A Comparative Study of Linear and Nonlinear Data-Driven Surrogate Models of Human Joints

Various linear feed-forward and recurrent data- driven models, as well as their nonlinear counterparts, are studied for dynamic musculoskeletal system identification. It is shown that dynamic neural networks are well suited for black- box modeling of biomechanical multi-body systems, as these nonlinear paradigms could capture human joint force- displacement dynamics with much lower computational complexity compared to traditional methods such as the finite element methods. This paper analyzes the performance of different surrogate model architectures using simulated knee data, and provides comparisons between their drawbacks and benefits such as computational efficiency. While linear models presented acceptable results, the non-linear implementations yielded substantial performance improvements with equal or shorter tapped delay lines over their linear counterparts.