Co-authors​ :

 Sun HUAIXIANG (CHINA), Xiaodong WANG (CHINA), Zhidong GUAN  

Damage identification in composite material structures is a focal and challenging area within structural health monitoring research. Beyond non-destructive testing methods like acoustic emission and ultrasonic inspection, data-driven machine learning algorithms, known for their rapid response and adaptive capabilities, are extensively used to unravel complex mapping relationships between data. This paper proposes a data-driven method for damage identification in structures, focusing on the correlation between strain responses and damage features in typical composite material reinforced panels under load. Structural response data under load are acquired through finite element simulation（FEM）. Using damage equivalency methods, various types of damages, such as panel/stringer debonding and stringer damage, are preset at different locations within the structure. An optimal strain measurement points arrangement is designed to extract structural responses and to create a structural damage-strain dataset. This dataset is enhanced through data augmentation, standardized, and damage labels are encoded using the onehot method. Five machine learning models - Graph Neural Network (GNN), Decision Tree (DT), Random Forest (RF), Artificial Neural Network (ANN), and K-Nearest Neighbors (KNN) - are developed and trained. These models predict the type and location of structural damage using different strain input processing methods. Among them, the GNN model, which inputs data in a graph structure and encodes the structure's geometric and topological information in node and edge features, stands out for its performance. The other four models input the original strain vectors. Network structures and model parameters are optimized using the Grid-Search method, and suitable classification metrics are used to evaluate model performance. The comparative analysis highlights that the GNN model, employing graph-based strain data as its input, demonstrates notably superior performance. This model achieves the highest accuracy and lowest variability in predicting debonding damage and stringer damage, especially demonstrating a significant performance advantage in predicting the presence of damage compared to the other four models. The robustness of the GNN model is further assessed by adding noise to the strain input to explore its capability in handling missing strain data, as verified by evaluating the confusion matrix of the prediction results. This results in a reliable and rapid method for identifying damage in reinforced panel structures. The deep learning-based method for damage identification in composite material structures established in this paper, considering the mode of strain data input, offers high accuracy and robustness, making it significantly valuable for the health monitoring of composite material structures.