Comparison of Markov chain and recurrent neural network in predicting bridge deterioration considering various factors

Abstract

Bridge deterioration reflects changes in bridge conditions and is caused by many factors. Predicting bridge deterioration is critical for implementing predictive interventions, but developing such credible models that consider multiple influencing factors remains difficult. This study uses Markov Chain (MC) and Recurrent Neural Network (RNN) to establish deterioration prediction models. Then, comparative studies of these two models are conducted on a censored database regarding the mean deterioration progress, deterioration progress of bridges with different deck types, deterioration progress of a single bridge, and the influence of each factor on deterioration progress. The results indicate that RNN-predicted deterioration to a worse state takes less time than MC-predicted deterioration, but the difference is less than 5 years. Deterioration progresses of bridges with different deck types show that PC deck and steel deck are more durable than RC deck and concrete-steel composite deck. Qualitative and quantitative analysis infers eight factors substantially affecting the deterioration progress. Overall, the deterioration progress provided by these two models can be utilized to determine bridge’s remaining life and analyze the effects of external factors affecting the deterioration. This information is extremely valuable for assisting in developing and implementing effective management and maintenance strategies.