A NOVEL STRUCTURAL PARAMETER EXTRACTION METHODOLOGY WITH NEURAL NETWORK BASED NONPARAMETRIC MODEL

Structural identification has become an increasingly important topic for health monitoring and performance assessment of engineering structures. On one hand, the auto-regressive and moving average (ARMA) model has been widely employed as a typical time-domain modeling method for dynamic systems. On the other hand, with the ability of mapping any complex function and its parallel computation characteristics, neural networks-based identification method has been widely employed as a nonparametric model for civil and mechanical engineering structures even the weights and thresholds in the trained neural network model do not necessarily have clear physical meaning. The equivalence of the two representative time-domain identification methodologies was studied firstly according to the discrete solution of structural dynamic responses. Then, a novel structural parameters extraction methodology was proposed by the use of excitation and dynamic response measurement time series. Finally, the accuracy and efficiency of the proposed approach were validated via a numerical model with three degree-of-freedoms (DOFs), and via a dynamic test on a four-story model structure with impulse excitation. Results show the structural parameters can be extracted from a nonparametric model with neural networks.