Support Vector Machine Classification of Ultrasonic Shaft Inspection Data Using Discrete Wavelet Transform

While many non-destructive ultrasonic signal test scenarios involve very shallow surfaces, signals for testing shafts are long and the new problem of mode-converted reflection emerges. They are echoes that do not correspond to cracks in the material, neither to characteristics of the shaft. Also, the length of the signals demands the application of efficient feature extraction mechanism to reduce the dimension of the pattern vectors and make classifier training feasible. The previous study by authors [9] established experimentally that Discrete Wavelet Transform (DWT) provided faster and more reliable feature extraction for Artificial Neural Networks (ANN) in these long signals in shafts. As an extended work, a new comparative experiment involving Support Vector Machines (SVM) approach instead of ANN models are made so that more confidence can be placed in the comparison result especially when there are only a limited number of training examples. The classification result from employing two different feature extraction schemes is also analyzed to investigate whether there exists any special relationship between feature types and class types. This investigation set the baseline for the future work of the authors, which is building a more advanced hybrid classifier model using multiple feature extraction schemes.