No abstract

An extensive discrimination study was conducted on ultrasonic signals very similar to each other obtained from artificial inserts in a carbon fibre reinforced epoxy plate. The performance of fifteen classification schemes consisting of non-parametric pattern recognition and Artificial Neural System (ANS) algorithms is assessed in this paper. The purpose of this study is to define an upper bound for the classification error expected when similar ultrasonic signals are processed, as well as to compare the different classification techniques. The results indicate that classification errors strongly depend upon feature space selection and problem complexity. In the test cases treated in this work, the Wilk's Λ criterion was proved efficient for descriptor selection. Algorithm groups, conventional pattern recognition and ANSs all exhibit comparable overall performance as far as the minimum classification error is concerned. It is the user's task to try several classification schemes and select the one most suited to the specific application.

No abstract

No abstract

No abstract