Abstract
The diagnosis with an acoustic impulse signal like hammering sound has been used to detect defects in mechanical structures. Since the acoustic impulse signal is of transient nature, it is generally difficult to process the signal and to detect the defects automatically. In this paper, the preliminary study of the signal processing technology for the acoustic impulse signal is described. Here the application of the pattern recognition technique used in the speech recognition is examined for the evaluation of the defect depth. The applied neural model is the multi-layered neural network which intends to imitate the superior processing performance in human brain and also has a good performance for pattern recognition. Experiments are performed to get the acoustic impulse signal for the various defect depths. The network architecture, the input feature to the network and the signal processing method appropriate for the evaluation of the defect depth with the acoustic impulse signal are examined. The result shows 99.4% discrimination accuracy in this experiment. Furthermore, the forgetting learning rule applied to the neural network is examined to remove the unnecessary weights to discriminate the defect depth. The discrimination accuracy is 99.4% and weights are removed over 90%. These results suggest that the neural network is effective for the evaluation of the defect depth with the acoustic impulse signal and the forgetting learning rule is able to remove the unnecessary weights to keep the high discrimination accuracy.
