On the performance of neural networks and pattern recognition paradigms for classifying ultrasonic transducers

Abstract

The authors study, analyze, and compare the performance of pattern recognition methods with various neural network techniques for ultrasonic transducer characterization. The characterization algorithms are discussed. A multilayer backpropagation neural network is developed for characterizing the transducers. It provided a misclassification rate of 6%. Two other multilayer neural networks, sum-of-products and a newly devised neural network called hybrid sum-of-products, had misclassification rates of 10% and 7%, respectively. The best pattern recognition technique for this application was found to be the perceptron, which provided a misclassification rate of 23%. The worst pattern recognition technique was found to be the Bayes theorem method, which provided a misclassification rate of 54%. The competitive learning technique provided poor results as compared to the K-means for preclustering.<>