Extraction of structural modal information using acoustic sensor measurements and machine learning

Abstract

Natural hazards could be very destructive. One challenge that society faces is how to rapidly, accurately and economically evaluate the damage of structures in the aftermath of a disaster. To overcome the limitations of visual inspections, a large body of research has focused on the design and deployment of the structural health monitoring systems, in which various sensors are deployed to obtain the necessary data for the specific health monitoring schemes. These sensors are typically installed with wired and/or wireless networks which the deployment and operation cost are obstacles to their usage. In this study, a novel systematic methodology, consisting of advanced cascading signal pre-processing techniques and a feature extraction method derived from machine learning, is proposed to identify low-frequency modal properties of structures using small amount of infrasound measurements obtained from microphones. In addition to providing the theoretical proofs, experiments on a vibrating structure subjected to ground excitations are conducted to verify the effectiveness and robustness of the proposed methodology. The study confirmed the feasibility of using microphone measurements as a means to perform non-contact and non-destructive structural health monitoring evaluation. To further investigate the effects of the several important factors, namely signal-to-noise ratio, number of reference sensors and the amount of collected data, a parametric study is performed to quantify their influences on the accuracy of identifying the desired modal property.