Prediction Model for Bridge Condition

Efficiency and sustainability of bridge infrastructures asset management are highly dependent on an accurate prediction of structural conditions, and it is essential for bridges safety. Many studies have been conducted on this matter, in which many researchers also sought the power of machine learning techniques. This work aims to build a machine learning classifier to predict the deck, superstructure, and substructure conditions of bridges, linking the weather data with bridge condition data. Therefore, the investigated dataset consists of the national bridge inventory data published by the Federal Highway Administration of the US and the climate data downloaded through the long-term pavement performance tool. Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology was adopted to achieve the objectives. Thus, a thorough investigation of the dataset was accomplished via the exploratory data analysis (EDA) approach at first. Data Visualization was used to first identify some of the easily identifiable trends in the data as well as have a preliminary idea about the most important parameters in this dataset. The whole dataset was also divided into three sub-datasets based on the bridge materials: concrete, prestressed concrete, and steel so the performance of whole data set and divided dataset can be assessed. Furthermore, four algorithms have been developed, namely Decision Trees (DTs), k-Nearest Neighbors (k-NN), Support Vector Machine (SVM), and Naive Bayes (NB). The results conclude that Decision Trees (DTs) outperform the other techniques in general. The different models exhibit different levels of accuracy depending on the dataset and the modifications and optimizations made. Finally, the best predictions have been delivered to the substructure conditions by all four techniques in terms of the structural components.