Application of neural networks in bridge health prediction based on acceleration and displacement data domain

Abstract

The health condition of the bridge can be predicted through sensors’ reading in bridge monitoring. The sensors measure the acceleration and displacement of bridge response. The data is sent to the local server through the data acquisition. Interpretation of the data applied neural network in the localized server system. This paper aims to define performance of the acceleration and displacement data domain as input in applied neural networks. The architecture of neural networks’ model used an input layer, one and two hidden layers with n neurons and an output layer. The input layer consists of time-acceleration domain and time-displacement domain of the bridge due to earthquake loads. Meanwhile, the output layer consists of bridge condition level which is determined using finite-element analysis software. The training activation used Gradient Descent Back-propagation and activation transfer function used Log Sigmoid function. The bridge condition is categorized in a range 0 to 3, which indicates the extent of bridge health condition ranging from safe to high-risk level. The case study is 3 spans of box-girder’s bridge subject to four earthquakes loads. The results showed that the prediction of bridge health condition based on displacement data domain with one hidden layer is more acceptable compared with based on acceleration. The comparison obtains the recommendation of the best of data reading from the sensors to predict the bridge health condition. The application neural networks in the bridge health prediction can help the authorities to know the condition of the bridge due to earthquake at monitoring time, as the repair and maintenance of bridges can be performed as early as possible before the bridge was damaged.