A Study on Concrete Crack Monitoring Method Based on Time–Frequency Characteristics of Electromagnetic Radiation Signals

Abstract

Concrete generates weak electromagnetic radiation (EMR) signals during the cracking process. The EMR monitoring method for concrete cracking is a noncontact and nondestructive monitoring approach. However, substantial interference from environmental EMR noise degrades the consistency of monitoring signals, resulting in low recognition accuracy and a limited monitoring range. To address this, this study introduces a resonance noise reduction method based on the structure of EMR monitoring sensors. The proposed approach enhances the clarity and interpretability of monitored EMR signals through frequency-domain amplitude transformation and discrete wavelet decomposition. The time–frequency characteristics of both EMR signals and noise are comprehensively analyzed by utilizing continuous wavelet transform. A parallel network architecture integrating a convolutional neural network (CNN) and long short-term memory (LSTM) is employed to extract features from EMR signals associated with concrete cracking. Additionally, multiscale convolutional kernels and attention mechanisms (AMs) are integrated according to the time–frequency characteristics of the signals. Experimental results reveal that the proposed method effectively distinguishes EMR signals from four common types of EMR noise, achieving an average signal recognition accuracy of over 98% on the validation and test datasets, with a crack identification accuracy rate of 94.6%. This approach significantly enhances the applicability and potential use of EMR monitoring in concrete cracking scenarios.