Premium
A simple neural network based model approach for nylon 66 fabrics used in safety restraint systems: A comparison of two training algorithms
Author(s) -
Keshavaraj Ramesh,
Tock Richard W.,
Nusholtz Guy S.
Publication year - 1995
Publication title -
journal of applied polymer science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.575
H-Index - 166
eISSN - 1097-4628
pISSN - 0021-8995
DOI - 10.1002/app.1995.070570911
Subject(s) - artificial neural network , weighting , nonlinear system , computer science , algorithm , artificial intelligence , medicine , physics , quantum mechanics , radiology
Airbag technology relies on woven fabrics as the material of construction and particularly on knowledge pertaining to the fabric's permeability as a function of pressure drop, inflation temperature of the gas, fabric weave, fiber denier, and biaxial stress–strain relationships under biaxial conditions. While fabric permeability can be quantified by actual experimental measurements, the number and nonlinearity of the variables involved make the experiments time‐ and cost‐intensive. Moreover, interpolations within a given data set can yield questionable results. In this study, a very simple feed forward neural network architecture was used with a training rule involving a nonlinear optimization routine for updating weights of the proposed network. This training was compared to the training with an error‐back propagation routine. During this training, the ANN is introduced to data that contain the actual cause and effect patterns, with adjustments being made by changes in weighting factors until the errors in the output variables are minimized. Once trained, ANN can ascertain the essentials of the relationships and assimilate henceforth. In this study, after the initial training, the ANN was tested on additional data which were not part of the training processes. The predictions of the proposed trained network agreed very well with the new experimental data. On this basis, the proposed ANN model appears to be an effective tool for modeling airbag fabric behavior. This ANN model can assimilate relationships between different variables from the real‐world data and does not require extensive normalizing of the process data like a back‐propagation algorithm. Once trained, only fractions of a second are needed for information assimilation and output generation. This coupled with simplicity of use and accuracy of predictions from the real‐world data make this ANN model attractive for on‐line applications. © 1995 John Wiley & Sons, Inc.