LSTM-based damage localization method for building structures using frequency feature damage index

This study presents an LSTM-based damage localization framework for building structures. The model is trained exclusively on acceleration responses measured in the healthy state, and damage localization is performed by analyzing, in the frequency domain, the discrepancy between the model-reconstructed response and the corresponding measured signal under the damaged state. To more accurately encode the structural dynamics, multistory acceleration responses from the healthy state are provided with selected story channels intentionally masked at the input, and the network is trained to reconstruct the masked channel at the output. When damaged-state acceleration responses are fed to the trained LSTM, a discrepancy arises between the model predictions and the measured responses. This discrepancy is transformed to the frequency domain via FFT and used to define the Frequency Feature Damage Index. The proposed procedure is implemented on a four-degree-of-freedom (4DOF) shear-frame numerical model subjected to white Gaussian noise excitation, where multiple damage scenarios are generated for training and validation. Its practical applicability is further examined using structural responses obtained from a shaking-table experiment on a three-story steel moment frame, through which the effectiveness and accuracy of the localization strategy are verified.