Least-Square Support Vector Regression for the Prognosis of the Deteriorating Structure Under the Seismic Excitations Using Autoregressive Model

Abstract

In the current work, the prognostic behavior of the degradation in the space frame is carried out through the proposed data-driven framework based on autoregressive modelling and least square support vector regression. The acceleration responses were obtained at certain intervals after introducing the time dependent damage in the building. These responses are used to develop the damage index through filtering and statistical measures. The series of damage index is nonstationary in nature and therefore, the time varying autoregression (TVAR) modelling is carried out to obtain the change point. Next, for the prognosis, the algorithm performs surrogate modeling of observed degradation through least square support vector regression (LS-SVR) and the same is used to predict the trend of degradation. To increase accuracy and obtain the confidence bounds, a few mini datasets are created through omitting and shuffling the observation in the training dataset. The LSSVR predicts the degradation trend through each mini dataset. The mean and covariance of the prediction provides the best fit value along with upper and lower bounds. The advantage of SVR is that it can provide high-order approximations with sparse availability of samples. However, SVR optimized through sequential minimal optimization is time consuming and iterative in nature. Therefore, in the present study least square SVR is used for the model fitting and regression. Overall, the results highlight the potential advantage of creating mini datasets for taking advantages of LSSVR for the prognosis and obtaining the confidence bounds.