An unsupervised learning approach to predict the deterioration of aging bridges using inspection data

Abstract

A common approach for developing degradation models for aging bridges involves fitting a stochastic process, such as a Markov or semi‐Markov chain, to condition data collected from visual inspections and stored within Bridge Management Systems. However, variations in environmental, structural, and operational factors result in different aging rates among bridges. Consequently, identifying groups of bridges exhibiting similar deterioration patterns and developing tailored deterioration models for each group can reduce the uncertainty in remaining useful life estimations and optimize the allocation of maintenance resources. This article presents an unsupervised learning approach to identify bridge populations with homogeneous degradation rates. The SNOB algorithm is applied to cluster bridge sojourn times across predefined degradation levels utilizing Weibull Mixture Models. Three distinct groups of bridges are identified, here referred to as fragile, normal, and robust bridges, each one characterized by a different degradation rate. For each group, a deterioration model based on a semi‐Markov process is developed, capturing the evolution of bridge conditions within the cluster. The proposed approach is applied to condition data from the US National Bridge Inventory (NBI) and the results are discussed by emphasizing a possible correlation between the identified clusters and climate conditions of bridge locations.