Pattern Recognition Algorithms for Automated Defect Classification of Ultrasonic Signals

This work compares the applications of different pattern recognition algorithms for the automated defect classification of ultrasonic signals in Carbon Fiber Reinforced Plastic (CFRP) laminates. Three CFRP standards were considered with the following simulated defects: impact damage, foreign object inclusion, and porosity. A parametric classification approach based on nearest mean (NM) and unsupervised artificial neural networks (ANN) based on Hopfield networks in conjunction with time-domain ultrasonic echo signals were developed to detect and classify the simulated defects. In order to test the classification strategy, one hundred (100) time-domain 5 MHz ultrasonic echo signals were acquired for each type of defect and from defect-free regions in CFRP panels. The recorded signals were start-point aligned with the means using cross-correlation prior to feature extraction process. Features were extracted from aligned A-scan signals based on mean, standard deviation, Discrete Fourier Transform (DFT), and 1-D Discrete Wavelet Transform (DWT) of A-scan signals. After extracting the features, the signals for each category were randomly partitioned into equal-size training and test sets. The partitioning process was repeated so that at least 1000 signals could be tested for each category and the results could be averaged to obtain a robust estimate of the performance. The results obtained demonstrated that classification accuracies exceeding 98% are possible by using different pattern recognition algorithms.