Semester of Graduation
Master of Science (MS)
1600 (12%) of Louisiana's 13,000 bridges that allow the passage of people, goods, and services are load posted, meaning they are deemed incapable of securely carrying all legal loads. A machine learning framework for estimating the number of load-posted bridges within a regional bridge portfolio would be extremely valuable in the future for allocating essential resources during long-term planning. A framework of this type is necessary since a significant proportion of the bridges are load posted. Furthermore, deterioration due to aging and future increases in legal loads can compound this situation. In this setting, there is a lack of data-driven frameworks to guide targeted maintenance and mitigation efforts necessary to reduce the future number of load-posted bridges significantly. As such, this research aims to quantify the number of load-posted bridges in Louisiana during the next 50 years using machine learning approaches. The need for the data-driven framework is because it's a laborious task to analyze each bridge of 13,000 bridges. Predicting load-posted bridges is complex and difficult to estimate using traditional statistical methods. Machine learning offers advantages and is effective over conventional methods because of its scalability and flexibility in dealing with huge data. To this end, the National Bridge Inventory (NBI) database and data from the Louisiana Department of Transportation and Development LADOTD were used to gather data on bridges. This research mainly focused on COSLAB (Concrete Slab), COPCSS (Concrete Precast Slab Units), LWPCSS (Lightweight Concrete Precast Slab Units), CONIBM (Steel I-Beam (Rolled), CODEKG (Concrete Deck Girder) category bridges. To identify important attributes in NBI data Random Forest feature selection method was used. Important features were found for every class of bridges. To analyze the important features and trends between the number of load posted bridges and various parameters, data tables were formed for all the classes of bridges using the NBI data. This study considered surrogate models such as Decision Trees, Random Forests, Logistic Regression, and Neural Networks as predictive models. These models were trained using bridge data and evaluated based on their accuracy and confusion matrix. The number of load-posted bridges for the next 50 years in Louisiana for different bridge classes were found using the Random Forest predictive model. Load posting decision was made for each bridge using the predictive model. In a particular year, load-posted bridges were predicted by replacing the time-varying attributes with their future predicted values. Concrete Slab (COSLAB) had the highest number of load-posted bridges in the next 50 years, and Concrete Deck Girder (CODEKG) had the least number. In the next step, uncertainty in the number of load posted bridges was estimated by training random forests using different bootstrap samples. In other words, different random forest models were trained with multiple random states, and the mean and standard deviation of the number of load posted bridges were obtained. Finally, an interactive web page was designed to showcase all the results.
Bandaru, Sai Naga Sasi Aditya, "MACHINE LEARNING-BASED BRIDGE LOAD POSTING PREDICTION" (2022). LSU Master's Theses. 5568.
Dr. Sabarethinam Kameshwar