Imperfect physics-guided neural networks

Abstract

Fault detection in complex systems can benefit greatly from advances in physics-guided machine learning. However, it is important to consider that physics-based engineering models often carry assumptions, simplifications, and general imperfections when representing real-world dynamics. This research article investigates how imperfections in physics-based engineering models affect the accuracy of Physics-guided Neural Networks (PGNNs). Specifically, the study examines neural networks applied in fault detection for nonlinear dynamical systems. The PGNN models used in this article combine simulated measurement data with physics-based model estimates to augment the feature space. This improves training and inference, enhancing the neural networks' fault detection capabilities. The results shared in this article demonstrate that ideal PGNNs can significantly outperform unguided neural networks in detecting faults. However, introducing imperfections in the physics-based model reveals that small inaccuracies can lead to a drastic reduction in fault detection performance. The effectiveness of the PGNN is highly dependent on the reliability of the physics-based model and the underlying dynamics of the system. We explain the likely cause of these findings and highlight the need for careful consideration when applying PGNNs in engineering applications.