Wavelet Transform Based Feature Extraction for Ultrasonic Flaw Signal Classification

In this paper, an automatic system is designed to classify the ultrasonic flaw signals from carbon fiber reinforced polymer (CFRP) specimens with void, delamination and debonding. In such system, different methods based on discrete wavelet transform (DWT) and wavelet packet transform (WPT) are first utilized for feature extraction. After that, the linear mapping is applied for dimensionality reduction. Artificial neural networks (ANNs) and support vector machines (SVMs) are trained to validate the effectiveness of different wavelet transform based features for flaw signal classification. Experimental results show that the normalized energy of WPT coefficients coupled with the statistical parameters of WPT representation of original signals can be taken as the reliable features to effectively classify different ultrasonic flaw signals with lower training elapsed time.