DLGP-Net: dual-branch local-global perception network for audio-based non-destructive detection of internal defects in wooden components of ancient buildings

Abstract It is important to evaluate the internal defects in wooden components for the maintenance of ancient buildings since ancient buildings in China are mainly built with wooden components. Recently, stress waves and x-rays have been used to detect internal defects of wooden components. Nevertheless, these methods usually rely on high-cost devices, which are difficult to operate and harmful to buildings and operators. To address these shortcomings of present methods, a novel audio-based strategy for non-destructive detection of internal defects in wooden components is proposed, which can largely reduce the equipment cost, operation complexity and avoid harming operators or buildings. Based on our strategy, a deep learning-based dual-branch local-global perception network (DLGP-Net) is proposed to achieve accurate non-destructive detection of internal defects in wooden components via extracting local-global features in parallel. The DLGP-Net contains four sub-networks: (1) the timbre extraction (TE) module, (2) the local feature encoder (LFE), (3) the global feature encoder (GFE), (4) the vote-based prediction (VP) module. The DLGP-Net leverages the complementary strengths of local and global feature extraction to enhance the detection accuracy of internal defects in wooden components. By integrating the TE module with the dual-branch LFEs and GFEs, the network can effectively capture both fine-grained details and broader contextual information from the knocking sound audio. The VP module further refines the final decision by aggregating insight from multiple feature representations, ensuring robust and reliable defect detection. To evaluate the performance of DLGP-Net, wooden components selected from two ancient buildings located in Hangzhou, China, and corresponding knocking sound audio are utilized as our self-constructed data set. According to the experimental results, the proposed DLGP-Net outperforms mainstream deep learning-based audio classification models, including PANNS, ResNet, CAM++, TDNN and ECAPA-TDNN.