A structure-preserving machine learning framework has been developed to accurately predict structural dynamics for systems with isolated nonlinearities. Nonlinearities in structural systems are often found in specific regions, such as joints or interfaces, and can complicate the development of reduced-order models by coupling the modes of the linear system. These isolated nonlinearities have a global impact on the system’s dynamics, particularly when there is evolving structural health due to accumulating damage.
This new data-driven formulation aims to identify and incorporate the contributions of isolated nonlinearities in the dynamics of the linear structure. A unique coordinate separation method decomposes the nonlinearities within the isolated subdomain from the known linear system across the entire domain. The influence of these isolated nonlinearities is reintroduced as a deviatoric force at the boundary of the isolated subdomain. This approach allows for accurate predictions of the deviatoric force using a structure-preserving multilayer perceptron based solely on measured responses at the subdomain’s boundary.
The machine learning system can predict the deviatoric force, enabling the ideal system to mimic the response of the original system beyond the isolated nonlinear subdomain. This approach proves robust in predicting responses under varying initial conditions and external excitation without requiring retraining. Overall, this data-driven strategy offers a comprehensive description of the structural dynamics of the system.
