Fault Assessment in Piezoelectric-Based Smart Strand Using 1D Convolutional Neural Network

Abstract

The smart strand technique has been recently developed as a cost-effective prestress load monitoring solution for post-tensioned engineering systems. Nonetheless, during its lifetime under various operational and environmental conditions, the sensing element of the smart strand has the potential to fail, threatening its functionality and resulting in inaccurate prestress load estimation. This study analyzes the effect of potential failures in the smart strand on impedance characteristics and develops a 1D convolutional neural network (1D CNN) for automated fault diagnosis. Instead of using a realistic experimental structure for which transducer faults can be hard to control accurately, we adopt a well-established finite element model to conduct all experiments. The results show that the impedance characteristics of a damaged smart strand are relatively different from other piezoelectric active sensing devices. While the slope of the susceptance response is widely accepted as a promising fault indicator, this study shows that the resistance response is more favorable for the smart strand. The developed network can accurately diagnose the potential faults in a damaged smart strand with the highest testing accuracy of 94.1%. Since the network can autonomously learn damage-sensitive features without pre-processing, it shows great potential for embedding in impedance-based damage identification systems for real-time structural health monitoring.

Wavelet Scattering Convolution Network-Based Detection Algorithm on Nondestructive Microcrack Electrical Signals of Eggs

“…Hu et al [24] designed a fault diagnosis method based on a one-dimensional convolutional neural network (1DCNN) and L2-support-vector machine(L2-SVM) for unbalanced data, which, compared with other intelligent methods, significantly improved the recognition accuracy and diagnostic performance of the model compared with other intelligent methods in processing unbalanced data. Le et al [25] developed a 1DCNN for automated fault diagnosis, which can autonomously learn damagesensitive features without pre-processing and can accurately diagnose potential faults that damage the smart chain. Relevant research [26][27][28] also shows that deep learning exhibits good performance in processing time series classifications.…”