"A Data Mining Study of Weld Quality Models Constructed with
MLP Neural Networks from Stratified Sampled Data"
Industrial Engineering Research Conference,
Dallas, TX, May 20-23, 2001, in the Conference's CD.
Liao, T.W., G. Wang, E. Triantaphyllou, and P.-C. Chang
Abstract:
A proportional stratified sampling method was implemented to
sample radiographic welding data. The sample size was varied at
different levels to study its effect on the model quality in
terms of training time and its prediction accuracy. The sampled
data of each size was then divided into training data and testing
data in the ratio of 9 to 1. The training data is used to obtain
multi-layer perceptron (MLP) neural network models. The quality
of each model was subsequently evaluated with unseen testing
data. Moreover, each sampled data set was characterized to show
its statistical representation of the population. The
correlation between the model quality and the sampled data
statistics is also discussed.
Data mining, MLP neural networks, Stratified sampling,
Radiography, Weld quality.
(Size = 53 KB)
