A diffusion model-based deep learning approach for denoising acoustic emission signals in concrete

Abstract

Accurately identifying acoustic emission (AE) signals in concrete is challenging due to material anisotropy and various noise sources like external vibrations and electromagnetic interference, leading to data misinterpretation and inaccuracies in damage detection. Effective noise filtering while preserving relevant AE signals is essential for reliable structural health monitoring (SHM). This study introduces a deep learning approach based on diffusion models, referred to as the acoustic emission denoising model (AEDM), to denoise AE signals in concrete. The proposed model utilizes a convolutional neural network (CNN) enhanced with attention mechanisms and a multi-scale convolutional structure. Additionally, the diffusion model is conditioned on noise observations, allowing the model to better learn and separate noise components from signals, thus effectively filtering out noise while retaining critical signal features necessary for accurate damage detection. Experimental results demonstrate that AEDM outperforms traditional denoising algorithms and other deep learning techniques in noise reduction and signal preservation. When the denoised data from AEDM is used to retrain the damage detection model, it achieves the most significant improvement in accuracy among all tested methods. Moreover, on-site experiments were conducted, and the results further validate the effectiveness of this approach in practical applications.