Bridge deck deterioration remains a major concern for owners because of its direct impact on road safety and usability. Existing deterioration models predict future conditions of the deck by linking deterioration to various factors (or features) identified through different engineering and statistical techniques. This led to a lack of a unified feature set, which mostly influences the deterioration of the bridge deck. Additionally, traditional deterioration models (e.g., linear regression) are unable to accurately predict the discretized values of the deck condition ratings, resulting in inaccuracies. For instance, a predicted value of 6.51 was approximated to be a condition rating of 7, which is inappropriate when using discrete data. This paper uniquely combines the bridge features identified in the literature and applies principal component analysis (PCA) to capture the relevant information in the feature set needed to predict the condition of the bridge deck. This case study analyzed 53,000 observations from 24,240 unique bridges in Maryland and Virginia, categorizing the bridges into five condition rating groups. Feedforward artificial neural network (ANN) models were developed using different combinations of principal components derived from the dataset. The performances of these models were compared to a base model that used 14 features collected from the literature. Analyses revealed that incorporating at least nine principal components resulted in a deterioration model with a prediction accuracy of 76%, surpassing the base model’s accuracy of 75%. The results demonstrate a lower prediction error compared to previous studies. Utilizing nine principal components reduces the feature-set’s dimensionality from 14 to nine, thereby minimizing the subjectivity associated with visual inspection data and enhancing model performance for bridge condition assessment.
Abiona et al. (Wed,) studied this question.