Automated Bayesian operational modal analysis of the long-span bridge using machine-learning algorithms

Abstract

The Bayesian Operational Modal Analysis (OMA) approach can both estimate the most probable value (MPV) of modal parameters and swiftly give identification uncertainty. However, the determination of certain parameters shall still require the expertise of professionals and process is labor-intensive. In light of the massively available data from the in-situ structural health monitoring (SHM) system of the long-span bridge, the proposal and development of an automated strategy for identifying and tracking varying bridge modal parameters are much desired. A novel machine-learning-based framework is proposed to automatically estimate the parameters required for the Bayesian OMA. The major contributions of the proposed method are: (i) The cross-modal assurance criterion (CMAC) matrix is proposed and adopted to obtain the correlation between the target modes from a global perspective. It is reconstructed using the convolutional autoencoder (CAE) to remove the embedded noise information and the modal resonant frequency band can be detected, and (ii) The Kohonen network is used to automatically determine the number of target modes by clustering the normalized singular value (SV) sequence. Collected data from numerical simulation and a full-scale long-span suspension bridge are adopted to validate the proposed framework. Results show that the proposed framework can successfully estimate the required parameters from the raw data in an automated manner. This feature transforms the conventional Bayesian OMA into an automated strategy for identifying and tracking the modal parameters of the long-span bridge.