Structural health monitoring of civil structures: A diagnostic framework powered by deep metric learning

Recent advances in learning systems and sensor technology have enabled powerful strategies for autonomous data-driven damage detection in structural systems. This work proposes a novel method for the real-time localization of damage relying on a Siamese convolutional neural network. The method exploits a learnable mapping of raw vibration measurements onto a low-dimensional space, wherein damage locations can be easily identified. The mapping is learned in a supervised pairwise fashion exploiting labelled data, to induce a task-specific metric that allows to encode the damage position in the structure. Training data are generated through a reduced-order numerical model of the monitored structure. The damage position is then identified by performing a regression in the resulting low-dimensional features space. The proposed method does not require to define a priori target classes and decision boundaries, thus requiring a limited amount of user-dependent assumptions. Results relevant to an L-shaped cantilever beam and a portal frame railway bridge demonstrate that the procedure can be effectively exploited for the purpose of damage localization. The method also proves to be insensitive to operational variability, measurement noise and modeling inaccuracies.