| The objective of this dissertation is to evaluate the relative performance of five classification algorithms for nondestructive damage detection. The classification algorithms investigated here are as follows: (1) a quadratic classifier obtained from Bayes' rule (i.e., using unequal damage and undamaged covariance matrices), (2) a linear classifier obtained from Bayes' rule (i.e., using equal damage and undamaged covariance matrices), (3) a linear classifier obtained from Bayes' rule: (i.e., assuming that the damaged and undamaged covariance matrices are equal to the identity matrix), (4) a classifier using Euclidean distance as a basis, and (5) a classifier using hypothesis testing. To meet this objective, an established theory of damage localization which yields information on the location of the damage directly from changes in mode shapes, is selected. Next, the application of sophisticated techniques from pattern recognition, in the form of the five classification algorithms, is performed to the existing theory of damage localization. Expressions for pattern classification using discriminant functions obtained from Bayes' rule, distance as similarity, and hypothesis testing are generated. Criteria for the evaluation of the proposed pattern recognition models are generated. Using the enhanced model, the locating of damage is attempted in: (1) a numerical model of a space structure with simulated damage at various locations, and (2) a real structure damaged at known locations. Finally, the accuracy and reliability of the pattern recognition models are evaluated using the established criteria. |
