Research On Multi-Model Prediction And Quantification Method Considering Model Uncertainty

In practical engineering,the mathematical model of the system is often used for analysis and designing,so the predictive performance of the model directly affects the analysis and design of engineering structures.In order to improve the accuracy and robustness of prediction analysis,in this paper,a multi-model prediction and quantification method considering model uncertainty is studied,and model-form uncertainty,parametric uncertainty and predictive error uncertainty are studied respectively.Firstly,in view of model form uncertainty,the advantages and disadvantages of model selection method and model combination method are compared and analyzed,an improved Bayesian combination forecasting method which considers the influence of prior probability on Bayesian combination forecasting and optimizes the setting of prior probability under the condition of no prior information is proposed.The accuracy of Bayesian combination forecasting is further improved.And then,the advantages of the proposed method are verified through engineering exanple analysis.Then,in view of parametric uncertainty and predictive error uncertainty,a method of improved Markov Chain Monte Carlo stratified resampling is proposed.The proposed method,which can reduce the number of samples for simulation and improve the efficiency of engineering analysis,fully combines the advantages of stratified sampling and Markov Chain Monte Carlo method.The feasibility of the proposed method is verified by given examples.Finally,the application of Bayesian combined model in the fatigue reliability analysis of aircraft structure is studied.The example shows that the Bayesian combined forecasting method can improve the robustness of the prediction results of aircraft structure fatigue reliability by considering the information of multi-model.The prediction results of Bayesian combined forecasting method are of higher credibility therefore.