ME290M Expert Systems in Mechanical Engineering Spring 1999, T-Th 12:30-2:00 pm 1165 Etcheverry Hall, Course Control No. 56369 http://best.me.berkeley.edu/~aagogino/me290m/s99

Abstract

Many studies have been performed in order to optimise the inspection plan such that a given structural reliability is maintained throughout the service life at the lowest possible cost. Most studies assume that the description of each individual component in the structure is independent of all other components. On an offshore platform however, many details are similar with respect to geometry, material properties and wave load.

The objective of the project is therefore to investigate the possibility of including the 'pool' of common information about the structural members of the structure. The model is designed to predict fatigue crack growth in the critical points of the structure, the so called hotspots.

The model for inspection planning is based on Bayesian networks. Bayesian networks have the advantage that new knowledge in a simple manner can be inserted in the model that can subsequently be updated accordingly. If any common knowledge of the system or the relationship between the variables is available, this information can also be included to yield an effective updating of the model.

Firstly a model of a single hotspot is constructed in order to establish a simple model describing the probability distribution of the crack depth as a function of time. This model is calibrated taking into account accuracy and computational efficiency. Calibrating the model, difficulties in finding appropriate discretisation schemes and in assigning conditional probability tables to the variables have been encountered. The probabilistic description of the crack depth distribution over time is compared to results obtained by more traditional methods namely the SORM approximation (second order reliability methods). The comparison shows that very good agreement between the two methods can be obtained.

Secondly the model is extended to include two hotspots and it is investigated to which extent the network is able to transfer information from one hotspot to the other. It shows that the model is able to update the reliability of one hotspot on the basis of an observed crack state in the other hotspot. This shows that the common knowledge about the hotspots enables transfer of information from one hotspot to an other. Even if this is still a rather simple model, the use of common information allows for less conservative inspection plans.

The model is extended in a number of respects and directions for further development of the model are given.

The present work is a step towards a model that can monitor the reliability of a structure by inspecting a minimum of the structures hotspots. This allows an economical optimisation of the maintenance work (inspection and repair) which, considering the proportions of the costs, plays a significant role for the working expenses for an operating platform.