Acoustic Signal‐Based Deep Learning Approach and Device for Detecting Interfacial Voids in Steel–Concrete Composite Structures

摘要

The reliable synergy between steel and concrete is an important evaluation criterion for the safety and long‐term use of steel–concrete composite structures (SCCSs). However, it is still a great challenge to realize fast and automated interface void detection for composite structures such as towers. Therefore, the experiments of void detection of the bridge tower full‐scale model were carried out with the background of Zhang Jinggao Yangtze River Bridge and other composite structure bridge towers. An automated inspection robotic device was developed, and a convolutional neural network (CNN) identification and classification model based on excitation vibration acoustic signals for the detection of void damage was proposed. In addition, the applicability of the response characteristics of the acoustic signal to the model and the visualization of the features were discussed. Developed tracked magnetic inspection robots can be applied to a variety of towering and complex confined areas, with real‐time inspection efficiency as high as 0.1 m 2 /s. The method based on excitation vibration acoustic signal analysis can be used as a new method for the detection of void damage in composite structures dependent on automatic inspection devices. The introduction of Mel spectrum features into the analysis of excitation vibration acoustic signals can improve the accuracy of identifying structural void damage. Compared with the conventional MLP and LSTM neural network models, the constructed Mel Spectrum with CNN model can realize high‐precision classification of damage, health, and invalid data of the composite structural interface, and the classification and recognition accuracy reaches 96.8%. The equipment and method can realize the accurate and fast detection of the interface void of the towering composite structure. It improves the automation of the void detection process and reduces the safety risk of the detection.