Transfer learning for structural health monitoring

• Ano de defesa: 2023
• Autor(a) principal: Yano, Marcus Omori
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Estadual Paulista (Unesp)
• Palavras-chave em Português: Transfer learning; Domain adaptation; Transfer component analysis; Joint distribution adaptation; Maximum independence domain adaptation; Structural health monitoring; Damage identification; Aprendizagem por transferência; Adaptação de domílnio; Análise de componentes de transferência; Adaptação de distribuição conjunta; Adaptação de domínio de independência máxima; Monitoramento de integridade estrutural; Identificação de dano
• Link de acesso: https://hdl.handle.net/11449/251197

Structural Health Monitoring (SHM) proposes the continuous assessment of the structural integrity to ensure the safety operation and increase the lifespan of structures from several areas. This traditional methodology still have generalization difficulties among structures, even when structures are nominally and topologically similar. The data sets present divergences between their underlying distributions that do not allow the generalization of the model estimated to different situations. In the last years, transfer learning has gained relevance due to extending the SHM concept to investigate different structures, while minimizing costs with monitoring systems and time associated with data acquisition. This methodology can change how SHM is currently proposed by leveraging knowledge gained from a well-monitored structure to improve the integrity assessment of other structures under unknown conditions. The main idea is to reuse the relevant knowledge from a labeled structure (source domain) to investigate another one (target domain) with limited data. This thesis intends to lay down the foundations of transfer learning for SHM by describing the motivations that led to its application in the analysis of structures. At first, transfer learning is applied to overcome difficulties inherent in the complexity of estimating an accurate finite element model, and unsupervised damage detection of a real bridge is performed using knowledge of its model. Then, transfer learning is combined with a stochastic model for damage detection in an experimental application and its underestimated numerical model, and the quantification of damage levels in the structure is carried out based on a requirement imposed by the transfer learning methodology. Finally, a comprehensive analysis using transfer learning is proposed to investigate three real bridges subjected to environmental and operational conditions, and the importance of the quality of knowledge transferred across different bridges for damage detection is highlighted. The transfer learning methodology indicated satisfactory results in all investigated cases and demonstrates the attractive benefits of its application, which can result in significant progress in the damage identification process considering different structures.