Research on anomaly detection methods for power inspection in UAV bearings based on graph-structured multiview feature fusion
Measurement Science and Technology · 15.05.2026 · 37 (20) · 206111
DOI: 10.1088/1361-6501/ae671d ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 19.05.2026. Verfasst von Yifei Shi und 1 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben.
Abstract (Originalsprache)
UAV bearings serve as critical components in power inspection systems. Their health directly impacts the grid inspection efficiency. However, electromagnetic interference and vibration coupling in high-voltage environments make conventional single-feature detection inadequate. Early fault signatures get masked by noise, whereas fixed models fail to adapt to dynamic data distributions. This paper addresses insufficient single vibration signal representation and poor robustness in complex environments. We propose a graph-structured multiview feature fusion framework for bearing anomaly detection in power inspection UAVs. The framework derives five complementary views from vibration signals: time-domain waveforms, frequency spectra, time-frequency images, envelope spectra, and statistical features. Dynamic graph structures enable deep correlation modeling among views. The framework first extracts multiple feature views using signal processing. View-specific encoders map heterogeneous features to unified graph node embeddings. Multirelational graphs and learnable adjacency matrix model various association patterns between views. Multiscale graph convolution networks extract fault features at different receptive fields. Hierarchical aggregation yields progressive fusion from the node level through subgraph level to graph level. The system combines reconstruction error, distribution modeling, and contrastive learning for comprehensive anomaly scoring. On the Case Western Reserve University bearing dataset, our method achieves 94.7% detection accuracy, 93.2% F 1 score, and 87.3% early fault identification rate. These metrics improve by 6.1%, 6.3%, and 11.7%, respectively, compared to the best existing method (1D-ResNet). In cross-dataset generalization experiments, the performance degradation remains at just 6.8%, lower than comparison methods’ range 14.8%–17.2%. These findings validate the framework’s effectiveness and robustness.
Zitieren
APA
Shi, Y., & Xiao, B. (2026). Research on anomaly detection methods for power inspection in UAV bearings based on graph-structured multiview feature fusion. Measurement Science and Technology, 37(20), 206111. https://doi.org/10.1088/1361-6501/ae671d
Vancouver
Shi Y, Xiao B. Research on anomaly detection methods for power inspection in UAV bearings based on graph-structured multiview feature fusion. Measurement Science and Technology. 2026;37(20):206111. doi:10.1088/1361-6501/ae671d
BibTeX
@article{Shi2026,
author = {Shi, Yifei and Xiao, Bin},
title = {Research on anomaly detection methods for power inspection in UAV bearings based on graph-structured multiview feature fusion},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {20},
pages = {206111},
doi = {10.1088/1361-6501/ae671d}
}
EUVIMED