Date of Award

12-2016

Document Type

Thesis

Degree Name

Master of Science (MS)

Legacy Department

Civil Engineering

Committee Member

Dr. Weichiang Pang, Committee Chair

Committee Member

Dr. Kalyan Piratla

Committee Member

Dr. Brandon Ross

Abstract

Given a database of approximately 8,000 culverts in South Carolina with varying sizes, types, configurations, and the associated ratings of different output categories, a predictive deterioration model was produced in an attempt to match the ratings of these output categories. These models used the physical culvert information given in the database of culverts and associated environmental characteristics including historical temperature, precipitation, pH, and estimated runoff coefficient as inputs for the model. The models used combinations of inputs that produced the model with the best performance measures. In addition, a separate group of models was created for each of the six culvert types commonly found in South Carolina. These models used different combinations of the input variables to produce a model that rated a culvert in ten categories: cracking, separation, corrosion, alignment, scour, sedimentation, vegetation, erosion, blockage, and piping. The scores for each of these categories were combined to give an overall composite score for each of the culverts. Two types of models were used for each of the culvert types and output categories, a logistic regression model and an artificial neural network model. The purpose of this model was to allow the user of the model to input a culvert or group of culverts and receive their expected culvert ratings in accordance with the SCDOT Field Inventory and Inspections Guidelines. The model also produced a composite rating, consisting of a combination of the ten input categories predicted by the model. There were several preset composite weights for these categories, but the model also adapted for a user input combination of output categories. The models produced were shown to have a coefficient of determination of between 0.25 for poorly correlated models to a coefficient of determination 0.80 for better correlated models when comparing the predicted culvert score with the actual culvert score. The models that were produced were meant to serve as both a tool to determine the approximate health of a group of culverts, and to compare the scores of a group of culverts allowing the SCDOT to make decisions about rehabilitation and repair without physically inspecting a culvert.

Share

COinS