On the application of generative adversarial networks for nonlinear modal analysis

Linear modal analysis is a useful and effective tool for the design of structures. However, comprehensive basis nonlinear remains to be developed. In current work, machine learning scheme proposed with view performing analysis. The focussed on defining one-to-one mapping from latent ‘modal’ space natural coordinate space, whilst also imposing orthogonality mode shapes. achieved via use recently-developed cycle-consistent generative adversarial network (cycle-GAN) an assembly neural networks targeted maintaining desired orthogonality. method tested simulated data structures cubic nonlinearities different numbers degrees freedom, experimental three-degree-of-freedom set-up column-bumper nonlinearity. results reveal method’s efficiency in separating ‘modes’. provides superposition function, which most cases has very good accuracy.