Simple Search:

Bayesian Network Classifiers for Damage Detection in Engineering Material


Mohamed Addin, Addin Osman (2007) Bayesian Network Classifiers for Damage Detection in Engineering Material. PhD thesis, Universiti Putra Malaysia.

Abstract / Synopsis

The automation of damage detection in engineering material using intelligent techniques (e.g. Neural networks) has not been matured enough to be practi- cable and needs more techniques to be implemented, improved, and developed. Nevertheless, the Neural networks have been implemented extensively for the damage detection, but in elementary ways. The damage detection and pre- diction are very important processes, since the damages have the potential of growing and leading to catastrophic loss of human life, and decrease in econ- omy (e.g. airline crashes, space shuttle explosions, and building collapses). Bayesian networks have been successfully implemented as classi¯ers in many research and industrial areas and they are used as models for representing un- certainty in knowledge domains. Nevertheless, they have not been thoroughly investigated and implemented such as Neural networks for the damage detec- tion. This thesis is dedicated to introduce them with the axiom of damage de- tection and implement them as a competitive probabilistic graphical model and as classi¯cation tools (Naijve bayes classi¯er and Bayesian network classi¯er) for the damage detection. The Bayesian networks have two-sided strengths: It is easy for humans to construct and to understand, and when communicated to a computer, they can easily be compiled. Changes in a system model should only induce local changes in a Bayesian network, where as system changes might require the design and training of an entirely new Neural network. The methodology used in the thesis to implement the Bayesian network for the damage detection provides a preliminary analysis used in proposing a novel fea- ture extraction algorithm (f-FFE: the f-folds feature extraction algorithm). The state-of-the-art shows that most of the feature reduction techniques, if not all, which have been implemented for the damage detection are feature selection not extraction. Feature selection is less °exible than feature extrac- tion in that feature selection is, in fact, a special case of feature extraction (with a coe±cient of one for each selected feature and a coe±cient of zero for any of the other features). This explains why an optimal feature set ob- tained by feature selection may or may not yield a good classi¯cation results. To validate the classi¯ers and the proposed algorithm, two data sets were used, the ¯rst set represents voltage amplitudes of Lamb-waves produced and col- lected by sensors and actuators mounted on the surface of laminates contain di®erent arti¯cial damages. The second set is a vibration data from a type of ball bearing operating under di®erent ¯ve fault conditions. The Bayesian net- work classi¯ers and the proposed algorithm have been tested using the second set. The studies conducted in this research have shown that Bayesian networks as one of the most successful machine learning classi¯ers for the damage detection in general and the Naijve bayes classi¯er in particular. They have also shown their e±ciency when compared to Neural networks in domains of uncertainty. The studies have also shown the e®ectiveness and e±ciency of the proposed algorithm in reducing the number of the input features while increasing the accuracy of the classi¯er. These techniques will play vital role in damage de- tection in engineering material, specially in the smart materials, which require continuous monitoring of the system for damages.

Download File


Download (213kB)

Additional Metadata

Item Type: Thesis (PhD)
Subject: Materials - Bayesian statistical decision theory
Call Number: ITMA 2007 6
Chairman Supervisor: Associate Professor Ir. Mohd. Sapuan Salit, PhD
Divisions: Institute of Advanced Technology
Depositing User: Rosmieza Mat Jusoh
Date Deposited: 12 Apr 2010 01:41
Last Modified: 27 May 2013 07:23
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item