Health monitoring method using committee of neural networks

Abstract

Preventive maintenance and structural safety of large structures such as bridges and buildings may be guaranteed by application of structural health monitoring systems. Damage assessment using structural identification technique is essential for the structural health monitoring. In this study, the committee technique for neural networks is applied to damage estimation of structures for the purpose of the health monitoring. The input to the neural networks consists of the modal parameters, and the output is composed of the element-level damage indices. In the committee technique, multiple neural networks are constructed and each individual networks is trained independently. Then, the estimated damage indices from different neural networks are averaged. Various committee techniques are possible. The architecture, the training patterns, and the input of each individual networks can be taken to be the same and/or different. In this study, the validity of the several committee methods for damage estimation was examined through numerical simulation study. Then, experiments were carried out to verify the effectiveness of the committee technique. It has been found that the estimated damage indices improve significantly by employing the committee of neural networks.