Bridge Damage Factor Recognition from Inspection Reports Using Active Recurrent Neural Network

Abstract

Extracting information from bridge inspection reports has recently attracted research interests, since the reports contain valuable information for predictive maintenance of bridges. However, a considerable amount of the reports has limited manually collecting such information from the inspection reports. The research objective was to propose a methodology for automatically recognizing information on bridge damages and the causal factors from the reports with less amount of training data. This study applied recurrent neural network to develop the recognition model, and active learning to train the model. In the active learning scheme, a human annotator was asked to label text data which the model had difficulty in recognizing damages and causal factors from. Experimental results showed that the developed model performs well using only 140 training data to get f-1 score of 0.778, 90.5% of the maximum performance of the model. The proposed methodology will be applied to develop a model for extracting valuable information from bridge inspection reports, and eventually enable efficient preventive maintenance of bridges.