Gravity dam displacement monitoring using in situ strain and deep learning

Recent studies in dam displacement monitoring primarily focus on single-response monitoring or model updating using advanced techniques. Few studies involve the combination analysis of displacement with other synchronized responses utilizing their monitoring characteristics. In situ strain data provide a strength-safety perspective for dam displacement monitoring. The challenge lies in that estimating displacement directly using limited discrete strain data may be misleading. This paper analyzes the relationship between displacement and global, and multipoint local strains from the perspective of the differences in load effects of gravity dams, and indicates that introducing appropriate state factors improves the estimation. A displacement estimation model driven by strain data and state factors is developed using stacked convolutional neural network, and the variable relationships within the model are interpretated via accumulated local effects. Incorporating specific strength criteria, a novel displacement monitoring indicator based on the tensile safety of the dam heel is proposed. A case study of a gravity dam showcases the effectiveness of the proposed approach in comparison with the solely strain-based model and the traditional hydrostatic-seasonal-time factors-based model.