Deep recurrent-convolutional neural network learning and physics Kalman filtering comparison in dynamic load identification

The article examines the effectiveness of different neural network architectures, including gated recurrent units and convolutional neural networks, in the task of dynamic load identification within civil engineering. It specifically compares these data-driven approaches to a physics-based residual Kalman filter method, aiming to assess their relative performance under realistic training conditions. The study focuses on the challenges associated with making accurate predictions in this application domain, highlighting the complexities involved in modeling dynamic loads. By evaluating both neural network models and traditional physics-based filtering, the research provides insights into the strengths and limitations of each approach. The comparison underscores the difficulty of achieving reliable load identification when training data and real-world conditions may vary. This work contributes to ongoing efforts to improve predictive modeling techniques in civil engineering, where precise load estimation is critical. The findings suggest that while neural networks offer promising capabilities, integrating physics-based methods remains important for robust performance.