Abstract:Structural health diagnosis has been investigated following a data-driven machine learning paradigm. However, the model accuracy and generalization capability highly rely on the quality and diversity of datasets. This study established a framework for structural health diagnosis under limited supervision. Firstly, an image augmentation algorithm of random elastic deformation, a novel neural network with self-attention and subnet modules, and a task-aware few-shot meta learning method were proposed for vision-based damage recognition. Secondly, deep learning networks were established to model intra- and inter-class temporal and probabilistic correlations of different quasi-static responses for condition assessment. Finally, a two-stage convergence criterion merging with the subset simulation and Kriging surrogate model was designed for reliability evaluation. Real-world applications on large-scale infrastructure demonstrated the effectiveness.

Structural Health Diagnosis under Limited Supervision