P2D-3 Performance Evaluation of Neural Network Based Ultrasonic Flaw Detection

In this study, a robust flaw detection algorithm using Neural Networks (NN) is presented for NDE applications. A three-layer feedforward NN which can perform a complex nonlinear mapping process has been used as a detection processor following the subband decomposition of the measured signal. The neural network architecture is trained to suppress the clutter echoes while maintaining the integrity of flaw echoes. The training process allows the neural network to learn about the statistics and the variation of the clutter signal. The robustness of the NN method is examined through testing materials with different grain sizes and multiple flaws. It has been shown that NN can improve the flaw-to-clutter (FCR) ratio significantly when the input experimental signal has FCR equal to 0 or less. Experimental results show that a typical FCR improvement of 40dB can be achieved using NN post detectors as opposed to 15dB with the conventional techniques including minimum, median, average, geometric mean and polarity detectors. The experimental results also confirm that the NN detector is capable of distinguishing two adjacent flaw echoes whereas the conventional techniques detect the presence of a single anomaly only. Furthermore, due its trainability, NN performs robustly when some of the subband signals used for detection have little or no flaw information.