Intelligent health inspection for road multipart covers based on vibration feature encoding and denoising diffusion model

Abstract

Road multipart covers (MPCs) are installed to seal the entrance ports of large drains. Due to the long-term impacts of traffic vehicles, MPC damages may occur. While manual inspection is effective, it can lead to traffic disruptions and is inefficient. In this article, we present a noninvasive approach that analyzes acoustic emissions generated by vehicle-MPC impacts. Specifically, we propose an effective deep learning-based method named vibration feature encoding and denoising diffusion model (VFEDDM), which includes three successive stages. First, the process of “peak window truncation-> scale normalization-> direction aligned RGB encoding” is proposed to appropriately form the key characteristics of vibrations in different propagation directions. This process ensures robustness against variations in vehicle running conditions and changes in measurement distances. Second, the recent generative AI, denoising diffusion model, is introduced to synthesize high-quality RGB feature images, achieving data augmentation. This can address the model training issue caused by data imbalances between normal and defective data. Third, a deep CNN is constructed and trained by utilizing the augmented RGB image set to learn MPC status-discriminative patterns, which are used to assess the health status of the test MPCs. The effectiveness of VFEDDM is verified in the dataset collected from real MPC sites in Hong Kong. It achieves an accuracy of 0.93 for the test MPCs, and the diagnostic results are well visualized by t-SNE. This would provide support for MPC maintenance decision-making and significantly improve inspection efficiency while reducing traffic interference rendered by road closures.