An automated method for interpreting fatigue fracture surfaces with marker load based on computer vision and artificial neural networks

The marker load method, recognized as one of the most effective approaches for reproducing the three-dimensional fatigue crack growth process in alloys, has been widely applied in theoretical and experimental studies of fatigue crack gross in reusable flight vehicle structures. However, this method suffers from low efficiency and poor accuracy in interpreting marker lines on fatigue fracture surfaces. In this study, an automatic method for marker line recognition and local defect completion is developed based on computer vision and artificial neural network techniques to enhance the efficiency and accuracy of crack interpretation, thereby strengthening the capability of the marker load method in extracting crack front (effective only for single-source cracks, multi-source cracks require further investigation.). Specifically, a convolutional neural network algorithm (constructed on the you only look once (YOLO) v8 framework) is first employed to identify continuous marker lines as a series of discrete coordinate points according to their geometric features. Subsequently, the density-based spatial clustering of applications with noise (DBSCAN) algorithm, combined with a newly developed scatter cluster matching algorithm, is used to cluster and match the points belonging to the same crack front. Finally, a long short term memory (LSTM) neural network model is utilized to reconstruct incomplete marker lines, establishing an automatic interpretation method for fatigue fracture marker lines and compensating for the loss of crack information caused by marker line defects (crack source need to be selected manually).