Support vector regression (SVR) has been widely used for reliability modeling and prediction in various engi- neering practices. In order to improve the accuracy and robustness of SVR models, this paper proposes a hier- archical Bayesian support vector regression (HBSVR) model, which can be used for dynamic high-dimensional reliability modeling with small data sets. First, a hierarchical Bayesian model is developed to obtain the posterior inference of HBSVR model parameters. Then, a step-size adaptive accelerated Markov Chain Monte Carlo (SAA- MCMC) method is proposed and combined with Sequential Minimal Optimization (SMO) for model parameter calibration. In addition, a HSBVR dynamic update algorithm is proposed to readjust the current parameters of HBSVR by using the updated dataset and SAA-MCMC to avoid repeated modeling during online prediction. Finally, an active learning algorithm is applied to guide the selection of new samples to improve the model accuracy continuously. Eight numerical examples and two real engineering cases are conducted. The results show that the proposed HBSVR outperforms other methods in terms of prediction accuracy and robustness in small sample reliability prediction tasks.
展开▼
