Feature extraction using auto-associative neural networks

Modal analysis is now mature and well accepted in the design of mechanical structures. It determines the vibration mode shapes and the corresponding natural frequencies. However, the validity of modal analysis is limited to structures showing a linear behaviour. In non-linear structural dynamics, it is well known that mode shapes are no longer useful for the characterization of the dynamic response. The purpose of the present paper is to define new features which efficiently capture the dynamics of a non-linear structure. The proposed methodology takes advantage of auto-associative neural networks to compute one-dimensional curves which allow for non-linear dependences between the coordinates. Synthetic data sampled from a non-linear normal mode motion are used to illustrate the method and to develop intuition about its implementation.