Dynamic early-warning model of dam deformation based on deep learning and fusion of spatiotemporal features

Abstract The reasonable formulation of early warning indicators is essential for dam deformation safety assessment. Previous studies on dam deformation early warning, however, failed to consider the spatiotemporal features, especially the spatial deep nonlinearity between the multiple environmental factors and multipoint displacements, and the randomness and fuzziness in the description of deformation time-domain distribution, limiting the updating ability of models and the reliability of early-warning indicators. To solve these issues, based on proper orthogonal decomposition (POD), the deep kernel extreme learning machine (DKELM) and a cloud model, a novel dynamic early-warning model using deep learning and spatiotemporal features fusion is proposed for dam deformation. In which, the POD and the correntropy-improved DKELM are coupled to extract the spatial principal component (SPC) of the deformation and to restrain the interference of outliers, and robust deep nonlinear mapping is established to enhance the dynamic updating ability. Moreover, embedding the concept transformation and fusion rule of the cloud model, the deformation early-warning indicator considering the randomness and fuzziness can be formulated and updated, and its reliability can be improved by the time-domain feature fusion of the SPC distribution. Hence, the dynamic warning of dam deformation is achieved by comparing the forecasted SPC and the early-warning indicators. A real dam application and model comparison demonstrate that the proposed model performs well in deformation SPC prediction and gives global and local early-warning indicators with the advantage of dynamic updating, which provides a more abundant and reliable basis for the dynamic safety warning of dam deformation.