Comparison of forecasting models to predict concrete bridge decks performance

Abstract

The accuracy of forecasting models for the prediction of an infrastructure's deterioration process plays a significant role in the estimation of optimal maintenance, rehabilitation, and replacement strategies. Numerous approaches have been developed to overcome the limitations of existing forecasting models. In this article, a direct comparison is made between different models using the same input data to derive conclusions of their distinct performance. The models selected for the comparison were Markov, semi‐Markov, and hidden Markov models together with artificial neural networks (ANNs), which have been reported in literature as reliable deterioration prediction models. A quality of fit was performed to measure how well the observed data corresponded to the predicted values, and therefore objectively compare the performance of each model. The results demonstrated that the most accurate prediction was accomplished by the ANN model. Nevertheless, all models presented differences with respect to typical values of concrete decks life expectancy, which is attributed to the inherent difficulties of the database. Additionally, the problem of the visual inspection subjectivity was also regarded as one of the potential causes for the found deviations. Therefore, this article also discusses the shortcomings of current condition assessment practices and encourages future bridge management systems to replace the classical methods by more sophisticated and objective tools.

Infrastructure deterioration modeling with an inhomogeneous continuous time Markov chain: A latent state approach with analytic transition probabilities

“…The next assumption examined was the conditional independence assumption, for which two extensions were introduced. One was the idea of a semi‐Markov chain in which the probability of transition from one state to another depended not only on the pair of states before and after the transition but also on the duration of the preceding state being stayed in (Ariza et al., 2020; Black et al., 2005; Manafpour et al., 2018; Nesbit et al., 1993; Thomas & Sobanjo, 2016; Yang et al., 2009). The other approach taken to relax the one‐step conditional independence assumption was due to Robelin and Madanat (2007), who proposed a history‐dependent Markov model using a state‐augmentation technique.…”

“…The other approach taken to relax the one‐step conditional independence assumption was due to Robelin and Madanat (2007), who proposed a history‐dependent Markov model using a state‐augmentation technique. Several comparative studies have been conducted to investigate the pros and cons of homogeneous and inhomogeneous Markov chains as well as the semi‐Markov chains (Ariza et al., 2020; Yamany et al., 2021). Related to the semi‐Markov model, Wellalage (2015) proposes a methodology to posteriorly estimate a Weibull hazard model from the estimation results of a homogeneous Markov chain.…”

IABSE Survey of Implemented Decision-making Models used by Public and Private Owners/Operators of Road- and Railway Infrastructures

“…24,25 One can cite in particular the semi-Markov (assuming a Weibull distribution for the sojourn time) and hidden Markov models together with Artificial Neural Networks (ANNs), which have been reported in the literature as reliable deterioration prediction models. [26][27][28] Gamma process representations also offer an alternative to discrete Markov models for the description of degradation processes. Gamma processes are continuous-time stochastic processes with independent, non-negative increments that follow gamma distributions with typically identical scale parameters and a time dependent shape parameter.…”

Zustandsprognose von Ingenieurbauwerken auf Basis von digitalen Zwillingen und Bestandsdaten

“…Cattan [9] entwickelte für Eisenbahnbrücken in Chicago ein MLP mit den Brückeneigenschaften und dem Brückenalter als Input. Weitere Autoren verwendeten MLPs [10–12], Case‐based Reasoning [13], Support Vector Machine [14] oder Random Forests [15] mit div. weiteren Brückeneigenschaften und dem Ziel, einen Zuwachs in der Prognosegenauigkeit zu erzielen.…”