Vibration-Based Structural Health Monitoring via Phase-Based Motion Estimation Using Deep Residual Networks
Feiyuan Xing, Ziquan Yan, Xiaoyu Ding
Abstract
In recent years, deep residual networks (ResNets) have been successfully used to detect structural damage. However, the performance of ResNets greatly depends on large high-quality datasets for learning, which can be difficult to obtain in actual engineering scenarios. In addition, in order to improve identification accuracy, vibration signals from multiple positions can be fused together to train the ResNets. On the one hand, vibration signals from multipositions can increase the quantity of the training data. On the other hand, hidden correlations among multivibration signals may also reflect some features improving identification accuracy. However, traditional vibration acquisition methods, such as accelerometers, cannot ensure that the vibration signals from multipositions are synchronized in the temporal domain, causing the correlations among multivibration signals not to be fully recorded. To solve the above problems, this study presents a novel structural health monitoring method by combining phase-based motion estimation (PME) with the use of ResNets. With the PME method, each pixel in a video can be regarded as a displacement sensor; thus, it is possible to obtain millions of vibration signals from a single video; in addition, all vibration signals obtained from a single video are perfectly synchronized, greatly facilitating ResNet applications. Here, an open gearbox mechanism was used for experimental validation of our approach. With the proposed method, only one video sample obtained under each structural condition was required to train a ResNet model for detecting gear misalignment (difference of only 0.2°) with 100% accuracy, indicating the outstanding performance of the proposed method.