Experimental Investigation of the Generalization Performance of Neural Network in Defect Localization System for Steel Pipe Health Monitoring

This study aimed to assess the generalization performance of a Metal Health Monitoring system, which is crucial for practical applications. Previous research has not thoroughly examined this aspect of performance. To enhance the system's performance, we conducted experiments using 90 metal pieces, anticipating improved results with increased sample size. The pieces were divided into nine classes, representing undamaged and damaged conditions at eight different positions. Vibration waveforms were obtained by attaching piezoelectric sensors to the pieces. The waveforms were then split into training and evaluation datasets, and a neural network (NN) was trained on the former to classify the latter. The findings revealed that the NN achieved a remarkable accuracy of up to 80.6% in classifying the damage positions, even for metal pieces not included in the training set. These results indicate a high level of generalization performance in the Metal Health Monitoring system.