Recognition method of internal concrete structure defects based on 1D-CNN

Abstract

Conventional intelligent recognition methods highly depend on artificial feature extraction and expert knowledge to recognize concrete structures’ internal defects. For solving this problem, one intelligent recognition method for internal concrete structure defects is proposed, based on a one-dimensional, convolutional neural network (1D-CNN). First, the impact echo detection signals are acquired, to establish the training and testing of samples for various internal concrete structure defects. Then, the convolutional network structure is used to achieve the adaptive, hierarchical extraction of the impact echo signals’ features. Finally, the Softmax classifier is used to provide the diagnosis result at the output end. The experimental results of four types of internal defects (including voids, water-filled, imperfect solids, and sound) show that the 1D-CNN classifier, with the predicted signals as the training set, enables the successful identification of the internal defects of the concrete structure and achieves more than 90% defect-recognition accuracy. In addition, the 1D-CNN classifier has strong anti-interference ability and feasibility in practical applications. This work improves the performance of ‘impact echo’ in identifying internal defects in concrete and realizes the intelligent analysis of impact echo signals.