Performance of a Hybrid Neural Network Reduced Flexibility Index Method for Localization and Severity Detection of Damages in a Gravity Dam

This paper presents a practical procedure based on a hybrid neural network-flexibility damage index technique to predict the severity level and location of damages in a gravity dam. Numerical models of the Beni-Haroun gravity dam in Algeria are elaborated and validated using ambient vibration tests. The measured frequency response functions are used to identify all the frequencies and mode shapes that can effectively be retrieved in order to limit the calculation of the flexibility matrix to these modes. Although the flexibility index is capable of capturing the damage signature in terms of location and severity, the extraction of such information, however, is difficult in most cases, particularly when reducing the number of degree of freedoms (DOFs) to those representing accessible positions of sensors on the downstream side of the dam. Hence, a feedforward neural network is trained using a database constituted of flexibility indices for numerically simulated damages. In the test phase, the neural network achieved 94% and 80% success in predicting the location and severity of the damage, respectively, showing the high potential of the hybrid method to be used in the practice of structural health monitoring.