IPDC Network-Based Image Reconstruction for Damage Detection of CFRP With Electrical Impedance Tomography

Abstract

The nonlinearity and ill-poseness of image reconstruction has hindered the application of electrical impedance tomography (EIT) in damage detection of Carbon Fiber-Reinforced Polymer (CFRP). To cope with this problem, a dual-channel Inception-Dense-Cbam (IPDC) network model integrated with attention mechanism is proposed. To combine the diversity of feature extraction with the sparseness of the true conductivity distribution, the whole frame is designed to be composed of two parallel branches: the Inception-resnet-cbam (IPRC) branch utilizing multi-scale feature fusion based on the Inception-resnet module and the Dense-Cbam (Dense-CB) branch utilizing feature reuse based on the dense module. Features extracted via the two branches are then fused as outputs. Additionally, to ensure extraction accuracy of damage boundary, convolutional block attention mechanism module (CBAM) is integrated into both branches. Both simulation results and prototype experiments showed that compared with traditional algorithms, the proposed method could effectively reduce the artifacts of damage images, improve accuracy of damage location, and enhance the definition of damage edges. Simulation data obtained with added noise and the data of prototype experiments demonstrated that the IPDC network had the good anti-noise performance in detecting different types of damages.