Output-only damage detection using neural network and sensor clustering under ambient vibration

Abstract

Time-series methods have become of interest in damage detection, particularly for automated and continuous structural health monitoring due to having no requirement for modal analysis or details of physical structural properties. Despite the success of the sensor clustering concept in improving the ability of time-series methods to detect, locate and quantify structural damage, most of the applications rely on free vibration response that can be obtained directly by impact testing, which is difficult to obtain for in-service structures, or indirectly by transforming the ambient vibration response. Therefore, the present study extends the use of sensor clustering for damage detection under ambient vibration by directly using the measured response. In this study, nonlinear autoregressive with exogenous inputs (NARX) system was modelled using artificial neural network for different sensor clusters using the acceleration response of the structure. The differences of the NARX neural network prediction errors are used as damage sensitive features to infer damage existence, location and severity. The applicability of the method is demonstrated using a numerical model of a two-span concrete slab under varying excitation conditions to simulate ambient vibration. The method performed successfully for single and multiple damage cases.