Detection of tunnel lining voids using multi-channel cochleagram coefficients with an attention mechanism

Tunnels play a crucial role in transportation infrastructure, and ensuring the safety of tunnel lining structures is critical for vehicles passing through. Acoustic detection of tunnel lining conditions is a viable approach to assessing structural integrity. In this study, acoustic signals are collected by striking the tunnel lining surface with a small hammer, and sound samples are generated in different reverberant environments to simulate tunnel conditions. A novel multi-channel cochleagram (MCCG) feature based on the multi-resolution cochleagram (MRCG) is introduced, which achieves data compression in reverberant environments. A convolutional network classification model based on the Convolutional Block Attention Module (CBAM), termed CBAM-SCNN3, is developed to classify hollow and solid sound samples under different reverberation. In the proposed method, the MCCG feature compresses and extracts multi-channel data, while the CBAM module adjusts weights of channels and spatial features. The depthwise separable network extracts features from MCCG in different directions. Experimental results show that the proposed method significantly outperform traditional features and exhibit improved robustness in reverberant environments while reducing network complexity. The average recognition accuracy over all reverberant environments is 98.4%, and the computational and parameter loads are reduced by 76.9% and 76.3%, respectively. Source code: https://github.com/EthanWu99/CBAM-SCNN3