Bayram Aygun

1

Using Neural Networks in Structural
Vulnerability Analysis

• Bayram Aygun
• Ryon Lab
• Rice University
• Civil Environmental Engineering
• ba2_at_rice.edu

2
Statement of the Original Problem

• Bridges are the most critical components of
  modern transportation networks. Although these
  networks provide the foundation for vibrant
  economies, a lack of understanding persists about
  the role that each bridge-soil system plays in
  network performance when subjected to unforeseen
  natural hazards such as earthquakes.

3
Why use neural networks (NNs)?

• Such a detailed bridge-soil model poses a
  computational challenge when one wants to derive
  its likelihood of failure or fragility when
  subjected to natural hazards.
• A neural network could be trained to approximate
  the nonlinear, dynamic behavior of the FE model
  and serve as a surrogate bridge model to evaluate
  its response to seismic hazards. This approach
  could save a lot of time.

4
Statement of the Problem Analyzed for the Project

• Because of the difficulties of gathering data the
  original problem is scaled down to the analysis
  of a 3-story truss building under static lateral
  loads. This problem is not trivial because it
  will give us insight on the potential of the
  usage of NNs in structural fragility analysis.

5
Objective

• The objective is to train a feed forward, fully
  connected neural network to mimic the behavior of
  an FE model of 3-story truss building. Once the
  network is trained it will be capable of
  producing fragility values for unseen lateral
  loads. These approximate fragility values will be
  used to generate system fragility curve. The
  analysis time is expected to decrease
  drastically, thus justifying the use of surrogate
  models.

6
Technical Approach on NN Training

• A three-layered, feed forward, fully connected NN
  with nonlinear transfer functions ( tanh(x) ) in
  every neuron is trained using error
  backpropagation algorithm.
• Bias neuron with constant input of 1 is
  introduced at the input layer.
• Momentum term is included in order to speed up
  the convergence.

7
Technical Approach on NN Training Contd

• Batch Training is used and hence the error is
  accumulated from all training data prior to
  weight update (180 instances 1 epoch 1
  iteration).
• Error is calculated as follows

8
Some Training and Test Results
9
Some Training and Test Results Contd
10
Some Training and Test Results Contd
11
Conclusions and Further Research

• The multi-layer NN did a good job approximating
  the fragility curve of the 3-story truss
  building.
• Further study should focus on getting enough data
  for the original bridge-soil system and instead
  of using static lateral loads earthquake
  time-histories shall be imposed to the structure.

12
References

• Kramer,S. L. (1996). Geotechnical Earthquake
  Engineering. Upper Saddle River, NJ Prentice
  Hall International Series
• Cosenza, E., Manfredi, G. (2000). Damage
  Indices and damage measures. Prog. Struct. Engng.
  Mater., 2, 50-59
• Marks, Roberts J. Artificial Neural Networks
  Supervised Models?" (n.d). 5 April 2008
  lthttp//cayman.globat.com/trademarksnet.com/Prese
  ntations/index.htmgt.
• Houmøller, Lars P. Neural Networks" (n.d). 4
  April 2008 ltwww.acabs.dk/submenus/courses/Files/Ne
  ural20networks.ppt gt.