Traditional time-dependent reliability analysis(TRA)often constructs a surrogate model through a large number of design of experiments(DoE)samples,thereby achieving the calculation of time-dependent reliability analysis.As the nonlinearity of the performance function and the difficulty of response solution increase,the calculation cost of DoE becomes increasingly expensive,making the reliability analysis time-consuming and cumbersome.To address this problem,a time-dependent reliability analysis method based on physics-informed neural network(PINN-Based TRA,PBTRA)is proposed in this paper.This method incorporates the partial differential equation that constrains the system response into the loss function of the PINN training process,which uses the PINN model to predict the system's response.Based on the PINN model,the system performance function is constructed to preform TRA,which can solve the problem that traditional TRA relies on a large number of simulation calculations to obtain DoE samples,effectively reducing the calculation cost.At the same time,for the problems of slow convergence and underfitting that occur in the traditional PINN training process,the proposed method judges whether the relevant region is close to the limit state of the system response according to the response of the PINN model in different sampling regions during the training process and divide the sensitive regions accordingly.The distribution of training point sampling is dynamically adjusted,and combined with the resampling method of neural network,a dynamic sampling training method based on regional reaction weight function that suitable for PINN model is proposed and applied to TRA.Compared with traditional PINN,this method has faster training speed and higher accuracy of performance response calculation,which can improve the accuracy and efficiency of TRA.Two cases are analyzed by PBTRA and compared with traditional TRA methods to verify the superiority of the method proposed in this paper.
time-dependent reliability analysisoptimal designphysics-informed neutral networkimportance samplingsurrogate model
NETL
NSTL
万方数据
