Detection of CFRP damages based on electrical impedance tomography by LP-ERes2NetV2

Abstract Owing to the self-conductive characteristics of carbon fiber reinforced polymer (CFRP), the electrical impedance tomography (EIT) technique can be used for damage detection. However, the inverse problem of EIT is underdetermined, nonlinear, and ill-posed, which leads to severe artifacts in image reconstruction, resulting in reduced image quality and fidelity. In recent years, deep learning algorithms have shown great potential for solving the EIT inverse problem. This paper proposes an EIT imaging algorithm based on LP-ERes2NetV2 deep network for damage detection in CFRP laminates. The algorithm consists of a pre-mapping module, a data reconstruction module, and a feature extraction module, which work together to achieve the reconstruction of CFRP damage images. The advantages of this algorithm include: (1) combining L1 regularization and principal component analysis dimensionality reduction effectively reduces redundant information, improves computational efficiency, and mitigates the underdetermined problem; (2) adopting a deep residual network structure and utilizing the nonlinear mapping ability of deep learning to deal with the nonlinear characteristics of the EIT inverse problem; (3) introducing attention mechanisms and multi-scale feature fusion, which enhances the ability to capture damage features and alleviates the ill-posed nature of the problem; (4) expanding channels and pruning the network structure maintains performance while reducing computational complexity, improving the algorithm’s practicality. The quality of the reconstructed images is evaluated by correlation coefficient (CC) and structural similarity index (SSIM). Simulation results show that the LP-Eres2NetV2 algorithm performs excellently in reconstructing all types of damage, with CC exceeding 98% and SSIM exceeding 86% in all cases. A 16-electrode EIT experimental platform was set up to detect damage in CFRP laminates. The experimental results are consistent with the simulation results, further confirming the superiority of the LP-ERes2NetV2 algorithm.