DEFORMATION DETECTION FOR CONCRETE BRIDGES USING A CONVOLUTIONAL NEURAL NETWORK WITH INFLUENCES THE NUMBER OF TRAINING DATA

Abstract

In recent years, there is demanded for both safety management and cost reduction as concrete bridges are aging. The deep neural network (DNN) is one of the machine learning which is expected to be an effective method to improve inspection efficiency. This study applied a DNN as a machine learning to build the learning model which is used as the detector for the deformation of concrete bridge. The learning models were built by training dataset integrated the two or three areas from our previous studies as two or three areas trained learning models, and the possibility of improvement the accuracy improvement for deformation detections in concrete members was investigated. The results showed that the two-areas trained learning model improved the classification accuracy by 2-6% with twice as much data as the single-area trained learning model, and the three-areas trained learning model improved the classification accuracy by 6-12% with thrice as much as the single-area trained learning model. Therefore, the expansion of training data was found to contribute to improve an accuracy of the detection for concrete deformation.