Structural damage assessment through a new generalized autoencoder with features in the quefrency domain

Abstract

Recently, the challenges in modeling complex dynamical systems, and the advancement in machine learning methodologies have indicated a new promising direction for damage assessment in civil and mechanical systems. Powerful and efficient data-driven approaches have been increasingly employed in Structural Health Monitoring (SHM) to extract Damage Sensitive Features (DSFs) from the monitored dynamic response of structures. In this study, a New Generalized Auto-Encoder (NGAE), integrated with a statistical-pattern-recognition-based approach that uses the power cepstral coefficients of structural acceleration responses as DSFs, is proposed for structural damage assessment. This NGAE is well-generalized in the components of cepstral coefficients that represent the structural properties of the entire system thanks to a newly defined encoder-decoder mapping, which largely reduces rid of the data variance attributed to different types of excitations and measurement noise. The cepstral coefficients, by virtue of a compact representation of the structural properties, can greatly simplify the structure of the NGAE, and therefore, significantly accelerate training and inference speeds with very few computational requirements. Two specific evaluation metrics that relates to the autoencoder signal reconstruction error are defined and used to assess the presence of damage. The proposed method has been validated through numerical simulations and experimental data, and shows better performance compared to a Traditional Auto-Encoder (TAE) and the Principal Component Analysis (PCA).