Unsupervised structural damage detection methodology using deep convolutional autoencoder and image clustering with wavelet transmissibility pattern spectra

Abstract

Vibration-based structural damage detection under non-stationary excitation remains challenging due to the complex time-varying characteristics of structural responses and the limited availability of labeled data. Although existing unsupervised deep learning approaches have demonstrated potential for extracting damage-sensitive features, many of them rely on reconstruction errors as damage indicators, which may not sufficiently characterize the intrinsic evolution of structural states in the latent feature space. To address this limitation, an unsupervised structural damage detection framework based on wavelet transmissibility pattern spectra (WTPS), deep convolutional autoencoder (DCAE), and density-based clustering is proposed in this study. First, WTPS is employed to transform non-stationary vibration responses into damage-sensitive time–frequency representations. Subsequently, a spliced WTPS representation is constructed by combining the retained intact baseline WTPS with the WTPS obtained from newly acquired unlabeled monitoring responses. The DCAE is then utilized to extract compact latent features from WTPS image patches, while density-based spatial clustering of applications with noise clustering is introduced to identify damage-induced feature distribution variations and emerging cluster patterns. Different from conventional reconstruction-error-based unsupervised approaches, the proposed method identifies structural state changes through latent feature clustering rather than empirical reconstruction-error thresholds. Numerical simulations on a simply supported beam and experimental validation on a full-scale steel tower benchmark model demonstrate that the proposed framework can effectively detect the investigated single and multiple damage scenarios under non-stationary excitation. In particular, the framework successfully identifies stiffness reductions ranging from slight damage (10%) to severe damage (30%) in numerical simulations and structural changes induced by bolt loosening in experimental tests. The results demonstrate the potential of the proposed method for unsupervised structural health monitoring under complex operational conditions.