Parameter Identification Approach to Vibration Loads Based on Regularizing Neural Networks

An identification algorithm for vibrating dynamic characterization by using artificial neural network is developed for multi-degree-of freedom systems. The over-fitting problem of classical back-propagation algorithm during neural network training is solved by using regularization procedure with regularized objective function. The practical application shows that the proposed training method is capable of enhancing the regularization procedure without getting stuck at these sub-optimal solutions, can be used to noisy data in order to omit an over-fitted neural approximation and has higher identification accuracy compared to the back-propagation algorithm