Unsupervised Damage Diagnostics with Data Normalization Using Experimental Data

Abstract

Environmental temperature variations can significantly affect modal parameters, complicating damage detection in vibration-based structural health monitoring (SHM) approaches. To mitigate these challenges, this study proposes a novel approach to transform temperature-variant data into a temperature-invariant domain for more effective data normalization and thus more accurate damage detection when the labeled data under temperature variation is unavailable. Experimental acceleration data collected from different structural health states are first categorized into temperature-variant and temperature-invariant domains, and modal parameters (modal eigenfrequencies, eigenforms, and damping ratios) are extracted accordingly. An adversarial domain adaptation technique is then employed to enhance the algorithm’s ability to generalize across domains while accurately distinguishing between damage states. The proposed approach is validated on a real field engineering structure realized as a 9 m high lattice mast under real environmental conditions. The approach consists of three main components: a general feature extractor, a classifier, and a domain discerner. The modal parameters are processed through the feature extractor to obtain domain-independent features, which are subsequently mapped to the corresponding damage states. Meanwhile, the domain discerner competes with the feature extractor by attempting to identify the domain origin of the features, ensuring robust feature extraction. The results demonstrate the potential of this approach to overcome temperature-induced challenges in SHM and improve damage detection accuracy.