ABSTRACT
A limitation of the current reliability approach has been that the failure probability was based on an assumed probabilistic model, which may not necessarily represent the real world. Observed performances could provide valuable information to calibrate this calculated probability to reflect the actual performance. This paper presents a Bayesian procedure in which failure/no failure performance records can be systematically incorporated to convert the calculated reliability index of a given system to an estimate of the real probability of failure. A method is also proposed to assess the uncertainty associated with the calibrated probability, resulting from various factors including the assumed probabilistic models and the availability of performance records for calibration. The method was applied to the stability performance of soil cut slopes in Hong Kong. Through this exercise, the reliability estimate of a given slope (of similar type and under similar environmental condition) and its associated confidence range can be realistically assessed to provide inputs for a more defensible decision.
