Semi-supervised networks integrated with autoencoder and pseudo-labels propagation for structural condition assessment

Insufficient labeled data of vibration measurement from unknown structural states cast great obstacle to attempt of extending the superiority of deep learning techniques into their rapid state evaluation. This study proposes novel semi-supervised networks for condition assessment integrated with deep autoencoder and pseudo-labels propagation. The architecture and mechanism of the workflow are elaborated. With sophisticated network design and novel strategies for improving performance, the proposed method succeeds to establish a systematic network embedded with simultaneously self-supervised autoencoder, optimized unsupervised fuzzy clustering and supervised classification algorithms in semi-supervised paradigm. Both numerical simulations and full-scale laboratory shaking table tests of a two-story building structure were conducted to validate its capability of identifying the post-earthquake damage scenarios. Driven by acceleration signals measured by merely single sensor, the prediction accuracy of proposed method achieved to be 0.95 in numerical validation and averagely 0.86 in laboratory case studies, respectively, confirming a powerful effectiveness and robustness.

Automated summary

This study proposes novel semi-supervised networks integrating deep autoencoder and pseudo-labels propagation for condition assessment with insufficient labeled data. The proposed method achieves a powerful effectiveness and robustness in identifying post-earthquake damage scenarios based on numerical simulations and full-scale laboratory tests.