Deterioration prediction of existing concrete bridges using a LSTM recurrent neural network

Abstract

Bridge censored databases can be used to analyze and assess structural deterioration conditions, but conducting the analysis is difficult. This difficulty occurs because many factors affect deterioration, and the time span of the data for these factors depends on the years in service of the respective bridge. In addition, the values of some factors are not regularly observed. The present study uses the long short-term memory (LSTM) to consider twelve potentially influencing factors to recognize the relationships between these factors and deterioration grades. Testing the model on an inspection database of 3,368 bridges indicates that the LSTM model obtained an accuracy of exceeding 80%, i.e., outperforms the performance of a multilayer perceptron model established using the same database. For four types of bridges, the LSTM model shows equivalent performance. In addition, the predictive ability of the LSTM model for coastal bridges is slightly superior to non-coastal bridges. No significant differences in accuracy are determined between different deck areas. Practically, the model can predict bridge deterioration paths, and could help decision-makers formulate predictive intervention strategies for improving the quality of maintenance management.