A Machine Learning Approach to Beam Structural Variation Detection Via Output-only Acceleration Measurements

This paper presents a two-stage machine learning method for beam structural variation detection using the responses measured on multiple acceleration sensors. In the first stage, a second-order model identification method through discrete Fourier transform (DFT) spectrum of the acceleration responses is developed to construct the structural model as well as its modal frequencies under unknown input force. In the second stage of the proposed method, a computationally effective multiclass support vector machine (MSVM) is first designed to solve a multiclass classification problem, which is shown to be a structural variation detection problem. Then, a damage indicator that combines the model information constructed in the first stage and the MSVM is derived to fulfill a multi-variation damage detection of the beam bridge structure. The usefulness of the proposed results is verified using an experimental example of an in-door simply supported beam bridge structure, which has unknown model parameters and input force. The proposed method could successfully detect the variations of the beam structure, even for unknown model parameters and input force magnitude.