Alternative Sequence Classification of Neural Networks for Bridge Deck Condition Rating
Setare Ghahri Saremi, Dimitrios Goulias, Yunpeng Zhao
Abstract
With the rapid deterioration of infrastructure, accurate predictions of bridge deck performance is critical in bridge management systems (BMS). The US Federal Highway Administration (FHWA) has developed the National Bridge Inventory (NBI) database, which has accumulated a vast repository of bridge performance data in the US. Although alternative methods have been used for harvesting such data, the use of machine learning (ML) has been explored to a lesser degree due to its modeling complexity. Objective of this study was to develop and assess alternative sequential models by employing long short-term memory (LSTM) and convolutional neural networks (CNN). The advantage of such modeling alternatives in relation to past ML studies is that consider (1) fixed or variable number of sequences of observations combined with (2) continuous or flexible sequences, producing better condition rating predictions. The results indicated that the models exceeded accuracy prediction from the mid 80% in past studies to levels of 95% for Maryland bridges, and all the way to 97% when Massachusetts data were added. Such results indicate that the proposed models are able to better capture the complex relationships between bridge-related features and condition rating. Transferability of the proposed approach was confirmed with the successful model response when data from the second state were used. It is expected that improved predictions of condition rating will promote more effective maintenance and rehabilitation strategies in BMS.