ABSTRACT
However, the ANNs applied in previous researches were all “shallow”, its performance is limited and highly dependent on the manually designed features due to shallow ANN’s drawbacks, such as overfitting problem, lack of data and large computational cost. Although the universal approximation theorem (Hornik et al., 1989) shows that a feedforward network with one hidden layer can approximate any mapping from any finite dimensional discrete space to another as long as it has enough hidden units, the number of hidden units required and the corresponding computational cost will increase exponentially with the increasing complexity of target mapping (Barron, 1993). Fortunately, in last decade, there is a revolution in the research field of ANN (Hinton et al., 2006), called Deep Learning. The term “deep” implies there are multiple hidden layers in an ANN instead of only one hidden layer. A deeper neural network has
1 INTRODUCTION
More and more large bridge structures, which play a vital role in a nation’s economic prosperity, have been demanded by modern society. Therefore, the safety issue of large bridge structures is an area of growing interest in recent decades. Structural Health Monitoring (SHM) is at the core location to solve this problem. It is defined as a process of implementing a damage identification strategy for aerospace, civil, and mechanical engineering infrastructures. A major current focus in SHM is how to establish an efficient and effective damage detection algorithm to recognize the structural damage from the monitoring sensor data.
