort to overcome such limitations, we propose an NN mapping constrained by known dynamic equations provided by a physics-based model of the operating system. Incorporating the system matrices into the loss function steers the learning process towards more physically reliable predictions while maintaining accuracy. Specifically, a surrogate model relying on physics-informed neural networks (PINN), is developed to enable accurate and fast ballistic impact damage identification on a helicopter transmission. At first, a physics-based model simulates the damage-induced vibration loads and eventually the system response. The PINN model receives the system response as input and predicts the external load and model parameters, eventually assessing the impact damage extent.