Automation of Engineering Design Aids using Neural Networks

1
Automation of Engineering Design Aids using
Neural Networks

• Siripong Malasri and Jittapong Malasri
• Christian Brothers University
• Kriangsiri Malasri
• Georgia Tech

MAESC 05 May 13, 2005
2
Presentation Overview

• Introduction
• Artificial Neural Networks
• The Stress Concentration Problem
• Software Development
• Data preparation
• Network training and validation
• Standalone application development
• Conclusions and Future Work

3
Introduction

• Traditional design aids
• Look-up tables
• Graphical plots
• Shortcomings
• Inaccurate interpolation/extrapolation
• Difficult to smoothly integrate with computer
  applications

4
Neural Networks

• Have been used to recognize patterns and project
  trends in data
• Backpropagation model can be trained to
  generate desired input-output relationships

5
Stress Concentration (1)

• Objective
• Calculate the peak stress in a notched beam
  cross-section subject to a bending moment
• Possible approaches
• Finite-element analysis
• Experimental procedures
• Determine a stress concentration factor from a
  design aid

6
Stress Concentration (2)

• Stress concentration factor, C
• Function of the ratios a/h2 and h1/h2
• Peak stress at notch
• M bending moment applied
• I cross-sectional moment of inertia
• c distance from N.A.

7
Software Data Preparation

• Training data obtained from a published graphical
  design aid

• Inputs a/h2 , h1/h2
• Output C
• 46 training pairs, 15 calibration pairs, 15
  validation pairs

8
Software Network Training

• NeuroShell 2 software
• Backpropagation network with 2 input neurons, 8
  hidden neurons, and 1 output neuron
• Excellent results from trained network

9
Software Standalone Program

• Interface developed in Visual Basic
• Network code generated from NeuroShell 2

10
Conclusions and Future Work

• Excellent network estimates of the stress
  concentration factor for this particular
  application
• Standalone executable is portable to any Windows
  computer
• Future work comprehensive stress analysis
  program with a variety of cross-sections