Damage detection in structures using modified back-propagation neural networks

A nonparametric structural damage detection methodology based on neural networks method is presented for health monitoring of structure-unknown systems. In this approach appropriate neural networks are trained by use of the modal test data from a 'healthy' structure. The trained networks which are subsequently fed with vibration measurements from the same structure in different stages have the capability of recognizing the location and the content of structural damage and thereby can monitor the health of the structure. A modified back-propagation neural network is proposed to solve the two practical problems encountered by the traditional back-propagation method, i.e., slow learning progress and convergence to a false local minimum. Various training algorithms, types of the input layer and numbers of the nodes in the input layer are considered. Numerical example results from a 5-degree-of-freedom spring-mass structure and analyses on the experimental data of an actual 5-storey-steel-frame demonstrate that neural-networks-based method is a robust procedure and a practical tool for the detection of structural damage, and that the modified back-propagation algorithm could improve the computational efficiency as well as the accuracy of detection.