An ordinal autoregressive deep learning model for probabilistic prediction of bridge deterioration condition levels

To address the need for more cost-effective and efficient bridge maintenance strategies beyond standard routine inspections for all bridges, this study introduces a novel probabilistic bridge condition forecasting method. The proposed method aids transportation agencies and decision-makers in prioritising bridges for inspection by predicting the likelihood of transitions to various condition levels in the following year. The method accounts for the ordinal nature of bridge deterioration levels and provides a full-spectrum probability distribution across all possible condition levels for each bridge in the following year. The model uses a novel autoregressive deep learning framework to capture complex temporal patterns among bridge attributes for forecasting likelihoods. It also incorporates an optimisation-based mechanism to adjust predicted probabilities, ensuring they adhere to the axioms of probability and maintain ordinal consistency. The implementation of the proposed method using the National Bridge Inventory (NBI) data from the neighbouring states of New Jersey, New York, and Connecticut, covering 19,500 bridges, demonstrated that the model could maintain a remarkable predictive accuracy. The model was trained on NBI data from 1992 to 2021 to predict bridge conditions in 2022, achieving a Rank Probability Score (RPS) of 0.005 and a Kendall's tau statistic of 98.7%.