An efficient method for time-dependent reliability prediction using domain adaptation

Due to dynamic uncertainties presence in service and performance conditions, time-dependent reliability prediction of a component or structure is a challenging problem. In this research, a transfer learning-based technique is proposed to predict the reliability in the future. The complete time interval is divided into two sub-intervals namely, present interval and future interval. It is assumed that the performance function information is available for the present interval only. Transfer learning, specifically domain adaptation is used to transform the stochastic processes to be represented in a way that their sample spaces in different time durations are made closer while maintaining some of their statistical properties such as variance. In order to transform the stochastic processes, correlated samples of stochastic processes are generated using a space-filling sampling technique for the complete time interval. An adaptive Kriging surrogate model is then built using the performance information available for the present interval only using transformed stochastic process samples. The built Kriging model is employed to estimate and predict the reliability for present and future intervals without retraining it using future data. Results show that the proposed method can predict the failure probability in present and future intervals accurately with significant efficiency improvement.