ANN-based averaging scheme for damage detection of high-rise buildings under noisy conditions

Abstract

Structural health monitoring (SHM) is crucial for assessing the condition of deteriorated high-rise buildings subjected to sudden and hazardous loads. This study proposes a novel averaging scheme to enhance the performance of an existing hybrid damage detection technique based on Artificial Neural Networks (ANNs). The proposed technique is validated numerically on a 30-storey building. The objective is to address the discrepancy between noise levels used during training and those present in generated modal data, thus mitigating the impact of measurement noise on damage predictions. By employing a series of ANNs trained with varying noise levels, a diverse range of predictions is obtained. These predictions are averaged to yield decisive conclusions, even when indecisive predictions outnumber decisive ones. This averaging scheme effectively reduces the influence of random noise, particularly when there is a notable disparity between the actual noise levels in measured data and statistical networks. Moreover, this study investigates the impact of the number of measurements on noise reduction, recommending approximately 100 measurements, in line with other experimental studies. Through the integration of the averaging scheme and increased measurement numbers, the ANN-based damage detection technique achieves remarkable accuracy in damage detection. Storey-level detection can be achieved when the noise levels in mode shapes reach up to 3.5 %. Additionally, the approach exhibits promising results in detecting damaged walls, with a noise threshold of up to 3 %. For damaged columns, a more modest threshold of 0.75 % suffices for light and complex damage scenarios.