Statistical methods for the risk -based design of civil structures

Probabilistic methods are with increasing frequency used in the design of civil structures such as dikes, storm surge barriers, bridges, buildings, etc. The methods are being applied directly, or are translated to relatively simple design rules with safety coefficients. In both cases the foundation of the calculations is given by the statistical distribution functions of the strength and load variables. In the application of probabilistic methods, the availability of useful calculation models and adequate statistical distributions is required. A larger problem in the application of probabilistic calculation techniques is the lack of good argued statistical models. It is on this area that the current thesis would like to give a contribution to the knowledge. The following three situations can be discerned: (1) there is an abundance of statistical data; (2) there are only few observations; (3) there is no data. The thesis focusses in particular on the second case: there are only few observations and it is not possible, or not economically feasible, to enlarge the amount of data. In civil engineering practice this is the most common case. The current thesis examines this case with different statistical methods, under which the Bayesian methods. These Bayesian methods are developed to optimally use background information in a decision problem with a scarce amount of data. There is an interaction with the final design decisions which have to be taken on the basis of the estimates: a too low estimate of a load parameter can have more severe consequences than a too high estimate. There exist also inverse situations. These kinds of interactions are treated in this thesis with Bayesian decision theory. Since small probabilities of failure are commonly required in civil engineering practice, the tails of the distribution are extremely crucial. Apart from Bayesian solutions, Entropy methods and L-Kurtosis approaches appear to perform satisfactory. Issues such as cleaning and homogenisation of data are mentioned. Which part of the dataset is useful, depends on the application. The thesis examines this aspect as well. Case studies such as the reliability analysis of dikes along Lake IJssel and the frequency analysis of more than 200 measurement stations.