| Uncertainties are widely encountered in practical engineering,including material properties,structural dimensions,external loads,and so on.Uncertainty is usually divided into two types: random uncertainty and epistemic uncertainty.Random uncertainty describes the uncertainty of design variables by probability model,but it is necessary to collect a large number of sample information to construct joint probability density function.The basic theories of epistemic uncertainty analysis include possibility theory,convex model theory,fuzzy theory,interval analysis theory and evidence theory.When it is difficult to get enough sample points in practical engineering,evidence theory can deal with multi-source information and various uncertain problems.However,due to the discontinuity of the interval of evidence variables and the discreteness of basic credibility distribution,evidence theory solve the problems with low efficiency.That is the reason limited its application in practical engineering.At present,most reliability analysis methods simplify the calculation process by approximate optimization method.Those methods improve the calculation efficiency but at the same time lose the partial calculation accuracy.With the development of computer system and cluster technology in recent years,parallel computing technology has made some achievements in scientific research,large-scale engineering and commercial computing.In order to take into account the calculation accuracy and efficiency,this paper introduces the parallel computing method,and carries out the related research on the reliability analysis method of evidence theory.The main contents are as follows:(1)A parallel reliability algorithm of evidence theory based on MPI is proposed to solve the single failure mode problem in structural reliability analysis.The extremum solution process of each focal element is independent to each other,and data it is not necessary to transfer data and information.The calculation process can be completed synchronously.Therefore,MPI master-slave mode can be used to design programs,and the performance of programs can be optimized by combining message aggregation,load balancing and increasing parallel granularity.Each slave process is responsible for the extreme value analysis of focal elements.The main process collects the results to calculate the cumulative reliability and likelihood of the structure,so as to efficiently complete the whole reliability analysis process.(2)For multi-failure mode structural system,a parallel algorithm for structural system reliability based on the combination of CUDA and evidence theory is proposed.Using parallel genetic algorithm to calculate the maximum and minimum of limit state function on the joint focal element.Then taking the advantages of GPU multi-threading and parallel reduction strategy,the calculation cost can be reduced to a certain extent and the reliability analysi s efficiency can be achieved. |
