Analysis on defect classification by deep neural networks and wavelet packet transform

Defect classification analysis in ultrasonic detection usually depends on the operator's professional knowledge and experience, which leads to the instability and individual differences of the analysis result. To solve the problem, a method for classification analysis of ultrasonic detection signals is presented. The method uses wavelet packet transform to extract the feature information from ultrasonic defection, and applies deep neural networks to classification. Experiments are tested by ultrasonic phased array system to collect detection signal which comes from via holes, inclined holes or flat-bottom holes. Then, classification analysis is done by the new method. Experimental results show two points. The first point is that using wavelet packet transform can increase the recognition accuracy and decrease the learning time of neural networks. The second point is that using deep neural networks can increase the recognition accuracy while increase acceptable learning time comparing to the common BP neural networks. By using the new method, the defect classification accuracy increases up 21.66% while the learning time increases only 91.9 s. Applying wavelet packet transform and deep neural networks to classification analysis in ultrasonic detection can exclude human interference and improve the recognition accuracy. As so, the new method has a great future in practical applications of ultrasonic detection.