Development of Data-based Hierarchical Learning Model for Predicting Condition Rating of Bridge Members over Time

Youngjin Choi, Jungsik Kong

Research output: Contribution to journalArticlepeer-review

1 Citation (Scopus)

Abstract

Bridge maintenance strategy is implemented to ensure that effective decisions are made based on the evaluation of the current performance and predicted future conditions of a bridge. The current bridge maintenance system relies on traditional manpower-based methods to determine bridge performance. However, the present techniques employ a material deterioration model that has uncertainty in predicting bridge performance. Moreover, the utilization of collected bridge inspection-related data is insignificant, and related research is lacking. Accordingly, a new type of bridge deterioration model is proposed using state performance data based on bridge inspection. To formulate the deterioration model, bridge condition inspection data accumulated over a long period were utilized. The data have been continuously updated since 1992 by the US Federal Highway Administration and consist of basic information and various influencing factors of bridges. The developed model uses the long short-term memory (LSTM) algorithm (a type of recurrent neural network) as well as layer normalization and label smoothing to improve the applicability of fundamental data. For the stable application of data, predicting the model performance for up to 30 years every 2 years is possible. By implementing the many-to-many type of the LSTM algorithm, the predicted probability of each grade derived for each sequence was weighted and averaged with the grade value to derive a continuous state grade result. Thus, the proposed model can express discrete historical degradation indices in continuous form according to the service life. In addition, it enables the prediction of bridge performance using only the fundamental information regarding new and existing bridges. For the effective use of basic data, an optimization model was derived using preprocessing and various regularization techniques. Additionally, a feed-forward deep learning model and stochastic model were developed using the same data. For performance assessment, a regression analysis evaluation method was applied, and comparative analysis was performed using the inspection data of an actual bridge. The use of a time-series model enabled the continuous and stable prediction of bridge performance.

Original languageEnglish
Pages (from-to)4406-4426
Number of pages21
JournalKSCE Journal of Civil Engineering
Volume27
Issue number10
DOIs
Publication statusPublished - 2023 Oct

Bibliographical note

Publisher Copyright:
© 2023, Korean Society of Civil Engineers.

Keywords

  • Bridge deterioration
  • Bridge maintenance
  • Deep learning
  • Deterioration model
  • National bridge inventory

ASJC Scopus subject areas

  • Civil and Structural Engineering

Fingerprint

Dive into the research topics of 'Development of Data-based Hierarchical Learning Model for Predicting Condition Rating of Bridge Members over Time'. Together they form a unique fingerprint.

Cite this