Support vector regression for structural identification

Structural identification based on the vibration data is still a challenging topic especially when the input and output (I/O) measurements are corrupted by high-level noise. In this paper, we propose a new structural parameter identification method based on the Support Vector Regression (SVR) which has been found working very well in many fields as an exclusively data based non-linear modeling method. Machine learning technologies such as Neural Networks has been applied widely in the field of health monitoring field. However, most papers just obtain the 'block-box' model of the studied structures from Neural Network training but the structural parameters are not identified actually. In our work, we not only generate the 'block-box' model but also identify the structural parameters by combining ARMA model together with SVR. Due to the “max-margin” idea used, SVR showed powerful properties in ARMA and structural identification under different kinds and amplitude noise. Furthermore, how to choose the parameters of SVR is also studied in this paper. Finally, numerical examples are given to demonstrate that the proposed method based on SVR is effective and powerful for identifying ARMA time series and structural models.