Keywords: C4.5, classification, composite, fuzzy clustering, pin-loaded hole, tensile stress.

A new application of data mining techniques for predicting the behaviour of pin-loaded composite joints is presented in this paper. The objective of this study is to evaluate the effectiveness of classification model to identify joint failure. In addition, we compare the performance of classification model with the model initialized by fuzzy c-means clustering. Our analysis indicates that the performance of classification model initiated by fuzzy c-means clustering in evaluating potential behaviour of joints is superior to the performance of classical classification.

To construct the model, the study used data sets obtained from experimental tests on pin-loaded composite as well as aluminium joints. It is desired to find: (a) the load-displacement tensile response for each specimen (b) the ultimate failure load.

A laminate composite (carbon fibre reinforced plastic) plate of length L, thickness t and width W with a hole of diameter D with a pin is used. In the tensile test, different edge distance/diameter (E/D) ratios were considered.

A classification algorithm, C4.5 has been applied to model the behaviour of composite and aluminium joints. In order to improve the performance of the above algorithm, fuzzy clustering has been applied for determination of the boundaries between different levels of the behaviour. The system is a four input-one output system. The input variables are material type, edge distance, load step and load value and the joint behaviour at that specific step is taken as the output variable. There are three possible behaviours for aluminium joints (Elastic, Plastic and Failure behaviour) while for composite there are only Non-Failure and Failure levels. In order to take the material properties into account, two models have been generated, one for each material type. In both cases, four different edge distances (i.e. 0.7, 1, 1.5 and 3 cm) have been used for training the systems while they were tested with edge distance 1.25 cm. For each edge distance, three sets of data have been applied.

The models usually have high accuracy when describing their own case. However, it might not be appropriate for a different material, even though the cases share similar steps of loading. We have implemented our models using Java under Windows 2000 on a COMPAQ Workstation for the C4.5 algorithm [1] and MATLAB (6.5.1 release 13, 1997) for fuzzy c-means clustering [2].

The performance of the prediction accuracy of the initial tree classifier, obtained from composite joints, verified that the classical classification tree is able to model the problem with around 95% accuracy. However, for aluminium joints, the comparison of the initial tree with the final fuzzy model indicated that the application of the fuzzy c-means provided much better results in terms of modelling and accuracy.

The results showed that fuzzy clustering could not improve much the composite model. The reason is due to the behaviour of composite that does not yield. This kind of behaviour makes the data not overlapping and therefore the boundaries of data are actually crisp rather than fuzzy.

1

J.R. Quinlan, .5: Programs for Machine Learning. Morgan Kaufman Publishers, San Mateo, California, 1993.

2

J.C.Bezdek, "A convergence theorem for the fuzzy ISODATA clustering algorithms", IEEE Trans.Pattern Anal..Machine Intell.PAMI, 2 (1), 1-8, 1981. doi:10.1109/TPAMI.1980.4766964

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £80 +P&P)