Multi-state deterioration prediction for infrastructure asset: Learning from uncertain data, knowledge and similar groups

Abstract

Infrastructure assets such as bridges need to be inspected regularly for signs of deterioration. Although a fixed inspection interval could be used, an estimate of the rate of deterioration allows us to schedule the next inspection more cost-effectively. Our earlier work outlined a Bayesian framework that uses both data and knowledge to predict the transition between assets, which has been extended and realised in this paper for asset deterioration prediction. In the Bayesian model, censorship is modelled to incorporate the uncertainty from inspection records and prior of the parameter is used to express expert knowledge. In particular, we also suggest how the prior probabilities of the parameters of a Weibull distribution can be set in practice using expert estimates such as the maximum and average times of a transition from one state to another.Furthermore, assets with similar characteristics may deteriorate similarly. We propose to separate related assets into groups and learn deterioration between these groups. This assumption allows us to tackle the challenge of limited data further and is experimented with the deck inspection records from the National Bridge Inventory database in Wyoming. This database includes over 100 features of each bridge such as structure type and average daily traffic: we use a modified random forest to select a subset of important features to separate assets into groups. The model is extended into hierarchical Bayesian models to learn between groups with the help of hyper-parameters and an aggregated variable from the feature levels. Performance of our method is compared with other existing approaches from various aspects.