iCMAD: Physics-Informed Deep Learning with Digital Image Correlation for Aircraft Crack Monitoring

Crack propagation as a consequence of fatigue in aircraft structure is a long-standing safety issue in the metallic framework, and the present system of checks is based on subjective evaluation and localized sensors, which cannot be extended to large aircraft. In this paper, the multi-task deep learning system iCMAD (Improved Crack Monitoring and Analysis Deep Network) that integrates full-field Digital Image Correlation (DIC) with fracture mechanics constraints to automatically detect cracks, segment them, and locate their tips and predict their growth is presented. iCMAD is based on the previous CMAD architecture and renders the Paris law into a physics-informed loss function, integrates temporal modules of variable amplitudes fatigue spectra, has stereo DIC-capable out-of-plane deformations, and caters to edge hardware via model compression. Comparison of datasets which consist of twelve aluminium 2024-T3 samples in various loading regimes demonstrates fidelity in prediction and close to real time inference on the embedded platforms.