Localization and imaging of internal hidden defects in concrete slabs based on deep learning of vibration signals

Abstract

Internal hidden defects such as delamination and voids are common distress of concrete slabs and the localization of these defects is important for structural safety. This paper develops a damage inspection method that is feasible for locating and imaging internal defects of concrete slabs based on deep learning of impact-caused vibration signals. To alleviate the problem of low efficiency of manual signal collection, a prototype of mobile detection system is developed for the automatic generation and acquisition of vibration signals. Two optimized deep learning models in the framework convolutional neural network (CNN) are proposed for 1D time series signal and 2D time-frequency spectrogram to address the problems of low recognition accuracy and long training time of the traditional network, where the principal component analysis (PCA) and squeeze and excitation (SE) modules are incorporated respectively for a better feature learning and extraction of vibration signals of internal defects. The softmax layer is added in the networks to calculate the state probability value for the completion of damage contour maps. The network proposed in this paper is used to systematically compare the training and recognition effects of the data collected by contact senor (accelerometer) and air coupled sensor (microphone). The experiments show that the detection accuracy of the one-dimensional (1D) and two-dimensional (2D) network based on the accelerometer-acquired data reaches 90.4% and 93.4%, respectively. The generalizability of trained models is also validated based on the independent dataset from other research and actual concrete floor of a multistorey building. In the future, the prototype of the hardware system can be further improved by embedding GPS information to realize the real-time automatic localization of hidden defects of engineering structures.